From 96125fd398a5bb988ab7e5e0df669efdec1d3546 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <sergey.kopienko@intel.com>
Date: Thu, 19 Oct 2023 14:20:00 +0200
Subject: [PATCH 001/566] Implement tag dispatching prototype

#Source commit from RA: aba39a7e22b025299e05302088261ef747ddcd90

# Conflicts:
#	include/oneapi/dpl/pstl/algorithm_impl.h
#	include/oneapi/dpl/pstl/execution_impl.h
#	include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h
#	include/oneapi/dpl/pstl/iterator_defs.h
#	include/oneapi/dpl/pstl/utils.h
---
 include/oneapi/dpl/pstl/algorithm_impl.h      | 74 +++++++++++++++++--
 include/oneapi/dpl/pstl/execution_impl.h      | 54 +++++++++++---
 include/oneapi/dpl/pstl/glue_algorithm_impl.h | 18 ++---
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 23 +++++-
 .../pstl/hetero/dpcpp/execution_sycl_defs.h   | 12 +++
 include/oneapi/dpl/pstl/iterator_defs.h       | 33 ++++-----
 6 files changed, 165 insertions(+), 49 deletions(-)

diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index f4e64d63b4b..6c79ca30b5c 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -134,6 +134,15 @@ __brick_walk1(_DifferenceType __n, _Function __f, ::std::true_type) noexcept
     oneapi::dpl::__internal::__brick_walk1(__n, __f, ::std::false_type{});
 }
 
+template <class _ExecutionPolicy, class _ForwardIterator, class _Function>
+void
+__pattern_walk1(__serial_tag<_ExecutionPolicy, _ForwardIterator> __tag, _ExecutionPolicy&&, _ForwardIterator __first,
+                _ForwardIterator __last, _Function __f) noexcept
+{
+    using __tag_type = decltype(__tag);
+    __internal::__brick_walk1(__first, __last, __f, typename __tag_type::__is_vector{});
+}
+
 template <class _ExecutionPolicy, class _ForwardIterator, class _Function, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>
 __pattern_walk1(_ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _Function __f,
@@ -143,6 +152,45 @@ __pattern_walk1(_ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator _
     __internal::__brick_walk1(__first, __last, __f, __is_vector);
 }
 
+template <class _ExecutionPolicy, class _ForwardIterator, class _Function>
+void
+__pattern_walk1(__parallel_forward_tag<_ExecutionPolicy, _ForwardIterator>, _ExecutionPolicy&& __exec,
+                _ForwardIterator __first, _ForwardIterator __last, _Function __f)
+{
+    typedef typename ::std::iterator_traits<_ForwardIterator>::reference _ReferenceType;
+    auto __func = [&__f](_ReferenceType arg) { __f(arg); };
+    __internal::__except_handler([&]() {
+        __par_backend::__parallel_for_each(::std::forward<_ExecutionPolicy>(__exec), __first, __last, __func);
+    });
+}
+
+template <class _ExecutionPolicy, class _ForwardIterator, class _Function>
+void
+__pattern_walk1(__parallel_tag<_ExecutionPolicy, _ForwardIterator> __tag, _ExecutionPolicy&& __exec,
+                _ForwardIterator __first, _ForwardIterator __last, _Function __f)
+{
+    using __tag_type = decltype(__tag);
+    __internal::__except_handler([&]() {
+        __par_backend::__parallel_for(::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+                                      [__f](_ForwardIterator __i, _ForwardIterator __j) {
+                                          __internal::__brick_walk1(__i, __j, __f, typename __tag_type::__is_vector{});
+                                      });
+    });
+}
+
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _UnaryPredicate, class _Tp>
+oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, void>
+__pattern_replace_if(_Tag __tag, _ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last,
+                     _UnaryPredicate __pred, const _Tp& __new_value)
+{
+    oneapi::dpl::__internal::__pattern_walk1(
+        __tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+        oneapi::dpl::__internal::__replace_functor<
+            oneapi::dpl::__internal::__ref_or_copy<_ExecutionPolicy, const _Tp>,
+            oneapi::dpl::__internal::__ref_or_copy<_ExecutionPolicy, _UnaryPredicate>>(__new_value, __pred));
+}
+
+template <class _ExecutionPolicy, class _ForwardIterator, class _Function, class _IsVector>
 template <class _ExecutionPolicy, class _RandomAccessIterator, class _Function, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy_conditional<
     _ExecutionPolicy, __is_random_access_iterator_v<_RandomAccessIterator>>
@@ -669,13 +717,12 @@ __brick_find_if(_RandomAccessIterator __first, _RandomAccessIterator __last, _Pr
         [&__pred](_RandomAccessIterator __it, _SizeType __i) { return __pred(__it[__i]); });
 }
 
-template <class _ExecutionPolicy, class _ForwardIterator, class _Predicate, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _ForwardIterator>
-__pattern_find_if(_ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _Predicate __pred,
-                  _IsVector __is_vector,
-                  /*is_parallel=*/::std::false_type) noexcept
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Predicate>
+_ForwardIterator
+__pattern_find_if(_Tag __tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
+                  _Predicate __pred) noexcept
 {
-    return __internal::__brick_find_if(__first, __last, __pred, __is_vector);
+    return __internal::__brick_find_if(__first, __last, __pred, typename _Tag::__is_vector{});
 }
 
 template <class _ExecutionPolicy, class _RandomAccessIterator, class _Predicate, class _IsVector>
@@ -693,6 +740,21 @@ __pattern_find_if(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _Ran
     });
 }
 
+template <class _ExecutionPolicy, class _ForwardIterator, class _Predicate>
+_ForwardIterator
+__pattern_find_if(__parallel_tag<_ExecutionPolicy, _ForwardIterator> __tag, _ExecutionPolicy&& __exec,
+                  _ForwardIterator __first, _ForwardIterator __last, _Predicate __pred)
+{
+    using __tag_type = decltype(__tag);
+    return __except_handler([&]() {
+        return __parallel_find(::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+                               [__pred](_ForwardIterator __i, _ForwardIterator __j) {
+                                   return __brick_find_if(__i, __j, __pred, typename __tag_type::__is_vector{});
+                               },
+                               ::std::true_type{});
+    });
+}
+
 //------------------------------------------------------------------------
 // find_end
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/execution_impl.h b/include/oneapi/dpl/pstl/execution_impl.h
index fb9d9682103..fc7d6bec1cd 100644
--- a/include/oneapi/dpl/pstl/execution_impl.h
+++ b/include/oneapi/dpl/pstl/execution_impl.h
@@ -127,22 +127,56 @@ __is_parallelization_preferred(_ExecutionPolicy& __exec)
                                   __internal::__is_random_access_iterator_t<_IteratorTypes...>());
 }
 
-template <typename policy, typename... _IteratorTypes>
-struct __prefer_unsequenced_tag
+//------------------------------------------------------------------------
+// backend selector with tags
+//------------------------------------------------------------------------
+
+template <class _Policy, class... _IteratorTypes>
+struct __vectorable_tag
+{
+    using __is_vector = __conjunction<__allow_unsequenced<_Policy>,
+                                      typename __internal::__is_random_access_iterator<_IteratorTypes...>>;
+};
+
+template <class _Policy, class... _IteratorTypes>
+struct __serial_tag : __vectorable_tag<_Policy, _IteratorTypes...>
+{
+};
+
+template <class _Policy, class... _IteratorTypes>
+struct __parallel_tag : __vectorable_tag<_Policy, _IteratorTypes...>
 {
-    static constexpr bool value =
-        __internal::__allow_unsequenced<policy>::value && __internal::__is_random_access_iterator_v<_IteratorTypes...>;
-    typedef ::std::integral_constant<bool, value> type;
 };
 
-template <typename policy, typename... _IteratorTypes>
-struct __prefer_parallel_tag
+template <class _Policy, class... _IteratorTypes>
+struct __parallel_forward_tag : __vectorable_tag<_Policy, _IteratorTypes...>
 {
-    static constexpr bool value =
-        __internal::__allow_parallel<policy>::value && __internal::__is_random_access_iterator_v<_IteratorTypes...>;
-    typedef ::std::integral_constant<bool, value> type;
 };
 
+template <typename _Policy, typename... _IteratorTypes>
+using __tag_type =
+    typename ::std::conditional<__internal::__is_random_access_iterator<_IteratorTypes...>::value,
+                                __parallel_tag<_Policy, _IteratorTypes...>,
+                                typename ::std::conditional<__is_forward_iterator<_IteratorTypes...>::value,
+                                                            __parallel_forward_tag<_Policy, _IteratorTypes...>,
+                                                            __serial_tag<_Policy, _IteratorTypes...>>::type>::type;
+
+template <typename _Policy, class... _IteratorTypes>
+typename ::std::enable_if<!__allow_parallel<_Policy>::value,
+                          __serial_tag<_Policy, typename ::std::decay<_IteratorTypes>::type...>>::type
+__select_backend(_Policy&&, _IteratorTypes&&...)
+{
+    return {};
+}
+
+template <typename _Policy, class... _IteratorTypes>
+typename ::std::enable_if<__allow_parallel<_Policy>::value,
+                          __tag_type<_Policy, typename ::std::decay<_IteratorTypes>::type...>>::type
+__select_backend(_Policy&&, _IteratorTypes&&...)
+{
+    return {};
+}
+
 } // namespace __internal
 } // namespace dpl
 } // namespace oneapi
diff --git a/include/oneapi/dpl/pstl/glue_algorithm_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
index 37d6f91a2a6..9c2186d6bda 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
@@ -97,10 +97,10 @@ template <class _ExecutionPolicy, class _ForwardIterator, class _Predicate>
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _ForwardIterator>
 find_if(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last, _Predicate __pred)
 {
-    return oneapi::dpl::__internal::__pattern_find_if(
-        ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __pred,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec));
+    auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(std::forward<_ExecutionPolicy>(__exec), __first);
+
+    return oneapi::dpl::__internal::__pattern_find_if(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first,
+                                                      __last, __pred);
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator, class _Predicate>
@@ -363,13 +363,9 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy>
 replace_if(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last, _UnaryPredicate __pred,
            const _Tp& __new_value)
 {
-    oneapi::dpl::__internal::__pattern_walk1(
-        ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
-        oneapi::dpl::__internal::__replace_functor<
-            oneapi::dpl::__internal::__ref_or_copy<_ExecutionPolicy, const _Tp>,
-            oneapi::dpl::__internal::__ref_or_copy<_ExecutionPolicy, _UnaryPredicate>>(__new_value, __pred),
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec));
+    auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(::std::forward<_ExecutionPolicy>(__exec), __first);
+    __pattern_replace_if(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __pred,
+                         __new_value);
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator, class _Tp>
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 87118eb27ce..4a3c9ce13b8 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -58,6 +58,24 @@ __pattern_walk1(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIte
         .wait();
 }
 
+template <typename _ExecutionPolicy, typename _ForwardIterator, typename _Function>
+void
+__pattern_walk1(__offload_tag, _ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last,
+                _Function __f)
+{
+    auto __n = __last - __first;
+    if (__n <= 0)
+        return;
+
+    auto __keep =
+        oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read_write, _ForwardIterator>();
+    auto __buf = __keep(__first, __last);
+
+    oneapi::dpl::__par_backend_hetero::__parallel_for(__exec, unseq_backend::walk_n<_ExecutionPolicy, _Function>{__f},
+                                                      __n, __buf.all_view())
+        .wait();
+}
+
 //------------------------------------------------------------------------
 // walk1_n
 //------------------------------------------------------------------------
@@ -639,9 +657,8 @@ __pattern_equal(_ExecutionPolicy&& __exec, _Iterator1 __first1, _Iterator1 __las
 //------------------------------------------------------------------------
 
 template <typename _ExecutionPolicy, typename _Iterator, typename _Pred>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, _Iterator>
-__pattern_find_if(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last, _Pred __pred,
-                  /*vector=*/::std::true_type, /*parallel=*/::std::true_type)
+_Iterator
+__pattern_find_if(__offload_tag, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last, _Pred __pred)
 {
     if (__first == __last)
         return __last;
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h b/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h
index b70e5518ff6..e8567f05442 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h
@@ -18,6 +18,7 @@
 
 #include "../../onedpl_config.h"
 #include "../../execution_defs.h"
+#include "../../iterator_defs.h"
 
 #include "sycl_defs.h"
 
@@ -312,6 +313,17 @@ using __enable_if_device_execution_policy_double_no_default =
                            oneapi::dpl::__internal::__is_convertible_to_event<_Events...>,
                        _T>;
 
+struct __offload_tag
+{
+};
+
+template <class... _IteratorTypes, typename... _PolicyParams>
+typename ::std::enable_if<__is_random_access_iterator<_IteratorTypes...>::value, __offload_tag>::type
+__select_backend(const execution::device_policy<_PolicyParams...>&, _IteratorTypes&&...)
+{
+    return {};
+}                       
+
 } // namespace __internal
 
 } // namespace dpl
diff --git a/include/oneapi/dpl/pstl/iterator_defs.h b/include/oneapi/dpl/pstl/iterator_defs.h
index 50c1f9a6507..f0aac046362 100644
--- a/include/oneapi/dpl/pstl/iterator_defs.h
+++ b/include/oneapi/dpl/pstl/iterator_defs.h
@@ -50,30 +50,25 @@ struct __iterator_traits<_Tp*, void> : ::std::iterator_traits<_Tp*>
 {
 };
 
-// Make is_random_access_iterator not to fail with a 'hard' error when it's used in SFINAE with
-// a non-iterator type by providing a default value.
-template <typename _IteratorType, typename = void>
-struct __is_random_access_iterator_impl : ::std::false_type
-{
-};
+// Make is_random_access_iterator and is_forward_iterator not to fail with a 'hard' error when it's used in
+//SFINAE with a non-iterator type by providing a default value.
+template <typename _IteratorTag, typename... _IteratorTypes>
+auto
+__is_needed_iter(int)
+    -> decltype(__conjunction<::std::is_base_of<_IteratorTag, typename __iterator_traits<typename ::std::decay<
+                                                                  _IteratorTypes>::type>::iterator_category>...>{});
 
-template <typename _IteratorType>
-struct __is_random_access_iterator_impl<_IteratorType,
-                                        __void_type<typename __iterator_traits<_IteratorType>::iterator_category>>
-    : ::std::is_same<typename __iterator_traits<_IteratorType>::iterator_category, ::std::random_access_iterator_tag>
-{
-};
+template <typename... _IteratorTypes>
+auto
+__is_needed_iter(...) -> ::std::false_type;
 
-/* iterator */
-template <typename _IteratorType, typename... _OtherIteratorTypes>
-struct __is_random_access_iterator
-    : ::std::conditional_t<__is_random_access_iterator_impl<_IteratorType>::value,
-                           __is_random_access_iterator<_OtherIteratorTypes...>, ::std::false_type>
+template <typename... _IteratorTypes>
+struct __is_random_access_iterator : decltype(__is_needed_iter<::std::random_access_iterator_tag, _IteratorTypes...>(0))
 {
 };
 
-template <typename _IteratorType>
-struct __is_random_access_iterator<_IteratorType> : __is_random_access_iterator_impl<_IteratorType>
+template <typename... _IteratorTypes>
+struct __is_forward_iterator : decltype(__is_needed_iter<::std::forward_iterator_tag, _IteratorTypes...>(0))
 {
 };
 

From 185023f5a741f962c7f17680401a3882bd2c3e63 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <sergey.kopienko@intel.com>
Date: Thu, 19 Oct 2023 14:25:22 +0200
Subject: [PATCH 002/566] Replace struct __conjunction -> std::conjunction

---
 include/oneapi/dpl/pstl/execution_impl.h |  4 ++--
 include/oneapi/dpl/pstl/iterator_defs.h  |  4 ++--
 include/oneapi/dpl/pstl/tuple_impl.h     |  4 ++--
 include/oneapi/dpl/pstl/utils.h          | 15 ---------------
 4 files changed, 6 insertions(+), 21 deletions(-)

diff --git a/include/oneapi/dpl/pstl/execution_impl.h b/include/oneapi/dpl/pstl/execution_impl.h
index fc7d6bec1cd..e75f7e6945d 100644
--- a/include/oneapi/dpl/pstl/execution_impl.h
+++ b/include/oneapi/dpl/pstl/execution_impl.h
@@ -134,8 +134,8 @@ __is_parallelization_preferred(_ExecutionPolicy& __exec)
 template <class _Policy, class... _IteratorTypes>
 struct __vectorable_tag
 {
-    using __is_vector = __conjunction<__allow_unsequenced<_Policy>,
-                                      typename __internal::__is_random_access_iterator<_IteratorTypes...>>;
+    using __is_vector = std::conjunction<__allow_unsequenced<_Policy>,
+                                         typename __internal::__is_random_access_iterator<_IteratorTypes...>>;
 };
 
 template <class _Policy, class... _IteratorTypes>
diff --git a/include/oneapi/dpl/pstl/iterator_defs.h b/include/oneapi/dpl/pstl/iterator_defs.h
index f0aac046362..210b7e0ca82 100644
--- a/include/oneapi/dpl/pstl/iterator_defs.h
+++ b/include/oneapi/dpl/pstl/iterator_defs.h
@@ -55,8 +55,8 @@ struct __iterator_traits<_Tp*, void> : ::std::iterator_traits<_Tp*>
 template <typename _IteratorTag, typename... _IteratorTypes>
 auto
 __is_needed_iter(int)
-    -> decltype(__conjunction<::std::is_base_of<_IteratorTag, typename __iterator_traits<typename ::std::decay<
-                                                                  _IteratorTypes>::type>::iterator_category>...>{});
+    -> decltype(std::conjunction<::std::is_base_of<_IteratorTag, typename __iterator_traits<typename ::std::decay<
+                                                                     _IteratorTypes>::type>::iterator_category>...>{});
 
 template <typename... _IteratorTypes>
 auto
diff --git a/include/oneapi/dpl/pstl/tuple_impl.h b/include/oneapi/dpl/pstl/tuple_impl.h
index b3e373a58fe..e5dc3bdf859 100644
--- a/include/oneapi/dpl/pstl/tuple_impl.h
+++ b/include/oneapi/dpl/pstl/tuple_impl.h
@@ -341,8 +341,8 @@ struct tuple<T1, T...>
     template <typename _U1, typename... _U,
               typename = ::std::enable_if_t<
                   (sizeof...(_U) == sizeof...(T) &&
-                   oneapi::dpl::__internal::__conjunction<::std::is_constructible<T1, _U1&&>,
-                                                          ::std::is_constructible<T, _U&&>...>::value)>>
+                   std::conjunction<::std::is_constructible<T1, _U1&&>,
+                                    ::std::is_constructible<T, _U&&>...>::value)>>
     tuple(_U1&& _value, _U&&... _next) : holder(::std::forward<_U1>(_value)), next(::std::forward<_U>(_next)...)
     {
     }
diff --git a/include/oneapi/dpl/pstl/utils.h b/include/oneapi/dpl/pstl/utils.h
index dec07f676a0..2f7edb08d31 100644
--- a/include/oneapi/dpl/pstl/utils.h
+++ b/include/oneapi/dpl/pstl/utils.h
@@ -553,21 +553,6 @@ struct __next_to_last
 template <typename _T, class _Enable = void>
 class __future;
 
-template <typename... _Bs>
-struct __conjunction : ::std::true_type
-{
-};
-
-template <typename _B1>
-struct __conjunction<_B1> : _B1
-{
-};
-
-template <typename _B1, typename... _Bs>
-struct __conjunction<_B1, _Bs...> : ::std::conditional_t<!bool(_B1::value), _B1, __conjunction<_Bs...>>
-{
-};
-
 // empty base class for type erasure
 struct __lifetime_keeper_base
 {

From 81e802eabe926042a903218134b1f19a89931174 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <sergey.kopienko@intel.com>
Date: Fri, 20 Oct 2023 10:29:53 +0200
Subject: [PATCH 003/566] Support FPGA backend. Adopt offload tags

(cherry picked from commit 42d44064d9a9102ac52c38d25ab81374ea864d25)

# Conflicts:
#	include/oneapi/dpl/pstl/algorithm_impl.h
#	include/oneapi/dpl/pstl/execution_impl.h
#	include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h
#	include/oneapi/dpl/pstl/iterator_defs.h
#	include/oneapi/dpl/pstl/parallel_backend_tbb.h
---
 include/oneapi/dpl/pstl/algorithm_impl.h      | 32 ++++-----
 include/oneapi/dpl/pstl/execution_impl.h      | 67 ++++++++++++-------
 include/oneapi/dpl/pstl/glue_algorithm_impl.h |  4 +-
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 12 ++--
 .../pstl/hetero/dpcpp/execution_sycl_defs.h   | 29 ++++++--
 .../pstl/hetero/dpcpp/parallel_backend_sycl.h | 41 ++++++++++++
 .../hetero/dpcpp/parallel_backend_sycl_fpga.h | 44 ++++++++++++
 include/oneapi/dpl/pstl/iterator_defs.h       | 10 +--
 .../oneapi/dpl/pstl/parallel_backend_tbb.h    | 18 +++++
 include/oneapi/dpl/pstl/parallel_impl.h       |  4 +-
 10 files changed, 201 insertions(+), 60 deletions(-)

diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 6c79ca30b5c..2d1f115a76e 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -134,13 +134,12 @@ __brick_walk1(_DifferenceType __n, _Function __f, ::std::true_type) noexcept
     oneapi::dpl::__internal::__brick_walk1(__n, __f, ::std::false_type{});
 }
 
-template <class _ExecutionPolicy, class _ForwardIterator, class _Function>
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Function>
 void
-__pattern_walk1(__serial_tag<_ExecutionPolicy, _ForwardIterator> __tag, _ExecutionPolicy&&, _ForwardIterator __first,
+__pattern_walk1(_Tag, _ExecutionPolicy&&, _ForwardIterator __first,
                 _ForwardIterator __last, _Function __f) noexcept
 {
-    using __tag_type = decltype(__tag);
-    __internal::__brick_walk1(__first, __last, __f, typename __tag_type::__is_vector{});
+    __internal::__brick_walk1(__first, __last, __f, typename _Tag::__is_vector{});
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator, class _Function, class _IsVector>
@@ -154,26 +153,26 @@ __pattern_walk1(_ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator _
 
 template <class _ExecutionPolicy, class _ForwardIterator, class _Function>
 void
-__pattern_walk1(__parallel_forward_tag<_ExecutionPolicy, _ForwardIterator>, _ExecutionPolicy&& __exec,
+__pattern_walk1(__parallel_forward_tag, _ExecutionPolicy&& __exec,
                 _ForwardIterator __first, _ForwardIterator __last, _Function __f)
 {
     typedef typename ::std::iterator_traits<_ForwardIterator>::reference _ReferenceType;
     auto __func = [&__f](_ReferenceType arg) { __f(arg); };
     __internal::__except_handler([&]() {
-        __par_backend::__parallel_for_each(::std::forward<_ExecutionPolicy>(__exec), __first, __last, __func);
+        __par_backend::__parallel_for_each(__parallel_forward_tag::__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __func);
     });
 }
 
-template <class _ExecutionPolicy, class _ForwardIterator, class _Function>
+template <class _IsVector, class _ExecutionPolicy, class _ForwardIterator, class _Function>
 void
-__pattern_walk1(__parallel_tag<_ExecutionPolicy, _ForwardIterator> __tag, _ExecutionPolicy&& __exec,
+__pattern_walk1(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec,
                 _ForwardIterator __first, _ForwardIterator __last, _Function __f)
 {
-    using __tag_type = decltype(__tag);
+    using __backend_tag = typename decltype(__tag)::__backend_tag;
     __internal::__except_handler([&]() {
-        __par_backend::__parallel_for(::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+        __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
                                       [__f](_ForwardIterator __i, _ForwardIterator __j) {
-                                          __internal::__brick_walk1(__i, __j, __f, typename __tag_type::__is_vector{});
+                                          __internal::__brick_walk1(__i, __j, __f, _IsVector{});
                                       });
     });
 }
@@ -725,6 +724,7 @@ __pattern_find_if(_Tag __tag, _ExecutionPolicy&&, _ForwardIterator __first, _For
     return __internal::__brick_find_if(__first, __last, __pred, typename _Tag::__is_vector{});
 }
 
+// KSATODO were deleted in prototype?
 template <class _ExecutionPolicy, class _RandomAccessIterator, class _Predicate, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator>
 __pattern_find_if(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIterator __last,
@@ -740,16 +740,16 @@ __pattern_find_if(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _Ran
     });
 }
 
-template <class _ExecutionPolicy, class _ForwardIterator, class _Predicate>
+template <class _IsVector, class _ExecutionPolicy, class _ForwardIterator, class _Predicate>
 _ForwardIterator
-__pattern_find_if(__parallel_tag<_ExecutionPolicy, _ForwardIterator> __tag, _ExecutionPolicy&& __exec,
+__pattern_find_if(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec,
                   _ForwardIterator __first, _ForwardIterator __last, _Predicate __pred)
 {
-    using __tag_type = decltype(__tag);
+    using __backend_tag = typename decltype(__tag)::__backend_tag;
     return __except_handler([&]() {
-        return __parallel_find(::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+        return __parallel_find(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
                                [__pred](_ForwardIterator __i, _ForwardIterator __j) {
-                                   return __brick_find_if(__i, __j, __pred, typename __tag_type::__is_vector{});
+                                   return __brick_find_if(__i, __j, __pred, _IsVector{});
                                },
                                ::std::true_type{});
     });
diff --git a/include/oneapi/dpl/pstl/execution_impl.h b/include/oneapi/dpl/pstl/execution_impl.h
index e75f7e6945d..d7743361f7b 100644
--- a/include/oneapi/dpl/pstl/execution_impl.h
+++ b/include/oneapi/dpl/pstl/execution_impl.h
@@ -131,48 +131,67 @@ __is_parallelization_preferred(_ExecutionPolicy& __exec)
 // backend selector with tags
 //------------------------------------------------------------------------
 
-template <class _Policy, class... _IteratorTypes>
-struct __vectorable_tag
-{
-    using __is_vector = std::conjunction<__allow_unsequenced<_Policy>,
-                                         typename __internal::__is_random_access_iterator<_IteratorTypes...>>;
-};
+struct __tbb_backend {};
+// KSATODO required create tag for onedpl
+// KSATODO required create tag for omp
 
-template <class _Policy, class... _IteratorTypes>
-struct __serial_tag : __vectorable_tag<_Policy, _IteratorTypes...>
+template <class _IsVector>
+struct __serial_tag
 {
+    using __is_vector = _IsVector;
 };
 
-template <class _Policy, class... _IteratorTypes>
-struct __parallel_tag : __vectorable_tag<_Policy, _IteratorTypes...>
+template <class _IsVector>
+struct __parallel_tag
 {
+    using __is_vector = _IsVector;
+    // backend tag can be change depending on
+    // TBB availability in the environment
+    using __backend_tag = __tbb_backend;
 };
 
-template <class _Policy, class... _IteratorTypes>
-struct __parallel_forward_tag : __vectorable_tag<_Policy, _IteratorTypes...>
+struct __parallel_forward_tag
 {
+    using __is_vector = ::std::false_type;
+    // backend tag can be change depending on
+    // TBB availability in the environment
+    using __backend_tag = __tbb_backend;
 };
 
-template <typename _Policy, typename... _IteratorTypes>
+template <class _IsVector, class... _IteratorTypes>
 using __tag_type =
     typename ::std::conditional<__internal::__is_random_access_iterator<_IteratorTypes...>::value,
-                                __parallel_tag<_Policy, _IteratorTypes...>,
+                                __parallel_tag<_IsVector>,
                                 typename ::std::conditional<__is_forward_iterator<_IteratorTypes...>::value,
-                                                            __parallel_forward_tag<_Policy, _IteratorTypes...>,
-                                                            __serial_tag<_Policy, _IteratorTypes...>>::type>::type;
+                                                            __parallel_forward_tag,
+                                                            __serial_tag<_IsVector>
+                                                            >::type
+                                >::type;
+
+template <class... _IteratorTypes>
+__serial_tag<std::false_type>
+__select_backend(oneapi::dpl::execution::sequenced_policy, _IteratorTypes&&...)
+{
+    return {};
+}
+
+template <class... _IteratorTypes>
+__serial_tag<__internal::__is_random_access_iterator<_IteratorTypes...>>
+__select_backend(oneapi::dpl::execution::unsequenced_policy, _IteratorTypes&&...)
+{
+    return {};
+}
 
-template <typename _Policy, class... _IteratorTypes>
-typename ::std::enable_if<!__allow_parallel<_Policy>::value,
-                          __serial_tag<_Policy, typename ::std::decay<_IteratorTypes>::type...>>::type
-__select_backend(_Policy&&, _IteratorTypes&&...)
+template <class... _IteratorTypes>
+__tag_type<std::false_type, _IteratorTypes...>
+__select_backend(oneapi::dpl::execution::parallel_policy, _IteratorTypes&&...)
 {
     return {};
 }
 
-template <typename _Policy, class... _IteratorTypes>
-typename ::std::enable_if<__allow_parallel<_Policy>::value,
-                          __tag_type<_Policy, typename ::std::decay<_IteratorTypes>::type...>>::type
-__select_backend(_Policy&&, _IteratorTypes&&...)
+template <class... _IteratorTypes>
+__tag_type<__internal::__is_random_access_iterator<_IteratorTypes...>, _IteratorTypes...>
+__select_backend(oneapi::dpl::execution::parallel_unsequenced_policy, _IteratorTypes&&...)
 {
     return {};
 }
diff --git a/include/oneapi/dpl/pstl/glue_algorithm_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
index 9c2186d6bda..86651375bee 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
@@ -97,7 +97,7 @@ template <class _ExecutionPolicy, class _ForwardIterator, class _Predicate>
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _ForwardIterator>
 find_if(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last, _Predicate __pred)
 {
-    auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(std::forward<_ExecutionPolicy>(__exec), __first);
+    auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first);
 
     return oneapi::dpl::__internal::__pattern_find_if(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first,
                                                       __last, __pred);
@@ -363,7 +363,7 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy>
 replace_if(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last, _UnaryPredicate __pred,
            const _Tp& __new_value)
 {
-    auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(::std::forward<_ExecutionPolicy>(__exec), __first);
+    auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first);
     __pattern_replace_if(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __pred,
                          __new_value);
 }
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 4a3c9ce13b8..1ec77c0d5ea 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -58,9 +58,9 @@ __pattern_walk1(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIte
         .wait();
 }
 
-template <typename _ExecutionPolicy, typename _ForwardIterator, typename _Function>
+template <typename _BackendTag, typename _ExecutionPolicy, typename _ForwardIterator, typename _Function>
 void
-__pattern_walk1(__offload_tag, _ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last,
+__pattern_walk1(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last,
                 _Function __f)
 {
     auto __n = __last - __first;
@@ -71,7 +71,7 @@ __pattern_walk1(__offload_tag, _ExecutionPolicy&& __exec, _ForwardIterator __fir
         oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read_write, _ForwardIterator>();
     auto __buf = __keep(__first, __last);
 
-    oneapi::dpl::__par_backend_hetero::__parallel_for(__exec, unseq_backend::walk_n<_ExecutionPolicy, _Function>{__f},
+    oneapi::dpl::__par_backend_hetero::__parallel_for(_BackendTag{}, __exec, unseq_backend::walk_n<_ExecutionPolicy, _Function>{__f},
                                                       __n, __buf.all_view())
         .wait();
 }
@@ -656,16 +656,16 @@ __pattern_equal(_ExecutionPolicy&& __exec, _Iterator1 __first1, _Iterator1 __las
 // find_if
 //------------------------------------------------------------------------
 
-template <typename _ExecutionPolicy, typename _Iterator, typename _Pred>
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator, typename _Pred>
 _Iterator
-__pattern_find_if(__offload_tag, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last, _Pred __pred)
+__pattern_find_if(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last, _Pred __pred)
 {
     if (__first == __last)
         return __last;
 
     using _Predicate = oneapi::dpl::unseq_backend::single_match_pred<_ExecutionPolicy, _Pred>;
 
-    return __par_backend_hetero::__parallel_find(
+    return __par_backend_hetero::__parallel_find(_BackendTag{},
         ::std::forward<_ExecutionPolicy>(__exec),
         __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__first),
         __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__last), _Predicate{__pred},
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h b/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h
index e8567f05442..6d4cad20373 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h
@@ -313,16 +313,35 @@ using __enable_if_device_execution_policy_double_no_default =
                            oneapi::dpl::__internal::__is_convertible_to_event<_Events...>,
                        _T>;
 
-struct __offload_tag
+template <typename _BackendTag>
+struct __hetero_tag
 {
+    using __backend_tag = _BackendTag;
 };
 
-template <class... _IteratorTypes, typename... _PolicyParams>
-typename ::std::enable_if<__is_random_access_iterator<_IteratorTypes...>::value, __offload_tag>::type
-__select_backend(const execution::device_policy<_PolicyParams...>&, _IteratorTypes&&...)
+struct __device_backend
+{
+};
+
+template <class... _IteratorTypes, typename _KernelName>
+typename ::std::enable_if<__is_random_access_iterator<_IteratorTypes...>::value, __hetero_tag<__device_backend>>::type
+__select_backend(const execution::device_policy<_KernelName>&, _IteratorTypes&&...)
 {
     return {};
-}                       
+}
+
+#if _ONEDPL_FPGA_DEVICE
+struct __fpga_backend : __device_backend
+{
+};
+
+template <class... _IteratorTypes, unsigned int _Factor, typename _KernelName>
+typename ::std::enable_if<__is_random_access_iterator<_IteratorTypes...>::value, __hetero_tag<__fpga_backend>>::type
+__select_backend(const execution::fpga_policy<_Factor, _KernelName>&, _IteratorTypes&&...)
+{
+    return {};
+}
+#endif
 
 } // namespace __internal
 
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
index 2335bad252e..4b23fc1d922 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
@@ -257,6 +257,31 @@ __parallel_for(_ExecutionPolicy&& __exec, _Fp __brick, _Index __count, _Ranges&&
                                                   ::std::forward<_Ranges>(__rngs)...);
 }
 
+template <typename _ExecutionPolicy, typename _Fp, typename _Index, typename... _Ranges>
+__future<void>
+__parallel_for(oneapi::dpl::__internal::__device_backend, _ExecutionPolicy&& __exec, _Fp __brick, _Index __count, _Ranges&&... __rngs)
+{
+    assert(__get_first_range_size(__rngs...) > 0);
+
+    using _Policy = typename ::std::decay<_ExecutionPolicy>::type;
+    using _CustomName = typename _Policy::kernel_name;
+    using _ForKernel = oneapi::dpl::__par_backend_hetero::__internal::_KernelName_t<__parallel_for_kernel, _CustomName,
+                                                                                    _Fp, _Ranges...>;
+
+    _PRINT_INFO_IN_DEBUG_MODE(__exec);
+    auto __event = __exec.queue().submit([&__rngs..., &__brick, __count](sycl::handler& __cgh) {
+        //get an access to data under SYCL buffer:
+        oneapi::dpl::__ranges::__require_access(__cgh, __rngs...);
+
+        __cgh.parallel_for<_ForKernel>(sycl::range</*dim=*/1>(__count), [=](sycl::item</*dim=*/1> __item_id) {
+            auto __idx = __item_id.get_linear_id();
+            __brick(__idx, __rngs...);
+        });
+    });
+
+    return __future<void>(__event);
+}
+
 //------------------------------------------------------------------------
 // parallel_transform_scan - async pattern
 //------------------------------------------------------------------------
@@ -1263,6 +1288,22 @@ __parallel_find(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last,
                          __f, _TagType{}, __buf.all_view());
 }
 
+template <typename _ExecutionPolicy, typename _Iterator, typename _Brick, typename _IsFirst>
+_Iterator
+__parallel_find(oneapi::dpl::__internal::__device_backend, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last, _Brick __f, _IsFirst)
+{
+    auto __keep = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _Iterator>();
+    auto __buf = __keep(__first, __last);
+
+    using _TagType =
+        typename ::std::conditional<_IsFirst::value, __parallel_find_forward_tag<decltype(__buf.all_view())>,
+                                    __parallel_find_backward_tag<decltype(__buf.all_view())>>::type;
+    return __first + oneapi::dpl::__par_backend_hetero::__parallel_find_or(
+                         __par_backend_hetero::make_wrapped_policy<__find_policy_wrapper>(
+                             ::std::forward<_ExecutionPolicy>(__exec)),
+                         __f, _TagType{}, __buf.all_view());
+}
+
 //------------------------------------------------------------------------
 // parallel_merge - async pattern
 //-----------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
index 8f9108c085e..4949aa48aaf 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
@@ -88,6 +88,38 @@ __parallel_for(_ExecutionPolicy&& __exec, _Fp __brick, _Index __count, _Ranges&&
                                                                 __count, std::forward<_Ranges>(__rngs)...);
 }
 
+template <typename _ExecutionPolicy, typename _Fp, typename _Index, typename... _Ranges>
+__future<void>
+__parallel_for(oneapi::dpl::__internal::__fpga_backend, _ExecutionPolicy&& __exec, _Fp __brick, _Index __count, _Ranges&&... __rngs)
+{
+    auto __n = __get_first_range_size(__rngs...);
+    assert(__n > 0);
+
+    using _Policy = typename ::std::decay<_ExecutionPolicy>::type;
+    using __kernel_name = typename _Policy::kernel_name;
+#if __SYCL_UNNAMED_LAMBDA__
+    using __kernel_name_t = __parallel_for_kernel<_Fp, __kernel_name, _Ranges...>;
+#else
+    using __kernel_name_t = __parallel_for_kernel<__kernel_name>;
+#endif
+
+    _PRINT_INFO_IN_DEBUG_MODE(__exec);
+    auto __event = __exec.queue().submit([&__rngs..., &__brick, __count](sycl::handler& __cgh) {
+        //get an access to data under SYCL buffer:
+        oneapi::dpl::__ranges::__require_access(__cgh, __rngs...);
+
+        __cgh.single_task<__kernel_name_t>([=]() {
+#pragma unroll(::std::decay <_ExecutionPolicy>::type::unroll_factor)
+            for (auto __idx = 0; __idx < __count; ++__idx)
+            {
+                __brick(__idx, __rngs...);
+            }
+        });
+    });
+
+    return __future<void>(__event);
+}
+
 //------------------------------------------------------------------------
 // parallel_transform_reduce
 //------------------------------------------------------------------------
@@ -228,6 +260,7 @@ oneapi::dpl::__internal::__enable_if_fpga_execution_policy<_ExecutionPolicy, _It
 __parallel_find(_ExecutionPolicy&& __exec, _Iterator1 __first, _Iterator1 __last, _Iterator2 __s_first,
                 _Iterator2 __s_last, _Brick __f, _IsFirst __is_first)
 {
+
     // workaround until we implement more performant version for patterns
     using _Policy = ::std::decay_t<_ExecutionPolicy>;
     using __kernel_name = typename _Policy::kernel_name;
@@ -247,6 +280,17 @@ __parallel_find(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last,
     return oneapi::dpl::__par_backend_hetero::__parallel_find(__device_policy, __first, __last, __f, __is_first);
 }
 
+template <typename _ExecutionPolicy, typename _Iterator, typename _Brick, typename _IsFirst>
+_Iterator
+__parallel_find(oneapi::dpl::__internal::__fpga_backend __tag, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last, _Brick __f, _IsFirst __is_first)
+{
+    // workaround until we implement more performant version for patterns
+    using _Policy = typename ::std::decay<_ExecutionPolicy>::type;
+    using __kernel_name = typename _Policy::kernel_name;
+    auto __device_policy = oneapi::dpl::execution::make_device_policy<__kernel_name>(__exec.queue());
+    return oneapi::dpl::__par_backend_hetero::__parallel_find(oneapi::dpl::__internal::__device_backend{}, __device_policy, __first, __last, __f, __is_first);
+}
+
 template <typename _ExecutionPolicy>
 auto
 __device_policy(_ExecutionPolicy&& __exec)
diff --git a/include/oneapi/dpl/pstl/iterator_defs.h b/include/oneapi/dpl/pstl/iterator_defs.h
index 210b7e0ca82..6cd2f6cb204 100644
--- a/include/oneapi/dpl/pstl/iterator_defs.h
+++ b/include/oneapi/dpl/pstl/iterator_defs.h
@@ -54,21 +54,21 @@ struct __iterator_traits<_Tp*, void> : ::std::iterator_traits<_Tp*>
 //SFINAE with a non-iterator type by providing a default value.
 template <typename _IteratorTag, typename... _IteratorTypes>
 auto
-__is_needed_iter(int)
-    -> decltype(std::conjunction<::std::is_base_of<_IteratorTag, typename __iterator_traits<typename ::std::decay<
-                                                                     _IteratorTypes>::type>::iterator_category>...>{});
+__is_iterator_of(int)
+    -> decltype(__conjunction<::std::is_base_of<_IteratorTag, typename __iterator_traits<typename ::std::decay<
+                                                                  _IteratorTypes>::type>::iterator_category>...>{});
 
 template <typename... _IteratorTypes>
 auto
 __is_needed_iter(...) -> ::std::false_type;
 
 template <typename... _IteratorTypes>
-struct __is_random_access_iterator : decltype(__is_needed_iter<::std::random_access_iterator_tag, _IteratorTypes...>(0))
+struct __is_random_access_iterator : decltype(__is_iterator_of<::std::random_access_iterator_tag, _IteratorTypes...>(0))
 {
 };
 
 template <typename... _IteratorTypes>
-struct __is_forward_iterator : decltype(__is_needed_iter<::std::forward_iterator_tag, _IteratorTypes...>(0))
+struct __is_forward_iterator : decltype(__is_iterator_of<::std::forward_iterator_tag, _IteratorTypes...>(0))
 {
 };
 
diff --git a/include/oneapi/dpl/pstl/parallel_backend_tbb.h b/include/oneapi/dpl/pstl/parallel_backend_tbb.h
index 556e305e1c7..67615fffc2b 100644
--- a/include/oneapi/dpl/pstl/parallel_backend_tbb.h
+++ b/include/oneapi/dpl/pstl/parallel_backend_tbb.h
@@ -125,6 +125,17 @@ __parallel_for(_ExecutionPolicy&&, _Index __first, _Index __last, _Fp __f)
     });
 }
 
+//! Evaluation of brick f[i,j) for each subrange [i,j) of [first,last)
+// wrapper over tbb::parallel_for
+template <class _ExecutionPolicy, class _Index, class _Fp>
+void
+__parallel_for(oneapi::dpl::__internal::__tbb_backend, _ExecutionPolicy&&, _Index __first, _Index __last, _Fp __f)
+{
+    tbb::this_task_arena::isolate([=]() {
+        tbb::parallel_for(tbb::blocked_range<_Index>(__first, __last), __parallel_for_body<_Index, _Fp>(__f));
+    });
+}
+
 //! Evaluation of brick f[i,j) for each subrange [i,j) of [first,last)
 // wrapper over tbb::parallel_reduce
 template <class _ExecutionPolicy, class _Value, class _Index, typename _RealBody, typename _Reduction>
@@ -1322,6 +1333,13 @@ __parallel_for_each(_ExecutionPolicy&&, _ForwardIterator __begin, _ForwardIterat
     tbb::this_task_arena::isolate([&]() { tbb::parallel_for_each(__begin, __end, __f); });
 }
 
+template <class _ExecutionPolicy, class _ForwardIterator, class _Fp>
+void
+__parallel_for_each(oneapi::dpl::__internal::__tbb_backend, _ExecutionPolicy&&, _ForwardIterator __begin, _ForwardIterator __end, _Fp __f)
+{
+    tbb::this_task_arena::isolate([&]() { tbb::parallel_for_each(__begin, __end, __f); });
+}
+
 } // namespace __tbb_backend
 } // namespace dpl
 } // namespace oneapi
diff --git a/include/oneapi/dpl/pstl/parallel_impl.h b/include/oneapi/dpl/pstl/parallel_impl.h
index a2d7d20e562..780a495897d 100644
--- a/include/oneapi/dpl/pstl/parallel_impl.h
+++ b/include/oneapi/dpl/pstl/parallel_impl.h
@@ -34,7 +34,7 @@ namespace __internal
 Each f[i,j) must return a value in [i,j). */
 template <class _ExecutionPolicy, class _Index, class _Brick, class _IsFirst>
 _Index
-__parallel_find(_ExecutionPolicy&& __exec, _Index __first, _Index __last, _Brick __f, _IsFirst)
+__parallel_find(__tbb_backend __tag, _ExecutionPolicy&& __exec, _Index __first, _Index __last, _Brick __f, _IsFirst)
 {
     typedef typename ::std::iterator_traits<_Index>::difference_type _DifferenceType;
     const _DifferenceType __n = __last - __first;
@@ -44,7 +44,7 @@ __parallel_find(_ExecutionPolicy&& __exec, _Index __first, _Index __last, _Brick
 
     ::std::atomic<_DifferenceType> __extremum(__initial_dist);
     // TODO: find out what is better here: parallel_for or parallel_reduce
-    __par_backend::__parallel_for(::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+    __par_backend::__parallel_for(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
                                   [__comp, __f, __first, &__extremum](_Index __i, _Index __j) {
                                       // See "Reducing Contention Through Priority Updates", PPoPP '13, for discussion of
                                       // why using a shared variable scales fairly well in this situation.

From 7005dc7558fa0c333e94cc4dbd0de408ec0b65f5 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <sergey.kopienko@intel.com>
Date: Thu, 19 Oct 2023 15:39:07 +0200
Subject: [PATCH 004/566] Replace struct __conjunction -> std::conjunction

---
 include/oneapi/dpl/pstl/iterator_defs.h | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/include/oneapi/dpl/pstl/iterator_defs.h b/include/oneapi/dpl/pstl/iterator_defs.h
index 6cd2f6cb204..524b3e887ff 100644
--- a/include/oneapi/dpl/pstl/iterator_defs.h
+++ b/include/oneapi/dpl/pstl/iterator_defs.h
@@ -55,8 +55,8 @@ struct __iterator_traits<_Tp*, void> : ::std::iterator_traits<_Tp*>
 template <typename _IteratorTag, typename... _IteratorTypes>
 auto
 __is_iterator_of(int)
-    -> decltype(__conjunction<::std::is_base_of<_IteratorTag, typename __iterator_traits<typename ::std::decay<
-                                                                  _IteratorTypes>::type>::iterator_category>...>{});
+    -> decltype(std::conjunction<::std::is_base_of<_IteratorTag, typename __iterator_traits<typename ::std::decay<
+                                                                     _IteratorTypes>::type>::iterator_category>...>{});
 
 template <typename... _IteratorTypes>
 auto

From 73d3ad3b8ea6afc844ff776e8f3a6bf5579de1bb Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <sergey.kopienko@intel.com>
Date: Thu, 19 Oct 2023 15:44:32 +0200
Subject: [PATCH 005/566] Fix compile error: cherry pick error #1

---
 include/oneapi/dpl/pstl/algorithm_impl.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 2d1f115a76e..e35f2465bc2 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -189,7 +189,7 @@ __pattern_replace_if(_Tag __tag, _ExecutionPolicy&& __exec, _ForwardIterator __f
             oneapi::dpl::__internal::__ref_or_copy<_ExecutionPolicy, _UnaryPredicate>>(__new_value, __pred));
 }
 
-template <class _ExecutionPolicy, class _ForwardIterator, class _Function, class _IsVector>
+//template <class _ExecutionPolicy, class _ForwardIterator, class _Function, class _IsVector>
 template <class _ExecutionPolicy, class _RandomAccessIterator, class _Function, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy_conditional<
     _ExecutionPolicy, __is_random_access_iterator_v<_RandomAccessIterator>>

From c5f1a11a7806f6c79a3b2bbfd5647fdd3efa29de Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <sergey.kopienko@intel.com>
Date: Thu, 19 Oct 2023 17:49:02 +0200
Subject: [PATCH 006/566] Fix compile error: __future<void> - too old code #2

---
 include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h  | 4 ++--
 .../oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h | 4 ++--
 2 files changed, 4 insertions(+), 4 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
index 4b23fc1d922..d2760f5ee17 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
@@ -258,7 +258,7 @@ __parallel_for(_ExecutionPolicy&& __exec, _Fp __brick, _Index __count, _Ranges&&
 }
 
 template <typename _ExecutionPolicy, typename _Fp, typename _Index, typename... _Ranges>
-__future<void>
+auto
 __parallel_for(oneapi::dpl::__internal::__device_backend, _ExecutionPolicy&& __exec, _Fp __brick, _Index __count, _Ranges&&... __rngs)
 {
     assert(__get_first_range_size(__rngs...) > 0);
@@ -279,7 +279,7 @@ __parallel_for(oneapi::dpl::__internal::__device_backend, _ExecutionPolicy&& __e
         });
     });
 
-    return __future<void>(__event);
+    return __future(__event);
 }
 
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
index 4949aa48aaf..36d755f5734 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
@@ -89,7 +89,7 @@ __parallel_for(_ExecutionPolicy&& __exec, _Fp __brick, _Index __count, _Ranges&&
 }
 
 template <typename _ExecutionPolicy, typename _Fp, typename _Index, typename... _Ranges>
-__future<void>
+auto
 __parallel_for(oneapi::dpl::__internal::__fpga_backend, _ExecutionPolicy&& __exec, _Fp __brick, _Index __count, _Ranges&&... __rngs)
 {
     auto __n = __get_first_range_size(__rngs...);
@@ -117,7 +117,7 @@ __parallel_for(oneapi::dpl::__internal::__fpga_backend, _ExecutionPolicy&& __exe
         });
     });
 
-    return __future<void>(__event);
+    return __future(__event);
 }
 
 //------------------------------------------------------------------------

From 0609db5d4b7797981326790c280c9a3732c0ae6f Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <sergey.kopienko@intel.com>
Date: Thu, 19 Oct 2023 17:57:45 +0200
Subject: [PATCH 007/566] Fix compile error: no template named '_KernelName_t'
 in namespace 'oneapi::dpl::__par_backend_hetero::__internal' #3

---
 .../oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h  | 8 ++++++--
 1 file changed, 6 insertions(+), 2 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
index d2760f5ee17..51854c08a1c 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
@@ -180,6 +180,9 @@ make_iter_mode(const _Iterator& __it) -> decltype(iter_mode<outMode>()(__it))
 
 // set of class templates to name kernels
 
+template <typename... _Name>
+class __parallel_for_kernel;
+
 template <typename... _Name>
 class __scan_local_kernel;
 
@@ -265,8 +268,9 @@ __parallel_for(oneapi::dpl::__internal::__device_backend, _ExecutionPolicy&& __e
 
     using _Policy = typename ::std::decay<_ExecutionPolicy>::type;
     using _CustomName = typename _Policy::kernel_name;
-    using _ForKernel = oneapi::dpl::__par_backend_hetero::__internal::_KernelName_t<__parallel_for_kernel, _CustomName,
-                                                                                    _Fp, _Ranges...>;
+    using _ForKernel =
+        oneapi::dpl::__par_backend_hetero::__internal::__kernel_name_generator<__parallel_for_kernel, _CustomName, _Fp,
+                                                                               _Ranges...>;
 
     _PRINT_INFO_IN_DEBUG_MODE(__exec);
     auto __event = __exec.queue().submit([&__rngs..., &__brick, __count](sycl::handler& __cgh) {

From 243c6620f4cb84831be796124276cd8675d8dcb9 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <sergey.kopienko@intel.com>
Date: Fri, 20 Oct 2023 10:14:46 +0200
Subject: [PATCH 008/566] include/oneapi/dpl/pstl/parallel_backend_tbb.h - fix
 compile error #4 in header_inclusion_order_algorithm_1.pass :
 parallel_backend_tbb.h:132:41: error: no type named '__tbb_backend' in
 namespace 'oneapi::dpl::__internal'

include/oneapi/dpl/pstl/parallel_backend_tbb.h:132:41: error: no type named '__tbb_backend' in namespace 'oneapi::dpl::__internal'
__parallel_for(oneapi::dpl::__internal::__tbb_backend, _ExecutionPolicy&&, _Index __first, _Index __last, _Fp __f)
               ~~~~~~~~~~~~~~~~~~~~~~~~~^
---
 include/oneapi/dpl/pstl/execution_impl.h       | 6 +++---
 include/oneapi/dpl/pstl/parallel_backend_tbb.h | 5 +++--
 include/oneapi/dpl/pstl/parallel_impl.h        | 2 +-
 3 files changed, 7 insertions(+), 6 deletions(-)

diff --git a/include/oneapi/dpl/pstl/execution_impl.h b/include/oneapi/dpl/pstl/execution_impl.h
index d7743361f7b..7e747db1223 100644
--- a/include/oneapi/dpl/pstl/execution_impl.h
+++ b/include/oneapi/dpl/pstl/execution_impl.h
@@ -131,7 +131,7 @@ __is_parallelization_preferred(_ExecutionPolicy& __exec)
 // backend selector with tags
 //------------------------------------------------------------------------
 
-struct __tbb_backend {};
+struct __tbb_backend_tag {};
 // KSATODO required create tag for onedpl
 // KSATODO required create tag for omp
 
@@ -147,7 +147,7 @@ struct __parallel_tag
     using __is_vector = _IsVector;
     // backend tag can be change depending on
     // TBB availability in the environment
-    using __backend_tag = __tbb_backend;
+    using __backend_tag = __tbb_backend_tag;
 };
 
 struct __parallel_forward_tag
@@ -155,7 +155,7 @@ struct __parallel_forward_tag
     using __is_vector = ::std::false_type;
     // backend tag can be change depending on
     // TBB availability in the environment
-    using __backend_tag = __tbb_backend;
+    using __backend_tag = __tbb_backend_tag;
 };
 
 template <class _IsVector, class... _IteratorTypes>
diff --git a/include/oneapi/dpl/pstl/parallel_backend_tbb.h b/include/oneapi/dpl/pstl/parallel_backend_tbb.h
index 67615fffc2b..d3663b2d301 100644
--- a/include/oneapi/dpl/pstl/parallel_backend_tbb.h
+++ b/include/oneapi/dpl/pstl/parallel_backend_tbb.h
@@ -23,6 +23,7 @@
 #include <type_traits>
 
 #include "parallel_backend_utils.h"
+#include "execution_impl.h"
 
 // Bring in minimal required subset of Intel(R) Threading Building Blocks (Intel(R) TBB)
 #include <tbb/blocked_range.h>
@@ -129,7 +130,7 @@ __parallel_for(_ExecutionPolicy&&, _Index __first, _Index __last, _Fp __f)
 // wrapper over tbb::parallel_for
 template <class _ExecutionPolicy, class _Index, class _Fp>
 void
-__parallel_for(oneapi::dpl::__internal::__tbb_backend, _ExecutionPolicy&&, _Index __first, _Index __last, _Fp __f)
+__parallel_for(oneapi::dpl::__internal::__tbb_backend_tag, _ExecutionPolicy&&, _Index __first, _Index __last, _Fp __f)
 {
     tbb::this_task_arena::isolate([=]() {
         tbb::parallel_for(tbb::blocked_range<_Index>(__first, __last), __parallel_for_body<_Index, _Fp>(__f));
@@ -1335,7 +1336,7 @@ __parallel_for_each(_ExecutionPolicy&&, _ForwardIterator __begin, _ForwardIterat
 
 template <class _ExecutionPolicy, class _ForwardIterator, class _Fp>
 void
-__parallel_for_each(oneapi::dpl::__internal::__tbb_backend, _ExecutionPolicy&&, _ForwardIterator __begin, _ForwardIterator __end, _Fp __f)
+__parallel_for_each(oneapi::dpl::__internal::__tbb_backend_tag, _ExecutionPolicy&&, _ForwardIterator __begin, _ForwardIterator __end, _Fp __f)
 {
     tbb::this_task_arena::isolate([&]() { tbb::parallel_for_each(__begin, __end, __f); });
 }
diff --git a/include/oneapi/dpl/pstl/parallel_impl.h b/include/oneapi/dpl/pstl/parallel_impl.h
index 780a495897d..cc92205b360 100644
--- a/include/oneapi/dpl/pstl/parallel_impl.h
+++ b/include/oneapi/dpl/pstl/parallel_impl.h
@@ -34,7 +34,7 @@ namespace __internal
 Each f[i,j) must return a value in [i,j). */
 template <class _ExecutionPolicy, class _Index, class _Brick, class _IsFirst>
 _Index
-__parallel_find(__tbb_backend __tag, _ExecutionPolicy&& __exec, _Index __first, _Index __last, _Brick __f, _IsFirst)
+__parallel_find(__tbb_backend_tag __tag, _ExecutionPolicy&& __exec, _Index __first, _Index __last, _Brick __f, _IsFirst)
 {
     typedef typename ::std::iterator_traits<_Index>::difference_type _DifferenceType;
     const _DifferenceType __n = __last - __first;

From 18ef3e7dfa695a5a7646e363516323748c398899 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <sergey.kopienko@intel.com>
Date: Fri, 20 Oct 2023 12:12:02 +0200
Subject: [PATCH 009/566] Fix cherry pick error #5: delete previously not
 removed code

---
 include/oneapi/dpl/pstl/algorithm_impl.h | 16 ----------------
 1 file changed, 16 deletions(-)

diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index e35f2465bc2..125c8ca312f 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -724,22 +724,6 @@ __pattern_find_if(_Tag __tag, _ExecutionPolicy&&, _ForwardIterator __first, _For
     return __internal::__brick_find_if(__first, __last, __pred, typename _Tag::__is_vector{});
 }
 
-// KSATODO were deleted in prototype?
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _Predicate, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator>
-__pattern_find_if(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIterator __last,
-                  _Predicate __pred, _IsVector __is_vector,
-                  /*is_parallel=*/::std::true_type)
-{
-    return __except_handler([&]() {
-        return __parallel_find(::std::forward<_ExecutionPolicy>(__exec), __first, __last,
-                               [__pred, __is_vector](_RandomAccessIterator __i, _RandomAccessIterator __j) {
-                                   return __brick_find_if(__i, __j, __pred, __is_vector);
-                               },
-                               ::std::true_type{});
-    });
-}
-
 template <class _IsVector, class _ExecutionPolicy, class _ForwardIterator, class _Predicate>
 _ForwardIterator
 __pattern_find_if(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec,

From 8d2efba6254c93edef996b4d7f8381e89f20e1b6 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <sergey.kopienko@intel.com>
Date: Fri, 20 Oct 2023 12:18:23 +0200
Subject: [PATCH 010/566] 
 include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h - using
 __is_random_access_iterator_v instead of
 __is_random_access_iterator<_IteratorTypes...>::value

---
 include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h b/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h
index 6d4cad20373..c77247e71eb 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h
@@ -324,7 +324,7 @@ struct __device_backend
 };
 
 template <class... _IteratorTypes, typename _KernelName>
-typename ::std::enable_if<__is_random_access_iterator<_IteratorTypes...>::value, __hetero_tag<__device_backend>>::type
+typename ::std::enable_if<__is_random_access_iterator_v<_IteratorTypes...>, __hetero_tag<__device_backend>>::type
 __select_backend(const execution::device_policy<_KernelName>&, _IteratorTypes&&...)
 {
     return {};
@@ -336,7 +336,7 @@ struct __fpga_backend : __device_backend
 };
 
 template <class... _IteratorTypes, unsigned int _Factor, typename _KernelName>
-typename ::std::enable_if<__is_random_access_iterator<_IteratorTypes...>::value, __hetero_tag<__fpga_backend>>::type
+typename ::std::enable_if<__is_random_access_iterator_v<_IteratorTypes...>, __hetero_tag<__fpga_backend>>::type
 __select_backend(const execution::fpga_policy<_Factor, _KernelName>&, _IteratorTypes&&...)
 {
     return {};

From e67b8709cbc05235523ef03cfb4d3bb7af6e7cbc Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <sergey.kopienko@intel.com>
Date: Fri, 20 Oct 2023 17:47:50 +0200
Subject: [PATCH 011/566] 
 include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h - using
 std::enable_if_t instead of typename ::std::enable_if<...>::type

---
 include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h b/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h
index c77247e71eb..75d1f54885d 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h
@@ -324,7 +324,7 @@ struct __device_backend
 };
 
 template <class... _IteratorTypes, typename _KernelName>
-typename ::std::enable_if<__is_random_access_iterator_v<_IteratorTypes...>, __hetero_tag<__device_backend>>::type
+::std::enable_if_t<__is_random_access_iterator_v<_IteratorTypes...>, __hetero_tag<__device_backend>>
 __select_backend(const execution::device_policy<_KernelName>&, _IteratorTypes&&...)
 {
     return {};
@@ -336,7 +336,7 @@ struct __fpga_backend : __device_backend
 };
 
 template <class... _IteratorTypes, unsigned int _Factor, typename _KernelName>
-typename ::std::enable_if<__is_random_access_iterator_v<_IteratorTypes...>, __hetero_tag<__fpga_backend>>::type
+::std::enable_if_t<__is_random_access_iterator_v<_IteratorTypes...>, __hetero_tag<__fpga_backend>>
 __select_backend(const execution::fpga_policy<_Factor, _KernelName>&, _IteratorTypes&&...)
 {
     return {};

From 0b27fc29626be303ff98852fd2b6ceaeb1d957ed Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <sergey.kopienko@intel.com>
Date: Wed, 25 Oct 2023 09:55:57 +0200
Subject: [PATCH 012/566] Rename __device_backend to __device_backend_tag

---
 include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h  | 6 +++---
 .../oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h    | 4 ++--
 .../dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h      | 2 +-
 3 files changed, 6 insertions(+), 6 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h b/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h
index 75d1f54885d..6fe53805c28 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h
@@ -319,19 +319,19 @@ struct __hetero_tag
     using __backend_tag = _BackendTag;
 };
 
-struct __device_backend
+struct __device_backend_tag
 {
 };
 
 template <class... _IteratorTypes, typename _KernelName>
-::std::enable_if_t<__is_random_access_iterator_v<_IteratorTypes...>, __hetero_tag<__device_backend>>
+::std::enable_if_t<__is_random_access_iterator_v<_IteratorTypes...>, __hetero_tag<__device_backend_tag>>
 __select_backend(const execution::device_policy<_KernelName>&, _IteratorTypes&&...)
 {
     return {};
 }
 
 #if _ONEDPL_FPGA_DEVICE
-struct __fpga_backend : __device_backend
+struct __fpga_backend : __device_backend_tag
 {
 };
 
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
index 51854c08a1c..7a9ce585b7a 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
@@ -262,7 +262,7 @@ __parallel_for(_ExecutionPolicy&& __exec, _Fp __brick, _Index __count, _Ranges&&
 
 template <typename _ExecutionPolicy, typename _Fp, typename _Index, typename... _Ranges>
 auto
-__parallel_for(oneapi::dpl::__internal::__device_backend, _ExecutionPolicy&& __exec, _Fp __brick, _Index __count, _Ranges&&... __rngs)
+__parallel_for(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&& __exec, _Fp __brick, _Index __count, _Ranges&&... __rngs)
 {
     assert(__get_first_range_size(__rngs...) > 0);
 
@@ -1294,7 +1294,7 @@ __parallel_find(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last,
 
 template <typename _ExecutionPolicy, typename _Iterator, typename _Brick, typename _IsFirst>
 _Iterator
-__parallel_find(oneapi::dpl::__internal::__device_backend, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last, _Brick __f, _IsFirst)
+__parallel_find(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last, _Brick __f, _IsFirst)
 {
     auto __keep = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _Iterator>();
     auto __buf = __keep(__first, __last);
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
index 36d755f5734..374bd4dfc14 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
@@ -288,7 +288,7 @@ __parallel_find(oneapi::dpl::__internal::__fpga_backend __tag, _ExecutionPolicy&
     using _Policy = typename ::std::decay<_ExecutionPolicy>::type;
     using __kernel_name = typename _Policy::kernel_name;
     auto __device_policy = oneapi::dpl::execution::make_device_policy<__kernel_name>(__exec.queue());
-    return oneapi::dpl::__par_backend_hetero::__parallel_find(oneapi::dpl::__internal::__device_backend{}, __device_policy, __first, __last, __f, __is_first);
+    return oneapi::dpl::__par_backend_hetero::__parallel_find(oneapi::dpl::__internal::__device_backend_tag{}, __device_policy, __first, __last, __f, __is_first);
 }
 
 template <typename _ExecutionPolicy>

From 29ab80bb129fa4acb4aa188587ee942a40e5a28e Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <sergey.kopienko@intel.com>
Date: Wed, 25 Oct 2023 09:57:46 +0200
Subject: [PATCH 013/566] Rename __fpga_backend to __fpga_backend_tag

---
 include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h    | 4 ++--
 .../oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h | 4 ++--
 2 files changed, 4 insertions(+), 4 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h b/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h
index 6fe53805c28..8a0e03f33f9 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h
@@ -331,12 +331,12 @@ __select_backend(const execution::device_policy<_KernelName>&, _IteratorTypes&&.
 }
 
 #if _ONEDPL_FPGA_DEVICE
-struct __fpga_backend : __device_backend_tag
+struct __fpga_backend_tag : __device_backend_tag
 {
 };
 
 template <class... _IteratorTypes, unsigned int _Factor, typename _KernelName>
-::std::enable_if_t<__is_random_access_iterator_v<_IteratorTypes...>, __hetero_tag<__fpga_backend>>
+::std::enable_if_t<__is_random_access_iterator_v<_IteratorTypes...>, __hetero_tag<__fpga_backend_tag>>
 __select_backend(const execution::fpga_policy<_Factor, _KernelName>&, _IteratorTypes&&...)
 {
     return {};
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
index 374bd4dfc14..4389e04132e 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
@@ -90,7 +90,7 @@ __parallel_for(_ExecutionPolicy&& __exec, _Fp __brick, _Index __count, _Ranges&&
 
 template <typename _ExecutionPolicy, typename _Fp, typename _Index, typename... _Ranges>
 auto
-__parallel_for(oneapi::dpl::__internal::__fpga_backend, _ExecutionPolicy&& __exec, _Fp __brick, _Index __count, _Ranges&&... __rngs)
+__parallel_for(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionPolicy&& __exec, _Fp __brick, _Index __count, _Ranges&&... __rngs)
 {
     auto __n = __get_first_range_size(__rngs...);
     assert(__n > 0);
@@ -282,7 +282,7 @@ __parallel_find(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last,
 
 template <typename _ExecutionPolicy, typename _Iterator, typename _Brick, typename _IsFirst>
 _Iterator
-__parallel_find(oneapi::dpl::__internal::__fpga_backend __tag, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last, _Brick __f, _IsFirst __is_first)
+__parallel_find(oneapi::dpl::__internal::__fpga_backend_tag __tag, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last, _Brick __f, _IsFirst __is_first)
 {
     // workaround until we implement more performant version for patterns
     using _Policy = typename ::std::decay<_ExecutionPolicy>::type;

From d5f80ee21e49c60fa9d3955cf2c42af992fb3bf0 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <sergey.kopienko@intel.com>
Date: Wed, 25 Oct 2023 10:22:52 +0200
Subject: [PATCH 014/566] 
 include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h - Please add
 overload for operator() of the __parallel_for_submitter similar to what we
 have on line 231. It should take the __device_backend_tag as the first
 argument

---
 .../pstl/hetero/dpcpp/parallel_backend_sycl.h | 19 +++++++++++++++++++
 1 file changed, 19 insertions(+)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
index 7a9ce585b7a..453bbd50782 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
@@ -243,6 +243,25 @@ struct __parallel_for_submitter<__internal::__optional_kernel_name<_Name...>>
         });
         return __future(__event);
     }
+
+    template <typename _ExecutionPolicy, typename _Fp, typename _Index, typename... _Ranges>
+    auto
+    operator()(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&& __exec, _Fp __brick, _Index __count,
+               _Ranges&&... __rngs) const
+    {
+        assert(oneapi::dpl::__ranges::__get_first_range_size(__rngs...) > 0);
+        _PRINT_INFO_IN_DEBUG_MODE(__exec);
+        auto __event = __exec.queue().submit([&__rngs..., &__brick, __count](sycl::handler& __cgh) {
+            //get an access to data under SYCL buffer:
+            oneapi::dpl::__ranges::__require_access(__cgh, __rngs...);
+
+            __cgh.parallel_for<_Name...>(sycl::range</*dim=*/1>(__count), [=](sycl::item</*dim=*/1> __item_id) {
+                auto __idx = __item_id.get_linear_id();
+                __brick(__idx, __rngs...);
+            });
+        });
+        return __future(__event);
+    }
 };
 
 //General version of parallel_for, one additional parameter - __count of iterations of loop __cgh.parallel_for,

From 7e11f557cd463ffef285710f5642742b58e3a4a0 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <sergey.kopienko@intel.com>
Date: Wed, 25 Oct 2023 10:24:34 +0200
Subject: [PATCH 015/566] 
 include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h - Please add
 overload for __parallel_for similar to what we have on line 253. It should
 take the __device_backend_tag as the first argument

---
 .../pstl/hetero/dpcpp/parallel_backend_sycl.h | 23 ++++---------------
 1 file changed, 5 insertions(+), 18 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
index 453bbd50782..98a010635a7 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
@@ -283,26 +283,13 @@ template <typename _ExecutionPolicy, typename _Fp, typename _Index, typename...
 auto
 __parallel_for(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&& __exec, _Fp __brick, _Index __count, _Ranges&&... __rngs)
 {
-    assert(__get_first_range_size(__rngs...) > 0);
-
-    using _Policy = typename ::std::decay<_ExecutionPolicy>::type;
+    using _Policy = ::std::decay_t<_ExecutionPolicy>;
     using _CustomName = typename _Policy::kernel_name;
-    using _ForKernel =
-        oneapi::dpl::__par_backend_hetero::__internal::__kernel_name_generator<__parallel_for_kernel, _CustomName, _Fp,
-                                                                               _Ranges...>;
-
-    _PRINT_INFO_IN_DEBUG_MODE(__exec);
-    auto __event = __exec.queue().submit([&__rngs..., &__brick, __count](sycl::handler& __cgh) {
-        //get an access to data under SYCL buffer:
-        oneapi::dpl::__ranges::__require_access(__cgh, __rngs...);
-
-        __cgh.parallel_for<_ForKernel>(sycl::range</*dim=*/1>(__count), [=](sycl::item</*dim=*/1> __item_id) {
-            auto __idx = __item_id.get_linear_id();
-            __brick(__idx, __rngs...);
-        });
-    });
+    using _ForKernel = oneapi::dpl::__par_backend_hetero::__internal::__kernel_name_provider<_CustomName>;
 
-    return __future(__event);
+    return __parallel_for_submitter<_ForKernel>()(oneapi::dpl::__internal::__device_backend_tag{},
+                                                  ::std::forward<_ExecutionPolicy>(__exec), __brick, __count,
+                                                  ::std::forward<_Ranges>(__rngs)...);
 }
 
 //------------------------------------------------------------------------

From a9cf52cda88032286999c272165cf7863ec184b8 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <sergey.kopienko@intel.com>
Date: Wed, 25 Oct 2023 10:25:44 +0200
Subject: [PATCH 016/566] 
 include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h - remove class
 __parallel_for_kernel as not required anymore

---
 include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h | 3 ---
 1 file changed, 3 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
index 98a010635a7..7c8530c004c 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
@@ -180,9 +180,6 @@ make_iter_mode(const _Iterator& __it) -> decltype(iter_mode<outMode>()(__it))
 
 // set of class templates to name kernels
 
-template <typename... _Name>
-class __parallel_for_kernel;
-
 template <typename... _Name>
 class __scan_local_kernel;
 

From 1622733cefb4425cc174f90e66653d0575c45be8 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <sergey.kopienko@intel.com>
Date: Wed, 25 Oct 2023 09:25:17 +0200
Subject: [PATCH 017/566] Support FPGA backend. Adopt offload tags

---
 include/oneapi/dpl/pstl/execution_impl.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/include/oneapi/dpl/pstl/execution_impl.h b/include/oneapi/dpl/pstl/execution_impl.h
index 7e747db1223..2b37babfae2 100644
--- a/include/oneapi/dpl/pstl/execution_impl.h
+++ b/include/oneapi/dpl/pstl/execution_impl.h
@@ -132,7 +132,7 @@ __is_parallelization_preferred(_ExecutionPolicy& __exec)
 //------------------------------------------------------------------------
 
 struct __tbb_backend_tag {};
-// KSATODO required create tag for onedpl
+// KSATODO required create tag for dpcpp -> already implemented: __device_backend_tag, __fpga_backend_tag
 // KSATODO required create tag for omp
 
 template <class _IsVector>

From 94e25c05516235e7d47e161ba9d0ea9abd8721fb Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <sergey.kopienko@intel.com>
Date: Wed, 25 Oct 2023 11:25:11 +0200
Subject: [PATCH 018/566] include/oneapi/dpl/pstl/parallel_backend_tbb.h -
 remove #include "execution_impl.h" as not required

---
 include/oneapi/dpl/pstl/parallel_backend_tbb.h | 1 -
 1 file changed, 1 deletion(-)

diff --git a/include/oneapi/dpl/pstl/parallel_backend_tbb.h b/include/oneapi/dpl/pstl/parallel_backend_tbb.h
index d3663b2d301..3525e47ca5c 100644
--- a/include/oneapi/dpl/pstl/parallel_backend_tbb.h
+++ b/include/oneapi/dpl/pstl/parallel_backend_tbb.h
@@ -23,7 +23,6 @@
 #include <type_traits>
 
 #include "parallel_backend_utils.h"
-#include "execution_impl.h"
 
 // Bring in minimal required subset of Intel(R) Threading Building Blocks (Intel(R) TBB)
 #include <tbb/blocked_range.h>

From 0f2a4ec2dc568c8ec7ca71c8e7eb58128994e3c5 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <sergey.kopienko@intel.com>
Date: Wed, 25 Oct 2023 17:13:21 +0200
Subject: [PATCH 019/566] include/oneapi/dpl/pstl/parallel_impl.h - restore
 __parallel_find(_ExecutionPolicy&& __exec, _Index __first, _Index __last,
 _Brick __f, _IsFirst)

---
 include/oneapi/dpl/pstl/parallel_impl.h | 34 +++++++++++++++++++++++++
 1 file changed, 34 insertions(+)

diff --git a/include/oneapi/dpl/pstl/parallel_impl.h b/include/oneapi/dpl/pstl/parallel_impl.h
index cc92205b360..752db370260 100644
--- a/include/oneapi/dpl/pstl/parallel_impl.h
+++ b/include/oneapi/dpl/pstl/parallel_impl.h
@@ -32,6 +32,40 @@ namespace __internal
 //-----------------------------------------------------------------------
 /** Return extremum value returned by brick f[i,j) for subranges [i,j) of [first,last)
 Each f[i,j) must return a value in [i,j). */
+template <class _ExecutionPolicy, class _Index, class _Brick, class _IsFirst>
+_Index
+__parallel_find(_ExecutionPolicy&& __exec, _Index __first, _Index __last, _Brick __f, _IsFirst)
+{
+    typedef typename ::std::iterator_traits<_Index>::difference_type _DifferenceType;
+    const _DifferenceType __n = __last - __first;
+    _DifferenceType __initial_dist = _IsFirst::value ? __n : -1;
+
+    constexpr auto __comp = ::std::conditional_t<_IsFirst::value, __pstl_less, __pstl_greater>{};
+
+    ::std::atomic<_DifferenceType> __extremum(__initial_dist);
+    // TODO: find out what is better here: parallel_for or parallel_reduce
+    __par_backend::__parallel_for(::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+                                  [__comp, __f, __first, &__extremum](_Index __i, _Index __j) {
+                                      // See "Reducing Contention Through Priority Updates", PPoPP '13, for discussion of
+                                      // why using a shared variable scales fairly well in this situation.
+                                      if (__comp(__i - __first, __extremum))
+                                      {
+                                          _Index __res = __f(__i, __j);
+                                          // If not '__last' returned then we found what we want so put this to extremum
+                                          if (__res != __j)
+                                          {
+                                              const _DifferenceType __k = __res - __first;
+                                              for (_DifferenceType __old = __extremum; __comp(__k, __old);
+                                                   __old = __extremum)
+                                              {
+                                                  __extremum.compare_exchange_weak(__old, __k);
+                                              }
+                                          }
+                                      }
+                                  });
+    return __extremum != __initial_dist ? __first + __extremum : __last;
+}
+
 template <class _ExecutionPolicy, class _Index, class _Brick, class _IsFirst>
 _Index
 __parallel_find(__tbb_backend_tag __tag, _ExecutionPolicy&& __exec, _Index __first, _Index __last, _Brick __f, _IsFirst)

From 4d1fb6a456c721d406854e3c2cfe264db3cbd27a Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <sergey.kopienko@intel.com>
Date: Wed, 25 Oct 2023 20:03:58 +0200
Subject: [PATCH 020/566] include/oneapi/dpl/pstl/iterator_defs.h - fix compile
 error in tbb backend and for_loop.pass

---
 include/oneapi/dpl/pstl/iterator_defs.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/include/oneapi/dpl/pstl/iterator_defs.h b/include/oneapi/dpl/pstl/iterator_defs.h
index 524b3e887ff..5e190289540 100644
--- a/include/oneapi/dpl/pstl/iterator_defs.h
+++ b/include/oneapi/dpl/pstl/iterator_defs.h
@@ -60,7 +60,7 @@ __is_iterator_of(int)
 
 template <typename... _IteratorTypes>
 auto
-__is_needed_iter(...) -> ::std::false_type;
+__is_iterator_of(...) -> ::std::false_type;
 
 template <typename... _IteratorTypes>
 struct __is_random_access_iterator : decltype(__is_iterator_of<::std::random_access_iterator_tag, _IteratorTypes...>(0))

From e30b32575ed4ae5eee06a6129fa2e5b8cbf5df27 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <sergey.kopienko@intel.com>
Date: Wed, 25 Oct 2023 20:28:24 +0200
Subject: [PATCH 021/566] include/oneapi/dpl/pstl/glue_numeric_impl.h - fix
 compile error in test/general/header_inclusion_order_numeric_1.pass.cpp: no
 type named '__tbb_backend_tag' in namespace 'oneapi::dpl::__internal'

[1/2] Building CXX object test/CMakeFiles/header_inclusion_order_numeric_1.pass.dir/general/header_inclusion_order_numeric_1.pass.cpp.o
FAILED: test/CMakeFiles/header_inclusion_order_numeric_1.pass.dir/general/header_inclusion_order_numeric_1.pass.cpp.o
/usr/bin/oneapi_compiler/linux_prod/compiler/linux/bin/icpx -DPSTL_USE_DEBUG -DTBB_USE_DEBUG=1 -D_PSTL_TEST_SUCCESSFUL_KEYWORD=1 -I/home/oneDPL/test -I/home/oneDPL/include -isystem /usr/lib/oneapi-tbb-2021.5.0/include -fsycl -fsycl-device-code-split=per_kernel -ftemplate-backtrace-limit=0 -O0 -g -fno-fast-math -fopenmp-simd -fsycl -std=c++17 -MD -MT test/CMakeFiles/header_inclusion_order_numeric_1.pass.dir/general/header_inclusion_order_numeric_1.pass.cpp.o -MF test/CMakeFiles/header_inclusion_order_numeric_1.pass.dir/general/header_inclusion_order_numeric_1.pass.cpp.o.d -o test/CMakeFiles/header_inclusion_order_numeric_1.pass.dir/general/header_inclusion_order_numeric_1.pass.cpp.o -c /home/oneDPL/test/general/header_inclusion_order_numeric_1.pass.cpp
In file included from /home/oneDPL/test/general/header_inclusion_order_numeric_1.pass.cpp:19:
In file included from /home/oneDPL/include/oneapi/dpl/execution:47:
In file included from /home/oneDPL/include/oneapi/dpl/pstl/glue_numeric_impl.h:24:
In file included from /home/oneDPL/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h:22:
In file included from /home/oneDPL/include/oneapi/dpl/pstl/hetero/../parallel_backend.h:21:
/home/oneDPL/include/oneapi/dpl/pstl/parallel_backend_tbb.h:132:41: error: no type named '__tbb_backend_tag' in namespace 'oneapi::dpl::__internal'
__parallel_for(oneapi::dpl::__internal::__tbb_backend_tag, _ExecutionPolicy&&, _Index __first, _Index __last, _Fp __f)
               ~~~~~~~~~~~~~~~~~~~~~~~~~^
/home/oneDPL/include/oneapi/dpl/pstl/parallel_backend_tbb.h:1338:46: error: no type named '__tbb_backend_tag' in namespace 'oneapi::dpl::__internal'
__parallel_for_each(oneapi::dpl::__internal::__tbb_backend_tag, _ExecutionPolicy&&, _ForwardIterator __begin, _ForwardIterator __end, _Fp __f)
                    ~~~~~~~~~~~~~~~~~~~~~~~~~^
---
 include/oneapi/dpl/pstl/glue_numeric_impl.h | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_numeric_impl.h b/include/oneapi/dpl/pstl/glue_numeric_impl.h
index f2564db3132..f9347a5f7ac 100644
--- a/include/oneapi/dpl/pstl/glue_numeric_impl.h
+++ b/include/oneapi/dpl/pstl/glue_numeric_impl.h
@@ -20,14 +20,14 @@
 
 #include "utils.h"
 
+#include "numeric_fwd.h"
+#include "execution_impl.h"
+
 #if _ONEDPL_HETERO_BACKEND
 #    include "hetero/algorithm_impl_hetero.h"
 #    include "hetero/numeric_impl_hetero.h"
 #endif
 
-#include "numeric_fwd.h"
-#include "execution_impl.h"
-
 namespace oneapi
 {
 namespace dpl

From e104737c06ad38f0cc9279ed87510961c87c1d4d Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <sergey.kopienko@intel.com>
Date: Wed, 25 Oct 2023 20:32:53 +0200
Subject: [PATCH 022/566] Revert "include/oneapi/dpl/pstl/glue_numeric_impl.h -
 fix compile error in test/general/header_inclusion_order_numeric_1.pass.cpp:
 no type named '__tbb_backend_tag' in namespace 'oneapi::dpl::__internal'"

This reverts commit e30b32575ed4ae5eee06a6129fa2e5b8cbf5df27.
---
 include/oneapi/dpl/pstl/glue_numeric_impl.h | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_numeric_impl.h b/include/oneapi/dpl/pstl/glue_numeric_impl.h
index f9347a5f7ac..f2564db3132 100644
--- a/include/oneapi/dpl/pstl/glue_numeric_impl.h
+++ b/include/oneapi/dpl/pstl/glue_numeric_impl.h
@@ -20,14 +20,14 @@
 
 #include "utils.h"
 
-#include "numeric_fwd.h"
-#include "execution_impl.h"
-
 #if _ONEDPL_HETERO_BACKEND
 #    include "hetero/algorithm_impl_hetero.h"
 #    include "hetero/numeric_impl_hetero.h"
 #endif
 
+#include "numeric_fwd.h"
+#include "execution_impl.h"
+
 namespace oneapi
 {
 namespace dpl

From 014d544183631866c602e3288a01c75a5de44037 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <sergey.kopienko@intel.com>
Date: Wed, 25 Oct 2023 20:33:56 +0200
Subject: [PATCH 023/566] include/oneapi/dpl/pstl/parallel_backend_tbb.h - fix
 compile error in test/general/header_inclusion_order_numeric_1.pass.cpp: no
 type named '__tbb_backend_tag' in namespace 'oneapi::dpl::__internal'

[1/2] Building CXX object test/CMakeFiles/header_inclusion_order_numeric_1.pass.dir/general/header_inclusion_order_numeric_1.pass.cpp.o
FAILED: test/CMakeFiles/header_inclusion_order_numeric_1.pass.dir/general/header_inclusion_order_numeric_1.pass.cpp.o
/usr/bin/oneapi_compiler/linux_prod/compiler/linux/bin/icpx -DPSTL_USE_DEBUG -DTBB_USE_DEBUG=1 -D_PSTL_TEST_SUCCESSFUL_KEYWORD=1 -I/home/oneDPL/test -I/home/oneDPL/include -isystem /usr/lib/oneapi-tbb-2021.5.0/include -fsycl -fsycl-device-code-split=per_kernel -ftemplate-backtrace-limit=0 -O0 -g -fno-fast-math -fopenmp-simd -fsycl -std=c++17 -MD -MT test/CMakeFiles/header_inclusion_order_numeric_1.pass.dir/general/header_inclusion_order_numeric_1.pass.cpp.o -MF test/CMakeFiles/header_inclusion_order_numeric_1.pass.dir/general/header_inclusion_order_numeric_1.pass.cpp.o.d -o test/CMakeFiles/header_inclusion_order_numeric_1.pass.dir/general/header_inclusion_order_numeric_1.pass.cpp.o -c /home/oneDPL/test/general/header_inclusion_order_numeric_1.pass.cpp
In file included from /home/oneDPL/test/general/header_inclusion_order_numeric_1.pass.cpp:19:
In file included from /home/oneDPL/include/oneapi/dpl/execution:47:
In file included from /home/oneDPL/include/oneapi/dpl/pstl/glue_numeric_impl.h:24:
In file included from /home/oneDPL/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h:22:
In file included from /home/oneDPL/include/oneapi/dpl/pstl/hetero/../parallel_backend.h:21:
/home/oneDPL/include/oneapi/dpl/pstl/parallel_backend_tbb.h:132:41: error: no type named '__tbb_backend_tag' in namespace 'oneapi::dpl::__internal'
__parallel_for(oneapi::dpl::__internal::__tbb_backend_tag, _ExecutionPolicy&&, _Index __first, _Index __last, _Fp __f)
               ~~~~~~~~~~~~~~~~~~~~~~~~~^
/home/oneDPL/include/oneapi/dpl/pstl/parallel_backend_tbb.h:1338:46: error: no type named '__tbb_backend_tag' in namespace 'oneapi::dpl::__internal'
__parallel_for_each(oneapi::dpl::__internal::__tbb_backend_tag, _ExecutionPolicy&&, _ForwardIterator __begin, _ForwardIterator __end, _Fp __f)
                    ~~~~~~~~~~~~~~~~~~~~~~~~~^
---
 include/oneapi/dpl/pstl/parallel_backend_tbb.h | 1 +
 1 file changed, 1 insertion(+)

diff --git a/include/oneapi/dpl/pstl/parallel_backend_tbb.h b/include/oneapi/dpl/pstl/parallel_backend_tbb.h
index 3525e47ca5c..d3663b2d301 100644
--- a/include/oneapi/dpl/pstl/parallel_backend_tbb.h
+++ b/include/oneapi/dpl/pstl/parallel_backend_tbb.h
@@ -23,6 +23,7 @@
 #include <type_traits>
 
 #include "parallel_backend_utils.h"
+#include "execution_impl.h"
 
 // Bring in minimal required subset of Intel(R) Threading Building Blocks (Intel(R) TBB)
 #include <tbb/blocked_range.h>

From f557579e2c2078d7074b441a13a05e0e24c7d5f3 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <sergey.kopienko@intel.com>
Date: Wed, 25 Oct 2023 20:48:29 +0200
Subject: [PATCH 024/566] 
 include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h - fix error
 in __parallel_for(oneapi::dpl::__internal::__fpga_backend_tag,
 _ExecutionPolicy&& __exec, _Fp __brick, _Index __count, _Ranges&&... __rngs)

---
 .../hetero/dpcpp/parallel_backend_sycl_fpga.h | 57 +++++++++++--------
 1 file changed, 32 insertions(+), 25 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
index 4389e04132e..dc0aafae8d1 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
@@ -74,6 +74,33 @@ struct __parallel_for_fpga_submitter<__internal::__optional_kernel_name<_Name...
         });
         return __future(__event);
     }
+
+// KSATODO is this define check really required here?
+#if _ONEDPL_FPGA_DEVICE
+    template <typename _ExecutionPolicy, typename _Fp, typename _Index, typename... _Ranges>
+    auto
+    operator()(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionPolicy&& __exec, _Fp __brick, _Index __count,
+               _Ranges&&... __rngs) const
+    {
+        auto __n = oneapi::dpl::__ranges::__get_first_range_size(__rngs...);
+        assert(__n > 0);
+
+        _PRINT_INFO_IN_DEBUG_MODE(__exec);
+        auto __event = __exec.queue().submit([&__rngs..., &__brick, __count](sycl::handler& __cgh) {
+            //get an access to data under SYCL buffer:
+            oneapi::dpl::__ranges::__require_access(__cgh, __rngs...);
+
+            __cgh.single_task<_Name...>([=]() {
+#pragma unroll(::std::decay <_ExecutionPolicy>::type::unroll_factor)
+                for (auto __idx = 0; __idx < __count; ++__idx)
+                {
+                    __brick(__idx, __rngs...);
+                }
+            });
+        });
+        return __future(__event);
+    }
+#endif // _ONEDPL_FPGA_DEVICE
 };
 
 template <typename _ExecutionPolicy, typename _Fp, typename _Index, typename... _Ranges,
@@ -92,32 +119,12 @@ template <typename _ExecutionPolicy, typename _Fp, typename _Index, typename...
 auto
 __parallel_for(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionPolicy&& __exec, _Fp __brick, _Index __count, _Ranges&&... __rngs)
 {
-    auto __n = __get_first_range_size(__rngs...);
-    assert(__n > 0);
-
-    using _Policy = typename ::std::decay<_ExecutionPolicy>::type;
-    using __kernel_name = typename _Policy::kernel_name;
-#if __SYCL_UNNAMED_LAMBDA__
-    using __kernel_name_t = __parallel_for_kernel<_Fp, __kernel_name, _Ranges...>;
-#else
-    using __kernel_name_t = __parallel_for_kernel<__kernel_name>;
-#endif
-
-    _PRINT_INFO_IN_DEBUG_MODE(__exec);
-    auto __event = __exec.queue().submit([&__rngs..., &__brick, __count](sycl::handler& __cgh) {
-        //get an access to data under SYCL buffer:
-        oneapi::dpl::__ranges::__require_access(__cgh, __rngs...);
-
-        __cgh.single_task<__kernel_name_t>([=]() {
-#pragma unroll(::std::decay <_ExecutionPolicy>::type::unroll_factor)
-            for (auto __idx = 0; __idx < __count; ++__idx)
-            {
-                __brick(__idx, __rngs...);
-            }
-        });
-    });
+    using _Policy = ::std::decay_t<_ExecutionPolicy>;
+    using __parallel_for_name = __internal::__kernel_name_provider<typename _Policy::kernel_name>;
 
-    return __future(__event);
+    return __parallel_for_fpga_submitter<__parallel_for_name>()(oneapi::dpl::__internal::__fpga_backend_tag{},
+                                                                std::forward<_ExecutionPolicy>(__exec), __brick,
+                                                                __count, std::forward<_Ranges>(__rngs)...);
 }
 
 //------------------------------------------------------------------------

From c01e222d556ddc8b44958fc2fcdeba4afbe9bf1c Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <sergey.kopienko@intel.com>
Date: Wed, 25 Oct 2023 20:55:59 +0200
Subject: [PATCH 025/566] Apply clang-format from gitHUB

---
 include/oneapi/dpl/pstl/algorithm_impl.h      | 29 ++++++++++---------
 include/oneapi/dpl/pstl/execution_impl.h      | 16 +++++-----
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 15 +++++-----
 .../pstl/hetero/dpcpp/parallel_backend_sycl.h |  6 ++--
 .../hetero/dpcpp/parallel_backend_sycl_fpga.h | 11 ++++---
 .../oneapi/dpl/pstl/parallel_backend_tbb.h    |  3 +-
 include/oneapi/dpl/pstl/tuple_impl.h          |  3 +-
 7 files changed, 44 insertions(+), 39 deletions(-)

diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 125c8ca312f..5d395d48754 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -136,8 +136,7 @@ __brick_walk1(_DifferenceType __n, _Function __f, ::std::true_type) noexcept
 
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Function>
 void
-__pattern_walk1(_Tag, _ExecutionPolicy&&, _ForwardIterator __first,
-                _ForwardIterator __last, _Function __f) noexcept
+__pattern_walk1(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _Function __f) noexcept
 {
     __internal::__brick_walk1(__first, __last, __f, typename _Tag::__is_vector{});
 }
@@ -153,20 +152,21 @@ __pattern_walk1(_ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator _
 
 template <class _ExecutionPolicy, class _ForwardIterator, class _Function>
 void
-__pattern_walk1(__parallel_forward_tag, _ExecutionPolicy&& __exec,
-                _ForwardIterator __first, _ForwardIterator __last, _Function __f)
+__pattern_walk1(__parallel_forward_tag, _ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last,
+                _Function __f)
 {
     typedef typename ::std::iterator_traits<_ForwardIterator>::reference _ReferenceType;
     auto __func = [&__f](_ReferenceType arg) { __f(arg); };
     __internal::__except_handler([&]() {
-        __par_backend::__parallel_for_each(__parallel_forward_tag::__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __func);
+        __par_backend::__parallel_for_each(__parallel_forward_tag::__backend_tag{},
+                                           ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __func);
     });
 }
 
 template <class _IsVector, class _ExecutionPolicy, class _ForwardIterator, class _Function>
 void
-__pattern_walk1(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec,
-                _ForwardIterator __first, _ForwardIterator __last, _Function __f)
+__pattern_walk1(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _ForwardIterator __first,
+                _ForwardIterator __last, _Function __f)
 {
     using __backend_tag = typename decltype(__tag)::__backend_tag;
     __internal::__except_handler([&]() {
@@ -726,16 +726,17 @@ __pattern_find_if(_Tag __tag, _ExecutionPolicy&&, _ForwardIterator __first, _For
 
 template <class _IsVector, class _ExecutionPolicy, class _ForwardIterator, class _Predicate>
 _ForwardIterator
-__pattern_find_if(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec,
-                  _ForwardIterator __first, _ForwardIterator __last, _Predicate __pred)
+__pattern_find_if(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _ForwardIterator __first,
+                  _ForwardIterator __last, _Predicate __pred)
 {
     using __backend_tag = typename decltype(__tag)::__backend_tag;
     return __except_handler([&]() {
-        return __parallel_find(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
-                               [__pred](_ForwardIterator __i, _ForwardIterator __j) {
-                                   return __brick_find_if(__i, __j, __pred, _IsVector{});
-                               },
-                               ::std::true_type{});
+        return __parallel_find(
+            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+            [__pred](_ForwardIterator __i, _ForwardIterator __j) {
+                return __brick_find_if(__i, __j, __pred, _IsVector{});
+            },
+            ::std::true_type{});
     });
 }
 
diff --git a/include/oneapi/dpl/pstl/execution_impl.h b/include/oneapi/dpl/pstl/execution_impl.h
index 2b37babfae2..a187ddb50cf 100644
--- a/include/oneapi/dpl/pstl/execution_impl.h
+++ b/include/oneapi/dpl/pstl/execution_impl.h
@@ -131,7 +131,9 @@ __is_parallelization_preferred(_ExecutionPolicy& __exec)
 // backend selector with tags
 //------------------------------------------------------------------------
 
-struct __tbb_backend_tag {};
+struct __tbb_backend_tag
+{
+};
 // KSATODO required create tag for dpcpp -> already implemented: __device_backend_tag, __fpga_backend_tag
 // KSATODO required create tag for omp
 
@@ -159,14 +161,10 @@ struct __parallel_forward_tag
 };
 
 template <class _IsVector, class... _IteratorTypes>
-using __tag_type =
-    typename ::std::conditional<__internal::__is_random_access_iterator<_IteratorTypes...>::value,
-                                __parallel_tag<_IsVector>,
-                                typename ::std::conditional<__is_forward_iterator<_IteratorTypes...>::value,
-                                                            __parallel_forward_tag,
-                                                            __serial_tag<_IsVector>
-                                                            >::type
-                                >::type;
+using __tag_type = typename ::std::conditional<
+    __internal::__is_random_access_iterator<_IteratorTypes...>::value, __parallel_tag<_IsVector>,
+    typename ::std::conditional<__is_forward_iterator<_IteratorTypes...>::value, __parallel_forward_tag,
+                                __serial_tag<_IsVector>>::type>::type;
 
 template <class... _IteratorTypes>
 __serial_tag<std::false_type>
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 1ec77c0d5ea..8dda2bb2088 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -60,8 +60,8 @@ __pattern_walk1(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIte
 
 template <typename _BackendTag, typename _ExecutionPolicy, typename _ForwardIterator, typename _Function>
 void
-__pattern_walk1(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last,
-                _Function __f)
+__pattern_walk1(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _ForwardIterator __first,
+                _ForwardIterator __last, _Function __f)
 {
     auto __n = __last - __first;
     if (__n <= 0)
@@ -71,8 +71,8 @@ __pattern_walk1(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _For
         oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read_write, _ForwardIterator>();
     auto __buf = __keep(__first, __last);
 
-    oneapi::dpl::__par_backend_hetero::__parallel_for(_BackendTag{}, __exec, unseq_backend::walk_n<_ExecutionPolicy, _Function>{__f},
-                                                      __n, __buf.all_view())
+    oneapi::dpl::__par_backend_hetero::__parallel_for(
+        _BackendTag{}, __exec, unseq_backend::walk_n<_ExecutionPolicy, _Function>{__f}, __n, __buf.all_view())
         .wait();
 }
 
@@ -658,15 +658,16 @@ __pattern_equal(_ExecutionPolicy&& __exec, _Iterator1 __first1, _Iterator1 __las
 
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator, typename _Pred>
 _Iterator
-__pattern_find_if(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last, _Pred __pred)
+__pattern_find_if(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last,
+                  _Pred __pred)
 {
     if (__first == __last)
         return __last;
 
     using _Predicate = oneapi::dpl::unseq_backend::single_match_pred<_ExecutionPolicy, _Pred>;
 
-    return __par_backend_hetero::__parallel_find(_BackendTag{},
-        ::std::forward<_ExecutionPolicy>(__exec),
+    return __par_backend_hetero::__parallel_find(
+        _BackendTag{}, ::std::forward<_ExecutionPolicy>(__exec),
         __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__first),
         __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__last), _Predicate{__pred},
         ::std::true_type{});
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
index 7c8530c004c..fe1f6e4c202 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
@@ -278,7 +278,8 @@ __parallel_for(_ExecutionPolicy&& __exec, _Fp __brick, _Index __count, _Ranges&&
 
 template <typename _ExecutionPolicy, typename _Fp, typename _Index, typename... _Ranges>
 auto
-__parallel_for(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&& __exec, _Fp __brick, _Index __count, _Ranges&&... __rngs)
+__parallel_for(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&& __exec, _Fp __brick, _Index __count,
+               _Ranges&&... __rngs)
 {
     using _Policy = ::std::decay_t<_ExecutionPolicy>;
     using _CustomName = typename _Policy::kernel_name;
@@ -1297,7 +1298,8 @@ __parallel_find(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last,
 
 template <typename _ExecutionPolicy, typename _Iterator, typename _Brick, typename _IsFirst>
 _Iterator
-__parallel_find(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last, _Brick __f, _IsFirst)
+__parallel_find(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&& __exec, _Iterator __first,
+                _Iterator __last, _Brick __f, _IsFirst)
 {
     auto __keep = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _Iterator>();
     auto __buf = __keep(__first, __last);
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
index dc0aafae8d1..d9552030354 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
@@ -91,7 +91,7 @@ struct __parallel_for_fpga_submitter<__internal::__optional_kernel_name<_Name...
             oneapi::dpl::__ranges::__require_access(__cgh, __rngs...);
 
             __cgh.single_task<_Name...>([=]() {
-#pragma unroll(::std::decay <_ExecutionPolicy>::type::unroll_factor)
+#    pragma unroll(::std::decay <_ExecutionPolicy>::type::unroll_factor)
                 for (auto __idx = 0; __idx < __count; ++__idx)
                 {
                     __brick(__idx, __rngs...);
@@ -117,7 +117,8 @@ __parallel_for(_ExecutionPolicy&& __exec, _Fp __brick, _Index __count, _Ranges&&
 
 template <typename _ExecutionPolicy, typename _Fp, typename _Index, typename... _Ranges>
 auto
-__parallel_for(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionPolicy&& __exec, _Fp __brick, _Index __count, _Ranges&&... __rngs)
+__parallel_for(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionPolicy&& __exec, _Fp __brick, _Index __count,
+               _Ranges&&... __rngs)
 {
     using _Policy = ::std::decay_t<_ExecutionPolicy>;
     using __parallel_for_name = __internal::__kernel_name_provider<typename _Policy::kernel_name>;
@@ -289,13 +290,15 @@ __parallel_find(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last,
 
 template <typename _ExecutionPolicy, typename _Iterator, typename _Brick, typename _IsFirst>
 _Iterator
-__parallel_find(oneapi::dpl::__internal::__fpga_backend_tag __tag, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last, _Brick __f, _IsFirst __is_first)
+__parallel_find(oneapi::dpl::__internal::__fpga_backend_tag __tag, _ExecutionPolicy&& __exec, _Iterator __first,
+                _Iterator __last, _Brick __f, _IsFirst __is_first)
 {
     // workaround until we implement more performant version for patterns
     using _Policy = typename ::std::decay<_ExecutionPolicy>::type;
     using __kernel_name = typename _Policy::kernel_name;
     auto __device_policy = oneapi::dpl::execution::make_device_policy<__kernel_name>(__exec.queue());
-    return oneapi::dpl::__par_backend_hetero::__parallel_find(oneapi::dpl::__internal::__device_backend_tag{}, __device_policy, __first, __last, __f, __is_first);
+    return oneapi::dpl::__par_backend_hetero::__parallel_find(oneapi::dpl::__internal::__device_backend_tag{},
+                                                              __device_policy, __first, __last, __f, __is_first);
 }
 
 template <typename _ExecutionPolicy>
diff --git a/include/oneapi/dpl/pstl/parallel_backend_tbb.h b/include/oneapi/dpl/pstl/parallel_backend_tbb.h
index d3663b2d301..67db87432e5 100644
--- a/include/oneapi/dpl/pstl/parallel_backend_tbb.h
+++ b/include/oneapi/dpl/pstl/parallel_backend_tbb.h
@@ -1336,7 +1336,8 @@ __parallel_for_each(_ExecutionPolicy&&, _ForwardIterator __begin, _ForwardIterat
 
 template <class _ExecutionPolicy, class _ForwardIterator, class _Fp>
 void
-__parallel_for_each(oneapi::dpl::__internal::__tbb_backend_tag, _ExecutionPolicy&&, _ForwardIterator __begin, _ForwardIterator __end, _Fp __f)
+__parallel_for_each(oneapi::dpl::__internal::__tbb_backend_tag, _ExecutionPolicy&&, _ForwardIterator __begin,
+                    _ForwardIterator __end, _Fp __f)
 {
     tbb::this_task_arena::isolate([&]() { tbb::parallel_for_each(__begin, __end, __f); });
 }
diff --git a/include/oneapi/dpl/pstl/tuple_impl.h b/include/oneapi/dpl/pstl/tuple_impl.h
index e5dc3bdf859..370751a19ae 100644
--- a/include/oneapi/dpl/pstl/tuple_impl.h
+++ b/include/oneapi/dpl/pstl/tuple_impl.h
@@ -341,8 +341,7 @@ struct tuple<T1, T...>
     template <typename _U1, typename... _U,
               typename = ::std::enable_if_t<
                   (sizeof...(_U) == sizeof...(T) &&
-                   std::conjunction<::std::is_constructible<T1, _U1&&>,
-                                    ::std::is_constructible<T, _U&&>...>::value)>>
+                   std::conjunction<::std::is_constructible<T1, _U1&&>, ::std::is_constructible<T, _U&&>...>::value)>>
     tuple(_U1&& _value, _U&&... _next) : holder(::std::forward<_U1>(_value)), next(::std::forward<_U>(_next)...)
     {
     }

From 8cb876c2a8f4a856ee8f14c0a3f71a95cb0f6338 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <sergey.kopienko@intel.com>
Date: Fri, 27 Oct 2023 12:43:03 +0200
Subject: [PATCH 026/566] Code improvements in tag dispatch prototype for
 current code base (#1249)

---
 include/oneapi/dpl/pstl/execution_impl.h      |  7 ++-
 .../experimental/internal/for_loop_impl.h     | 17 +++----
 include/oneapi/dpl/pstl/iterator_defs.h       | 49 +++++++------------
 3 files changed, 28 insertions(+), 45 deletions(-)

diff --git a/include/oneapi/dpl/pstl/execution_impl.h b/include/oneapi/dpl/pstl/execution_impl.h
index a187ddb50cf..06e7aa4d675 100644
--- a/include/oneapi/dpl/pstl/execution_impl.h
+++ b/include/oneapi/dpl/pstl/execution_impl.h
@@ -161,10 +161,9 @@ struct __parallel_forward_tag
 };
 
 template <class _IsVector, class... _IteratorTypes>
-using __tag_type = typename ::std::conditional<
-    __internal::__is_random_access_iterator<_IteratorTypes...>::value, __parallel_tag<_IsVector>,
-    typename ::std::conditional<__is_forward_iterator<_IteratorTypes...>::value, __parallel_forward_tag,
-                                __serial_tag<_IsVector>>::type>::type;
+using __tag_type = ::std::conditional_t<
+    __internal::__is_random_access_iterator_v<_IteratorTypes...>, __parallel_tag<_IsVector>,
+    ::std::conditional_t<__is_forward_iterator_v<_IteratorTypes...>, __parallel_forward_tag, __serial_tag<_IsVector>>>;
 
 template <class... _IteratorTypes>
 __serial_tag<std::false_type>
diff --git a/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h b/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
index d7f738036c3..52163caf0f8 100644
--- a/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
+++ b/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
@@ -65,7 +65,7 @@ struct __difference<_Ip, ::std::enable_if_t<::std::is_integral_v<_Ip>>>
 template <typename _Ip>
 struct __difference<_Ip, ::std::enable_if_t<!::std::is_integral_v<_Ip>>>
 {
-    using __type = typename oneapi::dpl::__internal::__iterator_traits<_Ip>::difference_type;
+    using __type = typename ::std::iterator_traits<_Ip>::difference_type;
 };
 
 // This type is used as a stride value when it's known that stride == 1 at compile time(the case of for_loop and for_loop_n).
@@ -232,9 +232,9 @@ __pattern_for_loop(_ExecutionPolicy&& __exec, _Ip __first, _Ip __last, _Function
 }
 
 template <typename _Ip, typename _Function, typename _Sp, typename _Pack, typename _IndexType>
-::std::enable_if_t<::std::is_same_v<typename oneapi::dpl::__internal::__iterator_traits<_Ip>::iterator_category,
-                                    ::std::bidirectional_iterator_tag>,
-                   _IndexType>
+::std::enable_if_t<
+    ::std::is_same_v<typename ::std::iterator_traits<_Ip>::iterator_category, ::std::bidirectional_iterator_tag>,
+    _IndexType>
 __execute_loop_strided(_Ip __first, _Ip __last, _Function __f, _Sp __stride, _Pack& __pack, _IndexType) noexcept
 {
     _IndexType __ordinal_position = 0;
@@ -269,11 +269,10 @@ __execute_loop_strided(_Ip __first, _Ip __last, _Function __f, _Sp __stride, _Pa
 }
 
 template <typename _Ip, typename _Function, typename _Sp, typename _Pack, typename _IndexType>
-::std::enable_if_t<::std::is_same_v<typename oneapi::dpl::__internal::__iterator_traits<_Ip>::iterator_category,
-                                    ::std::forward_iterator_tag> ||
-                       ::std::is_same_v<typename oneapi::dpl::__internal::__iterator_traits<_Ip>::iterator_category,
-                                        ::std::input_iterator_tag>,
-                   _IndexType>
+::std::enable_if_t<
+    ::std::is_same_v<typename ::std::iterator_traits<_Ip>::iterator_category, ::std::forward_iterator_tag> ||
+        ::std::is_same_v<typename ::std::iterator_traits<_Ip>::iterator_category, ::std::input_iterator_tag>,
+    _IndexType>
 __execute_loop_strided(_Ip __first, _Ip __last, _Function __f, _Sp __stride, _Pack& __pack, _IndexType) noexcept
 {
     _IndexType __ordinal_position = 0;
diff --git a/include/oneapi/dpl/pstl/iterator_defs.h b/include/oneapi/dpl/pstl/iterator_defs.h
index 5e190289540..7106ee19a4e 100644
--- a/include/oneapi/dpl/pstl/iterator_defs.h
+++ b/include/oneapi/dpl/pstl/iterator_defs.h
@@ -27,36 +27,13 @@ namespace dpl
 namespace __internal
 {
 
-// Internal wrapper around ::std::iterator_traits as it is required to be
-// SFINAE-friendly(not produce "hard" error when _Ip is not an iterator)
-// only starting with C++17. Although many standard library implementations
-// provide it for older versions, we cannot rely on that.
-template <typename _Ip, typename = void>
-struct __iterator_traits
-{
-};
-
-template <typename _Ip>
-struct __iterator_traits<_Ip,
-                         __void_type<typename _Ip::iterator_category, typename _Ip::value_type,
-                                     typename _Ip::difference_type, typename _Ip::pointer, typename _Ip::reference>>
-    : ::std::iterator_traits<_Ip>
-{
-};
-
-// Handles _Tp* and const _Tp* specializations
-template <typename _Tp>
-struct __iterator_traits<_Tp*, void> : ::std::iterator_traits<_Tp*>
-{
-};
-
 // Make is_random_access_iterator and is_forward_iterator not to fail with a 'hard' error when it's used in
-//SFINAE with a non-iterator type by providing a default value.
+// SFINAE with a non-iterator type by providing a default value.
 template <typename _IteratorTag, typename... _IteratorTypes>
 auto
-__is_iterator_of(int)
-    -> decltype(std::conjunction<::std::is_base_of<_IteratorTag, typename __iterator_traits<typename ::std::decay<
-                                                                     _IteratorTypes>::type>::iterator_category>...>{});
+__is_iterator_of(int) -> decltype(
+    std::conjunction<::std::is_base_of<
+        _IteratorTag, typename ::std::iterator_traits<::std::decay_t<_IteratorTypes>>::iterator_category>...>{});
 
 template <typename... _IteratorTypes>
 auto
@@ -78,28 +55,36 @@ using __is_random_access_iterator_t = typename __is_random_access_iterator<_Iter
 template <typename... _IteratorTypes>
 inline constexpr bool __is_random_access_iterator_v = __is_random_access_iterator<_IteratorTypes...>::value;
 
+template <typename... _IteratorTypes>
+inline constexpr bool __is_forward_iterator_v = __is_forward_iterator<_IteratorTypes...>::value;
+
 // struct for checking if iterator is heterogeneous or not
-template <typename Iter, typename Void = void> // for non-heterogeneous iterators
+// for non-heterogeneous iterators
+template <typename Iter, typename Void = void>
 struct is_hetero_iterator : ::std::false_type
 {
 };
 
-template <typename Iter> // for heterogeneous iterators
+// for heterogeneous iterators
+template <typename Iter>
 struct is_hetero_iterator<Iter, ::std::enable_if_t<Iter::is_hetero::value>> : ::std::true_type
 {
 };
 // struct for checking if iterator should be passed directly to device or not
-template <typename Iter, typename Void = void> // for iterators that should not be passed directly
+// for iterators that should not be passed directly
+template <typename Iter, typename Void = void>
 struct is_passed_directly : ::std::false_type
 {
 };
 
-template <typename Iter> // for iterators defined as direct pass
+// for iterators defined as direct pass
+template <typename Iter>
 struct is_passed_directly<Iter, ::std::enable_if_t<Iter::is_passed_directly::value>> : ::std::true_type
 {
 };
 
-template <typename Iter> // for pointers to objects on device
+// for pointers to objects on device
+template <typename Iter>
 struct is_passed_directly<Iter, ::std::enable_if_t<::std::is_pointer_v<Iter>>> : ::std::true_type
 {
 };

From 7edb8736f4ed3e8509e5051d430dd5d8ba5d0edf Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <sergey.kopienko@intel.com>
Date: Fri, 27 Oct 2023 12:47:01 +0200
Subject: [PATCH 027/566] include/oneapi/dpl/pstl/glue_algorithm_impl.h - pass
 empty iterator instance into __select_backend insttead of __first into
 __select_backend call

---
 include/oneapi/dpl/pstl/glue_algorithm_impl.h | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_algorithm_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
index 86651375bee..15782092570 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
@@ -97,7 +97,7 @@ template <class _ExecutionPolicy, class _ForwardIterator, class _Predicate>
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _ForwardIterator>
 find_if(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last, _Predicate __pred)
 {
-    auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first);
+    auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, _ForwardIterator{});
 
     return oneapi::dpl::__internal::__pattern_find_if(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first,
                                                       __last, __pred);
@@ -363,7 +363,7 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy>
 replace_if(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last, _UnaryPredicate __pred,
            const _Tp& __new_value)
 {
-    auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first);
+    auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, _ForwardIterator{});
     __pattern_replace_if(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __pred,
                          __new_value);
 }

From fb11f70eb94b9ede8d15f0149aaeba396a772d63 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 7 Feb 2024 16:41:04 +0100
Subject: [PATCH 028/566] include/oneapi/dpl/pstl/execution_impl.h - remove
 KSATODO

---
 include/oneapi/dpl/pstl/execution_impl.h | 2 --
 1 file changed, 2 deletions(-)

diff --git a/include/oneapi/dpl/pstl/execution_impl.h b/include/oneapi/dpl/pstl/execution_impl.h
index 0d237bac733..8069da72b0f 100644
--- a/include/oneapi/dpl/pstl/execution_impl.h
+++ b/include/oneapi/dpl/pstl/execution_impl.h
@@ -131,8 +131,6 @@ __is_parallelization_preferred(_ExecutionPolicy& __exec)
 struct __tbb_backend_tag
 {
 };
-// KSATODO required create tag for dpcpp -> already implemented: __device_backend_tag, __fpga_backend_tag
-// KSATODO required create tag for omp
 
 template <class _IsVector>
 struct __serial_tag

From 959a410f5082f9d9497fef1050c2a3bb10bf79ed Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 8 Feb 2024 09:20:38 +0100
Subject: [PATCH 029/566] Implementation of ability to check __is_backend_tag,
 __is_backend_tag_serial_v, __is_backend_tag_v

---
 include/oneapi/dpl/pstl/execution_impl.h      | 36 +++++++++++++++++++
 .../pstl/hetero/dpcpp/execution_sycl_defs.h   |  9 +++++
 2 files changed, 45 insertions(+)

diff --git a/include/oneapi/dpl/pstl/execution_impl.h b/include/oneapi/dpl/pstl/execution_impl.h
index 8069da72b0f..af425473bde 100644
--- a/include/oneapi/dpl/pstl/execution_impl.h
+++ b/include/oneapi/dpl/pstl/execution_impl.h
@@ -188,6 +188,42 @@ __select_backend(oneapi::dpl::execution::parallel_unsequenced_policy, _IteratorT
     return {};
 }
 
+template <class _Tag>
+struct __is_backend_tag : ::std::false_type
+{
+};
+
+template <class _Tag>
+struct __is_backend_tag_serial : ::std::false_type
+{
+};
+
+template <class _IsVector>
+struct __is_backend_tag<__serial_tag<_IsVector>> : ::std::true_type
+{
+};
+
+template <class _IsVector>
+struct __is_backend_tag_serial<__serial_tag<_IsVector>> : ::std::true_type
+{
+};
+
+template <class _IsVector>
+struct __is_backend_tag<__parallel_tag<_IsVector>> : ::std::true_type
+{
+};
+
+template <>
+struct __is_backend_tag<__parallel_forward_tag> : ::std::true_type
+{
+};
+
+template <class _Tag>
+inline constexpr bool __is_backend_tag_serial_v = __is_backend_tag_serial<::std::decay_t<_Tag>>::value;
+
+template <class _Tag>
+inline constexpr bool __is_backend_tag_v = __is_backend_tag<::std::decay_t<_Tag>>::value;
+
 } // namespace __internal
 } // namespace dpl
 } // namespace oneapi
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h b/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h
index 15fcfbb39f4..8c91c597c31 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h
@@ -342,6 +342,15 @@ __select_backend(const execution::fpga_policy<_Factor, _KernelName>&, _IteratorT
 }
 #endif
 
+
+template <class _Tag>
+struct __is_backend_tag;
+
+template <typename _BackendTag>
+struct __is_backend_tag<__hetero_tag<_BackendTag>> : ::std::true_type
+{
+};
+
 } // namespace __internal
 
 } // namespace dpl

From 35e74492a732674cfbf96c873f84b83b5a30dead Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Mon, 29 Jan 2024 16:53:38 +0100
Subject: [PATCH 030/566] Declare __is_vectorization_preferred as constexpr
 without runtime params

---
 include/oneapi/dpl/pstl/algorithm_impl.h      | 173 +++++++++---------
 include/oneapi/dpl/pstl/execution_impl.h      |   8 +-
 .../experimental/internal/for_loop_impl.h     |  21 +--
 include/oneapi/dpl/pstl/glue_algorithm_impl.h | 140 ++++++--------
 include/oneapi/dpl/pstl/glue_memory_impl.h    |  54 +++---
 include/oneapi/dpl/pstl/glue_numeric_impl.h   |  20 +-
 6 files changed, 191 insertions(+), 225 deletions(-)

diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 3b3759964ad..7b02f71722d 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -223,7 +223,7 @@ oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>
 __pattern_walk_brick(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last, _Brick __brick,
                      /*parallel=*/::std::false_type) noexcept
 {
-    const auto __is_vector = __internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec);
+    constexpr auto __is_vector = __internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>();
     __brick(__first, __last, __is_vector);
 }
 
@@ -233,7 +233,7 @@ __pattern_walk_brick(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _
                      _Brick __brick,
                      /*parallel=*/::std::true_type)
 {
-    const auto __is_vector = __internal::__is_vectorization_preferred<_ExecutionPolicy, _RandomAccessIterator>(__exec);
+    constexpr auto __is_vector = __internal::__is_vectorization_preferred<_ExecutionPolicy, _RandomAccessIterator>();
     __internal::__except_handler([&]() {
         __par_backend::__parallel_for(::std::forward<_ExecutionPolicy>(__exec), __first, __last,
                                       [__brick, __is_vector](_RandomAccessIterator __i, _RandomAccessIterator __j) {
@@ -285,7 +285,7 @@ oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _Fo
 __pattern_walk_brick_n(_ExecutionPolicy&& __exec, _ForwardIterator __first, _Size __n, _Brick __brick,
                        /*is_parallel=*/::std::false_type) noexcept
 {
-    const auto __is_vector = __internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec);
+    constexpr auto __is_vector = __internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>();
     return __brick(__first, __n, __is_vector);
 }
 
@@ -294,7 +294,7 @@ oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _Ra
 __pattern_walk_brick_n(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _Size __n, _Brick __brick,
                        /*is_parallel=*/::std::true_type)
 {
-    const auto __is_vector = __internal::__is_vectorization_preferred<_ExecutionPolicy, _RandomAccessIterator>(__exec);
+    constexpr auto __is_vector = __internal::__is_vectorization_preferred<_ExecutionPolicy, _RandomAccessIterator>();
     return __internal::__except_handler([&]() {
         __par_backend::__parallel_for(::std::forward<_ExecutionPolicy>(__exec), __first, __first + __n,
                                       [__brick, __is_vector](_RandomAccessIterator __i, _RandomAccessIterator __j) {
@@ -423,8 +423,8 @@ oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _Fo
 __pattern_walk2_brick(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
                       _ForwardIterator2 __first2, _Brick __brick, /*parallel=*/::std::false_type) noexcept
 {
-    const auto __is_vector =
-        __internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(__exec);
+    constexpr auto __is_vector =
+        __internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
     return __brick(__first1, __last1, __first2, __is_vector);
 }
 
@@ -435,9 +435,8 @@ oneapi::dpl::__internal::__enable_if_host_execution_policy_conditional<
 __pattern_walk2_brick(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1, _RandomAccessIterator1 __last1,
                       _RandomAccessIterator2 __first2, _Brick __brick, /*parallel=*/::std::true_type)
 {
-    const auto __is_vector =
-        __internal::__is_vectorization_preferred<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2>(
-            __exec);
+    constexpr auto __is_vector =
+        __internal::__is_vectorization_preferred<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2>();
 
     return __except_handler([&]() {
         __par_backend::__parallel_for(
@@ -483,9 +482,8 @@ oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _Ra
 __pattern_walk2_brick_n(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1, _Size __n,
                         _RandomAccessIterator2 __first2, _Brick __brick, /*parallel=*/::std::true_type)
 {
-    const auto __is_vector =
-        __internal::__is_vectorization_preferred<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2>(
-            __exec);
+    constexpr auto __is_vector =
+        __internal::__is_vectorization_preferred<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2>();
 
     return __except_handler([&]() {
         __par_backend::__parallel_for(
@@ -502,8 +500,8 @@ oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _Fo
 __pattern_walk2_brick_n(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _Size __n, _ForwardIterator2 __first2,
                         _Brick __brick, /*parallel=*/::std::false_type) noexcept
 {
-    const auto __is_vector =
-        __internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(__exec);
+    constexpr auto __is_vector =
+        __internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
     return __brick(__first1, __n, __first2, __is_vector);
 }
 
@@ -909,14 +907,14 @@ __pattern_find_end(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first, _R
     else
     {
         return __internal::__except_handler([&]() {
-            return __internal::__parallel_find(::std::forward<_ExecutionPolicy>(__exec), __first, __last,
-                                               [__last, __s_first, __s_last, __pred,
-                                                __is_vector](_RandomAccessIterator1 __i, _RandomAccessIterator1 __j) {
-                                                   return __internal::__find_subrange(__i, __j, __last, __s_first,
-                                                                                      __s_last, __pred, false,
-                                                                                      __is_vector);
-                                               },
-                                               ::std::false_type{});
+            return __internal::__parallel_find(
+                ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+                [__last, __s_first, __s_last, __pred, __is_vector](_RandomAccessIterator1 __i,
+                                                                   _RandomAccessIterator1 __j) {
+                    return __internal::__find_subrange(__i, __j, __last, __s_first, __s_last, __pred, false,
+                                                       __is_vector);
+                },
+                ::std::false_type{});
         });
     }
 }
@@ -1013,14 +1011,14 @@ __pattern_search(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first, _Ran
     else
     {
         return __internal::__except_handler([&]() {
-            return __internal::__parallel_find(::std::forward<_ExecutionPolicy>(__exec), __first, __last,
-                                               [__last, __s_first, __s_last, __pred,
-                                                __is_vector](_RandomAccessIterator1 __i, _RandomAccessIterator1 __j) {
-                                                   return __internal::__find_subrange(__i, __j, __last, __s_first,
-                                                                                      __s_last, __pred, true,
-                                                                                      __is_vector);
-                                               },
-                                               ::std::true_type{});
+            return __internal::__parallel_find(
+                ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+                [__last, __s_first, __s_last, __pred, __is_vector](_RandomAccessIterator1 __i,
+                                                                   _RandomAccessIterator1 __j) {
+                    return __internal::__find_subrange(__i, __j, __last, __s_first, __s_last, __pred, true,
+                                                       __is_vector);
+                },
+                ::std::true_type{});
         });
     }
 }
@@ -1492,8 +1490,8 @@ __remove_elements(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardI
                     return __local_min;
                 }
                 // find first iterator that should be removed
-                bool* __result = __internal::__brick_find_if(__mask + __i, __mask + __j,
-                                                             [](bool __val) { return !__val; }, __is_vector);
+                bool* __result = __internal::__brick_find_if(
+                    __mask + __i, __mask + __j, [](bool __val) { return !__val; }, __is_vector);
                 if (__result - __mask == __j)
                 {
                     return __local_min;
@@ -1520,8 +1518,8 @@ __remove_elements(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardI
         __par_backend::__parallel_strict_scan(
             ::std::forward<_ExecutionPolicy>(__exec), __n, _DifferenceType(0),
             [__mask, __is_vector](_DifferenceType __i, _DifferenceType __len) {
-                return __internal::__brick_count(__mask + __i, __mask + __i + __len, [](bool __val) { return __val; },
-                                                 __is_vector);
+                return __internal::__brick_count(
+                    __mask + __i, __mask + __i + __len, [](bool __val) { return __val; }, __is_vector);
             },
             ::std::plus<_DifferenceType>(),
             [=](_DifferenceType __i, _DifferenceType __len, _DifferenceType __initial) {
@@ -2405,10 +2403,12 @@ __pattern_sort(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _Random
                _IsVector /*is_vector*/, /*is_parallel=*/::std::true_type, /*is_move_constructible=*/::std::true_type)
 {
     __internal::__except_handler([&]() {
-        __par_backend::__parallel_stable_sort(::std::forward<_ExecutionPolicy>(__exec), __first, __last, __comp,
-                                              [](_RandomAccessIterator __first, _RandomAccessIterator __last,
-                                                 _Compare __comp) { ::std::sort(__first, __last, __comp); },
-                                              __last - __first);
+        __par_backend::__parallel_stable_sort(
+            ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __comp,
+            [](_RandomAccessIterator __first, _RandomAccessIterator __last, _Compare __comp) {
+                ::std::sort(__first, __last, __comp);
+            },
+            __last - __first);
     });
 }
 
@@ -2430,10 +2430,12 @@ __pattern_stable_sort(_ExecutionPolicy&& __exec, _RandomAccessIterator __first,
                       _Compare __comp, _IsVector /*is_vector*/, /*is_parallel=*/::std::true_type)
 {
     __internal::__except_handler([&]() {
-        __par_backend::__parallel_stable_sort(::std::forward<_ExecutionPolicy>(__exec), __first, __last, __comp,
-                                              [](_RandomAccessIterator __first, _RandomAccessIterator __last,
-                                                 _Compare __comp) { ::std::stable_sort(__first, __last, __comp); },
-                                              __last - __first);
+        __par_backend::__parallel_stable_sort(
+            ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __comp,
+            [](_RandomAccessIterator __first, _RandomAccessIterator __last, _Compare __comp) {
+                ::std::stable_sort(__first, __last, __comp);
+            },
+            __last - __first);
     });
 }
 
@@ -2450,8 +2452,9 @@ __pattern_sort_by_key(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __keys_f
 {
     auto __beg = oneapi::dpl::make_zip_iterator(__keys_first, __values_first);
     auto __end = __beg + (__keys_last - __keys_first);
-    auto __cmp_f =
-        [__comp](const auto& __a, const auto& __b) { return __comp(::std::get<0>(__a), ::std::get<0>(__b)); };
+    auto __cmp_f = [__comp](const auto& __a, const auto& __b) {
+        return __comp(::std::get<0>(__a), ::std::get<0>(__b));
+    };
 
     ::std::sort(__beg, __end, __cmp_f);
 }
@@ -2463,19 +2466,21 @@ __pattern_sort_by_key(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __keys_f
                       _RandomAccessIterator1 __keys_last, _RandomAccessIterator2 __values_first, _Compare __comp,
                       _IsVector /*vector=*/, /*is_parallel=*/::std::true_type)
 {
-    static_assert(::std::is_move_constructible_v<typename ::std::iterator_traits<_RandomAccessIterator1>::value_type>
-        && ::std::is_move_constructible_v<typename ::std::iterator_traits<_RandomAccessIterator2>::value_type>,
+    static_assert(
+        ::std::is_move_constructible_v<typename ::std::iterator_traits<_RandomAccessIterator1>::value_type> &&
+            ::std::is_move_constructible_v<typename ::std::iterator_traits<_RandomAccessIterator2>::value_type>,
         "The keys and values should be move constructible in case of parallel execution.");
 
     auto __beg = oneapi::dpl::make_zip_iterator(__keys_first, __values_first);
     auto __end = __beg + (__keys_last - __keys_first);
-    auto __cmp_f =
-        [__comp](const auto& __a, const auto& __b) { return __comp(::std::get<0>(__a), ::std::get<0>(__b)); };
+    auto __cmp_f = [__comp](const auto& __a, const auto& __b) {
+        return __comp(::std::get<0>(__a), ::std::get<0>(__b));
+    };
 
     __internal::__except_handler([&]() {
-        __par_backend::__parallel_stable_sort(::std::forward<_ExecutionPolicy>(__exec), __beg, __end, __cmp_f,
-                                              [](auto __first, auto __last, auto __cmp)
-                                                { ::std::sort(__first, __last, __cmp); },__end - __beg);
+        __par_backend::__parallel_stable_sort(
+            ::std::forward<_ExecutionPolicy>(__exec), __beg, __end, __cmp_f,
+            [](auto __first, auto __last, auto __cmp) { ::std::sort(__first, __last, __cmp); }, __end - __beg);
     });
 }
 
@@ -2565,23 +2570,24 @@ __pattern_partial_sort_copy(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __
             __par_backend::__buffer<_ExecutionPolicy, _T1> __buf(__n1);
             _T1* __r = __buf.get();
 
-            __par_backend::__parallel_stable_sort(::std::forward<_ExecutionPolicy>(__exec), __r, __r + __n1, __comp,
-                                                  [__n2, __first, __r](_T1* __i, _T1* __j, _Compare __comp) {
-                                                      _RandomAccessIterator1 __it = __first + (__i - __r);
+            __par_backend::__parallel_stable_sort(
+                ::std::forward<_ExecutionPolicy>(__exec), __r, __r + __n1, __comp,
+                [__n2, __first, __r](_T1* __i, _T1* __j, _Compare __comp) {
+                    _RandomAccessIterator1 __it = __first + (__i - __r);
 
-                                                      // 1. Copy elements from input to raw memory
-                                                      for (_T1* __k = __i; __k != __j; ++__k, ++__it)
-                                                      {
-                                                          ::new (__k) _T2(*__it);
-                                                      }
+                    // 1. Copy elements from input to raw memory
+                    for (_T1* __k = __i; __k != __j; ++__k, ++__it)
+                    {
+                        ::new (__k) _T2(*__it);
+                    }
 
-                                                      // 2. Sort elements in temporary buffer
-                                                      if (__n2 < __j - __i)
-                                                          ::std::partial_sort(__i, __i + __n2, __j, __comp);
-                                                      else
-                                                          ::std::sort(__i, __j, __comp);
-                                                  },
-                                                  __n2);
+                    // 2. Sort elements in temporary buffer
+                    if (__n2 < __j - __i)
+                        ::std::partial_sort(__i, __i + __n2, __j, __comp);
+                    else
+                        ::std::sort(__i, __j, __comp);
+                },
+                __n2);
 
             // 3. Move elements from temporary buffer to output
             __par_backend::__parallel_for(::std::forward<_ExecutionPolicy>(__exec), __r, __r + __n2,
@@ -2701,10 +2707,10 @@ __pattern_nth_element(_ExecutionPolicy&& __exec, _RandomAccessIterator __first,
     _RandomAccessIterator __x;
     do
     {
-        __x = __internal::__pattern_partition(::std::forward<_ExecutionPolicy>(__exec), __first + 1, __last,
-                                              [&__comp, __first](const _Tp& __x) { return __comp(__x, *__first); },
-                                              __is_vector,
-                                              /*is_parallel=*/::std::true_type());
+        __x = __internal::__pattern_partition(
+            ::std::forward<_ExecutionPolicy>(__exec), __first + 1, __last,
+            [&__comp, __first](const _Tp& __x) { return __comp(__x, *__first); }, __is_vector,
+            /*is_parallel=*/::std::true_type());
         --__x;
         if (__x != __first)
         {
@@ -2942,8 +2948,8 @@ __pattern_remove_if(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _R
     return __internal::__remove_elements(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
         [&__pred, __is_vector](bool* __b, bool* __e, _RandomAccessIterator __it) {
-            __internal::__brick_walk2(__b, __e, __it, [&__pred](bool& __x, _ReferenceType __y) { __x = !__pred(__y); },
-                                      __is_vector);
+            __internal::__brick_walk2(
+                __b, __e, __it, [&__pred](bool& __x, _ReferenceType __y) { __x = !__pred(__y); }, __is_vector);
         },
         __is_vector);
 }
@@ -3722,12 +3728,12 @@ __pattern_is_heap_until(_ExecutionPolicy&& __exec, _RandomAccessIterator __first
                         _Compare __comp, _IsVector __is_vector, /* is_parallel = */ ::std::true_type)
 {
     return __internal::__except_handler([&]() {
-        return __parallel_find(::std::forward<_ExecutionPolicy>(__exec), __first, __last,
-                               [__first, __comp, __is_vector](_RandomAccessIterator __i, _RandomAccessIterator __j) {
-                                   return __internal::__is_heap_until_local(__first, __i - __first, __j - __first,
-                                                                            __comp, __is_vector);
-                               },
-                               ::std::true_type{});
+        return __parallel_find(
+            ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+            [__first, __comp, __is_vector](_RandomAccessIterator __i, _RandomAccessIterator __j) {
+                return __internal::__is_heap_until_local(__first, __i - __first, __j - __first, __comp, __is_vector);
+            },
+            ::std::true_type{});
     });
 }
 
@@ -4077,11 +4083,12 @@ __pattern_lexicographical_compare(_ExecutionPolicy&& __exec, _RandomAccessIterat
         auto __result = __internal::__parallel_find(
             ::std::forward<_ExecutionPolicy>(__exec), __first1, __first1 + __n,
             [__first1, __first2, &__comp, __is_vector](_RandomAccessIterator1 __i, _RandomAccessIterator1 __j) {
-                return __internal::__brick_mismatch(__i, __j, __first2 + (__i - __first1), __first2 + (__j - __first1),
-                                                    [&__comp](const _RefType1 __x, const _RefType2 __y) {
-                                                        return !__comp(__x, __y) && !__comp(__y, __x);
-                                                    },
-                                                    __is_vector)
+                return __internal::__brick_mismatch(
+                           __i, __j, __first2 + (__i - __first1), __first2 + (__j - __first1),
+                           [&__comp](const _RefType1 __x, const _RefType2 __y) {
+                               return !__comp(__x, __y) && !__comp(__y, __x);
+                           },
+                           __is_vector)
                     .first;
             },
             ::std::true_type{});
diff --git a/include/oneapi/dpl/pstl/execution_impl.h b/include/oneapi/dpl/pstl/execution_impl.h
index af425473bde..828f703333d 100644
--- a/include/oneapi/dpl/pstl/execution_impl.h
+++ b/include/oneapi/dpl/pstl/execution_impl.h
@@ -105,12 +105,10 @@ template <typename _ExecutionPolicy>
 using __allow_parallel = typename __internal::__policy_traits<::std::decay_t<_ExecutionPolicy>>::__allow_parallel;
 
 template <typename _ExecutionPolicy, typename... _IteratorTypes>
-auto
-__is_vectorization_preferred(_ExecutionPolicy& __exec)
-    -> decltype(__internal::__lazy_and(__exec.__allow_vector(),
-                                       __internal::__is_random_access_iterator_t<_IteratorTypes...>()))
+constexpr decltype(auto)
+__is_vectorization_preferred()
 {
-    return __internal::__lazy_and(__exec.__allow_vector(),
+    return __internal::__lazy_and(::std::decay_t<_ExecutionPolicy>::__allow_vector(),
                                   __internal::__is_random_access_iterator_t<_IteratorTypes...>());
 }
 
diff --git a/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h b/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
index 52163caf0f8..d4827c11ea4 100644
--- a/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
+++ b/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
@@ -479,9 +479,9 @@ struct __use_par_vec_helper<_Ip, ::std::enable_if_t<::std::is_integral_v<_Ip>>>
 {
     template <typename _ExecutionPolicy>
     static constexpr auto
-    __use_vector(_ExecutionPolicy&& __exec) -> decltype(__exec.__allow_vector())
+    __use_vector()
     {
-        return __exec.__allow_vector();
+        return ::std::decay_t<_ExecutionPolicy>::__allow_vector();
     }
 
     template <typename _ExecutionPolicy>
@@ -497,12 +497,9 @@ struct __use_par_vec_helper<_Ip, ::std::enable_if_t<!::std::is_integral_v<_Ip>>>
 {
     template <typename _ExecutionPolicy>
     static constexpr auto
-    __use_vector(_ExecutionPolicy&& __exec)
-        -> decltype(oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _Ip>(
-            ::std::forward<_ExecutionPolicy>(__exec)))
+    __use_vector()
     {
-        return oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _Ip>(
-            ::std::forward<_ExecutionPolicy>(__exec));
+        return oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _Ip>();
     }
 
     template <typename _ExecutionPolicy>
@@ -519,10 +516,10 @@ struct __use_par_vec_helper<_Ip, ::std::enable_if_t<!::std::is_integral_v<_Ip>>>
 // Special versions for for_loop: handles both iterators and integral types(treated as random access iterators)
 template <typename _ExecutionPolicy, typename _Ip>
 auto
-__use_vectorization(_ExecutionPolicy&& __exec)
-    -> decltype(__use_par_vec_helper<_Ip>::__use_vector(::std::forward<_ExecutionPolicy>(__exec)))
+__use_vectorization()
+    -> decltype(__use_par_vec_helper<_Ip>::template __use_vector<_ExecutionPolicy>())
 {
-    return __use_par_vec_helper<_Ip>::__use_vector(::std::forward<_ExecutionPolicy>(__exec));
+    return __use_par_vec_helper<_Ip>::template __use_vector<_ExecutionPolicy>();
 }
 
 template <typename _ExecutionPolicy, typename _Ip>
@@ -541,7 +538,7 @@ __for_loop_impl(_ExecutionPolicy&& __exec, _Ip __start, _Ip __finish, _Fp&& __f,
 {
     oneapi::dpl::__internal::__pattern_for_loop(
         ::std::forward<_ExecutionPolicy>(__exec), __start, __finish, __f, __stride,
-        oneapi::dpl::__internal::__use_vectorization<_ExecutionPolicy, _Ip>(__exec),
+        oneapi::dpl::__internal::__use_vectorization<_ExecutionPolicy, _Ip>(),
         oneapi::dpl::__internal::__use_parallelization<_ExecutionPolicy, _Ip>(__exec),
         ::std::get<_Is>(::std::move(__t))...);
 }
@@ -554,7 +551,7 @@ __for_loop_n_impl(_ExecutionPolicy&& __exec, _Ip __start, _Size __n, _Fp&& __f,
 {
     oneapi::dpl::__internal::__pattern_for_loop_n(
         ::std::forward<_ExecutionPolicy>(__exec), __start, __n, __f, __stride,
-        oneapi::dpl::__internal::__use_vectorization<_ExecutionPolicy, _Ip>(__exec),
+        oneapi::dpl::__internal::__use_vectorization<_ExecutionPolicy, _Ip>(),
         oneapi::dpl::__internal::__use_parallelization<_ExecutionPolicy, _Ip>(__exec),
         ::std::get<_Is>(::std::move(__t))...);
 }
diff --git a/include/oneapi/dpl/pstl/glue_algorithm_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
index b7164977ff4..ac1ff531bce 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
@@ -45,7 +45,7 @@ any_of(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __l
 {
     return oneapi::dpl::__internal::__pattern_any_of(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __pred,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec),
+        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(),
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec));
 }
 
@@ -77,7 +77,7 @@ for_each(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator _
 {
     oneapi::dpl::__internal::__pattern_walk1(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __f,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec),
+        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(),
         __exec.__allow_parallel());
 }
 
@@ -87,7 +87,7 @@ for_each_n(_ExecutionPolicy&& __exec, _ForwardIterator __first, _Size __n, _Func
 {
     return oneapi::dpl::__internal::__pattern_walk1_n(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __n, __f,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec),
+        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(),
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec));
 }
 
@@ -131,8 +131,7 @@ find_end(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _ForwardIterator1
 {
     return oneapi::dpl::__internal::__pattern_find_end(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __s_first, __s_last, __pred,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(
-            __exec),
+        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(),
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(
             __exec));
 }
@@ -154,8 +153,7 @@ find_first_of(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _ForwardIter
 {
     return oneapi::dpl::__internal::__pattern_find_first_of(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __s_first, __s_last, __pred,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(
-            __exec),
+        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(),
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(
             __exec));
 }
@@ -178,7 +176,7 @@ adjacent_find(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardItera
     return oneapi::dpl::__internal::__pattern_adjacent_find(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last, ::std::equal_to<_ValueType>(),
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec),
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec),
+        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(),
         oneapi::dpl::__internal::__first_semantic());
 }
 
@@ -189,7 +187,7 @@ adjacent_find(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardItera
     return oneapi::dpl::__internal::__pattern_adjacent_find(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __pred,
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec),
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec),
+        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(),
         oneapi::dpl::__internal::__first_semantic());
 }
 
@@ -208,7 +206,7 @@ count(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __la
         oneapi::dpl::__internal::__equal_value<oneapi::dpl::__internal::__ref_or_copy<_ExecutionPolicy, const _Tp>>(
             __value),
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec),
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec));
+        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>());
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator, class _Predicate>
@@ -219,7 +217,7 @@ count_if(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator _
     return oneapi::dpl::__internal::__pattern_count(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __pred,
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec),
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec));
+        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>());
 }
 
 // [alg.search]
@@ -231,8 +229,7 @@ search(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _ForwardIterator1 _
 {
     return oneapi::dpl::__internal::__pattern_search(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __s_first, __s_last, __pred,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(
-            __exec),
+        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(),
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(
             __exec));
 }
@@ -253,7 +250,7 @@ search_n(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator _
 {
     return oneapi::dpl::__internal::__pattern_search_n(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __count, __value, __pred,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec),
+        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(),
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec));
 }
 
@@ -299,8 +296,7 @@ copy_if(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _ForwardIterator1
 {
     return oneapi::dpl::__internal::__pattern_copy_if(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result, __pred,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(
-            __exec),
+        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(),
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(
             __exec));
 }
@@ -320,8 +316,7 @@ swap_ranges(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardItera
             using ::std::swap;
             swap(__x, __y);
         },
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(
-            __exec),
+        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(),
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(
             __exec));
 }
@@ -336,8 +331,7 @@ transform(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _ForwardIterator
     return oneapi::dpl::__internal::__pattern_walk2(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result,
         oneapi::dpl::__internal::__transform_functor<_UnaryOperation>{::std::move(__op)},
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(
-            __exec),
+        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(),
         __exec.__allow_parallel());
 }
 
@@ -351,8 +345,7 @@ transform(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterato
     return oneapi::dpl::__internal::__pattern_walk3(
         ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __result,
         oneapi::dpl::__internal::__transform_functor<_BinaryOperation>(::std::move(__op)),
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2,
-                                                              _ForwardIterator>(__exec),
+        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2, _ForwardIterator>(),
         __exec.__allow_parallel());
 }
 
@@ -368,8 +361,7 @@ transform_if(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _ForwardItera
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result,
         oneapi::dpl::__internal::__transform_if_unary_functor<_UnaryOperation, _UnaryPredicate>(::std::move(__op),
                                                                                                 ::std::move(__pred)),
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(
-            __exec),
+        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(),
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(
             __exec));
 }
@@ -384,8 +376,7 @@ transform_if(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIter
         ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __result,
         oneapi::dpl::__internal::__transform_if_binary_functor<_BinaryOperation, _BinaryPredicate>(::std::move(__op),
                                                                                                    ::std::move(__pred)),
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2,
-                                                              _ForwardIterator3>(__exec),
+        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2, _ForwardIterator3>(),
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2,
                                                                 _ForwardIterator3>(__exec));
 }
@@ -427,8 +418,7 @@ replace_copy_if(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _ForwardIt
                                  _UnaryPredicate,
                                  oneapi::dpl::__internal::__ref_or_copy<_ExecutionPolicy, _UnaryPredicate>>>(
             __new_value, __pred),
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(
-            __exec),
+        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(),
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(
             __exec));
 }
@@ -454,7 +444,7 @@ fill(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __las
     oneapi::dpl::__internal::__pattern_fill(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __value,
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec),
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec));
+        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>());
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator, class _Size, class _Tp>
@@ -467,7 +457,7 @@ fill_n(_ExecutionPolicy&& __exec, _ForwardIterator __first, _Size __count, const
     return oneapi::dpl::__internal::__pattern_fill_n(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __count, __value,
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec),
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec));
+        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>());
 }
 
 // [alg.generate]
@@ -478,7 +468,7 @@ generate(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator _
     oneapi::dpl::__internal::__pattern_generate(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __g,
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec),
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec));
+        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>());
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator, class _Size, class _Generator>
@@ -491,7 +481,7 @@ generate_n(_ExecutionPolicy&& __exec, _ForwardIterator __first, _Size __count, _
     return oneapi::dpl::__internal::__pattern_generate_n(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __count, __g,
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec),
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec));
+        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>());
 }
 
 // [alg.remove]
@@ -524,7 +514,7 @@ remove_if(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator
 {
     return oneapi::dpl::__internal::__pattern_remove_if(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __pred,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec),
+        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(),
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec));
 }
 
@@ -546,7 +536,7 @@ unique(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __l
 {
     return oneapi::dpl::__internal::__pattern_unique(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __pred,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec),
+        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(),
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec));
 }
 
@@ -565,8 +555,7 @@ unique_copy(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _ForwardIterat
 {
     return oneapi::dpl::__internal::__pattern_unique_copy(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result, __pred,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(
-            __exec),
+        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(),
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(
             __exec));
 }
@@ -587,7 +576,7 @@ reverse(_ExecutionPolicy&& __exec, _BidirectionalIterator __first, _Bidirectiona
 {
     oneapi::dpl::__internal::__pattern_reverse(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _BidirectionalIterator>(__exec),
+        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _BidirectionalIterator>(),
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _BidirectionalIterator>(__exec));
 }
 
@@ -598,8 +587,7 @@ reverse_copy(_ExecutionPolicy&& __exec, _BidirectionalIterator __first, _Bidirec
 {
     return oneapi::dpl::__internal::__pattern_reverse_copy(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __d_first,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _BidirectionalIterator,
-                                                              _ForwardIterator>(__exec),
+        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _BidirectionalIterator, _ForwardIterator>(),
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _BidirectionalIterator,
                                                                 _ForwardIterator>(__exec));
 }
@@ -612,7 +600,7 @@ rotate(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __m
 {
     return oneapi::dpl::__internal::__pattern_rotate(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __middle, __last,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec),
+        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(),
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec));
 }
 
@@ -623,8 +611,7 @@ rotate_copy(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _ForwardIterat
 {
     return oneapi::dpl::__internal::__pattern_rotate_copy(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __middle, __last, __result,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(
-            __exec),
+        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(),
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(
             __exec));
 }
@@ -637,7 +624,7 @@ is_partitioned(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIter
 {
     return oneapi::dpl::__internal::__pattern_is_partitioned(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __pred,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec),
+        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(),
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec));
 }
 
@@ -647,7 +634,7 @@ partition(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator
 {
     return oneapi::dpl::__internal::__pattern_partition(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __pred,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec),
+        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(),
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec));
 }
 
@@ -658,7 +645,7 @@ stable_partition(_ExecutionPolicy&& __exec, _BidirectionalIterator __first, _Bid
 {
     return oneapi::dpl::__internal::__pattern_stable_partition(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __pred,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _BidirectionalIterator>(__exec),
+        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _BidirectionalIterator>(),
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _BidirectionalIterator>(__exec));
 }
 
@@ -671,8 +658,7 @@ partition_copy(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIter
 {
     return oneapi::dpl::__internal::__pattern_partition_copy(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __out_true, __out_false, __pred,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator, _ForwardIterator1,
-                                                              _ForwardIterator2>(__exec),
+        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator, _ForwardIterator1, _ForwardIterator2>(),
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator, _ForwardIterator1,
                                                                 _ForwardIterator2>(__exec));
 }
@@ -686,7 +672,7 @@ sort(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIter
     typedef typename ::std::iterator_traits<_RandomAccessIterator>::value_type _InputType;
     oneapi::dpl::__internal::__pattern_sort(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __comp,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _RandomAccessIterator>(__exec),
+        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _RandomAccessIterator>(),
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _RandomAccessIterator>(__exec),
         typename ::std::is_move_constructible<_InputType>::type());
 }
@@ -707,7 +693,7 @@ stable_sort(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAcc
 {
     oneapi::dpl::__internal::__pattern_stable_sort(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __comp,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _RandomAccessIterator>(__exec),
+        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _RandomAccessIterator>(),
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _RandomAccessIterator>(__exec));
 }
 
@@ -729,8 +715,7 @@ sort_by_key(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __keys_first, _Ran
 {
     oneapi::dpl::__internal::__pattern_sort_by_key(
         ::std::forward<_ExecutionPolicy>(__exec), __keys_first, __keys_last, __values_first, __comp,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _RandomAccessIterator1,
-            _RandomAccessIterator2>(__exec),
+        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2>(),
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _RandomAccessIterator1,
             _RandomAccessIterator2>(__exec));
 }
@@ -754,8 +739,7 @@ mismatch(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator
 {
     return oneapi::dpl::__internal::__pattern_mismatch(
         ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2, __pred,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(
-            __exec),
+        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(),
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(
             __exec));
 }
@@ -800,7 +784,7 @@ equal(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 _
 {
     return oneapi::dpl::__internal::__pattern_equal(
         ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __p,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1>(__exec),
+        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1>(),
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1>(__exec));
 }
 
@@ -819,7 +803,7 @@ equal(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 _
 {
     return oneapi::dpl::__internal::__pattern_equal(
         ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2, __p,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1>(__exec),
+        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1>(),
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1>(__exec));
 }
 
@@ -855,7 +839,7 @@ partial_sort(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAc
 {
     oneapi::dpl::__internal::__pattern_partial_sort(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __middle, __last, __comp,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _RandomAccessIterator>(__exec),
+        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _RandomAccessIterator>(),
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _RandomAccessIterator>(__exec));
 }
 
@@ -877,8 +861,7 @@ partial_sort_copy(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardI
 {
     return oneapi::dpl::__internal::__pattern_partial_sort_copy(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __d_first, __d_last, __comp,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator,
-                                                              _RandomAccessIterator>(__exec),
+        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator, _RandomAccessIterator>(),
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator,
                                                                 _RandomAccessIterator>(__exec));
 }
@@ -901,7 +884,7 @@ is_sorted_until(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIte
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
         oneapi::dpl::__internal::__reorder_pred<_Compare>(__comp),
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec),
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec),
+        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(),
         oneapi::dpl::__internal::__first_semantic());
     return __res == __last ? __last : oneapi::dpl::__internal::__pstl_next(__res);
 }
@@ -922,7 +905,7 @@ is_sorted(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator
                ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
                oneapi::dpl::__internal::__reorder_pred<_Compare>(__comp),
                oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec),
-               oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec),
+               oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(),
                oneapi::dpl::__internal::__or_semantic()) == __last;
 }
 
@@ -943,8 +926,7 @@ merge(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 _
 {
     return oneapi::dpl::__internal::__pattern_merge(
         ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2, __d_first, __comp,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2,
-                                                              _ForwardIterator>(__exec),
+        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2, _ForwardIterator>(),
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2,
                                                                 _ForwardIterator>(__exec));
 }
@@ -965,7 +947,7 @@ inplace_merge(_ExecutionPolicy&& __exec, _BidirectionalIterator __first, _Bidire
 {
     oneapi::dpl::__internal::__pattern_inplace_merge(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __middle, __last, __comp,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _BidirectionalIterator>(__exec),
+        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _BidirectionalIterator>(),
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _BidirectionalIterator>(__exec));
 }
 
@@ -987,8 +969,7 @@ includes(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator
 {
     return oneapi::dpl::__internal::__pattern_includes(
         ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2, __comp,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(
-            __exec),
+        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(),
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(
             __exec));
 }
@@ -1012,8 +993,7 @@ set_union(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterato
 {
     return oneapi::dpl::__internal::__pattern_set_union(
         ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2, __result, __comp,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2,
-                                                              _ForwardIterator>(__exec),
+        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2, _ForwardIterator>(),
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2,
                                                                 _ForwardIterator>(__exec));
 }
@@ -1037,8 +1017,7 @@ set_intersection(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _Forward
 {
     return oneapi::dpl::__internal::__pattern_set_intersection(
         ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2, __result, __comp,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2,
-                                                              _ForwardIterator>(__exec),
+        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2, _ForwardIterator>(),
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2,
                                                                 _ForwardIterator>(__exec));
 }
@@ -1062,8 +1041,7 @@ set_difference(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIt
 {
     return oneapi::dpl::__internal::__pattern_set_difference(
         ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2, __result, __comp,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2,
-                                                              _ForwardIterator>(__exec),
+        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2, _ForwardIterator>(),
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2,
                                                                 _ForwardIterator>(__exec));
 }
@@ -1088,8 +1066,7 @@ set_symmetric_difference(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1,
 {
     return oneapi::dpl::__internal::__pattern_set_symmetric_difference(
         ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2, __result, __comp,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2,
-                                                              _ForwardIterator>(__exec),
+        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2, _ForwardIterator>(),
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2,
                                                                 _ForwardIterator>(__exec));
 }
@@ -1110,7 +1087,7 @@ is_heap_until(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomA
 {
     return oneapi::dpl::__internal::__pattern_is_heap_until(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __comp,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _RandomAccessIterator>(__exec),
+        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _RandomAccessIterator>(),
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _RandomAccessIterator>(__exec));
 }
 
@@ -1128,7 +1105,7 @@ is_heap(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessI
 {
     return oneapi::dpl::__internal::__pattern_is_heap(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __comp,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _RandomAccessIterator>(__exec),
+        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _RandomAccessIterator>(),
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _RandomAccessIterator>(__exec));
 }
 
@@ -1148,7 +1125,7 @@ min_element(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterato
 {
     return oneapi::dpl::__internal::__pattern_min_element(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __comp,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec),
+        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(),
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec));
 }
 
@@ -1183,7 +1160,7 @@ minmax_element(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIter
 {
     return oneapi::dpl::__internal::__pattern_minmax_element(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __comp,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec),
+        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(),
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec));
 }
 
@@ -1204,7 +1181,7 @@ nth_element(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAcc
 {
     oneapi::dpl::__internal::__pattern_nth_element(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __nth, __last, __comp,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _RandomAccessIterator>(__exec),
+        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _RandomAccessIterator>(),
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _RandomAccessIterator>(__exec));
 }
 
@@ -1226,8 +1203,7 @@ lexicographical_compare(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _
 {
     return oneapi::dpl::__internal::__pattern_lexicographical_compare(
         ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2, __comp,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(
-            __exec),
+        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(),
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(
             __exec));
 }
@@ -1250,7 +1226,7 @@ shift_left(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator
 {
     return oneapi::dpl::__internal::__pattern_shift_left(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __n,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec),
+        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(),
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec));
 }
 
@@ -1263,7 +1239,7 @@ shift_right(_ExecutionPolicy&& __exec, _BidirectionalIterator __first, _Bidirect
 {
     return oneapi::dpl::__internal::__pattern_shift_right(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __n,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _BidirectionalIterator>(__exec),
+        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _BidirectionalIterator>(),
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _BidirectionalIterator>(__exec));
 }
 
diff --git a/include/oneapi/dpl/pstl/glue_memory_impl.h b/include/oneapi/dpl/pstl/glue_memory_impl.h
index 082856131e7..de59de54ebf 100644
--- a/include/oneapi/dpl/pstl/glue_memory_impl.h
+++ b/include/oneapi/dpl/pstl/glue_memory_impl.h
@@ -48,9 +48,8 @@ uninitialized_copy(_ExecutionPolicy&& __exec, _InputIterator __first, _InputIter
     const auto __is_parallel =
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _InputIterator, _ForwardIterator>(
             __exec);
-    const auto __is_vector =
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _InputIterator, _ForwardIterator>(
-            __exec);
+    constexpr auto __is_vector =
+        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _InputIterator, _ForwardIterator>();
 
     if constexpr (::std::is_trivial_v<_ValueType1> && ::std::is_trivial_v<_ValueType2>)
     {
@@ -78,9 +77,8 @@ uninitialized_copy_n(_ExecutionPolicy&& __exec, _InputIterator __first, _Size __
     const auto __is_parallel =
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _InputIterator, _ForwardIterator>(
             __exec);
-    const auto __is_vector =
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _InputIterator, _ForwardIterator>(
-            __exec);
+    constexpr auto __is_vector =
+        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _InputIterator, _ForwardIterator>();
 
     if constexpr (::std::is_trivial_v<_ValueType1> && ::std::is_trivial_v<_ValueType2>)
     {
@@ -110,9 +108,8 @@ uninitialized_move(_ExecutionPolicy&& __exec, _InputIterator __first, _InputIter
     const auto __is_parallel =
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _InputIterator, _ForwardIterator>(
             __exec);
-    const auto __is_vector =
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _InputIterator, _ForwardIterator>(
-            __exec);
+    constexpr auto __is_vector =
+        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _InputIterator, _ForwardIterator>();
 
     if constexpr (::std::is_trivial_v<_ValueType1> && ::std::is_trivial_v<_ValueType2>)
     {
@@ -140,9 +137,8 @@ uninitialized_move_n(_ExecutionPolicy&& __exec, _InputIterator __first, _Size __
     const auto __is_parallel =
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _InputIterator, _ForwardIterator>(
             __exec);
-    const auto __is_vector =
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _InputIterator, _ForwardIterator>(
-            __exec);
+    constexpr auto __is_vector =
+        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _InputIterator, _ForwardIterator>();
 
     if constexpr (::std::is_trivial_v<_ValueType1> && ::std::is_trivial_v<_ValueType2>)
     {
@@ -170,8 +166,8 @@ uninitialized_fill(_ExecutionPolicy&& __exec, _ForwardIterator __first, _Forward
 
     const auto __is_parallel =
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec);
-    const auto __is_vector =
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec);
+    constexpr auto __is_vector =
+        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>();
 
     if constexpr (::std::is_arithmetic_v<_ValueType>)
     {
@@ -198,8 +194,8 @@ uninitialized_fill_n(_ExecutionPolicy&& __exec, _ForwardIterator __first, _Size
 
     const auto __is_parallel =
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec);
-    const auto __is_vector =
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec);
+    constexpr auto __is_vector =
+        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>();
 
     if constexpr (::std::is_arithmetic_v<_ValueType>)
     {
@@ -233,10 +229,9 @@ destroy(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __
         ::std::conditional_t<
             oneapi::dpl::__internal::__is_host_execution_policy<::std::decay_t<_ExecutionPolicy>>::value,
             ::std::false_type,
-            decltype(oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(
-                __exec))>;
+            decltype(oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>())>;
 #else
-        decltype(oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec));
+        decltype(oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>());
 #endif // _PSTL_ICPX_OMP_SIMD_DESTROY_WINDOWS_BROKEN || _ONEDPL_ICPX_OMP_SIMD_DESTROY_WINDOWS_BROKEN
     constexpr _is_vector_type __is_vector;
 
@@ -262,10 +257,9 @@ destroy_n(_ExecutionPolicy&& __exec, _ForwardIterator __first, _Size __n)
         ::std::conditional_t<
             oneapi::dpl::__internal::__is_host_execution_policy<::std::decay_t<_ExecutionPolicy>>::value,
             ::std::false_type,
-            decltype(oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(
-                __exec))>;
+            decltype(oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>())>;
 #else
-        decltype(oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec));
+        decltype(oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>());
 #endif // _PSTL_ICPX_OMP_SIMD_DESTROY_WINDOWS_BROKEN || _ONEDPL_ICPX_OMP_SIMD_DESTROY_WINDOWS_BROKEN
     constexpr _is_vector_type __is_vector;
 
@@ -292,8 +286,8 @@ uninitialized_default_construct(_ExecutionPolicy&& __exec, _ForwardIterator __fi
 
     const auto __is_parallel =
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec);
-    const auto __is_vector =
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec);
+    constexpr auto __is_vector =
+        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>();
 
     if constexpr (!::std::is_trivial_v<_ValueType>)
     {
@@ -313,8 +307,8 @@ uninitialized_default_construct_n(_ExecutionPolicy&& __exec, _ForwardIterator __
 
     const auto __is_parallel =
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec);
-    const auto __is_vector =
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec);
+    constexpr auto __is_vector =
+        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>();
 
     if constexpr (::std::is_trivial_v<_ValueType>)
     {
@@ -340,8 +334,8 @@ uninitialized_value_construct(_ExecutionPolicy&& __exec, _ForwardIterator __firs
 
     const auto __is_parallel =
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec);
-    const auto __is_vector =
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec);
+    constexpr auto __is_vector =
+        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>();
 
     if constexpr (::std::is_trivial_v<_ValueType>)
     {
@@ -368,8 +362,8 @@ uninitialized_value_construct_n(_ExecutionPolicy&& __exec, _ForwardIterator __fi
 
     const auto __is_parallel =
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec);
-    const auto __is_vector =
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec);
+    constexpr auto __is_vector =
+        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>();
 
     if constexpr (::std::is_trivial_v<_ValueType>)
     {
diff --git a/include/oneapi/dpl/pstl/glue_numeric_impl.h b/include/oneapi/dpl/pstl/glue_numeric_impl.h
index f2564db3132..2a66a0e9743 100644
--- a/include/oneapi/dpl/pstl/glue_numeric_impl.h
+++ b/include/oneapi/dpl/pstl/glue_numeric_impl.h
@@ -73,8 +73,7 @@ transform_reduce(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _Forward
     return oneapi::dpl::__internal::__pattern_transform_reduce(
         ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __init, ::std::plus<_InputType>(),
         ::std::multiplies<_InputType>(),
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(
-            __exec),
+        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(),
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(
             __exec));
 }
@@ -87,8 +86,7 @@ transform_reduce(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _Forward
 {
     return oneapi::dpl::__internal::__pattern_transform_reduce(
         ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __init, __binary_op1, __binary_op2,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(
-            __exec),
+        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(),
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(
             __exec));
 }
@@ -100,7 +98,7 @@ transform_reduce(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIt
 {
     return oneapi::dpl::__internal::__pattern_transform_reduce(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __init, __binary_op, __unary_op,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec),
+        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(),
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec));
 }
 
@@ -228,8 +226,7 @@ transform_exclusive_scan(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _
     return oneapi::dpl::__internal::__pattern_transform_scan(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result, __unary_op, __init, __binary_op,
         /*inclusive=*/::std::false_type(),
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(
-            __exec),
+        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(),
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(
             __exec));
 }
@@ -246,8 +243,7 @@ transform_inclusive_scan(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _
     return oneapi::dpl::__internal::__pattern_transform_scan(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result, __unary_op, __init, __binary_op,
         /*inclusive=*/::std::true_type(),
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(
-            __exec),
+        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(),
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(
             __exec));
 }
@@ -261,8 +257,7 @@ transform_inclusive_scan(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _
     return oneapi::dpl::__internal::__pattern_transform_scan(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result, __unary_op, __binary_op,
         /*inclusive=*/::std::true_type(),
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(
-            __exec),
+        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(),
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(
             __exec));
 }
@@ -280,8 +275,7 @@ adjacent_difference(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _Forwa
 
     return oneapi::dpl::__internal::__pattern_adjacent_difference(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __d_first, __op,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(
-            __exec),
+        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(),
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(
             __exec));
 }

From 0b181a519a2c6d79d68b237ddaf223f9f8b60d84 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Mon, 29 Jan 2024 17:04:52 +0100
Subject: [PATCH 031/566] Declare __is_parallelization_preferred as constexpr
 without runtime params

---
 include/oneapi/dpl/pstl/execution_impl.h      |   8 +-
 .../experimental/internal/for_loop_impl.h     |  21 ++-
 include/oneapi/dpl/pstl/glue_algorithm_impl.h | 141 +++++++-----------
 include/oneapi/dpl/pstl/glue_memory_impl.h    |  52 +++----
 include/oneapi/dpl/pstl/glue_numeric_impl.h   |  20 +--
 5 files changed, 101 insertions(+), 141 deletions(-)

diff --git a/include/oneapi/dpl/pstl/execution_impl.h b/include/oneapi/dpl/pstl/execution_impl.h
index 828f703333d..27a955551f9 100644
--- a/include/oneapi/dpl/pstl/execution_impl.h
+++ b/include/oneapi/dpl/pstl/execution_impl.h
@@ -113,12 +113,10 @@ __is_vectorization_preferred()
 }
 
 template <typename _ExecutionPolicy, typename... _IteratorTypes>
-auto
-__is_parallelization_preferred(_ExecutionPolicy& __exec)
-    -> decltype(__internal::__lazy_and(__exec.__allow_parallel(),
-                                       __internal::__is_random_access_iterator_t<_IteratorTypes...>()))
+constexpr decltype(auto)
+__is_parallelization_preferred()
 {
-    return __internal::__lazy_and(__exec.__allow_parallel(),
+    return __internal::__lazy_and(::std::decay_t<_ExecutionPolicy>::__allow_parallel(),
                                   __internal::__is_random_access_iterator_t<_IteratorTypes...>());
 }
 
diff --git a/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h b/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
index d4827c11ea4..9f4314f20ee 100644
--- a/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
+++ b/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
@@ -486,9 +486,9 @@ struct __use_par_vec_helper<_Ip, ::std::enable_if_t<::std::is_integral_v<_Ip>>>
 
     template <typename _ExecutionPolicy>
     static constexpr auto
-    __use_parallel(_ExecutionPolicy&& __exec) -> decltype(__exec.__allow_parallel())
+    __use_parallel()
     {
-        return __exec.__allow_parallel();
+        return ::std::decay_t<_ExecutionPolicy>::__allow_parallel();
     }
 };
 
@@ -504,12 +504,9 @@ struct __use_par_vec_helper<_Ip, ::std::enable_if_t<!::std::is_integral_v<_Ip>>>
 
     template <typename _ExecutionPolicy>
     static constexpr auto
-    __use_parallel(_ExecutionPolicy&& __exec)
-        -> decltype(oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _Ip>(
-            ::std::forward<_ExecutionPolicy>(__exec)))
+    __use_parallel()
     {
-        return oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _Ip>(
-            ::std::forward<_ExecutionPolicy>(__exec));
+        return oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _Ip>();
     }
 };
 
@@ -524,10 +521,10 @@ __use_vectorization()
 
 template <typename _ExecutionPolicy, typename _Ip>
 auto
-__use_parallelization(_ExecutionPolicy&& __exec)
-    -> decltype(__use_par_vec_helper<_Ip>::__use_parallel(::std::forward<_ExecutionPolicy>(__exec)))
+__use_parallelization()
+    -> decltype(__use_par_vec_helper<_Ip>::template __use_parallel<_ExecutionPolicy>())
 {
-    return __use_par_vec_helper<_Ip>::__use_parallel(::std::forward<_ExecutionPolicy>(__exec));
+    return __use_par_vec_helper<_Ip>::template __use_parallel<_ExecutionPolicy>();
 }
 
 // Helper functions to extract to separate a Callable object from the pack of reductions and inductions
@@ -539,7 +536,7 @@ __for_loop_impl(_ExecutionPolicy&& __exec, _Ip __start, _Ip __finish, _Fp&& __f,
     oneapi::dpl::__internal::__pattern_for_loop(
         ::std::forward<_ExecutionPolicy>(__exec), __start, __finish, __f, __stride,
         oneapi::dpl::__internal::__use_vectorization<_ExecutionPolicy, _Ip>(),
-        oneapi::dpl::__internal::__use_parallelization<_ExecutionPolicy, _Ip>(__exec),
+        oneapi::dpl::__internal::__use_parallelization<_ExecutionPolicy, _Ip>(),
         ::std::get<_Is>(::std::move(__t))...);
 }
 
@@ -552,7 +549,7 @@ __for_loop_n_impl(_ExecutionPolicy&& __exec, _Ip __start, _Size __n, _Fp&& __f,
     oneapi::dpl::__internal::__pattern_for_loop_n(
         ::std::forward<_ExecutionPolicy>(__exec), __start, __n, __f, __stride,
         oneapi::dpl::__internal::__use_vectorization<_ExecutionPolicy, _Ip>(),
-        oneapi::dpl::__internal::__use_parallelization<_ExecutionPolicy, _Ip>(__exec),
+        oneapi::dpl::__internal::__use_parallelization<_ExecutionPolicy, _Ip>(),
         ::std::get<_Is>(::std::move(__t))...);
 }
 
diff --git a/include/oneapi/dpl/pstl/glue_algorithm_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
index ac1ff531bce..17c5e331710 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
@@ -46,7 +46,7 @@ any_of(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __l
     return oneapi::dpl::__internal::__pattern_any_of(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __pred,
         oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec));
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>());
 }
 
 // [alg.all_of]
@@ -88,7 +88,7 @@ for_each_n(_ExecutionPolicy&& __exec, _ForwardIterator __first, _Size __n, _Func
     return oneapi::dpl::__internal::__pattern_walk1_n(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __n, __f,
         oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec));
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>());
 }
 
 // [alg.find]
@@ -132,8 +132,7 @@ find_end(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _ForwardIterator1
     return oneapi::dpl::__internal::__pattern_find_end(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __s_first, __s_last, __pred,
         oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(
-            __exec));
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>());
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2>
@@ -154,8 +153,7 @@ find_first_of(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _ForwardIter
     return oneapi::dpl::__internal::__pattern_find_first_of(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __s_first, __s_last, __pred,
         oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(
-            __exec));
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>());
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2>
@@ -175,7 +173,7 @@ adjacent_find(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardItera
     typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _ValueType;
     return oneapi::dpl::__internal::__pattern_adjacent_find(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last, ::std::equal_to<_ValueType>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec),
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(),
         oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(),
         oneapi::dpl::__internal::__first_semantic());
 }
@@ -186,7 +184,7 @@ adjacent_find(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardItera
 {
     return oneapi::dpl::__internal::__pattern_adjacent_find(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __pred,
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec),
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(),
         oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(),
         oneapi::dpl::__internal::__first_semantic());
 }
@@ -205,7 +203,7 @@ count(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __la
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
         oneapi::dpl::__internal::__equal_value<oneapi::dpl::__internal::__ref_or_copy<_ExecutionPolicy, const _Tp>>(
             __value),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec),
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(),
         oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>());
 }
 
@@ -216,7 +214,7 @@ count_if(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator _
 {
     return oneapi::dpl::__internal::__pattern_count(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __pred,
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec),
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(),
         oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>());
 }
 
@@ -230,8 +228,7 @@ search(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _ForwardIterator1 _
     return oneapi::dpl::__internal::__pattern_search(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __s_first, __s_last, __pred,
         oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(
-            __exec));
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>());
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2>
@@ -251,7 +248,7 @@ search_n(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator _
     return oneapi::dpl::__internal::__pattern_search_n(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __count, __value, __pred,
         oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec));
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>());
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator, class _Size, class _Tp>
@@ -272,8 +269,7 @@ copy(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _ForwardIterator1 __l
     return oneapi::dpl::__internal::__pattern_walk2_brick(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result,
         oneapi::dpl::__internal::__brick_copy<_ExecutionPolicy>{},
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(
-            __exec));
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>());
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator1, class _Size, class _ForwardIterator2>
@@ -285,8 +281,7 @@ copy_n(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _Size __n, _Forward
     return oneapi::dpl::__internal::__pattern_walk2_brick_n(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __n, __result,
         oneapi::dpl::__internal::__brick_copy_n<_DecayedExecutionPolicy>{},
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(
-            __exec));
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>());
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Predicate>
@@ -297,8 +292,7 @@ copy_if(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _ForwardIterator1
     return oneapi::dpl::__internal::__pattern_copy_if(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result, __pred,
         oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(
-            __exec));
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>());
 }
 
 // [alg.swap]
@@ -317,8 +311,7 @@ swap_ranges(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardItera
             swap(__x, __y);
         },
         oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(
-            __exec));
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>());
 }
 
 // [alg.transform]
@@ -362,8 +355,7 @@ transform_if(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _ForwardItera
         oneapi::dpl::__internal::__transform_if_unary_functor<_UnaryOperation, _UnaryPredicate>(::std::move(__op),
                                                                                                 ::std::move(__pred)),
         oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(
-            __exec));
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>());
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _ForwardIterator3,
@@ -377,8 +369,7 @@ transform_if(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIter
         oneapi::dpl::__internal::__transform_if_binary_functor<_BinaryOperation, _BinaryPredicate>(::std::move(__op),
                                                                                                    ::std::move(__pred)),
         oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2, _ForwardIterator3>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2,
-                                                                _ForwardIterator3>(__exec));
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2, _ForwardIterator3>());
 }
 
 // [alg.replace]
@@ -419,8 +410,7 @@ replace_copy_if(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _ForwardIt
                                  oneapi::dpl::__internal::__ref_or_copy<_ExecutionPolicy, _UnaryPredicate>>>(
             __new_value, __pred),
         oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(
-            __exec));
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>());
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Tp>
@@ -443,7 +433,7 @@ fill(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __las
 {
     oneapi::dpl::__internal::__pattern_fill(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __value,
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec),
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(),
         oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>());
 }
 
@@ -456,7 +446,7 @@ fill_n(_ExecutionPolicy&& __exec, _ForwardIterator __first, _Size __count, const
 
     return oneapi::dpl::__internal::__pattern_fill_n(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __count, __value,
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec),
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(),
         oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>());
 }
 
@@ -467,7 +457,7 @@ generate(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator _
 {
     oneapi::dpl::__internal::__pattern_generate(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __g,
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec),
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(),
         oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>());
 }
 
@@ -480,7 +470,7 @@ generate_n(_ExecutionPolicy&& __exec, _ForwardIterator __first, _Size __count, _
 
     return oneapi::dpl::__internal::__pattern_generate_n(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __count, __g,
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec),
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(),
         oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>());
 }
 
@@ -515,7 +505,7 @@ remove_if(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator
     return oneapi::dpl::__internal::__pattern_remove_if(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __pred,
         oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec));
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>());
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator, class _Tp>
@@ -537,7 +527,7 @@ unique(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __l
     return oneapi::dpl::__internal::__pattern_unique(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __pred,
         oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec));
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>());
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator>
@@ -556,8 +546,7 @@ unique_copy(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _ForwardIterat
     return oneapi::dpl::__internal::__pattern_unique_copy(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result, __pred,
         oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(
-            __exec));
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>());
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2>
@@ -577,7 +566,7 @@ reverse(_ExecutionPolicy&& __exec, _BidirectionalIterator __first, _Bidirectiona
     oneapi::dpl::__internal::__pattern_reverse(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
         oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _BidirectionalIterator>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _BidirectionalIterator>(__exec));
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _BidirectionalIterator>());
 }
 
 template <class _ExecutionPolicy, class _BidirectionalIterator, class _ForwardIterator>
@@ -588,8 +577,7 @@ reverse_copy(_ExecutionPolicy&& __exec, _BidirectionalIterator __first, _Bidirec
     return oneapi::dpl::__internal::__pattern_reverse_copy(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __d_first,
         oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _BidirectionalIterator, _ForwardIterator>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _BidirectionalIterator,
-                                                                _ForwardIterator>(__exec));
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _BidirectionalIterator, _ForwardIterator>());
 }
 
 // [alg.rotate]
@@ -601,7 +589,7 @@ rotate(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __m
     return oneapi::dpl::__internal::__pattern_rotate(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __middle, __last,
         oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec));
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>());
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2>
@@ -612,8 +600,7 @@ rotate_copy(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _ForwardIterat
     return oneapi::dpl::__internal::__pattern_rotate_copy(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __middle, __last, __result,
         oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(
-            __exec));
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>());
 }
 
 // [alg.partitions]
@@ -625,7 +612,7 @@ is_partitioned(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIter
     return oneapi::dpl::__internal::__pattern_is_partitioned(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __pred,
         oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec));
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>());
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator, class _UnaryPredicate>
@@ -635,7 +622,7 @@ partition(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator
     return oneapi::dpl::__internal::__pattern_partition(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __pred,
         oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec));
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>());
 }
 
 template <class _ExecutionPolicy, class _BidirectionalIterator, class _UnaryPredicate>
@@ -646,7 +633,7 @@ stable_partition(_ExecutionPolicy&& __exec, _BidirectionalIterator __first, _Bid
     return oneapi::dpl::__internal::__pattern_stable_partition(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __pred,
         oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _BidirectionalIterator>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _BidirectionalIterator>(__exec));
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _BidirectionalIterator>());
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator, class _ForwardIterator1, class _ForwardIterator2,
@@ -659,8 +646,7 @@ partition_copy(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIter
     return oneapi::dpl::__internal::__pattern_partition_copy(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __out_true, __out_false, __pred,
         oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator, _ForwardIterator1, _ForwardIterator2>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator, _ForwardIterator1,
-                                                                _ForwardIterator2>(__exec));
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator, _ForwardIterator1, _ForwardIterator2>());
 }
 
 // [alg.sort]
@@ -673,7 +659,7 @@ sort(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIter
     oneapi::dpl::__internal::__pattern_sort(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __comp,
         oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _RandomAccessIterator>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _RandomAccessIterator>(__exec),
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _RandomAccessIterator>(),
         typename ::std::is_move_constructible<_InputType>::type());
 }
 
@@ -694,7 +680,7 @@ stable_sort(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAcc
     oneapi::dpl::__internal::__pattern_stable_sort(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __comp,
         oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _RandomAccessIterator>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _RandomAccessIterator>(__exec));
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _RandomAccessIterator>());
 }
 
 template <class _ExecutionPolicy, class _RandomAccessIterator>
@@ -716,8 +702,7 @@ sort_by_key(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __keys_first, _Ran
     oneapi::dpl::__internal::__pattern_sort_by_key(
         ::std::forward<_ExecutionPolicy>(__exec), __keys_first, __keys_last, __values_first, __comp,
         oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _RandomAccessIterator1,
-            _RandomAccessIterator2>(__exec));
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2>());
 }
 
 template <typename _ExecutionPolicy, typename _RandomAccessIterator1, typename _RandomAccessIterator2>
@@ -740,8 +725,7 @@ mismatch(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator
     return oneapi::dpl::__internal::__pattern_mismatch(
         ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2, __pred,
         oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(
-            __exec));
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>());
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _BinaryPredicate>
@@ -785,7 +769,7 @@ equal(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 _
     return oneapi::dpl::__internal::__pattern_equal(
         ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __p,
         oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1>(__exec));
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1>());
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2>
@@ -804,7 +788,7 @@ equal(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 _
     return oneapi::dpl::__internal::__pattern_equal(
         ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2, __p,
         oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1>(__exec));
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1>());
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2>
@@ -826,8 +810,7 @@ move(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _ForwardIterator1 __l
     return oneapi::dpl::__internal::__pattern_walk2_brick(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __d_first,
         oneapi::dpl::__internal::__brick_move<_DecayedExecutionPolicy>{},
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(
-            __exec));
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>());
 }
 
 // [partial.sort]
@@ -840,7 +823,7 @@ partial_sort(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAc
     oneapi::dpl::__internal::__pattern_partial_sort(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __middle, __last, __comp,
         oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _RandomAccessIterator>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _RandomAccessIterator>(__exec));
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _RandomAccessIterator>());
 }
 
 template <class _ExecutionPolicy, class _RandomAccessIterator>
@@ -862,8 +845,7 @@ partial_sort_copy(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardI
     return oneapi::dpl::__internal::__pattern_partial_sort_copy(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __d_first, __d_last, __comp,
         oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator, _RandomAccessIterator>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator,
-                                                                _RandomAccessIterator>(__exec));
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator, _RandomAccessIterator>());
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator, class _RandomAccessIterator>
@@ -883,7 +865,7 @@ is_sorted_until(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIte
     const _ForwardIterator __res = oneapi::dpl::__internal::__pattern_adjacent_find(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
         oneapi::dpl::__internal::__reorder_pred<_Compare>(__comp),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec),
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(),
         oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(),
         oneapi::dpl::__internal::__first_semantic());
     return __res == __last ? __last : oneapi::dpl::__internal::__pstl_next(__res);
@@ -904,7 +886,7 @@ is_sorted(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator
     return oneapi::dpl::__internal::__pattern_adjacent_find(
                ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
                oneapi::dpl::__internal::__reorder_pred<_Compare>(__comp),
-               oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec),
+               oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(),
                oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(),
                oneapi::dpl::__internal::__or_semantic()) == __last;
 }
@@ -927,8 +909,7 @@ merge(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 _
     return oneapi::dpl::__internal::__pattern_merge(
         ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2, __d_first, __comp,
         oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2, _ForwardIterator>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2,
-                                                                _ForwardIterator>(__exec));
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2, _ForwardIterator>());
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _ForwardIterator>
@@ -948,7 +929,7 @@ inplace_merge(_ExecutionPolicy&& __exec, _BidirectionalIterator __first, _Bidire
     oneapi::dpl::__internal::__pattern_inplace_merge(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __middle, __last, __comp,
         oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _BidirectionalIterator>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _BidirectionalIterator>(__exec));
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _BidirectionalIterator>());
 }
 
 template <class _ExecutionPolicy, class _BidirectionalIterator>
@@ -970,8 +951,7 @@ includes(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator
     return oneapi::dpl::__internal::__pattern_includes(
         ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2, __comp,
         oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(
-            __exec));
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>());
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2>
@@ -994,8 +974,7 @@ set_union(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterato
     return oneapi::dpl::__internal::__pattern_set_union(
         ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2, __result, __comp,
         oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2, _ForwardIterator>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2,
-                                                                _ForwardIterator>(__exec));
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2, _ForwardIterator>());
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _ForwardIterator>
@@ -1018,8 +997,7 @@ set_intersection(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _Forward
     return oneapi::dpl::__internal::__pattern_set_intersection(
         ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2, __result, __comp,
         oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2, _ForwardIterator>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2,
-                                                                _ForwardIterator>(__exec));
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2, _ForwardIterator>());
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _ForwardIterator>
@@ -1042,8 +1020,7 @@ set_difference(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIt
     return oneapi::dpl::__internal::__pattern_set_difference(
         ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2, __result, __comp,
         oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2, _ForwardIterator>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2,
-                                                                _ForwardIterator>(__exec));
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2, _ForwardIterator>());
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _ForwardIterator>
@@ -1067,8 +1044,7 @@ set_symmetric_difference(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1,
     return oneapi::dpl::__internal::__pattern_set_symmetric_difference(
         ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2, __result, __comp,
         oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2, _ForwardIterator>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2,
-                                                                _ForwardIterator>(__exec));
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2, _ForwardIterator>());
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _ForwardIterator>
@@ -1088,7 +1064,7 @@ is_heap_until(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomA
     return oneapi::dpl::__internal::__pattern_is_heap_until(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __comp,
         oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _RandomAccessIterator>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _RandomAccessIterator>(__exec));
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _RandomAccessIterator>());
 }
 
 template <class _ExecutionPolicy, class _RandomAccessIterator>
@@ -1106,7 +1082,7 @@ is_heap(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessI
     return oneapi::dpl::__internal::__pattern_is_heap(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __comp,
         oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _RandomAccessIterator>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _RandomAccessIterator>(__exec));
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _RandomAccessIterator>());
 }
 
 template <class _ExecutionPolicy, class _RandomAccessIterator>
@@ -1126,7 +1102,7 @@ min_element(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterato
     return oneapi::dpl::__internal::__pattern_min_element(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __comp,
         oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec));
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>());
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator>
@@ -1161,7 +1137,7 @@ minmax_element(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIter
     return oneapi::dpl::__internal::__pattern_minmax_element(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __comp,
         oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec));
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>());
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator>
@@ -1182,7 +1158,7 @@ nth_element(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAcc
     oneapi::dpl::__internal::__pattern_nth_element(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __nth, __last, __comp,
         oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _RandomAccessIterator>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _RandomAccessIterator>(__exec));
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _RandomAccessIterator>());
 }
 
 template <class _ExecutionPolicy, class _RandomAccessIterator>
@@ -1204,8 +1180,7 @@ lexicographical_compare(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _
     return oneapi::dpl::__internal::__pattern_lexicographical_compare(
         ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2, __comp,
         oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(
-            __exec));
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>());
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2>
@@ -1227,7 +1202,7 @@ shift_left(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator
     return oneapi::dpl::__internal::__pattern_shift_left(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __n,
         oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec));
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>());
 }
 
 // [shift.right]
@@ -1240,7 +1215,7 @@ shift_right(_ExecutionPolicy&& __exec, _BidirectionalIterator __first, _Bidirect
     return oneapi::dpl::__internal::__pattern_shift_right(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __n,
         oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _BidirectionalIterator>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _BidirectionalIterator>(__exec));
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _BidirectionalIterator>());
 }
 
 } // namespace dpl
diff --git a/include/oneapi/dpl/pstl/glue_memory_impl.h b/include/oneapi/dpl/pstl/glue_memory_impl.h
index de59de54ebf..6bdc66d510b 100644
--- a/include/oneapi/dpl/pstl/glue_memory_impl.h
+++ b/include/oneapi/dpl/pstl/glue_memory_impl.h
@@ -45,9 +45,8 @@ uninitialized_copy(_ExecutionPolicy&& __exec, _InputIterator __first, _InputIter
     typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _ValueType2;
     typedef ::std::decay_t<_ExecutionPolicy> _DecayedExecutionPolicy;
 
-    const auto __is_parallel =
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _InputIterator, _ForwardIterator>(
-            __exec);
+    constexpr auto __is_parallel =
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _InputIterator, _ForwardIterator>();
     constexpr auto __is_vector =
         oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _InputIterator, _ForwardIterator>();
 
@@ -74,9 +73,8 @@ uninitialized_copy_n(_ExecutionPolicy&& __exec, _InputIterator __first, _Size __
     typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _ValueType2;
     typedef ::std::decay_t<_ExecutionPolicy> _DecayedExecutionPolicy;
 
-    const auto __is_parallel =
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _InputIterator, _ForwardIterator>(
-            __exec);
+    constexpr auto __is_parallel =
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _InputIterator, _ForwardIterator>();
     constexpr auto __is_vector =
         oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _InputIterator, _ForwardIterator>();
 
@@ -105,9 +103,8 @@ uninitialized_move(_ExecutionPolicy&& __exec, _InputIterator __first, _InputIter
     typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _ValueType2;
     typedef ::std::decay_t<_ExecutionPolicy> _DecayedExecutionPolicy;
 
-    const auto __is_parallel =
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _InputIterator, _ForwardIterator>(
-            __exec);
+    constexpr auto __is_parallel =
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _InputIterator, _ForwardIterator>();
     constexpr auto __is_vector =
         oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _InputIterator, _ForwardIterator>();
 
@@ -134,9 +131,8 @@ uninitialized_move_n(_ExecutionPolicy&& __exec, _InputIterator __first, _Size __
     typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _ValueType2;
     typedef ::std::decay_t<_ExecutionPolicy> _DecayedExecutionPolicy;
 
-    const auto __is_parallel =
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _InputIterator, _ForwardIterator>(
-            __exec);
+    constexpr auto __is_parallel =
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _InputIterator, _ForwardIterator>();
     constexpr auto __is_vector =
         oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _InputIterator, _ForwardIterator>();
 
@@ -164,8 +160,8 @@ uninitialized_fill(_ExecutionPolicy&& __exec, _ForwardIterator __first, _Forward
     typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _ValueType;
     typedef ::std::decay_t<_ExecutionPolicy> _DecayedExecutionPolicy;
 
-    const auto __is_parallel =
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec);
+    constexpr auto __is_parallel =
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>();
     constexpr auto __is_vector =
         oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>();
 
@@ -192,8 +188,8 @@ uninitialized_fill_n(_ExecutionPolicy&& __exec, _ForwardIterator __first, _Size
     typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _ValueType;
     typedef ::std::decay_t<_ExecutionPolicy> _DecayedExecutionPolicy;
 
-    const auto __is_parallel =
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec);
+    constexpr auto __is_parallel =
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>();
     constexpr auto __is_vector =
         oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>();
 
@@ -222,8 +218,8 @@ destroy(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __
     typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _ValueType;
     typedef typename ::std::iterator_traits<_ForwardIterator>::reference _ReferenceType;
 
-    const auto __is_parallel =
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec);
+    constexpr auto __is_parallel =
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>();
     using _is_vector_type =
 #if (_PSTL_ICPX_OMP_SIMD_DESTROY_WINDOWS_BROKEN || _ONEDPL_ICPX_OMP_SIMD_DESTROY_WINDOWS_BROKEN)
         ::std::conditional_t<
@@ -250,8 +246,8 @@ destroy_n(_ExecutionPolicy&& __exec, _ForwardIterator __first, _Size __n)
     typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _ValueType;
     typedef typename ::std::iterator_traits<_ForwardIterator>::reference _ReferenceType;
 
-    const auto __is_parallel =
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec);
+    constexpr auto __is_parallel =
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>();
     using _is_vector_type =
 #if (_PSTL_ICPX_OMP_SIMD_DESTROY_WINDOWS_BROKEN || _ONEDPL_ICPX_OMP_SIMD_DESTROY_WINDOWS_BROKEN)
         ::std::conditional_t<
@@ -284,8 +280,8 @@ uninitialized_default_construct(_ExecutionPolicy&& __exec, _ForwardIterator __fi
     typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _ValueType;
     typedef ::std::decay_t<_ExecutionPolicy> _DecayedExecutionPolicy;
 
-    const auto __is_parallel =
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec);
+    constexpr auto __is_parallel =
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>();
     constexpr auto __is_vector =
         oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>();
 
@@ -305,8 +301,8 @@ uninitialized_default_construct_n(_ExecutionPolicy&& __exec, _ForwardIterator __
     typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _ValueType;
     typedef ::std::decay_t<_ExecutionPolicy> _DecayedExecutionPolicy;
 
-    const auto __is_parallel =
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec);
+    constexpr auto __is_parallel =
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>();
     constexpr auto __is_vector =
         oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>();
 
@@ -332,8 +328,8 @@ uninitialized_value_construct(_ExecutionPolicy&& __exec, _ForwardIterator __firs
     typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _ValueType;
     typedef ::std::decay_t<_ExecutionPolicy> _DecayedExecutionPolicy;
 
-    const auto __is_parallel =
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec);
+    constexpr auto __is_parallel =
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>();
     constexpr auto __is_vector =
         oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>();
 
@@ -360,8 +356,8 @@ uninitialized_value_construct_n(_ExecutionPolicy&& __exec, _ForwardIterator __fi
     typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _ValueType;
     typedef ::std::decay_t<_ExecutionPolicy> _DecayedExecutionPolicy;
 
-    const auto __is_parallel =
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec);
+    constexpr auto __is_parallel =
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>();
     constexpr auto __is_vector =
         oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>();
 
diff --git a/include/oneapi/dpl/pstl/glue_numeric_impl.h b/include/oneapi/dpl/pstl/glue_numeric_impl.h
index 2a66a0e9743..9c450f79a2b 100644
--- a/include/oneapi/dpl/pstl/glue_numeric_impl.h
+++ b/include/oneapi/dpl/pstl/glue_numeric_impl.h
@@ -74,8 +74,7 @@ transform_reduce(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _Forward
         ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __init, ::std::plus<_InputType>(),
         ::std::multiplies<_InputType>(),
         oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(
-            __exec));
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>());
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Tp, class _BinaryOperation1,
@@ -87,8 +86,7 @@ transform_reduce(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _Forward
     return oneapi::dpl::__internal::__pattern_transform_reduce(
         ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __init, __binary_op1, __binary_op2,
         oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(
-            __exec));
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>());
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator, class _Tp, class _BinaryOperation, class _UnaryOperation>
@@ -99,7 +97,7 @@ transform_reduce(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIt
     return oneapi::dpl::__internal::__pattern_transform_reduce(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __init, __binary_op, __unary_op,
         oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(__exec));
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>());
 }
 
 // [exclusive.scan]
@@ -227,8 +225,7 @@ transform_exclusive_scan(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result, __unary_op, __init, __binary_op,
         /*inclusive=*/::std::false_type(),
         oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(
-            __exec));
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>());
 }
 
 // [transform.inclusive.scan]
@@ -244,8 +241,7 @@ transform_inclusive_scan(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result, __unary_op, __init, __binary_op,
         /*inclusive=*/::std::true_type(),
         oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(
-            __exec));
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>());
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _UnaryOperation,
@@ -258,8 +254,7 @@ transform_inclusive_scan(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result, __unary_op, __binary_op,
         /*inclusive=*/::std::true_type(),
         oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(
-            __exec));
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>());
 }
 
 // [adjacent.difference]
@@ -276,8 +271,7 @@ adjacent_difference(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _Forwa
     return oneapi::dpl::__internal::__pattern_adjacent_difference(
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __d_first, __op,
         oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(
-            __exec));
+        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>());
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2>

From 382a14055756869a1586e22925512e0f5ec41a20 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Mon, 29 Jan 2024 19:49:39 +0100
Subject: [PATCH 032/566] include/oneapi/dpl/pstl/glue_memory_impl.h - move
 constexpr auto __is_vector under constexpr condition check

---
 include/oneapi/dpl/pstl/glue_memory_impl.h | 105 +++++++++++----------
 1 file changed, 57 insertions(+), 48 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_memory_impl.h b/include/oneapi/dpl/pstl/glue_memory_impl.h
index 6bdc66d510b..e810e20bf9a 100644
--- a/include/oneapi/dpl/pstl/glue_memory_impl.h
+++ b/include/oneapi/dpl/pstl/glue_memory_impl.h
@@ -47,8 +47,6 @@ uninitialized_copy(_ExecutionPolicy&& __exec, _InputIterator __first, _InputIter
 
     constexpr auto __is_parallel =
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _InputIterator, _ForwardIterator>();
-    constexpr auto __is_vector =
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _InputIterator, _ForwardIterator>();
 
     if constexpr (::std::is_trivial_v<_ValueType1> && ::std::is_trivial_v<_ValueType2>)
     {
@@ -58,6 +56,9 @@ uninitialized_copy(_ExecutionPolicy&& __exec, _InputIterator __first, _InputIter
     }
     else
     {
+        constexpr auto __is_vector =
+            oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _InputIterator, _ForwardIterator>();
+
         return oneapi::dpl::__internal::__pattern_walk2(
             ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result,
             oneapi::dpl::__internal::__op_uninitialized_copy<_DecayedExecutionPolicy>{}, __is_vector,
@@ -75,8 +76,6 @@ uninitialized_copy_n(_ExecutionPolicy&& __exec, _InputIterator __first, _Size __
 
     constexpr auto __is_parallel =
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _InputIterator, _ForwardIterator>();
-    constexpr auto __is_vector =
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _InputIterator, _ForwardIterator>();
 
     if constexpr (::std::is_trivial_v<_ValueType1> && ::std::is_trivial_v<_ValueType2>)
     {
@@ -86,6 +85,9 @@ uninitialized_copy_n(_ExecutionPolicy&& __exec, _InputIterator __first, _Size __
     }
     else
     {
+        constexpr auto __is_vector =
+            oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _InputIterator, _ForwardIterator>();
+
         return oneapi::dpl::__internal::__pattern_walk2_n(
             ::std::forward<_ExecutionPolicy>(__exec), __first, __n, __result,
             oneapi::dpl::__internal::__op_uninitialized_copy<_DecayedExecutionPolicy>{}, __is_vector,
@@ -105,8 +107,6 @@ uninitialized_move(_ExecutionPolicy&& __exec, _InputIterator __first, _InputIter
 
     constexpr auto __is_parallel =
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _InputIterator, _ForwardIterator>();
-    constexpr auto __is_vector =
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _InputIterator, _ForwardIterator>();
 
     if constexpr (::std::is_trivial_v<_ValueType1> && ::std::is_trivial_v<_ValueType2>)
     {
@@ -116,6 +116,9 @@ uninitialized_move(_ExecutionPolicy&& __exec, _InputIterator __first, _InputIter
     }
     else
     {
+        constexpr auto __is_vector =
+            oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _InputIterator, _ForwardIterator>();
+
         return oneapi::dpl::__internal::__pattern_walk2(
             ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result,
             oneapi::dpl::__internal::__op_uninitialized_move<_DecayedExecutionPolicy>{}, __is_vector,
@@ -133,8 +136,6 @@ uninitialized_move_n(_ExecutionPolicy&& __exec, _InputIterator __first, _Size __
 
     constexpr auto __is_parallel =
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _InputIterator, _ForwardIterator>();
-    constexpr auto __is_vector =
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _InputIterator, _ForwardIterator>();
 
     if constexpr (::std::is_trivial_v<_ValueType1> && ::std::is_trivial_v<_ValueType2>)
     {
@@ -144,6 +145,9 @@ uninitialized_move_n(_ExecutionPolicy&& __exec, _InputIterator __first, _Size __
     }
     else
     {
+        constexpr auto __is_vector =
+            oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _InputIterator, _ForwardIterator>();
+
         return oneapi::dpl::__internal::__pattern_walk2_n(
             ::std::forward<_ExecutionPolicy>(__exec), __first, __n, __result,
             oneapi::dpl::__internal::__op_uninitialized_move<_DecayedExecutionPolicy>{}, __is_vector,
@@ -162,8 +166,6 @@ uninitialized_fill(_ExecutionPolicy&& __exec, _ForwardIterator __first, _Forward
 
     constexpr auto __is_parallel =
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>();
-    constexpr auto __is_vector =
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>();
 
     if constexpr (::std::is_arithmetic_v<_ValueType>)
     {
@@ -174,6 +176,9 @@ uninitialized_fill(_ExecutionPolicy&& __exec, _ForwardIterator __first, _Forward
     }
     else
     {
+        constexpr auto __is_vector =
+            oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>();
+
         oneapi::dpl::__internal::__pattern_walk1(
             ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
             oneapi::dpl::__internal::__op_uninitialized_fill<_Tp, _DecayedExecutionPolicy>{__value}, __is_vector,
@@ -190,8 +195,6 @@ uninitialized_fill_n(_ExecutionPolicy&& __exec, _ForwardIterator __first, _Size
 
     constexpr auto __is_parallel =
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>();
-    constexpr auto __is_vector =
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>();
 
     if constexpr (::std::is_arithmetic_v<_ValueType>)
     {
@@ -202,6 +205,9 @@ uninitialized_fill_n(_ExecutionPolicy&& __exec, _ForwardIterator __first, _Size
     }
     else
     {
+        constexpr auto __is_vector =
+            oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>();
+
         return oneapi::dpl::__internal::__pattern_walk1_n(
             ::std::forward<_ExecutionPolicy>(__exec), __first, __n,
             oneapi::dpl::__internal::__op_uninitialized_fill<_Tp, _DecayedExecutionPolicy>{__value}, __is_vector,
@@ -220,19 +226,20 @@ destroy(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __
 
     constexpr auto __is_parallel =
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>();
-    using _is_vector_type =
+
+    if constexpr (!::std::is_trivially_destructible_v<_ValueType>)
+    {
+        using _is_vector_type =
 #if (_PSTL_ICPX_OMP_SIMD_DESTROY_WINDOWS_BROKEN || _ONEDPL_ICPX_OMP_SIMD_DESTROY_WINDOWS_BROKEN)
-        ::std::conditional_t<
-            oneapi::dpl::__internal::__is_host_execution_policy<::std::decay_t<_ExecutionPolicy>>::value,
-            ::std::false_type,
-            decltype(oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>())>;
+            ::std::conditional_t<
+                oneapi::dpl::__internal::__is_host_execution_policy<::std::decay_t<_ExecutionPolicy>>::value,
+                ::std::false_type,
+                decltype(oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>())>;
 #else
-        decltype(oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>());
+            decltype(oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>());
 #endif // _PSTL_ICPX_OMP_SIMD_DESTROY_WINDOWS_BROKEN || _ONEDPL_ICPX_OMP_SIMD_DESTROY_WINDOWS_BROKEN
-    constexpr _is_vector_type __is_vector;
+        constexpr _is_vector_type __is_vector;
 
-    if constexpr (!::std::is_trivially_destructible_v<_ValueType>)
-    {
         oneapi::dpl::__internal::__pattern_walk1(
             ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
             [](_ReferenceType __val) { __val.~_ValueType(); }, __is_vector, __is_parallel);
@@ -246,25 +253,25 @@ destroy_n(_ExecutionPolicy&& __exec, _ForwardIterator __first, _Size __n)
     typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _ValueType;
     typedef typename ::std::iterator_traits<_ForwardIterator>::reference _ReferenceType;
 
-    constexpr auto __is_parallel =
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>();
-    using _is_vector_type =
-#if (_PSTL_ICPX_OMP_SIMD_DESTROY_WINDOWS_BROKEN || _ONEDPL_ICPX_OMP_SIMD_DESTROY_WINDOWS_BROKEN)
-        ::std::conditional_t<
-            oneapi::dpl::__internal::__is_host_execution_policy<::std::decay_t<_ExecutionPolicy>>::value,
-            ::std::false_type,
-            decltype(oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>())>;
-#else
-        decltype(oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>());
-#endif // _PSTL_ICPX_OMP_SIMD_DESTROY_WINDOWS_BROKEN || _ONEDPL_ICPX_OMP_SIMD_DESTROY_WINDOWS_BROKEN
-    constexpr _is_vector_type __is_vector;
-
     if constexpr (::std::is_trivially_destructible_v<_ValueType>)
     {
         return oneapi::dpl::__internal::__pstl_next(__first, __n);
     }
     else
     {
+        constexpr auto __is_parallel =
+            oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>();
+        using _is_vector_type =
+#if (_PSTL_ICPX_OMP_SIMD_DESTROY_WINDOWS_BROKEN || _ONEDPL_ICPX_OMP_SIMD_DESTROY_WINDOWS_BROKEN)
+            ::std::conditional_t<
+                oneapi::dpl::__internal::__is_host_execution_policy<::std::decay_t<_ExecutionPolicy>>::value,
+                ::std::false_type,
+                decltype(oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>())>;
+#else
+            decltype(oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>());
+#endif // _PSTL_ICPX_OMP_SIMD_DESTROY_WINDOWS_BROKEN || _ONEDPL_ICPX_OMP_SIMD_DESTROY_WINDOWS_BROKEN
+        constexpr _is_vector_type __is_vector;
+
         return oneapi::dpl::__internal::__pattern_walk1_n(::std::forward<_ExecutionPolicy>(__exec), __first, __n,
                                                           [](_ReferenceType __val) { __val.~_ValueType(); },
                                                           __is_vector, __is_parallel);
@@ -280,13 +287,13 @@ uninitialized_default_construct(_ExecutionPolicy&& __exec, _ForwardIterator __fi
     typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _ValueType;
     typedef ::std::decay_t<_ExecutionPolicy> _DecayedExecutionPolicy;
 
-    constexpr auto __is_parallel =
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>();
-    constexpr auto __is_vector =
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>();
-
     if constexpr (!::std::is_trivial_v<_ValueType>)
     {
+        constexpr auto __is_parallel =
+            oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>();
+        constexpr auto __is_vector =
+            oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>();
+
         oneapi::dpl::__internal::__pattern_walk1(
             ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
             oneapi::dpl::__internal::__op_uninitialized_default_construct<_DecayedExecutionPolicy>{}, __is_vector,
@@ -301,17 +308,17 @@ uninitialized_default_construct_n(_ExecutionPolicy&& __exec, _ForwardIterator __
     typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _ValueType;
     typedef ::std::decay_t<_ExecutionPolicy> _DecayedExecutionPolicy;
 
-    constexpr auto __is_parallel =
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>();
-    constexpr auto __is_vector =
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>();
-
     if constexpr (::std::is_trivial_v<_ValueType>)
     {
         return oneapi::dpl::__internal::__pstl_next(__first, __n);
     }
     else
     {
+        constexpr auto __is_parallel =
+            oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>();
+        constexpr auto __is_vector =
+            oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>();
+
         return oneapi::dpl::__internal::__pattern_walk1_n(
             ::std::forward<_ExecutionPolicy>(__exec), __first, __n,
             oneapi::dpl::__internal::__op_uninitialized_default_construct<_DecayedExecutionPolicy>{}, __is_vector,
@@ -330,8 +337,6 @@ uninitialized_value_construct(_ExecutionPolicy&& __exec, _ForwardIterator __firs
 
     constexpr auto __is_parallel =
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>();
-    constexpr auto __is_vector =
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>();
 
     if constexpr (::std::is_trivial_v<_ValueType>)
     {
@@ -342,6 +347,9 @@ uninitialized_value_construct(_ExecutionPolicy&& __exec, _ForwardIterator __firs
     }
     else
     {
+        constexpr auto __is_vector =
+            oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>();
+
         oneapi::dpl::__internal::__pattern_walk1(
             ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
             oneapi::dpl::__internal::__op_uninitialized_value_construct<_DecayedExecutionPolicy>{}, __is_vector,
@@ -358,8 +366,6 @@ uninitialized_value_construct_n(_ExecutionPolicy&& __exec, _ForwardIterator __fi
 
     constexpr auto __is_parallel =
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>();
-    constexpr auto __is_vector =
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>();
 
     if constexpr (::std::is_trivial_v<_ValueType>)
     {
@@ -370,6 +376,9 @@ uninitialized_value_construct_n(_ExecutionPolicy&& __exec, _ForwardIterator __fi
     }
     else
     {
+        constexpr auto __is_vector =
+            oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>();
+
         return oneapi::dpl::__internal::__pattern_walk1_n(
             ::std::forward<_ExecutionPolicy>(__exec), __first, __n,
             oneapi::dpl::__internal::__op_uninitialized_value_construct<_DecayedExecutionPolicy>{}, __is_vector,

From 2ef6ba0edef1557f608064277306fbcfd0a37c04 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 8 Feb 2024 17:46:31 +0100
Subject: [PATCH 033/566] Implement __select_backend function as constexpr
 without params

---
 include/oneapi/dpl/pstl/execution_impl.h      | 32 ++++++++++++-------
 include/oneapi/dpl/pstl/glue_algorithm_impl.h |  5 +--
 .../pstl/hetero/dpcpp/execution_sycl_defs.h   | 29 +++++++++++------
 3 files changed, 42 insertions(+), 24 deletions(-)

diff --git a/include/oneapi/dpl/pstl/execution_impl.h b/include/oneapi/dpl/pstl/execution_impl.h
index 27a955551f9..5be0170150d 100644
--- a/include/oneapi/dpl/pstl/execution_impl.h
+++ b/include/oneapi/dpl/pstl/execution_impl.h
@@ -156,30 +156,38 @@ using __tag_type = ::std::conditional_t<
     __internal::__is_random_access_iterator_v<_IteratorTypes...>, __parallel_tag<_IsVector>,
     ::std::conditional_t<__is_forward_iterator_v<_IteratorTypes...>, __parallel_forward_tag, __serial_tag<_IsVector>>>;
 
-template <class... _IteratorTypes>
-__serial_tag<std::false_type>
-__select_backend(oneapi::dpl::execution::sequenced_policy, _IteratorTypes&&...)
+template <class _ExecutionPolicy, class... _IteratorTypes>
+constexpr ::std::enable_if_t<
+    ::std::is_same_v<::std::decay_t<_ExecutionPolicy>, oneapi::dpl::execution::sequenced_policy>,
+    __serial_tag<std::false_type>>
+__select_backend()
 {
     return {};
 }
 
-template <class... _IteratorTypes>
-__serial_tag<__internal::__is_random_access_iterator<_IteratorTypes...>>
-__select_backend(oneapi::dpl::execution::unsequenced_policy, _IteratorTypes&&...)
+template <class _ExecutionPolicy, class... _IteratorTypes>
+constexpr ::std::enable_if_t<
+    ::std::is_same_v<::std::decay_t<_ExecutionPolicy>, oneapi::dpl::execution::unsequenced_policy>,
+    __serial_tag<__internal::__is_random_access_iterator<_IteratorTypes...>>>
+__select_backend()
 {
     return {};
 }
 
-template <class... _IteratorTypes>
-__tag_type<std::false_type, _IteratorTypes...>
-__select_backend(oneapi::dpl::execution::parallel_policy, _IteratorTypes&&...)
+template <class _ExecutionPolicy, class... _IteratorTypes>
+constexpr ::std::enable_if_t<
+    ::std::is_same_v<::std::decay_t<_ExecutionPolicy>, oneapi::dpl::execution::parallel_policy>,
+    __tag_type<std::false_type, _IteratorTypes...>>
+__select_backend()
 {
     return {};
 }
 
-template <class... _IteratorTypes>
-__tag_type<__internal::__is_random_access_iterator<_IteratorTypes...>, _IteratorTypes...>
-__select_backend(oneapi::dpl::execution::parallel_unsequenced_policy, _IteratorTypes&&...)
+template <class _ExecutionPolicy, class... _IteratorTypes>
+constexpr ::std::enable_if_t<
+    ::std::is_same_v<::std::decay_t<_ExecutionPolicy>, oneapi::dpl::execution::parallel_unsequenced_policy>,
+    __tag_type<__internal::__is_random_access_iterator<_IteratorTypes...>, _IteratorTypes...>>
+__select_backend()
 {
     return {};
 }
diff --git a/include/oneapi/dpl/pstl/glue_algorithm_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
index 17c5e331710..c6b0956cb50 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
@@ -97,7 +97,7 @@ template <class _ExecutionPolicy, class _ForwardIterator, class _Predicate>
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _ForwardIterator>
 find_if(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last, _Predicate __pred)
 {
-    auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, _ForwardIterator{});
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
 
     return oneapi::dpl::__internal::__pattern_find_if(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first,
                                                       __last, __pred);
@@ -379,7 +379,8 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy>
 replace_if(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last, _UnaryPredicate __pred,
            const _Tp& __new_value)
 {
-    auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, _ForwardIterator{});
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+
     __pattern_replace_if(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __pred,
                          __new_value);
 }
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h b/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h
index 8c91c597c31..9f72220225a 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h
@@ -322,26 +322,35 @@ struct __device_backend_tag
 {
 };
 
-template <class... _IteratorTypes, typename _KernelName>
-::std::enable_if_t<__is_random_access_iterator_v<_IteratorTypes...>, __hetero_tag<__device_backend_tag>>
-__select_backend(const execution::device_policy<_KernelName>&, _IteratorTypes&&...)
+template <typename HeteroPolicy>
+struct __select_backend_for_hetero_policy_trait;
+
+template <typename KernelName>
+struct __select_backend_for_hetero_policy_trait<execution::device_policy<KernelName>>
 {
-    return {};
-}
+    using __backend_tag = __device_backend_tag;
+};
 
 #if _ONEDPL_FPGA_DEVICE
 struct __fpga_backend_tag : __device_backend_tag
 {
 };
 
-template <class... _IteratorTypes, unsigned int _Factor, typename _KernelName>
-::std::enable_if_t<__is_random_access_iterator_v<_IteratorTypes...>, __hetero_tag<__fpga_backend_tag>>
-__select_backend(const execution::fpga_policy<_Factor, _KernelName>&, _IteratorTypes&&...)
+template <unsigned int factor, typename KernelName>
+struct __select_backend_for_hetero_policy_trait<execution::fpga_policy<factor, KernelName>>
 {
-    return {};
-}
+    using __backend_tag = __fpga_backend_tag;
+};
 #endif
 
+template <class _ExecutionPolicy, class... _IteratorTypes>
+constexpr ::std::enable_if_t<
+    __is_random_access_iterator_v<_IteratorTypes...>,
+    __hetero_tag<typename __select_backend_for_hetero_policy_trait<::std::decay_t<_ExecutionPolicy>>::__backend_tag>>
+__select_backend()
+{
+    return {}; // return __hetero_tag<__device_backend_tag> or __hetero_tag<__fpga_backend_tag>
+}
 
 template <class _Tag>
 struct __is_backend_tag;

From a3dfcc00ec22d36d18a8f2ec85eda18fe5ba3f13 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 30 Jan 2024 12:56:51 +0100
Subject: [PATCH 034/566] Replace __internal::__lazy_and to
 ::std::conjunction_v; remove __internal::__lazy_and and __internal::__lazy_or
 as not required anymore

---
 include/oneapi/dpl/pstl/execution_impl.h | 40 ++++--------------------
 1 file changed, 6 insertions(+), 34 deletions(-)

diff --git a/include/oneapi/dpl/pstl/execution_impl.h b/include/oneapi/dpl/pstl/execution_impl.h
index 5be0170150d..ad29a81b0c5 100644
--- a/include/oneapi/dpl/pstl/execution_impl.h
+++ b/include/oneapi/dpl/pstl/execution_impl.h
@@ -29,34 +29,6 @@ namespace dpl
 namespace __internal
 {
 
-/* predicate */
-
-template <typename _Tp>
-::std::false_type __lazy_and(_Tp, ::std::false_type)
-{
-    return ::std::false_type{};
-}
-
-template <typename _Tp>
-inline _Tp
-__lazy_and(_Tp __a, ::std::true_type)
-{
-    return __a;
-}
-
-template <typename _Tp>
-::std::true_type __lazy_or(_Tp, ::std::true_type)
-{
-    return ::std::true_type{};
-}
-
-template <typename _Tp>
-inline _Tp
-__lazy_or(_Tp __a, ::std::false_type)
-{
-    return __a;
-}
-
 /* policy */
 template <typename Policy>
 struct __policy_traits
@@ -105,19 +77,19 @@ template <typename _ExecutionPolicy>
 using __allow_parallel = typename __internal::__policy_traits<::std::decay_t<_ExecutionPolicy>>::__allow_parallel;
 
 template <typename _ExecutionPolicy, typename... _IteratorTypes>
-constexpr decltype(auto)
+constexpr auto
 __is_vectorization_preferred()
 {
-    return __internal::__lazy_and(::std::decay_t<_ExecutionPolicy>::__allow_vector(),
-                                  __internal::__is_random_access_iterator_t<_IteratorTypes...>());
+    return ::std::conjunction<decltype(::std::decay_t<_ExecutionPolicy>::__allow_vector()),
+                              __internal::__is_random_access_iterator_t<_IteratorTypes...>>();
 }
 
 template <typename _ExecutionPolicy, typename... _IteratorTypes>
-constexpr decltype(auto)
+constexpr auto
 __is_parallelization_preferred()
 {
-    return __internal::__lazy_and(::std::decay_t<_ExecutionPolicy>::__allow_parallel(),
-                                  __internal::__is_random_access_iterator_t<_IteratorTypes...>());
+    return ::std::conjunction<decltype(::std::decay_t<_ExecutionPolicy>::__allow_parallel()),
+                              __internal::__is_random_access_iterator_t<_IteratorTypes...>>();
 }
 
 //------------------------------------------------------------------------

From cb79db2abec30ea77fb14c32e9ed26b30cace26c Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 31 Jan 2024 15:57:58 +0100
Subject: [PATCH 035/566] include/oneapi/dpl/pstl/algorithm_impl.h - formatting

---
 include/oneapi/dpl/pstl/algorithm_impl.h | 1 +
 1 file changed, 1 insertion(+)

diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 7b02f71722d..5bd488a0580 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -753,6 +753,7 @@ __pattern_find_if(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _F
                   _ForwardIterator __last, _Predicate __pred)
 {
     using __backend_tag = typename decltype(__tag)::__backend_tag;
+
     return __except_handler([&]() {
         return __parallel_find(
             __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,

From f075fe83bce8226b5f8968a0f20f9b93b5108050 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 8 Feb 2024 17:03:41 +0100
Subject: [PATCH 036/566] Add forward declarations for struct __parallel_tag,
 struct __parallel_forward_tag

---
 include/oneapi/dpl/pstl/algorithm_fwd.h | 5 +++++
 include/oneapi/dpl/pstl/numeric_fwd.h   | 2 ++
 2 files changed, 7 insertions(+)

diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index 34d09140fb0..97adcbe7d85 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -26,6 +26,11 @@ namespace dpl
 namespace __internal
 {
 
+template <class _IsVector>
+struct __parallel_tag;
+
+struct __parallel_forward_tag;
+
 //------------------------------------------------------------------------
 // any_of
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/numeric_fwd.h b/include/oneapi/dpl/pstl/numeric_fwd.h
index c663cd5c2d4..aab7edc0150 100644
--- a/include/oneapi/dpl/pstl/numeric_fwd.h
+++ b/include/oneapi/dpl/pstl/numeric_fwd.h
@@ -25,6 +25,8 @@ namespace dpl
 {
 namespace __internal
 {
+template <class _IsVector>
+struct __parallel_tag;
 
 //------------------------------------------------------------------------
 // transform_reduce (version with two binary functions, according to draft N4659)

From c1d389f62e27db110f58e4e24932fc07d7fec77d Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 9 Feb 2024 10:46:07 +0100
Subject: [PATCH 037/566] Implement tag dispatching prototype :
 __pattern_replace_if + tag impls

---
 include/oneapi/dpl/pstl/algorithm_impl.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 5bd488a0580..66e2a0c3231 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -178,7 +178,7 @@ __pattern_walk1(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _For
 }
 
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _UnaryPredicate, class _Tp>
-oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, void>
+void
 __pattern_replace_if(_Tag __tag, _ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last,
                      _UnaryPredicate __pred, const _Tp& __new_value)
 {

From d4c811406baee65f4554f3e2f6007f734bb398c7 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 9 Feb 2024 11:48:30 +0100
Subject: [PATCH 038/566] Implement tag dispatching prototype : __parallel_for
 + tag impls (omp, serial)

---
 include/oneapi/dpl/pstl/omp/parallel_for.h    | 19 +++++++++++++++++++
 .../oneapi/dpl/pstl/parallel_backend_serial.h |  7 +++++++
 2 files changed, 26 insertions(+)

diff --git a/include/oneapi/dpl/pstl/omp/parallel_for.h b/include/oneapi/dpl/pstl/omp/parallel_for.h
index 5b6ed66453a..296b24e282b 100644
--- a/include/oneapi/dpl/pstl/omp/parallel_for.h
+++ b/include/oneapi/dpl/pstl/omp/parallel_for.h
@@ -66,6 +66,25 @@ __parallel_for(_ExecutionPolicy&&, _Index __first, _Index __last, _Fp __f)
     }
 }
 
+template <class _ExecutionPolicy, class _Index, class _Fp>
+void
+__parallel_for(oneapi::dpl::__internal::__tbb_backend_tag, _ExecutionPolicy&&, _Index __first, _Index __last, _Fp __f)
+{
+    if (omp_in_parallel())
+    {
+        // we don't create a nested parallel region in an existing parallel
+        // region: just create tasks
+        oneapi::dpl::__omp_backend::__parallel_for_body(__first, __last, __f);
+    }
+    else
+    {
+        // in any case (nested or non-nested) one parallel region is created and
+        // only one thread creates a set of tasks
+        _PSTL_PRAGMA(omp parallel)
+        _PSTL_PRAGMA(omp single nowait) { oneapi::dpl::__omp_backend::__parallel_for_body(__first, __last, __f); }
+    }
+}
+
 } // namespace __omp_backend
 } // namespace dpl
 } // namespace oneapi
diff --git a/include/oneapi/dpl/pstl/parallel_backend_serial.h b/include/oneapi/dpl/pstl/parallel_backend_serial.h
index a2dd6468a34..a80f36734ed 100644
--- a/include/oneapi/dpl/pstl/parallel_backend_serial.h
+++ b/include/oneapi/dpl/pstl/parallel_backend_serial.h
@@ -71,6 +71,13 @@ __parallel_for(_ExecutionPolicy&&, _Index __first, _Index __last, _Fp __f)
     __f(__first, __last);
 }
 
+template <class _ExecutionPolicy, class _Index, class _Fp>
+void
+__parallel_for(oneapi::dpl::__internal::__tbb_backend_tag, _ExecutionPolicy&&, _Index __first, _Index __last, _Fp __f)
+{
+    __f(__first, __last);
+}
+
 template <class _ExecutionPolicy, class _Value, class _Index, typename _RealBody, typename _Reduction>
 _Value
 __parallel_reduce(_ExecutionPolicy&&, _Index __first, _Index __last, const _Value& __identity,

From f2962bcbe31262091aa8193a09311ac5a300ed22 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 9 Feb 2024 11:52:17 +0100
Subject: [PATCH 039/566] Implement tag dispatching prototype :
 __parallel_for_each + tag impls (omp, serial)

---
 .../oneapi/dpl/pstl/omp/parallel_for_each.h   | 20 +++++++++++++++++++
 .../oneapi/dpl/pstl/parallel_backend_serial.h |  9 +++++++++
 2 files changed, 29 insertions(+)

diff --git a/include/oneapi/dpl/pstl/omp/parallel_for_each.h b/include/oneapi/dpl/pstl/omp/parallel_for_each.h
index 7877ef095ef..510cd04c352 100644
--- a/include/oneapi/dpl/pstl/omp/parallel_for_each.h
+++ b/include/oneapi/dpl/pstl/omp/parallel_for_each.h
@@ -61,6 +61,26 @@ __parallel_for_each(_ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterat
     }
 }
 
+template <class _ExecutionPolicy, class _ForwardIterator, class _Fp>
+void
+__parallel_for_each(oneapi::dpl::__internal::__tbb_backend_tag, _ExecutionPolicy&&, _ForwardIterator __first,
+                    _ForwardIterator __last, _Fp __f)
+{
+    if (omp_in_parallel())
+    {
+        // we don't create a nested parallel region in an existing parallel
+        // region: just create tasks
+        oneapi::dpl::__omp_backend::__parallel_for_each_body(__first, __last, __f);
+    }
+    else
+    {
+        // in any case (nested or non-nested) one parallel region is created and
+        // only one thread creates a set of tasks
+        _PSTL_PRAGMA(omp parallel)
+        _PSTL_PRAGMA(omp single nowait) { oneapi::dpl::__omp_backend::__parallel_for_each_body(__first, __last, __f); }
+    }
+}
+
 } // namespace __omp_backend
 } // namespace dpl
 } // namespace oneapi
diff --git a/include/oneapi/dpl/pstl/parallel_backend_serial.h b/include/oneapi/dpl/pstl/parallel_backend_serial.h
index a80f36734ed..577fcbd36a0 100644
--- a/include/oneapi/dpl/pstl/parallel_backend_serial.h
+++ b/include/oneapi/dpl/pstl/parallel_backend_serial.h
@@ -155,6 +155,15 @@ __parallel_for_each(_ExecutionPolicy&&, _ForwardIterator __begin, _ForwardIterat
         __f(*__iter);
 }
 
+template <class _ExecutionPolicy, class _ForwardIterator, class _Fp>
+void
+__parallel_for_each(oneapi::dpl::__internal::__tbb_backend_tag, _ExecutionPolicy&&, _ForwardIterator __begin,
+                    _ForwardIterator __end, _Fp __f)
+{
+    for (auto __iter = __begin; __iter != __end; ++__iter)
+        __f(*__iter);
+}
+
 } // namespace __serial_backend
 } // namespace dpl
 } // namespace oneapi

From 5cb677f96a771533374cae3405b90de15e90a56c Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 31 Jan 2024 15:50:11 +0100
Subject: [PATCH 040/566] __pattern_walk1 + tag impls

---
 include/oneapi/dpl/pstl/algorithm_fwd.h  | 14 ++++++++++++++
 include/oneapi/dpl/pstl/algorithm_impl.h | 13 +++++++++----
 2 files changed, 23 insertions(+), 4 deletions(-)

diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index 97adcbe7d85..22cd38af956 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -74,6 +74,10 @@ oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>
 __pattern_walk1(_ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _Function, _IsVector,
                 /*parallel=*/::std::false_type) noexcept;
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Function, class _IsVector>
+void
+__pattern_walk1(_Tag, _ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _Function) noexcept;
+
 template <class _ExecutionPolicy, class _RandomAccessIterator, class _Function, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy_conditional<
     _ExecutionPolicy, __is_random_access_iterator_v<_RandomAccessIterator>>
@@ -81,6 +85,11 @@ __pattern_walk1(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _Rando
                 _IsVector __is_vector,
                 /*parallel=*/::std::true_type);
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Function>
+void
+__pattern_walk1(__parallel_forward_tag, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
+                _RandomAccessIterator __last, _Function __f);
+
 template <class _ExecutionPolicy, class _RandomAccessIterator, class _Function, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy_conditional<
     _ExecutionPolicy, !__is_random_access_iterator_v<_RandomAccessIterator>>
@@ -88,6 +97,11 @@ __pattern_walk1(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _Rando
                 _IsVector __is_vector,
                 /*parallel=*/::std::true_type);
 
+template <class _IsVector, class _ExecutionPolicy, class _ForwardIterator, class _Function>
+void
+__pattern_walk1(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last,
+                _Function __f);
+
 template <class _ExecutionPolicy, class _ForwardIterator, class _Brick>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>
 __pattern_walk_brick(_ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _Brick,
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 66e2a0c3231..0b1a538378e 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -138,6 +138,8 @@ template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Fun
 void
 __pattern_walk1(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _Function __f) noexcept
 {
+    static_assert(__is_backend_tag_serial_v<_Tag>);
+
     __internal::__brick_walk1(__first, __last, __f, typename _Tag::__is_vector{});
 }
 
@@ -155,20 +157,23 @@ void
 __pattern_walk1(__parallel_forward_tag, _ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last,
                 _Function __f)
 {
+    using __backend_tag = typename __parallel_forward_tag::__backend_tag;
+
     typedef typename ::std::iterator_traits<_ForwardIterator>::reference _ReferenceType;
     auto __func = [&__f](_ReferenceType arg) { __f(arg); };
     __internal::__except_handler([&]() {
-        __par_backend::__parallel_for_each(__parallel_forward_tag::__backend_tag{},
-                                           ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __func);
+        __par_backend::__parallel_for_each(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+                                           __func);
     });
 }
 
 template <class _IsVector, class _ExecutionPolicy, class _ForwardIterator, class _Function>
 void
-__pattern_walk1(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _ForwardIterator __first,
+__pattern_walk1(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _ForwardIterator __first,
                 _ForwardIterator __last, _Function __f)
 {
-    using __backend_tag = typename decltype(__tag)::__backend_tag;
+    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
+
     __internal::__except_handler([&]() {
         __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
                                       [__f](_ForwardIterator __i, _ForwardIterator __j) {

From 1fbeea54ca8441c17dfe65820f07a8d206e41ded Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 1 Feb 2024 18:19:04 +0100
Subject: [PATCH 041/566] __pattern_walk1 + tag calls

---
 include/oneapi/dpl/pstl/algorithm_impl.h      |  6 +-
 include/oneapi/dpl/pstl/glue_algorithm_impl.h |  8 +-
 include/oneapi/dpl/pstl/glue_memory_impl.h    | 78 ++++++++++---------
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 29 ++++---
 4 files changed, 70 insertions(+), 51 deletions(-)

diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 0b1a538378e..4a4396be8ee 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -280,8 +280,10 @@ __pattern_walk1_n(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _Siz
                   _IsVector __is_vector,
                   /*is_parallel=*/::std::true_type)
 {
-    oneapi::dpl::__internal::__pattern_walk1(::std::forward<_ExecutionPolicy>(__exec), __first, __first + __n, __f,
-                                             __is_vector, ::std::true_type());
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
+
+    oneapi::dpl::__internal::__pattern_walk1(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first,
+                                             __first + __n, __f);
     return __first + __n;
 }
 
diff --git a/include/oneapi/dpl/pstl/glue_algorithm_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
index c6b0956cb50..dd7eec8fcd2 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
@@ -75,10 +75,10 @@ template <class _ExecutionPolicy, class _ForwardIterator, class _Function>
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy>
 for_each(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last, _Function __f)
 {
-    oneapi::dpl::__internal::__pattern_walk1(
-        ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __f,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(),
-        __exec.__allow_parallel());
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+
+    oneapi::dpl::__internal::__pattern_walk1(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+                                             __f);
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator, class _Size, class _Function>
diff --git a/include/oneapi/dpl/pstl/glue_memory_impl.h b/include/oneapi/dpl/pstl/glue_memory_impl.h
index e810e20bf9a..667a129825b 100644
--- a/include/oneapi/dpl/pstl/glue_memory_impl.h
+++ b/include/oneapi/dpl/pstl/glue_memory_impl.h
@@ -164,11 +164,11 @@ uninitialized_fill(_ExecutionPolicy&& __exec, _ForwardIterator __first, _Forward
     typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _ValueType;
     typedef ::std::decay_t<_ExecutionPolicy> _DecayedExecutionPolicy;
 
-    constexpr auto __is_parallel =
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>();
-
     if constexpr (::std::is_arithmetic_v<_ValueType>)
     {
+        constexpr auto __is_parallel =
+            oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>();
+
         oneapi::dpl::__internal::__pattern_walk_brick(
             ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
             oneapi::dpl::__internal::__brick_fill<_ValueType, _DecayedExecutionPolicy>{_ValueType(__value)},
@@ -176,13 +176,12 @@ uninitialized_fill(_ExecutionPolicy&& __exec, _ForwardIterator __first, _Forward
     }
     else
     {
-        constexpr auto __is_vector =
-            oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>();
+        constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
 
         oneapi::dpl::__internal::__pattern_walk1(
+            __dispatch_tag,
             ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
-            oneapi::dpl::__internal::__op_uninitialized_fill<_Tp, _DecayedExecutionPolicy>{__value}, __is_vector,
-            __is_parallel);
+            oneapi::dpl::__internal::__op_uninitialized_fill<_Tp, _DecayedExecutionPolicy>{__value});
     }
 }
 
@@ -215,6 +214,26 @@ uninitialized_fill_n(_ExecutionPolicy&& __exec, _ForwardIterator __first, _Size
     }
 }
 
+#if (_PSTL_ICPX_OMP_SIMD_DESTROY_WINDOWS_BROKEN || _ONEDPL_ICPX_OMP_SIMD_DESTROY_WINDOWS_BROKEN)
+template <class _ExecutionPolicy>
+struct EvaDestroylWorkaroundPolicy
+{
+    using ExecutionPolicy = _ExecutionPolicy;
+};
+
+template <>
+struct EvaDestroylWorkaroundPolicy<oneapi::dpl::execution::parallel_unsequenced_policy>
+{
+    using ExecutionPolicy = oneapi::dpl::execution::parallel_policy;
+};
+
+template <>
+struct EvaDestroylWorkaroundPolicy<oneapi::dpl::execution::unsequenced_policy>
+{
+    using ExecutionPolicy = oneapi::dpl::execution::sequenced_policy;
+};
+#endif // _PSTL_ICPX_OMP_SIMD_DESTROY_WINDOWS_BROKEN || _ONEDPL_ICPX_OMP_SIMD_DESTROY_WINDOWS_BROKEN
+
 // [specialized.destroy]
 
 template <class _ExecutionPolicy, class _ForwardIterator>
@@ -224,25 +243,20 @@ destroy(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __
     typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _ValueType;
     typedef typename ::std::iterator_traits<_ForwardIterator>::reference _ReferenceType;
 
-    constexpr auto __is_parallel =
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>();
-
     if constexpr (!::std::is_trivially_destructible_v<_ValueType>)
     {
-        using _is_vector_type =
+        using _ExecutionPolicyDest =
 #if (_PSTL_ICPX_OMP_SIMD_DESTROY_WINDOWS_BROKEN || _ONEDPL_ICPX_OMP_SIMD_DESTROY_WINDOWS_BROKEN)
-            ::std::conditional_t<
-                oneapi::dpl::__internal::__is_host_execution_policy<::std::decay_t<_ExecutionPolicy>>::value,
-                ::std::false_type,
-                decltype(oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>())>;
+            typename EvaDestroylWorkaroundPolicy<::std::decay_t<_ExecutionPolicy>>::ExecutionPolicy;
 #else
-            decltype(oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>());
+            _ExecutionPolicy;
 #endif // _PSTL_ICPX_OMP_SIMD_DESTROY_WINDOWS_BROKEN || _ONEDPL_ICPX_OMP_SIMD_DESTROY_WINDOWS_BROKEN
-        constexpr _is_vector_type __is_vector;
 
-        oneapi::dpl::__internal::__pattern_walk1(
-            ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
-            [](_ReferenceType __val) { __val.~_ValueType(); }, __is_vector, __is_parallel);
+        constexpr auto __dispatch_tag =
+                oneapi::dpl::__internal::__select_backend<_ExecutionPolicyDest, _ForwardIterator>();
+
+        oneapi::dpl::__internal::__pattern_walk1(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first,
+                                                 __last, [](_ReferenceType __val) { __val.~_ValueType(); });
     }
 }
 
@@ -289,15 +303,11 @@ uninitialized_default_construct(_ExecutionPolicy&& __exec, _ForwardIterator __fi
 
     if constexpr (!::std::is_trivial_v<_ValueType>)
     {
-        constexpr auto __is_parallel =
-            oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>();
-        constexpr auto __is_vector =
-            oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>();
+        constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
 
         oneapi::dpl::__internal::__pattern_walk1(
-            ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
-            oneapi::dpl::__internal::__op_uninitialized_default_construct<_DecayedExecutionPolicy>{}, __is_vector,
-            __is_parallel);
+            __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+            oneapi::dpl::__internal::__op_uninitialized_default_construct<_DecayedExecutionPolicy>{});
     }
 }
 
@@ -335,11 +345,11 @@ uninitialized_value_construct(_ExecutionPolicy&& __exec, _ForwardIterator __firs
     typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _ValueType;
     typedef ::std::decay_t<_ExecutionPolicy> _DecayedExecutionPolicy;
 
-    constexpr auto __is_parallel =
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>();
-
     if constexpr (::std::is_trivial_v<_ValueType>)
     {
+        constexpr auto __is_parallel =
+            oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>();
+
         oneapi::dpl::__internal::__pattern_walk_brick(
             ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
             oneapi::dpl::__internal::__brick_fill<_ValueType, _DecayedExecutionPolicy>{_ValueType()},
@@ -347,13 +357,11 @@ uninitialized_value_construct(_ExecutionPolicy&& __exec, _ForwardIterator __firs
     }
     else
     {
-        constexpr auto __is_vector =
-            oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>();
+        constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
 
         oneapi::dpl::__internal::__pattern_walk1(
-            ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
-            oneapi::dpl::__internal::__op_uninitialized_value_construct<_DecayedExecutionPolicy>{}, __is_vector,
-            __is_parallel);
+            __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+            oneapi::dpl::__internal::__op_uninitialized_value_construct<_DecayedExecutionPolicy>{});
     }
 }
 
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 244980ecff8..dac24712d4d 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -85,8 +85,9 @@ oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, _
 __pattern_walk1_n(_ExecutionPolicy&& __exec, _ForwardIterator __first, _Size __n, _Function __f,
                   /*vector=*/::std::true_type, /*parallel=*/::std::true_type)
 {
-    __pattern_walk1(::std::forward<_ExecutionPolicy>(__exec), __first, __first + __n, __f,
-                    /*vector=*/::std::true_type(), /*parallel=*/::std::true_type());
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+
+    __pattern_walk1(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __first + __n, __f);
     return __first + __n;
 }
 
@@ -201,10 +202,12 @@ __pattern_walk_brick(_ExecutionPolicy&& __exec, _ForwardIterator __first, _Forwa
     if (__last - __first <= 0)
         return;
 
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+
     __pattern_walk1(
+        __dispatch_tag,
         __par_backend_hetero::make_wrapped_policy<__walk_brick_wrapper>(::std::forward<_ExecutionPolicy>(__exec)),
-        __first, __last, __f,
-        /*vector=*/::std::true_type{}, /*parallel=*/::std::true_type{});
+        __first, __last, __f);
 }
 
 template <typename _Name>
@@ -217,10 +220,12 @@ oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, _
 __pattern_walk_brick_n(_ExecutionPolicy&& __exec, _ForwardIterator __first, _Size __n, _Function __f,
                        /*parallel=*/::std::true_type)
 {
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+
     __pattern_walk1(
+        __dispatch_tag,
         __par_backend_hetero::make_wrapped_policy<__walk_brick_n_wrapper>(::std::forward<_ExecutionPolicy>(__exec)),
-        __first, __first + __n, __f,
-        /*vector=*/::std::true_type{}, /*parallel=*/::std::true_type{});
+        __first, __first + __n, __f);
     return __first + __n;
 }
 
@@ -332,10 +337,12 @@ oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, _
 __pattern_fill(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last, const _T& __value,
                /*vector=*/::std::true_type, /*parallel=*/::std::true_type)
 {
-    __pattern_walk1(::std::forward<_ExecutionPolicy>(__exec),
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+
+    __pattern_walk1(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
                     __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::write>(__first),
                     __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::write>(__last),
-                    fill_functor<_T>{__value}, ::std::true_type{}, ::std::true_type{});
+                    fill_functor<_T>{__value});
     return __last;
 }
 
@@ -361,10 +368,12 @@ oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, _
 __pattern_generate(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last, _Generator __g,
                    /*vector=*/::std::true_type, /*parallel=*/::std::true_type)
 {
-    __pattern_walk1(::std::forward<_ExecutionPolicy>(__exec),
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+
+    __pattern_walk1(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
                     __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::write>(__first),
                     __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::write>(__last),
-                    generate_functor<_Generator>{__g}, ::std::true_type{}, ::std::true_type{});
+                    generate_functor<_Generator>{__g});
     return __last;
 }
 

From cbb14735ff79a4dbd3e1d1ac033124a9cd30bc98 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 9 Feb 2024 12:26:30 +0100
Subject: [PATCH 042/566] __pattern_walk2 + tag impls

---
 include/oneapi/dpl/pstl/algorithm_fwd.h       | 15 ++++++
 include/oneapi/dpl/pstl/algorithm_impl.h      | 53 +++++++++++++++++++
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 28 ++++++++++
 3 files changed, 96 insertions(+)

diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index 22cd38af956..bc9d87b767a 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -171,6 +171,10 @@ oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _Fo
 __pattern_walk2(_ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2, _Function, _IsVector,
                 /*parallel=*/::std::false_type) noexcept;
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Function>
+_ForwardIterator2
+__pattern_walk2(_Tag, _ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2, _Function) noexcept;
+
 template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _Function,
           class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy_conditional<
@@ -179,12 +183,23 @@ oneapi::dpl::__internal::__enable_if_host_execution_policy_conditional<
 __pattern_walk2(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1, _RandomAccessIterator1 __last1,
                 _RandomAccessIterator2 __first2, _Function __f, _IsVector __is_vector, /*parallel=*/::std::true_type);
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
+          class _Function>
+_RandomAccessIterator2
+__pattern_walk2(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator1, _RandomAccessIterator1,
+                _RandomAccessIterator2, _Function);
+
 template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Function, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy_conditional<
     _ExecutionPolicy, !__is_random_access_iterator_v<_ForwardIterator1, _ForwardIterator2>, _ForwardIterator2>
 __pattern_walk2(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
                 _ForwardIterator2 __first2, _Function __f, _IsVector __is_vector, /*parallel=*/::std::true_type);
 
+template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Function>
+_ForwardIterator2
+__pattern_walk2(__parallel_forward_tag, _ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2,
+                _Function);
+
 template <class _ExecutionPolicy, class _ForwardIterator1, class _Size, class _ForwardIterator2, class _Function,
           class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _ForwardIterator2>
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 4a4396be8ee..782a0763047 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -361,6 +361,16 @@ __pattern_walk2(_ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator
     return __internal::__brick_walk2(__first1, __last1, __first2, __f, __is_vector);
 }
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Function>
+_ForwardIterator2
+__pattern_walk2(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
+                _ForwardIterator2 __first2, _Function __f) noexcept
+{
+    static_assert(__is_backend_tag_serial_v<_Tag>);
+
+    return __internal::__brick_walk2(__first1, __last1, __first2, __f, typename _Tag::__is_vector{});
+}
+
 template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _Function,
           class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy_conditional<
@@ -379,6 +389,21 @@ __pattern_walk2(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1, _Ran
     });
 }
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _Function>
+_RandomAccessIterator2
+__pattern_walk2(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1,
+                _RandomAccessIterator1 __last1, _RandomAccessIterator2 __first2, _Function __f)
+{
+    return __internal::__except_handler([&]() {
+        __par_backend::__parallel_for(
+            ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1,
+            [__f, __first1, __first2](_RandomAccessIterator1 __i, _RandomAccessIterator1 __j) {
+                __internal::__brick_walk2(__i, __j, __first2 + (__i - __first1), __f, _IsVector{});
+            });
+        return __first2 + (__last1 - __first1);
+    });
+}
+
 template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Function, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy_conditional<
     _ExecutionPolicy, !__is_random_access_iterator_v<_ForwardIterator1, _ForwardIterator2>, _ForwardIterator2>
@@ -406,6 +431,34 @@ __pattern_walk2(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardI
     });
 }
 
+template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Function>
+_ForwardIterator2
+__pattern_walk2(__parallel_forward_tag, _ExecutionPolicy&& __exec, _ForwardIterator1 __first1,
+                _ForwardIterator1 __last1, _ForwardIterator2 __first2, _Function __f)
+{
+    using __backend_tag = typename __parallel_forward_tag::__backend_tag;
+
+    return __internal::__except_handler([&]() {
+        using _iterator_tuple = zip_forward_iterator<_ForwardIterator1, _ForwardIterator2>;
+        auto __begin = _iterator_tuple(__first1, __first2);
+        auto __end = _iterator_tuple(__last1, /*dummy parameter*/ _ForwardIterator2());
+
+        typedef typename ::std::iterator_traits<_ForwardIterator1>::reference _ReferenceType1;
+        typedef typename ::std::iterator_traits<_ForwardIterator2>::reference _ReferenceType2;
+
+        __par_backend::__parallel_for_each(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __begin, __end,
+                                           [&__f](::std::tuple<_ReferenceType1, _ReferenceType2> __val) {
+                                               __f(::std::get<0>(__val), ::std::get<1>(__val));
+                                           });
+
+        //TODO: parallel_for_each does not allow to return correct iterator value according to the ::std::transform
+        // implementation. Therefore, iterator value is calculated separately.
+        for (; __begin != __end; ++__begin)
+            ;
+        return ::std::get<1>(__begin.base());
+    });
+}
+
 template <class _ExecutionPolicy, class _ForwardIterator1, class _Size, class _ForwardIterator2, class _Function,
           class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _ForwardIterator2>
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index dac24712d4d..7c9782e0acf 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -126,6 +126,34 @@ __pattern_walk2(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardI
     return __first2 + __n;
 }
 
+template <typename _BackendTag, typename _IsSync = ::std::true_type,
+          __par_backend_hetero::access_mode __acc_mode1 = __par_backend_hetero::access_mode::read,
+          __par_backend_hetero::access_mode __acc_mode2 = __par_backend_hetero::access_mode::write,
+          typename _ExecutionPolicy, typename _ForwardIterator1, typename _ForwardIterator2, typename _Function>
+_ForwardIterator2
+__pattern_walk2(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _ForwardIterator1 __first1,
+                _ForwardIterator1 __last1, _ForwardIterator2 __first2, _Function __f)
+{
+    auto __n = __last1 - __first1;
+    if (__n <= 0)
+        return __first2;
+
+    auto __keep1 = oneapi::dpl::__ranges::__get_sycl_range<__acc_mode1, _ForwardIterator1>();
+    auto __buf1 = __keep1(__first1, __last1);
+
+    auto __keep2 = oneapi::dpl::__ranges::__get_sycl_range<__acc_mode2, _ForwardIterator2>();
+    auto __buf2 = __keep2(__first2, __first2 + __n);
+
+    auto __future_obj = oneapi::dpl::__par_backend_hetero::__parallel_for(
+        ::std::forward<_ExecutionPolicy>(__exec), unseq_backend::walk_n<_ExecutionPolicy, _Function>{__f}, __n,
+        __buf1.all_view(), __buf2.all_view());
+
+    if constexpr (_IsSync())
+        __future_obj.wait();
+
+    return __first2 + __n;
+}
+
 template <typename _ExecutionPolicy, typename _ForwardIterator1, typename _Size, typename _ForwardIterator2,
           typename _Function>
 oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, _ForwardIterator2>

From 9ed3e7f0ed4a1bac5131eb5b24727659d179393c Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 1 Feb 2024 09:58:00 +0100
Subject: [PATCH 043/566] __pattern_walk2 + tag calls

---
 include/oneapi/dpl/pstl/algorithm_impl.h      |  6 ++-
 include/oneapi/dpl/pstl/glue_algorithm_impl.h | 15 +++---
 include/oneapi/dpl/pstl/glue_memory_impl.h    | 28 +++++-----
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 54 +++++++++++++------
 4 files changed, 62 insertions(+), 41 deletions(-)

diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 782a0763047..b1f7c2c01ae 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -474,8 +474,10 @@ oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _Ra
 __pattern_walk2_n(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1, _Size __n,
                   _RandomAccessIterator2 __first2, _Function __f, _IsVector __is_vector, /*parallel=*/::std::true_type)
 {
-    return __internal::__pattern_walk2(::std::forward<_ExecutionPolicy>(__exec), __first1, __first1 + __n, __first2,
-                                       __f, __is_vector, ::std::true_type());
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2>();
+
+    return __internal::__pattern_walk2(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first1,
+                                       __first1 + __n, __first2, __f);
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Brick>
diff --git a/include/oneapi/dpl/pstl/glue_algorithm_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
index dd7eec8fcd2..905708d9fe9 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
@@ -321,11 +321,12 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _Forward
 transform(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _ForwardIterator1 __last, _ForwardIterator2 __result,
           _UnaryOperation __op)
 {
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
+
     return oneapi::dpl::__internal::__pattern_walk2(
+        __dispatch_tag, 
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result,
-        oneapi::dpl::__internal::__transform_functor<_UnaryOperation>{::std::move(__op)},
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(),
-        __exec.__allow_parallel());
+        oneapi::dpl::__internal::__transform_functor<_UnaryOperation>{::std::move(__op)});
 }
 
 // we can't use non-const __op here
@@ -402,16 +403,16 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _Forward
 replace_copy_if(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _ForwardIterator1 __last,
                 _ForwardIterator2 __result, _UnaryPredicate __pred, const _Tp& __new_value)
 {
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
+
     return oneapi::dpl::__internal::__pattern_walk2(
-        ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result,
+        __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result,
         oneapi::dpl::__internal::__replace_copy_functor<
             oneapi::dpl::__internal::__ref_or_copy<_ExecutionPolicy, const _Tp>,
             ::std::conditional_t<oneapi::dpl::__internal::__is_const_callable_object_v<_UnaryPredicate>,
                                  _UnaryPredicate,
                                  oneapi::dpl::__internal::__ref_or_copy<_ExecutionPolicy, _UnaryPredicate>>>(
-            __new_value, __pred),
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>());
+            __new_value, __pred));
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Tp>
diff --git a/include/oneapi/dpl/pstl/glue_memory_impl.h b/include/oneapi/dpl/pstl/glue_memory_impl.h
index 667a129825b..4d68973c814 100644
--- a/include/oneapi/dpl/pstl/glue_memory_impl.h
+++ b/include/oneapi/dpl/pstl/glue_memory_impl.h
@@ -45,24 +45,22 @@ uninitialized_copy(_ExecutionPolicy&& __exec, _InputIterator __first, _InputIter
     typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _ValueType2;
     typedef ::std::decay_t<_ExecutionPolicy> _DecayedExecutionPolicy;
 
-    constexpr auto __is_parallel =
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _InputIterator, _ForwardIterator>();
-
     if constexpr (::std::is_trivial_v<_ValueType1> && ::std::is_trivial_v<_ValueType2>)
     {
+        constexpr auto __is_parallel =
+            oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _InputIterator, _ForwardIterator>();
+
         return oneapi::dpl::__internal::__pattern_walk2_brick(
             ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result,
             oneapi::dpl::__internal::__brick_copy<_DecayedExecutionPolicy>{}, __is_parallel);
     }
     else
     {
-        constexpr auto __is_vector =
-            oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _InputIterator, _ForwardIterator>();
+        constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _InputIterator, _ForwardIterator>();
 
         return oneapi::dpl::__internal::__pattern_walk2(
-            ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result,
-            oneapi::dpl::__internal::__op_uninitialized_copy<_DecayedExecutionPolicy>{}, __is_vector,
-            __is_parallel);
+            __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result,
+            oneapi::dpl::__internal::__op_uninitialized_copy<_DecayedExecutionPolicy>{});
     }
 }
 
@@ -105,24 +103,22 @@ uninitialized_move(_ExecutionPolicy&& __exec, _InputIterator __first, _InputIter
     typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _ValueType2;
     typedef ::std::decay_t<_ExecutionPolicy> _DecayedExecutionPolicy;
 
-    constexpr auto __is_parallel =
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _InputIterator, _ForwardIterator>();
-
     if constexpr (::std::is_trivial_v<_ValueType1> && ::std::is_trivial_v<_ValueType2>)
     {
+        constexpr auto __is_parallel =
+            oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _InputIterator, _ForwardIterator>();
+
         return oneapi::dpl::__internal::__pattern_walk2_brick(
             ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result,
             oneapi::dpl::__internal::__brick_copy<_DecayedExecutionPolicy>{}, __is_parallel);
     }
     else
     {
-        constexpr auto __is_vector =
-            oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _InputIterator, _ForwardIterator>();
+        const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _InputIterator, _ForwardIterator>();
 
         return oneapi::dpl::__internal::__pattern_walk2(
-            ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result,
-            oneapi::dpl::__internal::__op_uninitialized_move<_DecayedExecutionPolicy>{}, __is_vector,
-            __is_parallel);
+            __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result,
+            oneapi::dpl::__internal::__op_uninitialized_move<_DecayedExecutionPolicy>{});
     }
 }
 
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 7c9782e0acf..d9a0e37908b 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -174,10 +174,15 @@ __pattern_swap(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIt
                _ForwardIterator2 __first2, _Function __f, /*is_vector=*/::std::true_type,
                /*is_parallel=*/::std::true_type)
 {
-    return __pattern_walk2</*_IsSync=*/::std::true_type, __par_backend_hetero::access_mode::read_write,
-                           __par_backend_hetero::access_mode::read_write>(::std::forward<_ExecutionPolicy>(__exec),
-                                                                          __first1, __last1, __first2, __f,
-                                                                          ::std::true_type(), ::std::true_type());
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
+
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
+    return __pattern_walk2<__backend_tag, /*_IsSync=*/::std::true_type,
+                           __par_backend_hetero::access_mode::read_write,
+                           __par_backend_hetero::access_mode::read_write>(
+        __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __f);
 }
 
 //------------------------------------------------------------------------
@@ -311,14 +316,18 @@ __pattern_walk2_transform_if(_ExecutionPolicy&& __exec, _ForwardIterator1 __firs
                              /*vector=*/::std::true_type,
                              /*parallel=*/::std::true_type)
 {
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
+
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     // Require `read_write` access mode for output sequence to force a copy in for host iterators to capture incoming
     // values of the output sequence for elements where the predicate is false.
-    return __pattern_walk2</*_IsSync=*/::std::true_type, __par_backend_hetero::access_mode::read,
+    return __pattern_walk2<__backend_tag, /*_IsSync=*/::std::true_type, __par_backend_hetero::access_mode::read,
                            __par_backend_hetero::access_mode::read_write>(
         __par_backend_hetero::make_wrapped_policy<__walk2_transform_if_wrapper>(
-            ::std::forward<_ExecutionPolicy>(__exec)),
-        __first1, __last1, __first2, __func,
-        /*vector=*/::std::true_type{}, /*parallel*/ ::std::true_type{});
+            __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec)),
+        __first1, __last1, __first2, __func);
 }
 
 template <typename _Name>
@@ -1098,11 +1107,17 @@ __pattern_unique(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last,
     auto __copy_last = __pattern_unique_copy(__exec, __first, __last, __copy_first, __pred,
                                              /*vector=*/::std::true_type{}, /*parallel*/ ::std::true_type{});
 
+    constexpr auto __dispatch_tag1 =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, decltype(__copy_first), decltype(__copy_last),
+                                                  decltype(__first)>();
+    using __backend_tag1 = typename decltype(__dispatch_tag1)::__backend_tag;
+
     //TODO: optimize copy back depending on Iterator, i.e. set_final_data for host iterator/pointer
-    return __pattern_walk2</*_IsSync=*/::std::true_type, __par_backend_hetero::access_mode::read_write,
+    return __pattern_walk2<__backend_tag1, /*_IsSync=*/::std::true_type, __par_backend_hetero::access_mode::read_write,
                            __par_backend_hetero::access_mode::read_write>(
+        __dispatch_tag1,
         __par_backend_hetero::make_wrapped_policy<copy_back_wrapper>(::std::forward<_ExecutionPolicy>(__exec)),
-        __copy_first, __copy_last, __first, __brick_copy<_ExecutionPolicy>{}, ::std::true_type{}, ::std::true_type{});
+        __copy_first, __copy_last, __first, __brick_copy<_ExecutionPolicy>{});
 }
 
 //------------------------------------------------------------------------
@@ -1363,14 +1378,21 @@ __pattern_stable_partition(_ExecutionPolicy&& __exec, _Iterator __first, _Iterat
     auto true_count = copy_result.first - __true_result;
 
     //TODO: optimize copy back if possible (inplace, decrease number of submits)
-    __pattern_walk2</*_IsSync=*/::std::false_type>(
-        __par_backend_hetero::make_wrapped_policy<copy_back_wrapper>(__exec),
-        __true_result, copy_result.first, __first, __brick_move<_ExecutionPolicy>{}, ::std::true_type{},
-        ::std::true_type{});
+    constexpr auto __dispatch_tag1 =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, decltype(__true_result),
+                                                  decltype(copy_result.first), decltype(__first)>();
+    using __backend_tag1 = typename decltype(__dispatch_tag1)::__backend_tag;
+    __pattern_walk2<__backend_tag1, /*_IsSync=*/::std::false_type>(
+        __dispatch_tag1, __par_backend_hetero::make_wrapped_policy<copy_back_wrapper>(__exec), __true_result,
+        copy_result.first, __first, __brick_move<_ExecutionPolicy>{});
+
+    constexpr auto __dispatch_tag2 =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, decltype(__false_result),
+                                                  decltype(copy_result.second), decltype(__first + true_count)>();
     __pattern_walk2(
+        __dispatch_tag2,
         __par_backend_hetero::make_wrapped_policy<copy_back_wrapper2>(::std::forward<_ExecutionPolicy>(__exec)),
-        __false_result, copy_result.second, __first + true_count, __brick_move<_ExecutionPolicy>{}, ::std::true_type{},
-        ::std::true_type{});
+        __false_result, copy_result.second, __first + true_count, __brick_move<_ExecutionPolicy>{});
 
     return __first + true_count;
 }

From 06a7cefdb7235468f08be6731268fc345312f15d Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 1 Feb 2024 11:58:52 +0100
Subject: [PATCH 044/566] __pattern_walk3 + tag impls

---
 include/oneapi/dpl/pstl/algorithm_fwd.h       | 18 ++++++
 include/oneapi/dpl/pstl/algorithm_impl.h      | 59 +++++++++++++++++++
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 29 +++++++++
 3 files changed, 106 insertions(+)

diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index bc9d87b767a..24db58ed68d 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -263,6 +263,12 @@ __pattern_walk3(_ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _Forwa
                 _Function, _IsVector,
                 /*parallel=*/::std::false_type) noexcept;
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _ForwardIterator3,
+          class _Function>
+_ForwardIterator3
+__pattern_walk3(_Tag, _ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2, _ForwardIterator3,
+                _Function) noexcept;
+
 template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
           class _RandomAccessIterator3, class _Function, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy_conditional<
@@ -273,6 +279,12 @@ __pattern_walk3(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1, _Ran
                 _RandomAccessIterator2 __first2, _RandomAccessIterator3 __first3, _Function __f, _IsVector __is_vector,
                 /*parallel=*/::std::true_type);
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
+          class _RandomAccessIterator3, class _Function>
+_RandomAccessIterator3
+__pattern_walk3(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator1, _RandomAccessIterator1,
+                _RandomAccessIterator2, _RandomAccessIterator3, _Function);
+
 template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
           class _RandomAccessIterator3, class _Function, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy_conditional<
@@ -283,6 +295,12 @@ __pattern_walk3(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1, _Ran
                 _RandomAccessIterator2 __first2, _RandomAccessIterator3 __first3, _Function __f, _IsVector __is_vector,
                 /*parallel=*/::std::true_type);
 
+template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _ForwardIterator3,
+          class _Function>
+_ForwardIterator3
+__pattern_walk3(__parallel_forward_tag, _ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2,
+                _ForwardIterator3, _Function);
+
 //------------------------------------------------------------------------
 // transform_if
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index b1f7c2c01ae..2bea3b3417e 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -600,6 +600,17 @@ __pattern_walk3(_ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator
     return __internal::__brick_walk3(__first1, __last1, __first2, __first3, __f, __is_vector);
 }
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _ForwardIterator3,
+          class _Function>
+_ForwardIterator3
+__pattern_walk3(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
+                _ForwardIterator2 __first2, _ForwardIterator3 __first3, _Function __f) noexcept
+{
+    static_assert(__is_backend_tag_serial_v<_Tag>);
+
+    return __internal::__brick_walk3(__first1, __last1, __first2, __first3, __f, typename _Tag::__is_vector{});
+}
+
 template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
           class _RandomAccessIterator3, class _Function, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy_conditional<
@@ -621,6 +632,23 @@ __pattern_walk3(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1, _Ran
     });
 }
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
+          class _RandomAccessIterator3, class _Function>
+_RandomAccessIterator3
+__pattern_walk3(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1, _RandomAccessIterator1 __last1,
+                _RandomAccessIterator2 __first2, _RandomAccessIterator3 __first3, _Function __f)
+{
+    return __internal::__except_handler([&]() {
+        __par_backend::__parallel_for(
+            ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1,
+            [__f, __first1, __first2, __first3](_RandomAccessIterator1 __i, _RandomAccessIterator1 __j) {
+                __internal::__brick_walk3(__i, __j, __first2 + (__i - __first1), __first3 + (__i - __first1), __f,
+                                          _IsVector{});
+            });
+        return __first3 + (__last1 - __first1);
+    });
+}
+
 template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _ForwardIterator3,
           class _Function, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy_conditional<
@@ -653,6 +681,37 @@ __pattern_walk3(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardI
     });
 }
 
+template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _ForwardIterator3,
+          class _Function>
+_ForwardIterator3
+__pattern_walk3(__parallel_forward_tag, _ExecutionPolicy&& __exec, _ForwardIterator1 __first1,
+                _ForwardIterator1 __last1, _ForwardIterator2 __first2, _ForwardIterator3 __first3, _Function __f)
+{
+    using __backend_tag = typename __parallel_forward_tag::__backend_tag;
+
+    return __internal::__except_handler([&]() {
+        using _iterator_tuple = zip_forward_iterator<_ForwardIterator1, _ForwardIterator2, _ForwardIterator3>;
+        auto __begin = _iterator_tuple(__first1, __first2, __first3);
+        auto __end = _iterator_tuple(__last1, /*dummy parameter*/ _ForwardIterator2(),
+                                     /*dummy parameter*/ _ForwardIterator3());
+
+        typedef typename ::std::iterator_traits<_ForwardIterator1>::reference _ReferenceType1;
+        typedef typename ::std::iterator_traits<_ForwardIterator2>::reference _ReferenceType2;
+        typedef typename ::std::iterator_traits<_ForwardIterator3>::reference _ReferenceType3;
+
+        __par_backend::__parallel_for_each(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __begin, __end,
+                                           [&](::std::tuple<_ReferenceType1, _ReferenceType2, _ReferenceType3> __val) {
+                                               __f(::std::get<0>(__val), ::std::get<1>(__val), ::std::get<2>(__val));
+                                           });
+
+        //TODO: parallel_for_each does not allow to return correct iterator value according to the ::std::transform
+        // implementation. Therefore, iterator value is calculated separately.
+        for (; __begin != __end; ++__begin)
+            ;
+        return ::std::get<2>(__begin.base());
+    });
+}
+
 //------------------------------------------------------------------------
 // transform_if
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index d9a0e37908b..9fb0b1c9bf7 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -218,6 +218,35 @@ __pattern_walk3(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardI
     return __first3 + __n;
 }
 
+template <typename _BackendTag,
+          __par_backend_hetero::access_mode __acc_mode1 = __par_backend_hetero::access_mode::read,
+          __par_backend_hetero::access_mode __acc_mode2 = __par_backend_hetero::access_mode::read,
+          __par_backend_hetero::access_mode __acc_mode3 = __par_backend_hetero::access_mode::write,
+          typename _ExecutionPolicy, typename _ForwardIterator1, typename _ForwardIterator2, typename _ForwardIterator3,
+          typename _Function>
+_ForwardIterator3
+__pattern_walk3(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _ForwardIterator1 __first1,
+                _ForwardIterator1 __last1, _ForwardIterator2 __first2, _ForwardIterator3 __first3, _Function __f)
+{
+    auto __n = __last1 - __first1;
+    if (__n <= 0)
+        return __first3;
+
+    auto __keep1 = oneapi::dpl::__ranges::__get_sycl_range<__acc_mode1, _ForwardIterator1>();
+    auto __buf1 = __keep1(__first1, __last1);
+    auto __keep2 = oneapi::dpl::__ranges::__get_sycl_range<__acc_mode2, _ForwardIterator2>();
+    auto __buf2 = __keep2(__first2, __first2 + __n);
+    auto __keep3 = oneapi::dpl::__ranges::__get_sycl_range<__acc_mode3, _ForwardIterator3>();
+    auto __buf3 = __keep3(__first3, __first3 + __n);
+
+    oneapi::dpl::__par_backend_hetero::__parallel_for(::std::forward<_ExecutionPolicy>(__exec),
+                                                      unseq_backend::walk_n<_ExecutionPolicy, _Function>{__f}, __n,
+                                                      __buf1.all_view(), __buf2.all_view(), __buf3.all_view())
+        .wait();
+
+    return __first3 + __n;
+}
+
 //------------------------------------------------------------------------
 // walk_brick, walk_brick_n
 //------------------------------------------------------------------------

From 0b8b028ff27623929b15600c383563ba99ebd332 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 1 Feb 2024 12:33:58 +0100
Subject: [PATCH 045/566] __pattern_walk3 + tag calls

---
 include/oneapi/dpl/pstl/glue_algorithm_impl.h | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_algorithm_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
index 905708d9fe9..207967a3a3a 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
@@ -336,11 +336,12 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _Forward
 transform(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 __last1, _ForwardIterator2 __first2,
           _ForwardIterator __result, _BinaryOperation __op)
 {
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2, _ForwardIterator>();
+
     return oneapi::dpl::__internal::__pattern_walk3(
+        __dispatch_tag,
         ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __result,
-        oneapi::dpl::__internal::__transform_functor<_BinaryOperation>(::std::move(__op)),
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2, _ForwardIterator>(),
-        __exec.__allow_parallel());
+        oneapi::dpl::__internal::__transform_functor<_BinaryOperation>(::std::move(__op)));
 }
 
 // [alg.transform_if]

From d3c200ab9a30c81c7b79d0a3d1a6cc88527d33f1 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 1 Feb 2024 16:16:17 +0100
Subject: [PATCH 046/566] __pattern_walk1_n + tag impls

---
 include/oneapi/dpl/pstl/algorithm_fwd.h       | 14 ++++++++++----
 include/oneapi/dpl/pstl/algorithm_impl.h      | 19 +++++++++++++++++++
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 14 ++++++++++++--
 3 files changed, 41 insertions(+), 6 deletions(-)

diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index 24db58ed68d..7903e0fadc3 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -87,8 +87,7 @@ __pattern_walk1(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _Rando
 
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Function>
 void
-__pattern_walk1(__parallel_forward_tag, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
-                _RandomAccessIterator __last, _Function __f);
+__pattern_walk1(__parallel_forward_tag, _ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _Function);
 
 template <class _ExecutionPolicy, class _RandomAccessIterator, class _Function, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy_conditional<
@@ -99,8 +98,7 @@ __pattern_walk1(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _Rando
 
 template <class _IsVector, class _ExecutionPolicy, class _ForwardIterator, class _Function>
 void
-__pattern_walk1(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last,
-                _Function __f);
+__pattern_walk1(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _Function);
 
 template <class _ExecutionPolicy, class _ForwardIterator, class _Brick>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>
@@ -129,11 +127,19 @@ oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _Fo
 __pattern_walk1_n(_ExecutionPolicy&&, _ForwardIterator, _Size, _Function, _IsVector,
                   /*is_parallel=*/::std::false_type) noexcept;
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Size, class _Function>
+_ForwardIterator
+__pattern_walk1_n(_Tag, _ExecutionPolicy&&, _ForwardIterator, _Size, _Function) noexcept;
+
 template <class _ExecutionPolicy, class _RandomAccessIterator, class _Size, class _Function, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator>
 __pattern_walk1_n(_ExecutionPolicy&&, _RandomAccessIterator, _Size, _Function, _IsVector,
                   /*is_parallel=*/::std::true_type);
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Size, class _Function>
+_RandomAccessIterator
+__pattern_walk1_n(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator, _Size, _Function);
+
 template <class _ExecutionPolicy, class _ForwardIterator, class _Size, class _Brick>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _ForwardIterator>
 __pattern_walk_brick_n(_ExecutionPolicy&&, _ForwardIterator, _Size, _Brick,
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 2bea3b3417e..3e20b86469b 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -274,6 +274,15 @@ __pattern_walk1_n(_ExecutionPolicy&&, _ForwardIterator __first, _Size __n, _Func
     return __internal::__brick_walk1_n(__first, __n, __f, __is_vector);
 }
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Size, class _Function>
+_ForwardIterator
+__pattern_walk1_n(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _Size __n, _Function __f) noexcept
+{
+    static_assert(__is_backend_tag_v<_Tag>);
+
+    return __internal::__brick_walk1_n(__first, __n, __f, typename _Tag::__is_vector{});
+}
+
 template <class _ExecutionPolicy, class _RandomAccessIterator, class _Size, class _Function, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator>
 __pattern_walk1_n(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _Size __n, _Function __f,
@@ -287,6 +296,16 @@ __pattern_walk1_n(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _Siz
     return __first + __n;
 }
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Size, class _Function>
+_RandomAccessIterator
+__pattern_walk1_n(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _RandomAccessIterator __first, _Size __n,
+                  _Function __f)
+{
+    oneapi::dpl::__internal::__pattern_walk1(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first,
+                                             __first + __n, __f);
+    return __first + __n;
+}
+
 template <class _ExecutionPolicy, class _ForwardIterator, class _Size, class _Brick>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _ForwardIterator>
 __pattern_walk_brick_n(_ExecutionPolicy&& __exec, _ForwardIterator __first, _Size __n, _Brick __brick,
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 9fb0b1c9bf7..19be5cd634f 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -60,8 +60,8 @@ __pattern_walk1(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIte
 
 template <typename _BackendTag, typename _ExecutionPolicy, typename _ForwardIterator, typename _Function>
 void
-__pattern_walk1(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _ForwardIterator __first,
-                _ForwardIterator __last, _Function __f)
+__pattern_walk1(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last,
+                _Function __f)
 {
     auto __n = __last - __first;
     if (__n <= 0)
@@ -91,6 +91,16 @@ __pattern_walk1_n(_ExecutionPolicy&& __exec, _ForwardIterator __first, _Size __n
     return __first + __n;
 }
 
+template <typename _BackendTag, typename _ExecutionPolicy, typename _ForwardIterator, typename _Size,
+          typename _Function>
+_ForwardIterator
+__pattern_walk1_n(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _ForwardIterator __first, _Size __n,
+                  _Function __f)
+{
+    __pattern_walk1(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __first + __n, __f);
+    return __first + __n;
+}
+
 //------------------------------------------------------------------------
 // walk2
 //------------------------------------------------------------------------

From 6ba79f93f5f27191f1c27e615da843d6adbcccbe Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 1 Feb 2024 16:22:48 +0100
Subject: [PATCH 047/566] __pattern_walk1_n + tag calls

---
 include/oneapi/dpl/pstl/glue_algorithm_impl.h |  8 +--
 include/oneapi/dpl/pstl/glue_memory_impl.h    | 59 ++++++++-----------
 2 files changed, 28 insertions(+), 39 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_algorithm_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
index 207967a3a3a..2b77e0c44b7 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
@@ -85,10 +85,10 @@ template <class _ExecutionPolicy, class _ForwardIterator, class _Size, class _Fu
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _ForwardIterator>
 for_each_n(_ExecutionPolicy&& __exec, _ForwardIterator __first, _Size __n, _Function __f)
 {
-    return oneapi::dpl::__internal::__pattern_walk1_n(
-        ::std::forward<_ExecutionPolicy>(__exec), __first, __n, __f,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>());
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+
+    return oneapi::dpl::__internal::__pattern_walk1_n(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first,
+                                                      __n, __f);
 }
 
 // [alg.find]
diff --git a/include/oneapi/dpl/pstl/glue_memory_impl.h b/include/oneapi/dpl/pstl/glue_memory_impl.h
index 4d68973c814..1e450a4efa5 100644
--- a/include/oneapi/dpl/pstl/glue_memory_impl.h
+++ b/include/oneapi/dpl/pstl/glue_memory_impl.h
@@ -188,11 +188,11 @@ uninitialized_fill_n(_ExecutionPolicy&& __exec, _ForwardIterator __first, _Size
     typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _ValueType;
     typedef ::std::decay_t<_ExecutionPolicy> _DecayedExecutionPolicy;
 
-    constexpr auto __is_parallel =
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>();
-
     if constexpr (::std::is_arithmetic_v<_ValueType>)
     {
+        constexpr auto __is_parallel =
+            oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>();
+
         return oneapi::dpl::__internal::__pattern_walk_brick_n(
             ::std::forward<_ExecutionPolicy>(__exec), __first, __n,
             oneapi::dpl::__internal::__brick_fill_n<_ValueType, _DecayedExecutionPolicy>{_ValueType(__value)},
@@ -200,13 +200,11 @@ uninitialized_fill_n(_ExecutionPolicy&& __exec, _ForwardIterator __first, _Size
     }
     else
     {
-        constexpr auto __is_vector =
-            oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>();
+        constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
 
         return oneapi::dpl::__internal::__pattern_walk1_n(
-            ::std::forward<_ExecutionPolicy>(__exec), __first, __n,
-            oneapi::dpl::__internal::__op_uninitialized_fill<_Tp, _DecayedExecutionPolicy>{__value}, __is_vector,
-            __is_parallel);
+            __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __n,
+            oneapi::dpl::__internal::__op_uninitialized_fill<_Tp, _DecayedExecutionPolicy>{__value});
     }
 }
 
@@ -269,22 +267,19 @@ destroy_n(_ExecutionPolicy&& __exec, _ForwardIterator __first, _Size __n)
     }
     else
     {
-        constexpr auto __is_parallel =
-            oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>();
-        using _is_vector_type =
+        using _ExecutionPolicyDest =
 #if (_PSTL_ICPX_OMP_SIMD_DESTROY_WINDOWS_BROKEN || _ONEDPL_ICPX_OMP_SIMD_DESTROY_WINDOWS_BROKEN)
-            ::std::conditional_t<
-                oneapi::dpl::__internal::__is_host_execution_policy<::std::decay_t<_ExecutionPolicy>>::value,
-                ::std::false_type,
-                decltype(oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>())>;
+            typename EvaDestroylWorkaroundPolicy<::std::decay_t<_ExecutionPolicy>>::ExecutionPolicy;
 #else
-            decltype(oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>());
+            _ExecutionPolicy;
 #endif // _PSTL_ICPX_OMP_SIMD_DESTROY_WINDOWS_BROKEN || _ONEDPL_ICPX_OMP_SIMD_DESTROY_WINDOWS_BROKEN
-        constexpr _is_vector_type __is_vector;
 
-        return oneapi::dpl::__internal::__pattern_walk1_n(::std::forward<_ExecutionPolicy>(__exec), __first, __n,
-                                                          [](_ReferenceType __val) { __val.~_ValueType(); },
-                                                          __is_vector, __is_parallel);
+        constexpr auto __dispatch_tag =
+            oneapi::dpl::__internal::__select_backend<_ExecutionPolicyDest, _ForwardIterator>();
+
+        return oneapi::dpl::__internal::__pattern_walk1_n(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
+                                                          __first, __n,
+                                                          [](_ReferenceType __val) { __val.~_ValueType(); });
     }
 }
 
@@ -320,15 +315,11 @@ uninitialized_default_construct_n(_ExecutionPolicy&& __exec, _ForwardIterator __
     }
     else
     {
-        constexpr auto __is_parallel =
-            oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>();
-        constexpr auto __is_vector =
-            oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>();
+        constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
 
         return oneapi::dpl::__internal::__pattern_walk1_n(
-            ::std::forward<_ExecutionPolicy>(__exec), __first, __n,
-            oneapi::dpl::__internal::__op_uninitialized_default_construct<_DecayedExecutionPolicy>{}, __is_vector,
-            __is_parallel);
+            __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __n,
+            oneapi::dpl::__internal::__op_uninitialized_default_construct<_DecayedExecutionPolicy>{});
     }
 }
 
@@ -368,11 +359,11 @@ uninitialized_value_construct_n(_ExecutionPolicy&& __exec, _ForwardIterator __fi
     typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _ValueType;
     typedef ::std::decay_t<_ExecutionPolicy> _DecayedExecutionPolicy;
 
-    constexpr auto __is_parallel =
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>();
-
     if constexpr (::std::is_trivial_v<_ValueType>)
     {
+        constexpr auto __is_parallel =
+            oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>();
+
         return oneapi::dpl::__internal::__pattern_walk_brick_n(
             ::std::forward<_ExecutionPolicy>(__exec), __first, __n,
             oneapi::dpl::__internal::__brick_fill_n<_ValueType, _DecayedExecutionPolicy>{_ValueType()},
@@ -380,13 +371,11 @@ uninitialized_value_construct_n(_ExecutionPolicy&& __exec, _ForwardIterator __fi
     }
     else
     {
-        constexpr auto __is_vector =
-            oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>();
+        constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
 
         return oneapi::dpl::__internal::__pattern_walk1_n(
-            ::std::forward<_ExecutionPolicy>(__exec), __first, __n,
-            oneapi::dpl::__internal::__op_uninitialized_value_construct<_DecayedExecutionPolicy>{}, __is_vector,
-            __is_parallel);
+            __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __n,
+            oneapi::dpl::__internal::__op_uninitialized_value_construct<_DecayedExecutionPolicy>{});
     }
 }
 

From 825391c27696320743eb4a32d6e8944c06659805 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 2 Feb 2024 10:55:24 +0100
Subject: [PATCH 048/566] __pattern_find_if + tag impls

---
 include/oneapi/dpl/pstl/algorithm_fwd.h  | 14 ++++++--------
 include/oneapi/dpl/pstl/algorithm_impl.h |  8 +++++---
 2 files changed, 11 insertions(+), 11 deletions(-)

diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index 7903e0fadc3..9faa0bde46d 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -381,15 +381,13 @@ template <class _RandomAccessIterator, class _Predicate>
 _RandomAccessIterator __brick_find_if(_RandomAccessIterator, _RandomAccessIterator, _Predicate,
                                       /*is_vector=*/::std::true_type) noexcept;
 
-template <class _ExecutionPolicy, class _ForwardIterator, class _Predicate, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _ForwardIterator>
-__pattern_find_if(_ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _Predicate, _IsVector,
-                  /*is_parallel=*/::std::false_type) noexcept;
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Predicate>
+_ForwardIterator
+__pattern_find_if(_Tag, _ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _Predicate) noexcept;
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _Predicate, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator>
-__pattern_find_if(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _Predicate, _IsVector,
-                  /*is_parallel=*/::std::true_type);
+template <class _IsVector, class _ExecutionPolicy, class _ForwardIterator, class _Predicate>
+_ForwardIterator
+__pattern_find_if(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _Predicate);
 
 //------------------------------------------------------------------------
 // find_end
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 3e20b86469b..75b2f0fa87f 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -881,18 +881,20 @@ __brick_find_if(_RandomAccessIterator __first, _RandomAccessIterator __last, _Pr
 
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Predicate>
 _ForwardIterator
-__pattern_find_if(_Tag __tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
+__pattern_find_if(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                   _Predicate __pred) noexcept
 {
+    static_assert(__is_backend_tag_v<_Tag>);
+
     return __internal::__brick_find_if(__first, __last, __pred, typename _Tag::__is_vector{});
 }
 
 template <class _IsVector, class _ExecutionPolicy, class _ForwardIterator, class _Predicate>
 _ForwardIterator
-__pattern_find_if(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _ForwardIterator __first,
+__pattern_find_if(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _ForwardIterator __first,
                   _ForwardIterator __last, _Predicate __pred)
 {
-    using __backend_tag = typename decltype(__tag)::__backend_tag;
+    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
 
     return __except_handler([&]() {
         return __parallel_find(

From bc10cdf618a5f4f7df9cc8046039134483191002 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 8 Feb 2024 09:25:13 +0100
Subject: [PATCH 049/566] __pattern_replace_if + tag impls

---
 include/oneapi/dpl/pstl/algorithm_impl.h | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 75b2f0fa87f..3059748bd0d 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -187,6 +187,8 @@ void
 __pattern_replace_if(_Tag __tag, _ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last,
                      _UnaryPredicate __pred, const _Tp& __new_value)
 {
+    static_assert(__is_backend_tag_v<_Tag>);
+
     oneapi::dpl::__internal::__pattern_walk1(
         __tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
         oneapi::dpl::__internal::__replace_functor<

From f3bf9acd54e9d77284feab966c99337580fc5eb8 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 2 Feb 2024 11:49:44 +0100
Subject: [PATCH 050/566] __pattern_walk_brick + tag impls

---
 include/oneapi/dpl/pstl/algorithm_fwd.h       |  9 +++++++
 include/oneapi/dpl/pstl/algorithm_impl.h      | 24 +++++++++++++++++++
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 14 +++++++++++
 3 files changed, 47 insertions(+)

diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index 9faa0bde46d..700d8ca5541 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -105,11 +105,20 @@ oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>
 __pattern_walk_brick(_ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _Brick,
                      /*parallel=*/::std::false_type) noexcept;
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Brick>
+void
+__pattern_walk_brick(_Tag, _ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _Brick) noexcept;
+
 template <class _ExecutionPolicy, class _RandomAccessIterator, class _Brick>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>
 __pattern_walk_brick(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _Brick,
                      /*parallel=*/::std::true_type);
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Brick>
+void
+__pattern_walk_brick(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator,
+                     _Brick);
+
 //------------------------------------------------------------------------
 // walk1_n
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 3059748bd0d..5fd7d2d87cb 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -234,6 +234,16 @@ __pattern_walk_brick(_ExecutionPolicy&& __exec, _ForwardIterator __first, _Forwa
     __brick(__first, __last, __is_vector);
 }
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Brick>
+void
+__pattern_walk_brick(_Tag, _ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last,
+                     _Brick __brick) noexcept
+{
+    static_assert(__is_backend_tag_v<_Tag>);
+
+    __brick(__first, __last, typename _Tag::__is_vector{});
+}
+
 template <class _ExecutionPolicy, class _RandomAccessIterator, class _Brick>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>
 __pattern_walk_brick(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIterator __last,
@@ -249,6 +259,20 @@ __pattern_walk_brick(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _
     });
 }
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Brick>
+void
+__pattern_walk_brick(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
+                     _RandomAccessIterator __last, _Brick __brick)
+{
+    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
+
+    __internal::__except_handler([&]() {
+        __par_backend::__parallel_for(
+            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+            [__brick](_RandomAccessIterator __i, _RandomAccessIterator __j) { __brick(__i, __j, _IsVector{}); });
+    });
+}
+
 //------------------------------------------------------------------------
 // walk1_n
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 19be5cd634f..dd045d3bb04 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -282,6 +282,20 @@ __pattern_walk_brick(_ExecutionPolicy&& __exec, _ForwardIterator __first, _Forwa
         __first, __last, __f);
 }
 
+template <typename _BackendTag, typename _ExecutionPolicy, typename _ForwardIterator, typename _Function>
+void
+__pattern_walk_brick(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _ForwardIterator __first,
+                     _ForwardIterator __last, _Function __f)
+{
+    if (__last - __first <= 0)
+        return;
+
+    __pattern_walk1(
+        __tag,
+        __par_backend_hetero::make_wrapped_policy<__walk_brick_wrapper>(::std::forward<_ExecutionPolicy>(__exec)),
+        __first, __last, __f);
+}
+
 template <typename _Name>
 struct __walk_brick_n_wrapper
 {

From 20349f66c0dde96e447ad1f97aa502743613cf23 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 2 Feb 2024 11:50:19 +0100
Subject: [PATCH 051/566] __pattern_walk_brick + tag calls

---
 include/oneapi/dpl/pstl/glue_memory_impl.h | 23 ++++++++--------------
 1 file changed, 8 insertions(+), 15 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_memory_impl.h b/include/oneapi/dpl/pstl/glue_memory_impl.h
index 1e450a4efa5..1c6a3bdd6ab 100644
--- a/include/oneapi/dpl/pstl/glue_memory_impl.h
+++ b/include/oneapi/dpl/pstl/glue_memory_impl.h
@@ -160,20 +160,17 @@ uninitialized_fill(_ExecutionPolicy&& __exec, _ForwardIterator __first, _Forward
     typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _ValueType;
     typedef ::std::decay_t<_ExecutionPolicy> _DecayedExecutionPolicy;
 
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+
     if constexpr (::std::is_arithmetic_v<_ValueType>)
     {
-        constexpr auto __is_parallel =
-            oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>();
-
         oneapi::dpl::__internal::__pattern_walk_brick(
+            __dispatch_tag,
             ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
-            oneapi::dpl::__internal::__brick_fill<_ValueType, _DecayedExecutionPolicy>{_ValueType(__value)},
-            __is_parallel);
+            oneapi::dpl::__internal::__brick_fill<_ValueType, _DecayedExecutionPolicy>{_ValueType(__value)});
     }
     else
     {
-        constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
-
         oneapi::dpl::__internal::__pattern_walk1(
             __dispatch_tag,
             ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
@@ -332,20 +329,16 @@ uninitialized_value_construct(_ExecutionPolicy&& __exec, _ForwardIterator __firs
     typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _ValueType;
     typedef ::std::decay_t<_ExecutionPolicy> _DecayedExecutionPolicy;
 
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+
     if constexpr (::std::is_trivial_v<_ValueType>)
     {
-        constexpr auto __is_parallel =
-            oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>();
-
         oneapi::dpl::__internal::__pattern_walk_brick(
-            ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
-            oneapi::dpl::__internal::__brick_fill<_ValueType, _DecayedExecutionPolicy>{_ValueType()},
-            __is_parallel);
+            __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+            oneapi::dpl::__internal::__brick_fill<_ValueType, _DecayedExecutionPolicy>{_ValueType()});
     }
     else
     {
-        constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
-
         oneapi::dpl::__internal::__pattern_walk1(
             __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
             oneapi::dpl::__internal::__op_uninitialized_value_construct<_DecayedExecutionPolicy>{});

From 674dc00c5df805f4f16dfee8e0affb0819cdca7b Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 2 Feb 2024 12:38:31 +0100
Subject: [PATCH 052/566] __pattern_walk2_n + tag impls

---
 include/oneapi/dpl/pstl/algorithm_fwd.h       | 11 ++++++++++
 include/oneapi/dpl/pstl/algorithm_impl.h      | 21 +++++++++++++++++++
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   |  9 ++++++++
 3 files changed, 41 insertions(+)

diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index 700d8ca5541..6d42e90d927 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -221,12 +221,23 @@ oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _Fo
 __pattern_walk2_n(_ExecutionPolicy&&, _ForwardIterator1, _Size, _ForwardIterator2, _Function, _IsVector,
                   /*parallel=*/::std::false_type) noexcept;
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _Size, class _ForwardIterator2,
+          class _Function>
+_ForwardIterator2
+__pattern_walk2_n(_Tag, _ExecutionPolicy&&, _ForwardIterator1, _Size, _ForwardIterator2, _Function) noexcept;
+
 template <class _ExecutionPolicy, class _RandomAccessIterator1, class _Size, class _RandomAccessIterator2,
           class _Function, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator2>
 __pattern_walk2_n(_ExecutionPolicy&&, _RandomAccessIterator1, _Size, _RandomAccessIterator2, _Function, _IsVector,
                   /*parallel=*/::std::true_type);
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _Size,
+          class _RandomAccessIterator2, class _Function>
+_RandomAccessIterator2
+__pattern_walk2_n(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator1, _Size, _RandomAccessIterator2,
+                  _Function);
+
 template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Brick>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _ForwardIterator2>
 __pattern_walk2_brick(_ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2, _Brick,
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 5fd7d2d87cb..12479e8c9fd 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -513,6 +513,17 @@ __pattern_walk2_n(_ExecutionPolicy&&, _ForwardIterator1 __first1, _Size __n, _Fo
     return __internal::__brick_walk2_n(__first1, __n, __first2, __f, __is_vector);
 }
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _Size, class _ForwardIterator2,
+          class _Function>
+_ForwardIterator2
+__pattern_walk2_n(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _Size __n, _ForwardIterator2 __first2,
+                  _Function __f) noexcept
+{
+    static_assert(__is_backend_tag_v<_Tag>);
+
+    return __internal::__brick_walk2_n(__first1, __n, __first2, __f, typename _Tag::__is_vector{});
+}
+
 template <class _ExecutionPolicy, class _RandomAccessIterator1, class _Size, class _RandomAccessIterator2,
           class _Function, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator2>
@@ -525,6 +536,16 @@ __pattern_walk2_n(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1, _S
                                        __first1 + __n, __first2, __f);
 }
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _Size,
+          class _RandomAccessIterator2, class _Function>
+_RandomAccessIterator2
+__pattern_walk2_n(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1,
+                  _Size __n, _RandomAccessIterator2 __first2, _Function __f)
+{
+    return __internal::__pattern_walk2(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __first1 + __n,
+                                       __first2, __f);
+}
+
 template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Brick>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _ForwardIterator2>
 __pattern_walk2_brick(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index dd045d3bb04..69eba6177fc 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -174,6 +174,15 @@ __pattern_walk2_n(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _Size _
                            ::std::true_type(), ::std::true_type());
 }
 
+template <typename _BackendTag, typename _ExecutionPolicy, typename _ForwardIterator1, typename _Size,
+          typename _ForwardIterator2, typename _Function>
+_ForwardIterator2
+__pattern_walk2_n(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _Size __n,
+                  _ForwardIterator2 __first2, _Function __f)
+{
+    return __pattern_walk2(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __first1 + __n, __first2, __f);
+}
+
 //------------------------------------------------------------------------
 // swap
 //------------------------------------------------------------------------

From 4338b96ce996eefb94fe22de840f2d6efb960b59 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 2 Feb 2024 12:39:00 +0100
Subject: [PATCH 053/566] __pattern_walk2_n + tag calls

---
 include/oneapi/dpl/pstl/glue_memory_impl.h | 20 ++++++++------------
 1 file changed, 8 insertions(+), 12 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_memory_impl.h b/include/oneapi/dpl/pstl/glue_memory_impl.h
index 1c6a3bdd6ab..cb0a86fa036 100644
--- a/include/oneapi/dpl/pstl/glue_memory_impl.h
+++ b/include/oneapi/dpl/pstl/glue_memory_impl.h
@@ -72,6 +72,8 @@ uninitialized_copy_n(_ExecutionPolicy&& __exec, _InputIterator __first, _Size __
     typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _ValueType2;
     typedef ::std::decay_t<_ExecutionPolicy> _DecayedExecutionPolicy;
 
+     const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _InputIterator, _ForwardIterator>();
+
     constexpr auto __is_parallel =
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _InputIterator, _ForwardIterator>();
 
@@ -83,13 +85,9 @@ uninitialized_copy_n(_ExecutionPolicy&& __exec, _InputIterator __first, _Size __
     }
     else
     {
-        constexpr auto __is_vector =
-            oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _InputIterator, _ForwardIterator>();
-
         return oneapi::dpl::__internal::__pattern_walk2_n(
-            ::std::forward<_ExecutionPolicy>(__exec), __first, __n, __result,
-            oneapi::dpl::__internal::__op_uninitialized_copy<_DecayedExecutionPolicy>{}, __is_vector,
-            __is_parallel);
+            __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __n, __result,
+            oneapi::dpl::__internal::__op_uninitialized_copy<_DecayedExecutionPolicy>{});
     }
 }
 
@@ -130,6 +128,8 @@ uninitialized_move_n(_ExecutionPolicy&& __exec, _InputIterator __first, _Size __
     typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _ValueType2;
     typedef ::std::decay_t<_ExecutionPolicy> _DecayedExecutionPolicy;
 
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _InputIterator, _ForwardIterator>();
+
     constexpr auto __is_parallel =
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _InputIterator, _ForwardIterator>();
 
@@ -141,13 +141,9 @@ uninitialized_move_n(_ExecutionPolicy&& __exec, _InputIterator __first, _Size __
     }
     else
     {
-        constexpr auto __is_vector =
-            oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _InputIterator, _ForwardIterator>();
-
         return oneapi::dpl::__internal::__pattern_walk2_n(
-            ::std::forward<_ExecutionPolicy>(__exec), __first, __n, __result,
-            oneapi::dpl::__internal::__op_uninitialized_move<_DecayedExecutionPolicy>{}, __is_vector,
-            __is_parallel);
+            __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __n, __result,
+            oneapi::dpl::__internal::__op_uninitialized_move<_DecayedExecutionPolicy>{});
     }
 }
 

From 846f7d96458282dd57454ef27216080b32c7d8b7 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 2 Feb 2024 12:54:09 +0100
Subject: [PATCH 054/566] __pattern_walk_brick_n + tag impls

---
 include/oneapi/dpl/pstl/algorithm_fwd.h       |  8 +++++++
 include/oneapi/dpl/pstl/algorithm_impl.h      | 24 +++++++++++++++++++
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 13 ++++++++++
 3 files changed, 45 insertions(+)

diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index 6d42e90d927..591553e4f9b 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -154,11 +154,19 @@ oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _Fo
 __pattern_walk_brick_n(_ExecutionPolicy&&, _ForwardIterator, _Size, _Brick,
                        /*is_parallel=*/::std::false_type) noexcept;
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Size, class _Brick>
+_ForwardIterator
+__pattern_walk_brick_n(_Tag, _ExecutionPolicy&&, _ForwardIterator, _Size, _Brick) noexcept;
+
 template <class _ExecutionPolicy, class _RandomAccessIterator, class _Size, class _Brick>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator>
 __pattern_walk_brick_n(_ExecutionPolicy&&, _RandomAccessIterator, _Size, _Brick,
                        /*is_parallel=*/::std::true_type);
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Size, class _Brick>
+_RandomAccessIterator
+__pattern_walk_brick_n(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator, _Size, _Brick);
+
 //------------------------------------------------------------------------
 // walk2 (pseudo)
 //
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 12479e8c9fd..90806aa392f 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -341,6 +341,15 @@ __pattern_walk_brick_n(_ExecutionPolicy&& __exec, _ForwardIterator __first, _Siz
     return __brick(__first, __n, __is_vector);
 }
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Size, class _Brick>
+_ForwardIterator
+__pattern_walk_brick_n(_Tag, _ExecutionPolicy&& __exec, _ForwardIterator __first, _Size __n, _Brick __brick) noexcept
+{
+    static_assert(__is_backend_tag_v<_Tag>);
+
+    return __brick(__first, __n, typename _Tag::__is_vector{});
+}
+
 template <class _ExecutionPolicy, class _RandomAccessIterator, class _Size, class _Brick>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator>
 __pattern_walk_brick_n(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _Size __n, _Brick __brick,
@@ -356,6 +365,21 @@ __pattern_walk_brick_n(_ExecutionPolicy&& __exec, _RandomAccessIterator __first,
     });
 }
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Size, class _Brick>
+_RandomAccessIterator
+__pattern_walk_brick_n(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator __first, _Size __n,
+                       _Brick __brick)
+{
+    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
+
+    return __internal::__except_handler([&]() {
+        __par_backend::__parallel_for(
+            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __first + __n,
+            [__brick](_RandomAccessIterator __i, _RandomAccessIterator __j) { __brick(__i, __j - __i, _IsVector{}); });
+        return __first + __n;
+    });
+}
+
 //------------------------------------------------------------------------
 // walk2 (pseudo)
 //
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 69eba6177fc..a430ee845cd 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -324,6 +324,19 @@ __pattern_walk_brick_n(_ExecutionPolicy&& __exec, _ForwardIterator __first, _Siz
     return __first + __n;
 }
 
+template <typename _BackendTag, typename _ExecutionPolicy, typename _ForwardIterator, typename _Size,
+          typename _Function>
+_ForwardIterator
+__pattern_walk_brick_n(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _ForwardIterator __first, _Size __n,
+                       _Function __f)
+{
+    __pattern_walk1(
+        __tag,
+        __par_backend_hetero::make_wrapped_policy<__walk_brick_n_wrapper>(::std::forward<_ExecutionPolicy>(__exec)),
+        __first, __first + __n, __f);
+    return __first + __n;
+}
+
 //------------------------------------------------------------------------
 // walk2_brick, walk2_brick_n
 //------------------------------------------------------------------------

From 6139218326bfe14c185c8f80f83d1de174c753aa Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 2 Feb 2024 12:54:38 +0100
Subject: [PATCH 055/566] __pattern_walk_brick_n + tag calls

---
 include/oneapi/dpl/pstl/glue_memory_impl.h | 24 ++++++++--------------
 1 file changed, 8 insertions(+), 16 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_memory_impl.h b/include/oneapi/dpl/pstl/glue_memory_impl.h
index cb0a86fa036..72e2f6cc1ad 100644
--- a/include/oneapi/dpl/pstl/glue_memory_impl.h
+++ b/include/oneapi/dpl/pstl/glue_memory_impl.h
@@ -181,20 +181,16 @@ uninitialized_fill_n(_ExecutionPolicy&& __exec, _ForwardIterator __first, _Size
     typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _ValueType;
     typedef ::std::decay_t<_ExecutionPolicy> _DecayedExecutionPolicy;
 
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+
     if constexpr (::std::is_arithmetic_v<_ValueType>)
     {
-        constexpr auto __is_parallel =
-            oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>();
-
         return oneapi::dpl::__internal::__pattern_walk_brick_n(
-            ::std::forward<_ExecutionPolicy>(__exec), __first, __n,
-            oneapi::dpl::__internal::__brick_fill_n<_ValueType, _DecayedExecutionPolicy>{_ValueType(__value)},
-            __is_parallel);
+            __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __n,
+            oneapi::dpl::__internal::__brick_fill_n<_ValueType, _DecayedExecutionPolicy>{_ValueType(__value)});
     }
     else
     {
-        constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
-
         return oneapi::dpl::__internal::__pattern_walk1_n(
             __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __n,
             oneapi::dpl::__internal::__op_uninitialized_fill<_Tp, _DecayedExecutionPolicy>{__value});
@@ -348,20 +344,16 @@ uninitialized_value_construct_n(_ExecutionPolicy&& __exec, _ForwardIterator __fi
     typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _ValueType;
     typedef ::std::decay_t<_ExecutionPolicy> _DecayedExecutionPolicy;
 
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+
     if constexpr (::std::is_trivial_v<_ValueType>)
     {
-        constexpr auto __is_parallel =
-            oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>();
-
         return oneapi::dpl::__internal::__pattern_walk_brick_n(
-            ::std::forward<_ExecutionPolicy>(__exec), __first, __n,
-            oneapi::dpl::__internal::__brick_fill_n<_ValueType, _DecayedExecutionPolicy>{_ValueType()},
-            __is_parallel);
+            __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __n,
+            oneapi::dpl::__internal::__brick_fill_n<_ValueType, _DecayedExecutionPolicy>{_ValueType()});
     }
     else
     {
-        constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
-
         return oneapi::dpl::__internal::__pattern_walk1_n(
             __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __n,
             oneapi::dpl::__internal::__op_uninitialized_value_construct<_DecayedExecutionPolicy>{});

From 84aae3b0be50cd8568d3a3f05e0522f17a82ee77 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 2 Feb 2024 16:27:57 +0100
Subject: [PATCH 056/566] __pattern_walk2_brick + tag impls

---
 include/oneapi/dpl/pstl/algorithm_fwd.h       | 16 +++++
 include/oneapi/dpl/pstl/algorithm_impl.h      | 58 +++++++++++++++++++
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 12 ++++
 3 files changed, 86 insertions(+)

diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index 591553e4f9b..d6743445c26 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -251,6 +251,11 @@ oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _Fo
 __pattern_walk2_brick(_ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2, _Brick,
                       /*parallel=*/::std::false_type) noexcept;
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Brick>
+_ForwardIterator2
+__pattern_walk2_brick(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1, _ForwardIterator2,
+                      _Brick) noexcept;
+
 template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _Brick>
 oneapi::dpl::__internal::__enable_if_host_execution_policy_conditional<
     _ExecutionPolicy, __is_random_access_iterator_v<_RandomAccessIterator1, _RandomAccessIterator2>,
@@ -258,6 +263,12 @@ oneapi::dpl::__internal::__enable_if_host_execution_policy_conditional<
 __pattern_walk2_brick(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1, _RandomAccessIterator1 __last1,
                       _RandomAccessIterator2 __first2, _Brick __brick, /*parallel=*/::std::true_type);
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
+          class _Brick>
+_RandomAccessIterator2
+__pattern_walk2_brick(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator1, _RandomAccessIterator1,
+                      _RandomAccessIterator2, _Brick);
+
 template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _Brick>
 oneapi::dpl::__internal::__enable_if_host_execution_policy_conditional<
     _ExecutionPolicy, !__is_random_access_iterator_v<_RandomAccessIterator1, _RandomAccessIterator2>,
@@ -265,6 +276,11 @@ oneapi::dpl::__internal::__enable_if_host_execution_policy_conditional<
 __pattern_walk2_brick(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1, _RandomAccessIterator1 __last1,
                       _RandomAccessIterator2 __first2, _Brick __brick, /*parallel=*/::std::true_type);
 
+template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Brick>
+_ForwardIterator2
+__pattern_walk2_brick(__parallel_forward_tag, _ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1,
+                      _ForwardIterator2, _Brick);
+
 template <class _ExecutionPolicy, class _ForwardIterator1, class _Size, class _ForwardIterator2, class _Brick>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _ForwardIterator2>
 __pattern_walk2_brick_n(_ExecutionPolicy&&, _ForwardIterator1, _Size, _ForwardIterator2, _Brick,
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 90806aa392f..097d1e5f3db 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -580,6 +580,16 @@ __pattern_walk2_brick(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _Fo
     return __brick(__first1, __last1, __first2, __is_vector);
 }
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Brick>
+_ForwardIterator2
+__pattern_walk2_brick(_Tag, _ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
+                      _ForwardIterator2 __first2, _Brick __brick) noexcept
+{
+    static_assert(__is_backend_tag_serial_v<_Tag>);
+
+    return __brick(__first1, __last1, __first2, typename _Tag::__is_vector{});
+}
+
 template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _Brick>
 oneapi::dpl::__internal::__enable_if_host_execution_policy_conditional<
     _ExecutionPolicy, __is_random_access_iterator_v<_RandomAccessIterator1, _RandomAccessIterator2>,
@@ -600,6 +610,24 @@ __pattern_walk2_brick(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1
     });
 }
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
+          class _Brick>
+_RandomAccessIterator2
+__pattern_walk2_brick(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1,
+                      _RandomAccessIterator1 __last1, _RandomAccessIterator2 __first2, _Brick __brick)
+{
+    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
+
+    return __except_handler([&]() {
+        __par_backend::__parallel_for(
+            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1,
+            [__first1, __first2, __brick](_RandomAccessIterator1 __i, _RandomAccessIterator1 __j) {
+                __brick(__i, __j, __first2 + (__i - __first1), _IsVector{});
+            });
+        return __first2 + (__last1 - __first1);
+    });
+}
+
 //TODO: it postponed till adding more or less effective parallel implementation
 template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Brick>
 oneapi::dpl::__internal::__enable_if_host_execution_policy_conditional<
@@ -629,6 +657,36 @@ __pattern_walk2_brick(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _Fo
     });
 }
 
+//TODO: it postponed till adding more or less effective parallel implementation
+template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Brick>
+_ForwardIterator2
+__pattern_walk2_brick(__parallel_forward_tag, _ExecutionPolicy&& __exec, _ForwardIterator1 __first1,
+                      _ForwardIterator1 __last1, _ForwardIterator2 __first2, _Brick __brick)
+{
+    using __backend_tag = typename __parallel_forward_tag::__backend_tag;
+
+    using _iterator_tuple = zip_forward_iterator<_ForwardIterator1, _ForwardIterator2>;
+    auto __begin = _iterator_tuple(__first1, __first2);
+    auto __end = _iterator_tuple(__last1, /*dummy parameter*/ _ForwardIterator2());
+
+    typedef typename ::std::iterator_traits<_ForwardIterator1>::reference _ReferenceType1;
+    typedef typename ::std::iterator_traits<_ForwardIterator2>::reference _ReferenceType2;
+
+    return __except_handler([&]() {
+        __par_backend::__parallel_for_each(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __begin, __end,
+                                           [__brick](::std::tuple<_ReferenceType1, _ReferenceType2> __val) {
+                                               __brick(::std::get<0>(__val),
+                                                       ::std::forward<_ReferenceType2>(::std::get<1>(__val)));
+                                           });
+
+        //TODO: parallel_for_each does not allow to return correct iterator value according to the ::std::transform
+        // implementation. Therefore, iterator value is calculated separately.
+        for (; __begin != __end; ++__begin)
+            ;
+        return ::std::get<1>(__begin.base());
+    });
+}
+
 template <class _ExecutionPolicy, class _RandomAccessIterator1, class _Size, class _RandomAccessIterator2, class _Brick>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator2>
 __pattern_walk2_brick_n(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1, _Size __n,
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index a430ee845cd..3aca94ad160 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -357,6 +357,18 @@ __pattern_walk2_brick(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _Fo
         /*vector=*/::std::true_type{}, /*parallel*/ ::std::true_type{});
 }
 
+template <typename _BackendTag, typename _ExecutionPolicy, typename _ForwardIterator1, typename _ForwardIterator2,
+          typename _Brick>
+_ForwardIterator2
+__pattern_walk2_brick(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _ForwardIterator1 __first1,
+                      _ForwardIterator1 __last1, _ForwardIterator2 __first2, _Brick __brick)
+{
+    return __pattern_walk2(
+        __tag,
+        __par_backend_hetero::make_wrapped_policy<__walk2_brick_wrapper>(::std::forward<_ExecutionPolicy>(__exec)),
+        __first1, __last1, __first2, __brick);
+}
+
 template <typename _Name>
 struct __walk2_brick_n_wrapper
 {

From 79c3f1215f234044c722878016772007e3d11d73 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 2 Feb 2024 13:26:57 +0100
Subject: [PATCH 057/566] __pattern_walk2_brick + tag calls

---
 include/oneapi/dpl/pstl/algorithm_impl.h      | 48 ++++++++++---------
 include/oneapi/dpl/pstl/glue_algorithm_impl.h | 16 ++++---
 include/oneapi/dpl/pstl/glue_memory_impl.h    | 26 ++++------
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 31 ++++++++----
 .../dpl/pstl/hetero/numeric_impl_hetero.h     | 15 +++---
 5 files changed, 76 insertions(+), 60 deletions(-)

diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 097d1e5f3db..874e2883f37 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -3523,6 +3523,8 @@ __parallel_set_union_op(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _
 {
     typedef typename ::std::iterator_traits<_ForwardIterator1>::difference_type _DifferenceType;
 
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2, _OutputIterator>();
+
     const auto __n1 = __last1 - __first1;
     const auto __n2 = __last2 - __first2;
 
@@ -3530,13 +3532,13 @@ __parallel_set_union_op(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _
 
     // {1} {}: parallel copying just first sequence
     if (__n2 == 0)
-        return __internal::__pattern_walk2_brick(::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __result,
-                                                 __copy_range, ::std::true_type());
+        return __internal::__pattern_walk2_brick(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first1,
+                                                 __last1, __result, __copy_range);
 
     // {} {2}: parallel copying justmake  second sequence
     if (__n1 == 0)
-        return __internal::__pattern_walk2_brick(::std::forward<_ExecutionPolicy>(__exec), __first2, __last2, __result,
-                                                 __copy_range, ::std::true_type());
+        return __internal::__pattern_walk2_brick(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first2,
+                                                 __last2, __result, __copy_range);
 
     // testing  whether the sequences are intersected
     _ForwardIterator1 __left_bound_seq_1 = ::std::lower_bound(__first1, __last1, *__first2, __comp);
@@ -3547,12 +3549,12 @@ __parallel_set_union_op(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _
         __par_backend::__parallel_invoke(
             ::std::forward<_ExecutionPolicy>(__exec),
             [=] {
-                __internal::__pattern_walk2_brick(::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __result,
-                                                  __copy_range, ::std::true_type());
+                __internal::__pattern_walk2_brick(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first1,
+                                                  __last1, __result, __copy_range);
             },
             [=] {
-                __internal::__pattern_walk2_brick(::std::forward<_ExecutionPolicy>(__exec), __first2, __last2,
-                                                  __result + __n1, __copy_range, ::std::true_type());
+                __internal::__pattern_walk2_brick(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first2,
+                                                  __last2, __result + __n1, __copy_range);
             });
         return __result + __n1 + __n2;
     }
@@ -3566,12 +3568,12 @@ __parallel_set_union_op(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _
         __par_backend::__parallel_invoke(
             ::std::forward<_ExecutionPolicy>(__exec),
             [=] {
-                __internal::__pattern_walk2_brick(::std::forward<_ExecutionPolicy>(__exec), __first2, __last2, __result,
-                                                  __copy_range, ::std::true_type());
+                __internal::__pattern_walk2_brick(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first2,
+                                                  __last2, __result, __copy_range);
             },
             [=] {
-                __internal::__pattern_walk2_brick(::std::forward<_ExecutionPolicy>(__exec), __first1, __last1,
-                                                  __result + __n2, __copy_range, ::std::true_type());
+                __internal::__pattern_walk2_brick(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first1,
+                                                  __last1, __result + __n2, __copy_range);
             });
         return __result + __n1 + __n2;
     }
@@ -3585,8 +3587,8 @@ __parallel_set_union_op(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _
             ::std::forward<_ExecutionPolicy>(__exec),
             //do parallel copying of [first1; left_bound_seq_1)
             [=] {
-                __internal::__pattern_walk2_brick(::std::forward<_ExecutionPolicy>(__exec), __first1,
-                                                  __left_bound_seq_1, __res_or, __copy_range, ::std::true_type());
+                __internal::__pattern_walk2_brick(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first1,
+                                                  __left_bound_seq_1, __res_or, __copy_range);
             },
             [=, &__result] {
                 __result = __internal::__parallel_set_op(
@@ -3607,8 +3609,8 @@ __parallel_set_union_op(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _
             ::std::forward<_ExecutionPolicy>(__exec),
             //do parallel copying of [first2; left_bound_seq_2)
             [=] {
-                __internal::__pattern_walk2_brick(::std::forward<_ExecutionPolicy>(__exec), __first2,
-                                                  __left_bound_seq_2, __res_or, __copy_range, ::std::true_type());
+                __internal::__pattern_walk2_brick(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first2,
+                                                  __left_bound_seq_2, __res_or, __copy_range);
             },
             [=, &__result] {
                 __result = __internal::__parallel_set_op(
@@ -3838,6 +3840,8 @@ __pattern_set_difference(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __fir
     typedef typename ::std::iterator_traits<_RandomAccessIterator3>::value_type _T;
     typedef typename ::std::iterator_traits<_RandomAccessIterator1>::difference_type _DifferenceType;
 
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2, _RandomAccessIterator3>();
+
     const auto __n1 = __last1 - __first1;
     const auto __n2 = __last2 - __first2;
 
@@ -3847,22 +3851,22 @@ __pattern_set_difference(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __fir
 
     // {1} \ {}: parallel copying just first sequence
     if (__n2 == 0)
-        return __pattern_walk2_brick(::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __result,
-                                     __internal::__brick_copy<_ExecutionPolicy>{}, ::std::true_type());
+        return __pattern_walk2_brick(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1,
+                                     __result, __internal::__brick_copy<_ExecutionPolicy>{});
 
     // testing  whether the sequences are intersected
     _RandomAccessIterator1 __left_bound_seq_1 = ::std::lower_bound(__first1, __last1, *__first2, __comp);
     //{1} < {2}: seq 2 is wholly greater than seq 1, so, parallel copying just first sequence
     if (__left_bound_seq_1 == __last1)
-        return __pattern_walk2_brick(::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __result,
-                                     __internal::__brick_copy<_ExecutionPolicy>{}, ::std::true_type());
+        return __pattern_walk2_brick(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1,
+                                     __result, __internal::__brick_copy<_ExecutionPolicy>{});
 
     // testing  whether the sequences are intersected
     _RandomAccessIterator2 __left_bound_seq_2 = ::std::lower_bound(__first2, __last2, *__first1, __comp);
     //{2} < {1}: seq 1 is wholly greater than seq 2, so, parallel copying just first sequence
     if (__left_bound_seq_2 == __last2)
-        return __internal::__pattern_walk2_brick(::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __result,
-                                                 __brick_copy<_ExecutionPolicy>{}, ::std::true_type());
+        return __internal::__pattern_walk2_brick(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first1,
+                                                 __last1, __result, __brick_copy<_ExecutionPolicy>{});
 
     if (__n1 + __n2 > __set_algo_cut_off)
         return __parallel_set_op(
diff --git a/include/oneapi/dpl/pstl/glue_algorithm_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
index 2b77e0c44b7..023264dd732 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
@@ -266,10 +266,11 @@ template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterato
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _ForwardIterator2>
 copy(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _ForwardIterator1 __last, _ForwardIterator2 __result)
 {
-    return oneapi::dpl::__internal::__pattern_walk2_brick(
-        ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result,
-        oneapi::dpl::__internal::__brick_copy<_ExecutionPolicy>{},
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>());
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
+
+    return oneapi::dpl::__internal::__pattern_walk2_brick(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
+                                                          __first, __last, __result,
+                                                          oneapi::dpl::__internal::__brick_copy<_ExecutionPolicy>{});
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator1, class _Size, class _ForwardIterator2>
@@ -810,10 +811,11 @@ move(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _ForwardIterator1 __l
 {
     using _DecayedExecutionPolicy = ::std::decay_t<_ExecutionPolicy>;
 
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
+
     return oneapi::dpl::__internal::__pattern_walk2_brick(
-        ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __d_first,
-        oneapi::dpl::__internal::__brick_move<_DecayedExecutionPolicy>{},
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>());
+        __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __d_first,
+        oneapi::dpl::__internal::__brick_move<_DecayedExecutionPolicy>{});
 }
 
 // [partial.sort]
diff --git a/include/oneapi/dpl/pstl/glue_memory_impl.h b/include/oneapi/dpl/pstl/glue_memory_impl.h
index 72e2f6cc1ad..d96cc8a62bc 100644
--- a/include/oneapi/dpl/pstl/glue_memory_impl.h
+++ b/include/oneapi/dpl/pstl/glue_memory_impl.h
@@ -45,19 +45,16 @@ uninitialized_copy(_ExecutionPolicy&& __exec, _InputIterator __first, _InputIter
     typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _ValueType2;
     typedef ::std::decay_t<_ExecutionPolicy> _DecayedExecutionPolicy;
 
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _InputIterator, _ForwardIterator>();
+
     if constexpr (::std::is_trivial_v<_ValueType1> && ::std::is_trivial_v<_ValueType2>)
     {
-        constexpr auto __is_parallel =
-            oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _InputIterator, _ForwardIterator>();
-
         return oneapi::dpl::__internal::__pattern_walk2_brick(
-            ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result,
-            oneapi::dpl::__internal::__brick_copy<_DecayedExecutionPolicy>{}, __is_parallel);
+            __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result,
+            oneapi::dpl::__internal::__brick_copy<_DecayedExecutionPolicy>{});
     }
     else
     {
-        constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _InputIterator, _ForwardIterator>();
-
         return oneapi::dpl::__internal::__pattern_walk2(
             __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result,
             oneapi::dpl::__internal::__op_uninitialized_copy<_DecayedExecutionPolicy>{});
@@ -72,7 +69,7 @@ uninitialized_copy_n(_ExecutionPolicy&& __exec, _InputIterator __first, _Size __
     typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _ValueType2;
     typedef ::std::decay_t<_ExecutionPolicy> _DecayedExecutionPolicy;
 
-     const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _InputIterator, _ForwardIterator>();
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _InputIterator, _ForwardIterator>();
 
     constexpr auto __is_parallel =
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _InputIterator, _ForwardIterator>();
@@ -101,19 +98,16 @@ uninitialized_move(_ExecutionPolicy&& __exec, _InputIterator __first, _InputIter
     typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _ValueType2;
     typedef ::std::decay_t<_ExecutionPolicy> _DecayedExecutionPolicy;
 
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _InputIterator, _ForwardIterator>();
+
     if constexpr (::std::is_trivial_v<_ValueType1> && ::std::is_trivial_v<_ValueType2>)
     {
-        constexpr auto __is_parallel =
-            oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _InputIterator, _ForwardIterator>();
-
         return oneapi::dpl::__internal::__pattern_walk2_brick(
-            ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result,
-            oneapi::dpl::__internal::__brick_copy<_DecayedExecutionPolicy>{}, __is_parallel);
+            __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result,
+            oneapi::dpl::__internal::__brick_copy<_DecayedExecutionPolicy>{});
     }
     else
     {
-        const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _InputIterator, _ForwardIterator>();
-
         return oneapi::dpl::__internal::__pattern_walk2(
             __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result,
             oneapi::dpl::__internal::__op_uninitialized_move<_DecayedExecutionPolicy>{});
@@ -128,7 +122,7 @@ uninitialized_move_n(_ExecutionPolicy&& __exec, _InputIterator __first, _Size __
     typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _ValueType2;
     typedef ::std::decay_t<_ExecutionPolicy> _DecayedExecutionPolicy;
 
-    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _InputIterator, _ForwardIterator>();
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _InputIterator, _ForwardIterator>();
 
     constexpr auto __is_parallel =
         oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _InputIterator, _ForwardIterator>();
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 3aca94ad160..d47a621ccb5 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -1320,19 +1320,21 @@ __pattern_merge(_ExecutionPolicy&& __exec, _Iterator1 __first1, _Iterator1 __las
     if (__n == 0)
         return __d_first;
 
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator1, _Iterator2, _Iterator3>();
+
     //To consider the direct copying pattern call in case just one of sequences is empty.
     if (__n1 == 0)
         oneapi::dpl::__internal::__pattern_walk2_brick(
+            __dispatch_tag,
             oneapi::dpl::__par_backend_hetero::make_wrapped_policy<copy_back_wrapper>(
                 ::std::forward<_ExecutionPolicy>(__exec)),
-            __first2, __last2, __d_first, oneapi::dpl::__internal::__brick_copy<_ExecutionPolicy>{},
-            ::std::true_type());
+            __first2, __last2, __d_first, oneapi::dpl::__internal::__brick_copy<_ExecutionPolicy>{});
     else if (__n2 == 0)
         oneapi::dpl::__internal::__pattern_walk2_brick(
+            __dispatch_tag,
             oneapi::dpl::__par_backend_hetero::make_wrapped_policy<copy_back_wrapper2>(
                 ::std::forward<_ExecutionPolicy>(__exec)),
-            __first1, __last1, __d_first, oneapi::dpl::__internal::__brick_copy<_ExecutionPolicy>{},
-            ::std::true_type());
+            __first1, __last1, __d_first, oneapi::dpl::__internal::__brick_copy<_ExecutionPolicy>{});
     else
     {
         auto __keep1 = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _Iterator1>();
@@ -1930,10 +1932,13 @@ __pattern_set_difference(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1,
     // {1} \ {}: the difference is {1}
     if (__first2 == __last2)
     {
+        const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2, _OutputIterator>();
+
         return oneapi::dpl::__internal::__pattern_walk2_brick(
+            __dispatch_tag,
             oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__set_difference_copy_case_1>(
                 ::std::forward<_ExecutionPolicy>(__exec)),
-            __first1, __last1, __result, oneapi::dpl::__internal::__brick_copy<_ExecutionPolicy>{}, ::std::true_type());
+            __first1, __last1, __result, oneapi::dpl::__internal::__brick_copy<_ExecutionPolicy>{});
     }
 
     return __pattern_hetero_set_op(::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2,
@@ -1961,22 +1966,26 @@ __pattern_set_union(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _Forw
     if (__first1 == __last1 && __first2 == __last2)
         return __result;
 
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2, _OutputIterator>();
+
     //{1} is empty
     if (__first1 == __last1)
     {
         return oneapi::dpl::__internal::__pattern_walk2_brick(
+            __dispatch_tag,
             oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__set_union_copy_case_1>(
                 ::std::forward<_ExecutionPolicy>(__exec)),
-            __first2, __last2, __result, oneapi::dpl::__internal::__brick_copy<_ExecutionPolicy>{}, ::std::true_type());
+            __first2, __last2, __result, oneapi::dpl::__internal::__brick_copy<_ExecutionPolicy>{});
     }
 
     //{2} is empty
     if (__first2 == __last2)
     {
         return oneapi::dpl::__internal::__pattern_walk2_brick(
+            __dispatch_tag,
             oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__set_union_copy_case_2>(
                 ::std::forward<_ExecutionPolicy>(__exec)),
-            __first1, __last1, __result, oneapi::dpl::__internal::__brick_copy<_ExecutionPolicy>{}, ::std::true_type());
+            __first1, __last1, __result, oneapi::dpl::__internal::__brick_copy<_ExecutionPolicy>{});
     }
 
     typedef typename ::std::iterator_traits<_OutputIterator>::value_type _ValueType;
@@ -2037,22 +2046,26 @@ __pattern_set_symmetric_difference(_ExecutionPolicy&& __exec, _ForwardIterator1
     if (__first1 == __last1 && __first2 == __last2)
         return __result;
 
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2, _OutputIterator>();
+
     //{1} is empty
     if (__first1 == __last1)
     {
         return oneapi::dpl::__internal::__pattern_walk2_brick(
+            __dispatch_tag,
             oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__set_symmetric_difference_copy_case_1>(
                 ::std::forward<_ExecutionPolicy>(__exec)),
-            __first2, __last2, __result, oneapi::dpl::__internal::__brick_copy<_ExecutionPolicy>{}, ::std::true_type());
+            __first2, __last2, __result, oneapi::dpl::__internal::__brick_copy<_ExecutionPolicy>{});
     }
 
     //{2} is empty
     if (__first2 == __last2)
     {
         return oneapi::dpl::__internal::__pattern_walk2_brick(
+            __dispatch_tag,
             oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__set_symmetric_difference_copy_case_2>(
                 ::std::forward<_ExecutionPolicy>(__exec)),
-            __first1, __last1, __result, oneapi::dpl::__internal::__brick_copy<_ExecutionPolicy>{}, ::std::true_type());
+            __first1, __last1, __result, oneapi::dpl::__internal::__brick_copy<_ExecutionPolicy>{});
     }
 
     typedef typename ::std::iterator_traits<_OutputIterator>::value_type _ValueType;
diff --git a/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
index 60c1001f5b8..90cf7764f4d 100644
--- a/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
@@ -131,6 +131,8 @@ __pattern_transform_scan_base(_ExecutionPolicy&& __exec, _Iterator1 __first, _It
     if (__first == __last)
         return __result;
 
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator1, _Iterator2>();
+
     const auto __n = __last - __first;
 
     auto __keep1 = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _Iterator1>();
@@ -173,9 +175,9 @@ __pattern_transform_scan_base(_ExecutionPolicy&& __exec, _Iterator1 __first, _It
             .wait();
 
         // Move data from temporary buffer into results
-        oneapi::dpl::__internal::__pattern_walk2_brick(::std::move(__policy), __first_tmp, __last_tmp, __result,
-                                                       oneapi::dpl::__internal::__brick_move<_NewExecutionPolicy>{},
-                                                       ::std::true_type{});
+        oneapi::dpl::__internal::__pattern_walk2_brick(__dispatch_tag, ::std::move(__policy), __first_tmp, __last_tmp,
+                                                       __result,
+                                                       oneapi::dpl::__internal::__brick_move<_NewExecutionPolicy>{});
 
         //TODO: optimize copy back depending on Iterator, i.e. set_final_data for host iterator/pointer
     }
@@ -233,6 +235,8 @@ __pattern_adjacent_difference(_ExecutionPolicy&& __exec, _ForwardIterator1 __fir
     if (__n <= 0)
         return __d_first;
 
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
+
     using _It1ValueT = typename ::std::iterator_traits<_ForwardIterator1>::value_type;
     using _It2ValueTRef = typename ::std::iterator_traits<_ForwardIterator2>::reference;
 
@@ -246,9 +250,8 @@ __pattern_adjacent_difference(_ExecutionPolicy&& __exec, _ForwardIterator1 __fir
             auto __wrapped_policy = __par_backend_hetero::make_wrapped_policy<adjacent_difference_wrapper>(
                 ::std::forward<_ExecutionPolicy>(__exec));
 
-            __internal::__pattern_walk2_brick(__wrapped_policy, __first, __last, __d_first,
-                                              __internal::__brick_copy<decltype(__wrapped_policy)>{},
-                                              ::std::true_type{});
+            __internal::__pattern_walk2_brick(__dispatch_tag, __wrapped_policy, __first, __last, __d_first,
+                                              __internal::__brick_copy<decltype(__wrapped_policy)>{});
 
             return __d_last;
         });

From 845409b26126e45b6da33116fe06711bf940f56c Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Mon, 5 Feb 2024 10:31:40 +0100
Subject: [PATCH 058/566] include/oneapi/dpl/pstl/algorithm_impl.h - fix
 compile error in struct __brick_move - implement operator()(_ReferenceType1&&
 __val, _ReferenceType2&& __result) const

---
 include/oneapi/dpl/pstl/algorithm_impl.h | 7 +++++++
 1 file changed, 7 insertions(+)

diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 874e2883f37..26da51f89f5 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -1437,6 +1437,13 @@ struct __brick_move<_ExecutionPolicy, oneapi::dpl::__internal::__enable_if_host_
     {
         return ::std::move(__first, __last, __result);
     }
+
+    template <typename _ReferenceType1, typename _ReferenceType2>
+    void
+    operator()(_ReferenceType1&& __val, _ReferenceType2&& __result) const
+    {
+        ::std::forward<_ReferenceType2>(__result) = ::std::move(::std::forward<_ReferenceType1>(__val));
+    }
 };
 
 template <typename _ExecutionPolicy, typename _Dummy = void>

From 25c5cdecdec92430e86f2cabfcf16c8c2174db38 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 2 Feb 2024 17:53:41 +0100
Subject: [PATCH 059/566] __pattern_walk2_brick_n + tag impls

---
 include/oneapi/dpl/pstl/algorithm_fwd.h       | 13 ++++++++-
 include/oneapi/dpl/pstl/algorithm_impl.h      | 29 +++++++++++++++++++
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 12 ++++++++
 3 files changed, 53 insertions(+), 1 deletion(-)

diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index d6743445c26..e03bc5f95bd 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -253,7 +253,7 @@ __pattern_walk2_brick(_ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1,
 
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Brick>
 _ForwardIterator2
-__pattern_walk2_brick(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1, _ForwardIterator2,
+__pattern_walk2_brick(_Tag, _ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2,
                       _Brick) noexcept;
 
 template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _Brick>
@@ -286,11 +286,22 @@ oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _Fo
 __pattern_walk2_brick_n(_ExecutionPolicy&&, _ForwardIterator1, _Size, _ForwardIterator2, _Brick,
                         /*parallel=*/::std::false_type) noexcept;
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _Size, class _ForwardIterator2,
+          class _Brick>
+_ForwardIterator2
+__pattern_walk2_brick_n(_Tag, _ExecutionPolicy&&, _ForwardIterator1, _Size, _ForwardIterator2, _Brick) noexcept;
+
 template <class _ExecutionPolicy, class _RandomAccessIterator1, class _Size, class _RandomAccessIterator2, class _Brick>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator2>
 __pattern_walk2_brick_n(_ExecutionPolicy&&, _RandomAccessIterator1, _Size, _RandomAccessIterator2, _Brick,
                         /*parallel=*/::std::true_type);
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _Size,
+          class _RandomAccessIterator2, class _Brick>
+_RandomAccessIterator2
+__pattern_walk2_brick_n(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator1, _Size,
+                        _RandomAccessIterator2, _Brick);
+
 //------------------------------------------------------------------------
 // walk3 (pseudo)
 //
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 26da51f89f5..9d1deab8f41 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -705,6 +705,24 @@ __pattern_walk2_brick_n(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __firs
     });
 }
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _Size, class _RandomAccessIterator2, class _Brick>
+_RandomAccessIterator2
+__pattern_walk2_brick_n(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1,
+                        _Size __n, _RandomAccessIterator2 __first2, _Brick __brick)
+{
+    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
+
+    return __except_handler([&]() {
+        __par_backend::__parallel_for(
+            __backend_tag{},
+            ::std::forward<_ExecutionPolicy>(__exec), __first1, __first1 + __n,
+            [__first1, __first2, __brick](_RandomAccessIterator1 __i, _RandomAccessIterator1 __j) {
+                __brick(__i, __j - __i, __first2 + (__i - __first1), _IsVector{});
+            });
+        return __first2 + __n;
+    });
+}
+
 template <class _ExecutionPolicy, class _ForwardIterator1, class _Size, class _ForwardIterator2, class _Brick>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _ForwardIterator2>
 __pattern_walk2_brick_n(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _Size __n, _ForwardIterator2 __first2,
@@ -715,6 +733,17 @@ __pattern_walk2_brick_n(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _
     return __brick(__first1, __n, __first2, __is_vector);
 }
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _Size, class _ForwardIterator2,
+          class _Brick>
+_ForwardIterator2
+__pattern_walk2_brick_n(_Tag, _ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _Size __n,
+                        _ForwardIterator2 __first2, _Brick __brick) noexcept
+{
+    static_assert(__is_backend_tag_v<_Tag>);
+
+    return __brick(__first1, __n, __first2, typename _Tag::__is_vector{});
+}
+
 //------------------------------------------------------------------------
 // walk3 (pseudo)
 //
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index d47a621ccb5..d53d2518c3b 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -387,6 +387,18 @@ __pattern_walk2_brick_n(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _
         /*vector=*/::std::true_type{}, /*parallel*/ ::std::true_type{});
 }
 
+template <typename _BackendTag, typename _ExecutionPolicy, typename _ForwardIterator1, typename _Size,
+          typename _ForwardIterator2, typename _Brick>
+_ForwardIterator2
+__pattern_walk2_brick_n(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _ForwardIterator1 __first1,
+                        _Size __n, _ForwardIterator2 __first2, _Brick __brick)
+{
+    return __pattern_walk2(
+        __tag,
+        __par_backend_hetero::make_wrapped_policy<__walk2_brick_n_wrapper>(::std::forward<_ExecutionPolicy>(__exec)),
+        __first1, __first1 + __n, __first2, __brick);
+}
+
 //------------------------------------------------------------------------
 // transform_if
 //------------------------------------------------------------------------

From 7261791f695bb2a71fc7258566c44bbf42476545 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 2 Feb 2024 17:54:16 +0100
Subject: [PATCH 060/566] __pattern_walk2_brick_n + tag calls

---
 include/oneapi/dpl/pstl/glue_algorithm_impl.h |  6 ++++--
 include/oneapi/dpl/pstl/glue_memory_impl.h    | 14 ++++----------
 2 files changed, 8 insertions(+), 12 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_algorithm_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
index 023264dd732..afd5e9483e7 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
@@ -279,10 +279,12 @@ copy_n(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _Size __n, _Forward
 {
     using _DecayedExecutionPolicy = ::std::decay_t<_ExecutionPolicy>;
 
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
+
     return oneapi::dpl::__internal::__pattern_walk2_brick_n(
+        __dispatch_tag,
         ::std::forward<_ExecutionPolicy>(__exec), __first, __n, __result,
-        oneapi::dpl::__internal::__brick_copy_n<_DecayedExecutionPolicy>{},
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>());
+        oneapi::dpl::__internal::__brick_copy_n<_DecayedExecutionPolicy>{});
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Predicate>
diff --git a/include/oneapi/dpl/pstl/glue_memory_impl.h b/include/oneapi/dpl/pstl/glue_memory_impl.h
index d96cc8a62bc..cc46b189b0f 100644
--- a/include/oneapi/dpl/pstl/glue_memory_impl.h
+++ b/include/oneapi/dpl/pstl/glue_memory_impl.h
@@ -71,14 +71,11 @@ uninitialized_copy_n(_ExecutionPolicy&& __exec, _InputIterator __first, _Size __
 
     constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _InputIterator, _ForwardIterator>();
 
-    constexpr auto __is_parallel =
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _InputIterator, _ForwardIterator>();
-
     if constexpr (::std::is_trivial_v<_ValueType1> && ::std::is_trivial_v<_ValueType2>)
     {
         return oneapi::dpl::__internal::__pattern_walk2_brick_n(
-            ::std::forward<_ExecutionPolicy>(__exec), __first, __n, __result,
-            oneapi::dpl::__internal::__brick_copy_n<_DecayedExecutionPolicy>{}, __is_parallel);
+            __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __n, __result,
+            oneapi::dpl::__internal::__brick_copy_n<_DecayedExecutionPolicy>{});
     }
     else
     {
@@ -124,14 +121,11 @@ uninitialized_move_n(_ExecutionPolicy&& __exec, _InputIterator __first, _Size __
 
     constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _InputIterator, _ForwardIterator>();
 
-    constexpr auto __is_parallel =
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _InputIterator, _ForwardIterator>();
-
     if constexpr (::std::is_trivial_v<_ValueType1> && ::std::is_trivial_v<_ValueType2>)
     {
         return oneapi::dpl::__internal::__pattern_walk2_brick_n(
-            ::std::forward<_ExecutionPolicy>(__exec), __first, __n, __result,
-            oneapi::dpl::__internal::__brick_copy_n<_DecayedExecutionPolicy>{}, __is_parallel);
+            __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __n, __result,
+            oneapi::dpl::__internal::__brick_copy_n<_DecayedExecutionPolicy>{});
     }
     else
     {

From 53d7c309225c76e6c15e38ca8faf627f7de57fdf Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Mon, 5 Feb 2024 11:06:55 +0100
Subject: [PATCH 061/566] __pattern_walk2_transform_if + tag impls

---
 include/oneapi/dpl/pstl/algorithm_fwd.h         |  6 ++++++
 include/oneapi/dpl/pstl/algorithm_impl.h        | 11 +++++++++++
 .../dpl/pstl/hetero/algorithm_impl_hetero.h     | 17 +++++++++++++++++
 3 files changed, 34 insertions(+)

diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index e03bc5f95bd..2f8d2d953b8 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -373,6 +373,12 @@ __pattern_walk2_transform_if(_ExecutionPolicy&& __exec, _ForwardIterator1 __firs
                              _ForwardIterator2 __first2, _Function __func, _IsVector __is_vector,
                              _IsParallel __is_parallel) noexcept;
 
+template <class _Tag, typename _ExecutionPolicy, typename _ForwardIterator1, typename _ForwardIterator2,
+          typename _Function>
+_ForwardIterator2
+__pattern_walk2_transform_if(_Tag, _ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2,
+                             _Function) noexcept;
+
 template <typename _ExecutionPolicy, typename _ForwardIterator1, typename _ForwardIterator2, typename _ForwardIterator3,
           typename _Function, class _IsVector, class _IsParallel>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _ForwardIterator3>
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 9d1deab8f41..64d814a7049 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -904,6 +904,17 @@ __pattern_walk2_transform_if(_ExecutionPolicy&& __exec, _ForwardIterator1 __firs
                            __is_parallel);
 }
 
+template <class _Tag, typename _ExecutionPolicy, typename _ForwardIterator1, typename _ForwardIterator2,
+          typename _Function>
+_ForwardIterator2
+__pattern_walk2_transform_if(_Tag __tag, _ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
+                             _ForwardIterator2 __first2, _Function __func) noexcept
+{
+    static_assert(__is_backend_tag_v<_Tag>);
+
+    return __pattern_walk2(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __func);
+}
+
 template <typename _ExecutionPolicy, typename _ForwardIterator1, typename _ForwardIterator2, typename _ForwardIterator3,
           typename _Function, class _IsVector, class _IsParallel>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _ForwardIterator3>
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index d53d2518c3b..39b2f4c032a 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -429,6 +429,23 @@ __pattern_walk2_transform_if(_ExecutionPolicy&& __exec, _ForwardIterator1 __firs
         __first1, __last1, __first2, __func);
 }
 
+template <typename _BackendTag, typename _ExecutionPolicy, typename _ForwardIterator1, typename _ForwardIterator2,
+          typename _Function>
+_ForwardIterator2
+__pattern_walk2_transform_if(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _ForwardIterator1 __first1,
+                             _ForwardIterator1 __last1, _ForwardIterator2 __first2, _Function __func)
+{
+    // Require `read_write` access mode for output sequence to force a copy in for host iterators to capture incoming
+    // values of the output sequence for elements where the predicate is false.
+    return __pattern_walk2<_BackendTag, /*_IsSync=*/::std::true_type,
+                           __par_backend_hetero::access_mode::read,
+                           __par_backend_hetero::access_mode::read_write>(
+        __tag,
+        __par_backend_hetero::make_wrapped_policy<__walk2_transform_if_wrapper>(
+            ::std::forward<_ExecutionPolicy>(__exec)),
+        __first1, __last1, __first2, __func);
+}
+
 template <typename _Name>
 struct __walk3_transform_if_wrapper
 {

From e3bf163021fda673d818a000b2b9a2edd03d241a Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Mon, 5 Feb 2024 11:07:10 +0100
Subject: [PATCH 062/566] __pattern_walk2_transform_if + tag calls

---
 include/oneapi/dpl/pstl/glue_algorithm_impl.h | 9 +++++----
 1 file changed, 5 insertions(+), 4 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_algorithm_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
index afd5e9483e7..c119a978671 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
@@ -355,12 +355,13 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _Forward
 transform_if(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _ForwardIterator1 __last, _ForwardIterator2 __result,
              _UnaryOperation __op, _UnaryPredicate __pred)
 {
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
+
     return oneapi::dpl::__internal::__pattern_walk2_transform_if(
-        ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result,
+        __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result,
         oneapi::dpl::__internal::__transform_if_unary_functor<_UnaryOperation, _UnaryPredicate>(::std::move(__op),
-                                                                                                ::std::move(__pred)),
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>());
+                                                                                                ::std::move(__pred)));
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _ForwardIterator3,

From 888364df0f05ba0cc2512157e7aff7d0e478f713 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Mon, 5 Feb 2024 11:35:33 +0100
Subject: [PATCH 063/566] __pattern_fill + tag impls

---
 include/oneapi/dpl/pstl/algorithm_fwd.h       |  8 ++++++
 include/oneapi/dpl/pstl/algorithm_impl.h      | 26 +++++++++++++++++++
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 12 +++++++++
 3 files changed, 46 insertions(+)

diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index 2f8d2d953b8..02ae8f1bc92 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -1050,11 +1050,19 @@ oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>
 __pattern_fill(_ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, const _Tp&,
                /*is_parallel=*/::std::false_type, _IsVector) noexcept;
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Tp>
+void
+__pattern_fill(_Tag, _ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, const _Tp&) noexcept;
+
 template <class _ExecutionPolicy, class _RandomAccessIterator, class _Tp, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator>
 __pattern_fill(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, const _Tp&,
                /*is_parallel=*/::std::true_type, _IsVector);
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Tp>
+_RandomAccessIterator
+__pattern_fill(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, const _Tp&);
+
 template <typename _Tp, typename _ExecutionPolicy, typename = void>
 struct __brick_fill_n;
 
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 64d814a7049..edc2c5af00f 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -3087,6 +3087,15 @@ __pattern_fill(_ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __
     __internal::__brick_fill<_Tp, _ExecutionPolicy>{__value}(__first, __last, __is_vector);
 }
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Tp>
+void
+__pattern_fill(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, const _Tp& __value) noexcept
+{
+    static_assert(__is_backend_tag_v<_Tag>);
+
+    __internal::__brick_fill<_Tp, _ExecutionPolicy>{__value}(__first, __last, typename _Tag::__is_vector{});
+}
+
 template <class _ExecutionPolicy, class _RandomAccessIterator, class _Tp, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator>
 __pattern_fill(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIterator __last,
@@ -3103,6 +3112,23 @@ __pattern_fill(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _Random
     });
 }
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Tp>
+_RandomAccessIterator
+__pattern_fill(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
+               _RandomAccessIterator __last, const _Tp& __value)
+{
+    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
+
+    return __internal::__except_handler([&__exec, __first, __last, &__value]() {
+        __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+                                      [&__value](_RandomAccessIterator __begin, _RandomAccessIterator __end) {
+                                          __internal::__brick_fill<_Tp, _ExecutionPolicy>{__value}(__begin, __end,
+                                                                                                   _IsVector{});
+                                      });
+        return __last;
+    });
+}
+
 template <typename _Tp, typename _ExecutionPolicy>
 struct __brick_fill_n<_Tp, _ExecutionPolicy,
                       oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>>
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 39b2f4c032a..665d2e472e3 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -499,6 +499,18 @@ __pattern_fill(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIter
     return __last;
 }
 
+template <typename _BackendTag, typename _ExecutionPolicy, typename _ForwardIterator, typename _T>
+_ForwardIterator
+__pattern_fill(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _ForwardIterator __first,
+               _ForwardIterator __last, const _T& __value)
+{
+    __pattern_walk1(__tag, ::std::forward<_ExecutionPolicy>(__exec),
+                    __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::write>(__first),
+                    __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::write>(__last),
+                    fill_functor<_T>{__value});
+    return __last;
+}
+
 //------------------------------------------------------------------------
 // generate
 //------------------------------------------------------------------------

From d32a2077c3a9c6ad4e898088bce4625cc3a4dac9 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Mon, 5 Feb 2024 11:35:54 +0100
Subject: [PATCH 064/566] __pattern_fill + tag calls

---
 include/oneapi/dpl/pstl/algorithm_impl.h      | 8 +++++---
 include/oneapi/dpl/pstl/glue_algorithm_impl.h | 8 ++++----
 2 files changed, 9 insertions(+), 7 deletions(-)

diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index edc2c5af00f..1b0d5b6d0d3 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -3163,10 +3163,12 @@ __pattern_fill_n(_ExecutionPolicy&&, _OutputIterator __first, _Size __count, con
 template <class _ExecutionPolicy, class _RandomAccessIterator, class _Size, class _Tp, class _IsVector>
 _RandomAccessIterator
 __pattern_fill_n(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _Size __count, const _Tp& __value,
-                 /*is_parallel=*/::std::true_type, _IsVector __is_vector)
+                 /*is_parallel=*/::std::true_type, _IsVector /*__is_vector*/)
 {
-    return __internal::__pattern_fill(::std::forward<_ExecutionPolicy>(__exec), __first, __first + __count, __value,
-                                      ::std::true_type(), __is_vector);
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
+
+    return __internal::__pattern_fill(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first,
+                                      __first + __count, __value);
 }
 
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/glue_algorithm_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
index c119a978671..c7ad0f1db2a 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
@@ -438,10 +438,10 @@ template <class _ExecutionPolicy, class _ForwardIterator, class _Tp>
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy>
 fill(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last, const _Tp& __value)
 {
-    oneapi::dpl::__internal::__pattern_fill(
-        ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __value,
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(),
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>());
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+
+    oneapi::dpl::__internal::__pattern_fill(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+                                            __value);
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator, class _Size, class _Tp>

From b0afc8f17347ceb57e95628c84b23780428c12be Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Mon, 5 Feb 2024 11:49:40 +0100
Subject: [PATCH 065/566] __pattern_fill_n + tag impls

---
 include/oneapi/dpl/pstl/algorithm_fwd.h  |  8 ++++++++
 include/oneapi/dpl/pstl/algorithm_impl.h | 18 ++++++++++++++++++
 2 files changed, 26 insertions(+)

diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index 02ae8f1bc92..070ef60d2dc 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -1071,11 +1071,19 @@ _OutputIterator
 __pattern_fill_n(_ExecutionPolicy&&, _OutputIterator, _Size, const _Tp&,
                  /*is_parallel=*/::std::false_type, _IsVector) noexcept;
 
+template <class _Tag, class _ExecutionPolicy, class _OutputIterator, class _Size, class _Tp>
+_OutputIterator
+__pattern_fill_n(_Tag, _ExecutionPolicy&&, _OutputIterator, _Size, const _Tp&) noexcept;
+
 template <class _ExecutionPolicy, class _RandomAccessIterator, class _Size, class _Tp, class _IsVector>
 _RandomAccessIterator
 __pattern_fill_n(_ExecutionPolicy&&, _RandomAccessIterator, _Size, const _Tp&,
                  /*is_parallel=*/::std::true_type, _IsVector);
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Size, class _Tp>
+_RandomAccessIterator
+__pattern_fill_n(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator, _Size, const _Tp&);
+
 //------------------------------------------------------------------------
 // generate, generate_n
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 1b0d5b6d0d3..3c6d72b5a1b 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -3160,6 +3160,15 @@ __pattern_fill_n(_ExecutionPolicy&&, _OutputIterator __first, _Size __count, con
     return __internal::__brick_fill_n<_Tp, _ExecutionPolicy>{__value}(__first, __count, __is_vector);
 }
 
+template <class _Tag, class _ExecutionPolicy, class _OutputIterator, class _Size, class _Tp>
+_OutputIterator
+__pattern_fill_n(_Tag, _ExecutionPolicy&&, _OutputIterator __first, _Size __count, const _Tp& __value) noexcept
+{
+    static_assert(__is_backend_tag_v<_Tag>);
+
+    return __internal::__brick_fill_n<_Tp, _ExecutionPolicy>{__value}(__first, __count, typename _Tag::__is_vector{});
+}
+
 template <class _ExecutionPolicy, class _RandomAccessIterator, class _Size, class _Tp, class _IsVector>
 _RandomAccessIterator
 __pattern_fill_n(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _Size __count, const _Tp& __value,
@@ -3171,6 +3180,15 @@ __pattern_fill_n(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _Size
                                       __first + __count, __value);
 }
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Size, class _Tp>
+_RandomAccessIterator
+__pattern_fill_n(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
+                 _Size __count, const _Tp& __value)
+{
+    return __internal::__pattern_fill(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __first + __count,
+                                      __value);
+}
+
 //------------------------------------------------------------------------
 // generate, generate_n
 //------------------------------------------------------------------------

From d87c2f2b47f4258da9c0d247752a7b226f306c27 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Mon, 5 Feb 2024 11:51:23 +0100
Subject: [PATCH 066/566] __pattern_fill_n + tag calls

---
 include/oneapi/dpl/pstl/glue_algorithm_impl.h | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_algorithm_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
index c7ad0f1db2a..128d0c1b35b 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
@@ -451,10 +451,10 @@ fill_n(_ExecutionPolicy&& __exec, _ForwardIterator __first, _Size __count, const
     if (__count <= 0)
         return __first;
 
-    return oneapi::dpl::__internal::__pattern_fill_n(
-        ::std::forward<_ExecutionPolicy>(__exec), __first, __count, __value,
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(),
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>());
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+
+    return oneapi::dpl::__internal::__pattern_fill_n(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first,
+                                                     __count, __value);
 }
 
 // [alg.generate]

From 26b6dcd4a4e2914a0cfc03bd96765941c7365112 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Mon, 5 Feb 2024 12:01:29 +0100
Subject: [PATCH 067/566] __pattern_copy_if + tag impls

---
 include/oneapi/dpl/pstl/algorithm_fwd.h       | 11 +++++
 include/oneapi/dpl/pstl/algorithm_impl.h      | 45 +++++++++++++++++++
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 25 +++++++++++
 3 files changed, 81 insertions(+)

diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index 070ef60d2dc..e81b38d31e1 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -632,12 +632,23 @@ oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _Ou
 __pattern_copy_if(_ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _OutputIterator, _UnaryPredicate, _IsVector,
                   /*parallel=*/::std::false_type) noexcept;
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _OutputIterator, class _UnaryPredicate>
+_OutputIterator
+__pattern_copy_if(_Tag, _ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _OutputIterator,
+                  _UnaryPredicate) noexcept;
+
 template <class _ExecutionPolicy, class _RandomAccessIterator, class _OutputIterator, class _UnaryPredicate,
           class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _OutputIterator>
 __pattern_copy_if(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _OutputIterator, _UnaryPredicate,
                   _IsVector, /*parallel=*/::std::true_type);
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
+          class _UnaryPredicate>
+_RandomAccessIterator2
+__pattern_copy_if(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator1, _RandomAccessIterator1,
+                  _RandomAccessIterator2, _UnaryPredicate);
+
 //------------------------------------------------------------------------
 // count
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 3c6d72b5a1b..3870be47fdd 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -1667,6 +1667,16 @@ __pattern_copy_if(_ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator
     return __internal::__brick_copy_if(__first, __last, __result, __pred, __is_vector);
 }
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _OutputIterator, class _UnaryPredicate>
+_OutputIterator
+__pattern_copy_if(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _OutputIterator __result,
+                  _UnaryPredicate __pred) noexcept
+{
+    static_assert(__is_backend_tag_v<_Tag>);
+
+    return __internal::__brick_copy_if(__first, __last, __result, __pred, typename _Tag::__is_vector{});
+}
+
 template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _UnaryPredicate,
           class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator2>
@@ -1703,6 +1713,41 @@ __pattern_copy_if(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first, _Ra
     return __internal::__brick_copy_if(__first, __last, __result, __pred, __is_vector);
 }
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
+          class _UnaryPredicate>
+_RandomAccessIterator2
+__pattern_copy_if(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator1 __first,
+                  _RandomAccessIterator1 __last, _RandomAccessIterator2 __result, _UnaryPredicate __pred)
+{
+    typedef typename ::std::iterator_traits<_RandomAccessIterator1>::difference_type _DifferenceType;
+    const _DifferenceType __n = __last - __first;
+    if (_DifferenceType(1) < __n)
+    {
+        __par_backend::__buffer<_ExecutionPolicy, bool> __mask_buf(__n);
+        return __internal::__except_handler([&__exec, __n, __first, __result, __pred, &__mask_buf]() {
+            bool* __mask = __mask_buf.get();
+            _DifferenceType __m{};
+            __par_backend::__parallel_strict_scan(
+                ::std::forward<_ExecutionPolicy>(__exec), __n, _DifferenceType(0),
+                [=](_DifferenceType __i, _DifferenceType __len) { // Reduce
+                    return __internal::__brick_calc_mask_1<_DifferenceType>(__first + __i, __first + (__i + __len),
+                                                                            __mask + __i, __pred, _IsVector{})
+                        .first;
+                },
+                ::std::plus<_DifferenceType>(),                                              // Combine
+                [=](_DifferenceType __i, _DifferenceType __len, _DifferenceType __initial) { // Scan
+                    __internal::__brick_copy_by_mask(
+                        __first + __i, __first + (__i + __len), __result + __initial, __mask + __i,
+                        [](_RandomAccessIterator1 __x, _RandomAccessIterator2 __z) { *__z = *__x; }, _IsVector{});
+                },
+                [&__m](_DifferenceType __total) { __m = __total; });
+            return __result + __m;
+        });
+    }
+    // trivial sequence - use serial algorithm
+    return __internal::__brick_copy_if(__first, __last, __result, __pred, _IsVector{});
+}
+
 //------------------------------------------------------------------------
 // count
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 665d2e472e3..d1cfa28d933 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -1138,6 +1138,31 @@ __pattern_copy_if(_ExecutionPolicy&& __exec, _Iterator1 __first, _Iterator1 __la
     return __result_first + __num_copied;
 }
 
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2,
+          typename _Predicate>
+_Iterator2
+__pattern_copy_if(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator1 __first, _Iterator1 __last,
+                  _Iterator2 __result_first, _Predicate __pred)
+{
+    using _It1DifferenceType = typename ::std::iterator_traits<_Iterator1>::difference_type;
+
+    if (__first == __last)
+        return __result_first;
+
+    _It1DifferenceType __n = __last - __first;
+
+    auto __keep1 = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _Iterator1>();
+    auto __buf1 = __keep1(__first, __last);
+    auto __keep2 = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::write, _Iterator2>();
+    auto __buf2 = __keep2(__result_first, __result_first + __n);
+
+    auto __res = __par_backend_hetero::__parallel_copy_if(::std::forward<_ExecutionPolicy>(__exec), __buf1.all_view(),
+                                                          __buf2.all_view(), __n, __pred);
+
+    ::std::size_t __num_copied = __res.get();
+    return __result_first + __num_copied;
+}
+
 //------------------------------------------------------------------------
 // partition_copy
 //------------------------------------------------------------------------

From 46ba9c41ff49e937186c00aa7a273bc026fc2b8d Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Mon, 5 Feb 2024 12:01:42 +0100
Subject: [PATCH 068/566] __pattern_copy_if + tag calls

---
 include/oneapi/dpl/pstl/glue_algorithm_impl.h          | 9 +++++----
 include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h | 6 ++++--
 2 files changed, 9 insertions(+), 6 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_algorithm_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
index 128d0c1b35b..c1c507d029a 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
@@ -292,10 +292,11 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _Forward
 copy_if(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _ForwardIterator1 __last, _ForwardIterator2 __result,
         _Predicate __pred)
 {
-    return oneapi::dpl::__internal::__pattern_copy_if(
-        ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result, __pred,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>());
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
+
+    return oneapi::dpl::__internal::__pattern_copy_if(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first,
+                                                      __last, __result, __pred);
 }
 
 // [alg.swap]
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index d1cfa28d933..6c1202a5c11 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -1234,10 +1234,12 @@ __pattern_remove_if(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __la
 
     using _ValueType = typename ::std::iterator_traits<_Iterator>::value_type;
 
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator>();
+
     oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __buf(__exec, __last - __first);
     auto __copy_first = __buf.get();
-    auto __copy_last = __pattern_copy_if(__exec, __first, __last, __copy_first, __not_pred<_Predicate>{__pred},
-                                         /*vector=*/::std::true_type{}, /*parallel*/ ::std::true_type{});
+    auto __copy_last =
+        __pattern_copy_if(__dispatch_tag, __exec, __first, __last, __copy_first, __not_pred<_Predicate>{__pred});
 
     //TODO: optimize copy back depending on Iterator, i.e. set_final_data for host iterator/pointer
     return __pattern_walk2(

From cc9148d38965f7b79faa99dd198c394e01821009 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Mon, 5 Feb 2024 16:11:08 +0100
Subject: [PATCH 069/566] 
 include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h - fix compiler error:
 implement previously absent impl of __pattern_fill_n for hetero policy

---
 .../oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h | 14 ++++++++++++++
 1 file changed, 14 insertions(+)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 6c1202a5c11..2a1c0a0ca4e 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -511,6 +511,20 @@ __pattern_fill(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Forw
     return __last;
 }
 
+template <typename _BackendTag, typename _ExecutionPolicy, typename _ForwardIterator, class _Size, typename _T>
+_ForwardIterator
+__pattern_fill_n(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _ForwardIterator __first, _Size __count,
+                 const _T& __value)
+{
+    // TODO: is this new implementation are correct?
+    // Previously we hadn't hetero impl for __pattern_fill_n
+
+    return __pattern_walk1_n(__tag, ::std::forward<_ExecutionPolicy>(__exec),
+                             __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::write>(__first),
+                             __count, fill_functor<_T>{__value});
+}
+
+
 //------------------------------------------------------------------------
 // generate
 //------------------------------------------------------------------------

From 4553fff012e9391404f0469b3e901fa6a4995269 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Mon, 5 Feb 2024 16:33:55 +0100
Subject: [PATCH 070/566] __pattern_adjacent_find + tag impls

---
 include/oneapi/dpl/pstl/algorithm_fwd.h       | 10 ++++
 include/oneapi/dpl/pstl/algorithm_impl.h      | 54 ++++++++++++++++++
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 57 +++++++++++++++++++
 3 files changed, 121 insertions(+)

diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index e81b38d31e1..b7202b5cf28 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -1030,11 +1030,21 @@ oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _Fo
 __pattern_adjacent_find(_ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _BinaryPredicate,
                         /* is_parallel */ ::std::false_type, _IsVector, _Semantic) noexcept;
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _BinaryPredicate, class _Semantic>
+_ForwardIterator
+__pattern_adjacent_find(_Tag, _ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _BinaryPredicate,
+                        _Semantic) noexcept;
+
 template <class _ExecutionPolicy, class _RandomAccessIterator, class _BinaryPredicate, class _IsVector, class _Semantic>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator>
 __pattern_adjacent_find(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _BinaryPredicate,
                         /* is_parallel */ ::std::true_type, _IsVector, _Semantic);
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _BinaryPredicate, class _Semantic>
+_RandomAccessIterator
+__pattern_adjacent_find(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator,
+                        _BinaryPredicate, _Semantic);
+
 //------------------------------------------------------------------------
 // nth_element
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 3870be47fdd..e6b68c55df4 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -2997,6 +2997,16 @@ __pattern_adjacent_find(_ExecutionPolicy&&, _ForwardIterator __first, _ForwardIt
     return __internal::__brick_adjacent_find(__first, __last, __pred, __is_vector, _Semantic::value);
 }
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _BinaryPredicate, class _Semantic>
+_ForwardIterator
+__pattern_adjacent_find(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
+                        _BinaryPredicate __pred, _Semantic) noexcept
+{
+    static_assert(__is_backend_tag_v<_Tag>);
+
+    return __internal::__brick_adjacent_find(__first, __last, __pred, typename _Tag::__is_vector{}, _Semantic::value);
+}
+
 template <class _ExecutionPolicy, class _RandomAccessIterator, class _BinaryPredicate, class _IsVector, class _Semantic>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator>
 __pattern_adjacent_find(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIterator __last,
@@ -3042,6 +3052,50 @@ __pattern_adjacent_find(_ExecutionPolicy&& __exec, _RandomAccessIterator __first
     });
 }
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _BinaryPredicate, class _Semantic>
+_RandomAccessIterator
+__pattern_adjacent_find(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
+                        _RandomAccessIterator __last, _BinaryPredicate __pred, _Semantic __or_semantic)
+{
+    if (__last - __first < 2)
+        return __last;
+
+    return __internal::__except_handler([&]() {
+        return __par_backend::__parallel_reduce(
+            ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __last,
+            [__last, __pred, __or_semantic](_RandomAccessIterator __begin, _RandomAccessIterator __end,
+                                            _RandomAccessIterator __value) -> _RandomAccessIterator {
+                // TODO: investigate performance benefits from the use of shared variable for the result,
+                // checking (compare_and_swap idiom) its __value at __first.
+                if (__or_semantic && __value < __last)
+                { //found
+                    return __value;
+                }
+
+                if (__value > __begin)
+                {
+                    // modify __end to check the predicate on the boundary __values;
+                    // TODO: to use a custom range with boundaries overlapping
+                    // TODO: investigate what if we remove "if" below and run algorithm on range [__first, __last-1)
+                    // then check the pair [__last-1, __last)
+                    if (__end != __last)
+                        ++__end;
+
+                    //correct the global result iterator if the "brick" returns a local "__last"
+                    const _RandomAccessIterator __res =
+                        __internal::__brick_adjacent_find(__begin, __end, __pred, _IsVector{}, __or_semantic);
+                    if (__res < __end)
+                        __value = __res;
+                }
+                return __value;
+            },
+            [](_RandomAccessIterator __x, _RandomAccessIterator __y) -> _RandomAccessIterator {
+                return __x < __y ? __x : __y;
+            } //reduce a __value
+        );
+    });
+}
+
 //------------------------------------------------------------------------
 // nth_element
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 2a1c0a0ca4e..8fc80e800e7 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -782,6 +782,34 @@ __pattern_adjacent_find(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator
     return result ? __first : __last;
 }
 
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator, typename _BinaryPredicate>
+_Iterator
+__pattern_adjacent_find(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last,
+                        _BinaryPredicate __predicate, oneapi::dpl::__internal::__or_semantic)
+{
+    if (__last - __first < 2)
+        return __last;
+
+    using _Predicate =
+        oneapi::dpl::unseq_backend::single_match_pred<_ExecutionPolicy, adjacent_find_fn<_BinaryPredicate>>;
+
+    auto __keep1 = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _Iterator>();
+    auto __buf1 = __keep1(__first, __last - 1);
+    auto __keep2 = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _Iterator>();
+    auto __buf2 = __keep2(__first + 1, __last);
+
+    // TODO: in case of confilicting names
+    // __par_backend_hetero::make_wrapped_policy<__par_backend_hetero::__or_policy_wrapper>()
+    bool result = __par_backend_hetero::__parallel_find_or(
+        ::std::forward<_ExecutionPolicy>(__exec), _Predicate{adjacent_find_fn<_BinaryPredicate>{__predicate}},
+        __par_backend_hetero::__parallel_or_tag{},
+        oneapi::dpl::__ranges::make_zip_view(__buf1.all_view(), __buf2.all_view()));
+
+    // inverted conditional because of
+    // reorder_predicate in glue_algorithm_impl.h
+    return result ? __first : __last;
+}
+
 template <typename _ExecutionPolicy, typename _Iterator, typename _BinaryPredicate>
 oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, _Iterator>
 __pattern_adjacent_find(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last, _BinaryPredicate __predicate,
@@ -811,6 +839,35 @@ __pattern_adjacent_find(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator
     return (__result_iterator == __last - 1) ? __last : __result_iterator;
 }
 
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator, typename _BinaryPredicate>
+_Iterator
+__pattern_adjacent_find(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last,
+                        _BinaryPredicate __predicate,
+                        oneapi::dpl::__internal::__first_semantic)
+{
+    if (__last - __first < 2)
+        return __last;
+
+    using _Predicate =
+        oneapi::dpl::unseq_backend::single_match_pred<_ExecutionPolicy, adjacent_find_fn<_BinaryPredicate>>;
+
+    auto __result = __par_backend_hetero::__parallel_find(
+        ::std::forward<_ExecutionPolicy>(__exec),
+        __par_backend_hetero::zip(
+            __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__first),
+            __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__first + 1)),
+        __par_backend_hetero::zip(
+            __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__last - 1),
+            __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__last)),
+        _Predicate{adjacent_find_fn<_BinaryPredicate>{__predicate}}, ::std::true_type{});
+
+    auto __zip_at_first = __par_backend_hetero::zip(
+        __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__first),
+        __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__first + 1));
+    _Iterator __result_iterator = __first + (__result - __zip_at_first);
+    return (__result_iterator == __last - 1) ? __last : __result_iterator;
+}
+
 //------------------------------------------------------------------------
 // count, count_if
 //------------------------------------------------------------------------

From 749d90176f4743bcdaf510222acb51d6c77aefd9 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Mon, 5 Feb 2024 16:35:07 +0100
Subject: [PATCH 071/566] __pattern_adjacent_find + tag calls

---
 include/oneapi/dpl/pstl/glue_algorithm_impl.h | 37 ++++++++++---------
 1 file changed, 19 insertions(+), 18 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_algorithm_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
index c1c507d029a..070c30b36b0 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
@@ -171,21 +171,23 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _Forward
 adjacent_find(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last)
 {
     typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _ValueType;
-    return oneapi::dpl::__internal::__pattern_adjacent_find(
-        ::std::forward<_ExecutionPolicy>(__exec), __first, __last, ::std::equal_to<_ValueType>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(),
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(),
-        oneapi::dpl::__internal::__first_semantic());
+
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+
+    return oneapi::dpl::__internal::__pattern_adjacent_find(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
+                                                            __first, __last, ::std::equal_to<_ValueType>(),
+                                                            oneapi::dpl::__internal::__first_semantic());
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator, class _BinaryPredicate>
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _ForwardIterator>
 adjacent_find(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last, _BinaryPredicate __pred)
 {
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+
     return oneapi::dpl::__internal::__pattern_adjacent_find(
+        __dispatch_tag,
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __pred,
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(),
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(),
         oneapi::dpl::__internal::__first_semantic());
 }
 
@@ -871,12 +873,11 @@ template <class _ExecutionPolicy, class _ForwardIterator, class _Compare>
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _ForwardIterator>
 is_sorted_until(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last, _Compare __comp)
 {
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+
     const _ForwardIterator __res = oneapi::dpl::__internal::__pattern_adjacent_find(
-        ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
-        oneapi::dpl::__internal::__reorder_pred<_Compare>(__comp),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(),
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(),
-        oneapi::dpl::__internal::__first_semantic());
+        __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+        oneapi::dpl::__internal::__reorder_pred<_Compare>(__comp), oneapi::dpl::__internal::__first_semantic());
     return __res == __last ? __last : oneapi::dpl::__internal::__pstl_next(__res);
 }
 
@@ -892,12 +893,12 @@ template <class _ExecutionPolicy, class _ForwardIterator, class _Compare>
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, bool>
 is_sorted(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last, _Compare __comp)
 {
-    return oneapi::dpl::__internal::__pattern_adjacent_find(
-               ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
-               oneapi::dpl::__internal::__reorder_pred<_Compare>(__comp),
-               oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(),
-               oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(),
-               oneapi::dpl::__internal::__or_semantic()) == __last;
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+
+    return oneapi::dpl::__internal::__pattern_adjacent_find(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
+                                                            __first, __last,
+                                                            oneapi::dpl::__internal::__reorder_pred<_Compare>(__comp),
+                                                            oneapi::dpl::__internal::__or_semantic()) == __last;
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator>

From 27e59d50ece9a69bf527a6c264d86f9024d3a5c4 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Mon, 5 Feb 2024 16:48:32 +0100
Subject: [PATCH 072/566] __pattern_any_of + tag impls

---
 include/oneapi/dpl/pstl/algorithm_fwd.h       |  8 +++++++
 include/oneapi/dpl/pstl/algorithm_impl.h      | 22 +++++++++++++++++++
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 19 ++++++++++++++++
 3 files changed, 49 insertions(+)

diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index b7202b5cf28..f39eee7ef32 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -50,11 +50,19 @@ oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, boo
 __pattern_any_of(_ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _Pred, _IsVector,
                  /*parallel=*/::std::false_type) noexcept;
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Pred>
+bool
+__pattern_any_of(_Tag, _ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _Pred) noexcept;
+
 template <class _ExecutionPolicy, class _RandomAccessIterator, class _Pred, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, bool>
 __pattern_any_of(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _Pred, _IsVector,
                  /*parallel=*/::std::true_type);
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Pred>
+bool
+__pattern_any_of(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _Pred);
+
 //------------------------------------------------------------------------
 // walk1 (pseudo)
 //
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index e6b68c55df4..61d0995a4bb 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -71,6 +71,15 @@ __pattern_any_of(_ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator
     return __internal::__brick_any_of(__first, __last, __pred, __is_vector);
 }
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Pred>
+bool
+__pattern_any_of(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _Pred __pred) noexcept
+{
+    static_assert(__is_backend_tag_v<_Tag>);
+
+    return __internal::__brick_any_of(__first, __last, __pred, typename _Tag::__is_vector{});
+}
+
 template <class _ExecutionPolicy, class _RandomAccessIterator, class _Pred, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, bool>
 __pattern_any_of(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIterator __last, _Pred __pred,
@@ -84,6 +93,19 @@ __pattern_any_of(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _Rand
     });
 }
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Pred>
+bool
+__pattern_any_of(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
+                 _RandomAccessIterator __last, _Pred __pred)
+{
+    return __internal::__except_handler([&]() {
+        return __internal::__parallel_or(::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+                                         [__pred](_RandomAccessIterator __i, _RandomAccessIterator __j) {
+                                             return __internal::__brick_any_of(__i, __j, __pred, _IsVector{});
+                                         });
+    });
+}
+
 // [alg.foreach]
 // for_each_n with no policy
 
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 8fc80e800e7..f1288fdb078 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -924,6 +924,25 @@ __pattern_any_of(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last,
         _Predicate{__pred}, __par_backend_hetero::__parallel_or_tag{}, __buf.all_view());
 }
 
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator, typename _Pred>
+bool
+__pattern_any_of(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last,
+                 _Pred __pred)
+{
+    if (__first == __last)
+        return false;
+
+    using _Predicate = oneapi::dpl::unseq_backend::single_match_pred<_ExecutionPolicy, _Pred>;
+
+    auto __keep = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _Iterator>();
+    auto __buf = __keep(__first, __last);
+
+    return oneapi::dpl::__par_backend_hetero::__parallel_find_or(
+        __par_backend_hetero::make_wrapped_policy<__par_backend_hetero::__or_policy_wrapper>(
+            ::std::forward<_ExecutionPolicy>(__exec)),
+        _Predicate{__pred}, __par_backend_hetero::__parallel_or_tag{}, __buf.all_view());
+}
+
 //------------------------------------------------------------------------
 // equal
 //------------------------------------------------------------------------

From 3fd9c6577b9f14911222caa4040ba43c6a1dfc48 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Mon, 5 Feb 2024 16:49:24 +0100
Subject: [PATCH 073/566] __pattern_any_of + tag calls

---
 include/oneapi/dpl/pstl/glue_algorithm_impl.h      |  8 ++++----
 .../oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h | 14 +++++++++-----
 2 files changed, 13 insertions(+), 9 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_algorithm_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
index 070c30b36b0..6925816b17b 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
@@ -43,10 +43,10 @@ template <class _ExecutionPolicy, class _ForwardIterator, class _Predicate>
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, bool>
 any_of(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last, _Predicate __pred)
 {
-    return oneapi::dpl::__internal::__pattern_any_of(
-        ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __pred,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>());
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+
+    return oneapi::dpl::__internal::__pattern_any_of(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first,
+                                                     __last, __pred);
 }
 
 // [alg.all_of]
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index f1288fdb078..c2353b9ceb1 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -1132,9 +1132,10 @@ __pattern_search_n(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __las
 
     if (__last - __first == __count)
     {
-        return (!__internal::__pattern_any_of(::std::forward<_ExecutionPolicy>(__exec), __first, __last,
-                                              __search_n_unary_predicate<_Tp, _BinaryPredicate>{__value, __pred},
-                                              ::std::true_type{}, ::std::true_type{}))
+        constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator>();
+
+        return (!__internal::__pattern_any_of(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+                                              __search_n_unary_predicate<_Tp, _BinaryPredicate>{__value, __pred}))
                    ? __first
                    : __last;
     }
@@ -1606,10 +1607,13 @@ __pattern_stable_partition(_ExecutionPolicy&& __exec, _Iterator __first, _Iterat
     if (__last == __first)
         return __last;
     else if (__last - __first < 2)
-        return __pattern_any_of(::std::forward<_ExecutionPolicy>(__exec), __first, __last, __pred, ::std::true_type(),
-                                ::std::true_type())
+    {
+        constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator>();
+
+        return __pattern_any_of(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __pred)
                    ? __last
                    : __first;
+    }
 
     using _ValueType = typename ::std::iterator_traits<_Iterator>::value_type;
 

From 74735909dfde368d860dfd693a077be1e364e6bf Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Mon, 5 Feb 2024 16:56:20 +0100
Subject: [PATCH 074/566] __pattern_count + tag impls

---
 include/oneapi/dpl/pstl/algorithm_fwd.h       |  9 ++++++
 include/oneapi/dpl/pstl/algorithm_impl.h      | 30 +++++++++++++++++++
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 28 +++++++++++++++++
 3 files changed, 67 insertions(+)

diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index f39eee7ef32..35882ecc71d 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -677,12 +677,21 @@ oneapi::dpl::__internal::__enable_if_host_execution_policy<
 __pattern_count(_ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _Predicate,
                 /* is_parallel */ ::std::false_type, _IsVector) noexcept;
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Predicate>
+typename ::std::iterator_traits<_ForwardIterator>::difference_type
+__pattern_count(_Tag, _ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _Predicate) noexcept;
+
 template <class _ExecutionPolicy, class _RandomAccessIterator, class _Predicate, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<
     _ExecutionPolicy, typename ::std::iterator_traits<_RandomAccessIterator>::difference_type>
 __pattern_count(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _Predicate,
                 /* is_parallel */ ::std::true_type, _IsVector);
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Predicate>
+typename ::std::iterator_traits<_RandomAccessIterator>::difference_type
+__pattern_count(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator,
+                _Predicate);
+
 //------------------------------------------------------------------------
 // unique
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 61d0995a4bb..b2d66091175 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -1798,6 +1798,15 @@ __pattern_count(_ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator _
     return __internal::__brick_count(__first, __last, __pred, __is_vector);
 }
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Predicate>
+typename ::std::iterator_traits<_ForwardIterator>::difference_type
+__pattern_count(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _Predicate __pred) noexcept
+{
+    static_assert(__is_backend_tag_v<_Tag>);
+
+    return __internal::__brick_count(__first, __last, __pred, typename _Tag::__is_vector{});
+}
+
 template <class _ExecutionPolicy, class _RandomAccessIterator, class _Predicate, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<
     _ExecutionPolicy, typename ::std::iterator_traits<_RandomAccessIterator>::difference_type>
@@ -1820,6 +1829,27 @@ __pattern_count(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _Rando
     });
 }
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Predicate>
+typename ::std::iterator_traits<_RandomAccessIterator>::difference_type
+__pattern_count(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
+                _RandomAccessIterator __last, _Predicate __pred)
+{
+    typedef typename ::std::iterator_traits<_RandomAccessIterator>::difference_type _SizeType;
+
+    //trivial pre-checks
+    if (__first == __last)
+        return _SizeType(0);
+
+    return __internal::__except_handler([&]() {
+        return __par_backend::__parallel_reduce(
+            ::std::forward<_ExecutionPolicy>(__exec), __first, __last, _SizeType(0),
+            [__pred](_RandomAccessIterator __begin, _RandomAccessIterator __end, _SizeType __value) -> _SizeType {
+                return __value + __internal::__brick_count(__begin, __end, __pred, _IsVector{});
+            },
+            ::std::plus<_SizeType>());
+    });
+}
+
 //------------------------------------------------------------------------
 // unique
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index c2353b9ceb1..a414fffbb35 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -901,6 +901,34 @@ __pattern_count(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last,
         .get();
 }
 
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator, typename _Predicate>
+typename ::std::iterator_traits<_Iterator>::difference_type
+__pattern_count(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last,
+                _Predicate __predicate)
+{
+    if (__first == __last)
+        return 0;
+
+    using _ReduceValueType = typename ::std::iterator_traits<_Iterator>::difference_type;
+
+    auto __reduce_fn = ::std::plus<_ReduceValueType>{};
+    // int is being implicitly casted to difference_type
+    // otherwise we can only pass the difference_type as a functor template parameter
+    auto __transform_fn = [__predicate](auto __gidx, auto __acc) -> int {
+        return (__predicate(__acc[__gidx]) ? 1 : 0);
+    };
+
+    auto __keep = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _Iterator>();
+    auto __buf = __keep(__first, __last);
+
+    return oneapi::dpl::__par_backend_hetero::__parallel_transform_reduce<_ReduceValueType,
+                                                                          ::std::true_type /*is_commutative*/>(
+               ::std::forward<_ExecutionPolicy>(__exec), __reduce_fn, __transform_fn,
+               unseq_backend::__no_init_value{}, // no initial value
+               __buf.all_view())
+        .get();
+}
+
 //------------------------------------------------------------------------
 // any_of
 //------------------------------------------------------------------------

From 9807a462ca805acf16e0859d03d3e0b47f8ad1b8 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Mon, 5 Feb 2024 16:56:42 +0100
Subject: [PATCH 075/566] __pattern_count + tag calls

---
 include/oneapi/dpl/pstl/glue_algorithm_impl.h | 16 ++++++++--------
 1 file changed, 8 insertions(+), 8 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_algorithm_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
index 6925816b17b..7a3582ac073 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
@@ -201,12 +201,12 @@ oneapi::dpl::__internal::__enable_if_execution_policy<
     _ExecutionPolicy, typename ::std::iterator_traits<_ForwardIterator>::difference_type>
 count(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last, const _Tp& __value)
 {
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+
     return oneapi::dpl::__internal::__pattern_count(
-        ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+        __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
         oneapi::dpl::__internal::__equal_value<oneapi::dpl::__internal::__ref_or_copy<_ExecutionPolicy, const _Tp>>(
-            __value),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(),
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>());
+            __value));
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator, class _Predicate>
@@ -214,10 +214,10 @@ oneapi::dpl::__internal::__enable_if_execution_policy<
     _ExecutionPolicy, typename ::std::iterator_traits<_ForwardIterator>::difference_type>
 count_if(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last, _Predicate __pred)
 {
-    return oneapi::dpl::__internal::__pattern_count(
-        ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __pred,
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(),
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>());
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+
+    return oneapi::dpl::__internal::__pattern_count(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first,
+                                                    __last, __pred);
 }
 
 // [alg.search]

From 5ae36d448a48cbf2eb4bc0878c519e4476593550 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Mon, 5 Feb 2024 17:18:05 +0100
Subject: [PATCH 076/566] __pattern_equal + tag impls

---
 include/oneapi/dpl/pstl/algorithm_fwd.h       | 22 +++++++
 include/oneapi/dpl/pstl/algorithm_impl.h      | 57 +++++++++++++++++++
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 33 +++++++++++
 3 files changed, 112 insertions(+)

diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index 35882ecc71d..73bb1bc2f79 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -412,12 +412,23 @@ oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, boo
 __pattern_equal(_ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2, _BinaryPredicate,
                 _IsVector, /* is_parallel = */ ::std::false_type) noexcept;
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _BinaryPredicate>
+bool
+__pattern_equal(_Tag, _ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2, _ForwardIterator2,
+                _BinaryPredicate) noexcept;
+
 template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _BinaryPredicate,
           class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, bool>
 __pattern_equal(_ExecutionPolicy&&, _RandomAccessIterator1, _RandomAccessIterator1, _RandomAccessIterator2,
                 _BinaryPredicate, _IsVector, /* is_parallel = */ ::std::true_type);
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
+          class _BinaryPredicate>
+bool
+__pattern_equal(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator1, _RandomAccessIterator1,
+                _RandomAccessIterator2, _RandomAccessIterator2, _BinaryPredicate);
+
 template <class _ForwardIterator1, class _ForwardIterator2, class _BinaryPredicate>
 bool __brick_equal(_ForwardIterator1, _ForwardIterator1, _ForwardIterator2, _ForwardIterator2, _BinaryPredicate,
                    /* is_vector = */ ::std::false_type) noexcept;
@@ -432,12 +443,23 @@ oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, boo
 __pattern_equal(_ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2, _ForwardIterator2,
                 _BinaryPredicate, _IsVector, /* is_parallel = */ ::std::false_type) noexcept;
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _BinaryPredicate>
+bool
+__pattern_equal(_Tag, _ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2,
+                _BinaryPredicate) noexcept;
+
 template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _BinaryPredicate,
           class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, bool>
 __pattern_equal(_ExecutionPolicy&&, _RandomAccessIterator1, _RandomAccessIterator1, _RandomAccessIterator2,
                 _RandomAccessIterator2, _BinaryPredicate, _IsVector, /* is_parallel = */ ::std::true_type);
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
+          class _BinaryPredicate>
+bool
+__pattern_equal(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator1, _RandomAccessIterator1,
+                _RandomAccessIterator2, _BinaryPredicate);
+
 //------------------------------------------------------------------------
 // find_if
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index b2d66091175..1d9003e41bb 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -982,6 +982,17 @@ __pattern_equal(_ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator
     return __internal::__brick_equal(__first1, __last1, __first2, __last2, __p, __is_vector);
 }
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _BinaryPredicate>
+bool
+__pattern_equal(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
+                _ForwardIterator2 __first2,
+                _ForwardIterator2 __last2, _BinaryPredicate __p) noexcept
+{
+    static_assert(__is_backend_tag_v<_Tag>);
+
+    return __internal::__brick_equal(__first1, __last1, __first2, __last2, __p, typename _Tag::__is_vector{});
+}
+
 template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _BinaryPredicate,
           class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, bool>
@@ -1002,6 +1013,27 @@ __pattern_equal(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1, _Ran
     });
 }
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
+          class _BinaryPredicate>
+bool
+__pattern_equal(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1,
+                _RandomAccessIterator1 __last1, _RandomAccessIterator2 __first2, _RandomAccessIterator2 __last2,
+                _BinaryPredicate __p)
+{
+    if (__last1 - __first1 != __last2 - __first2)
+        return false;
+
+    return __internal::__except_handler([&]() {
+        return !__internal::__parallel_or(
+            ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1,
+            [__first1, __first2, __p](_RandomAccessIterator1 __i, _RandomAccessIterator1 __j) {
+                return !__internal::__brick_equal(__i, __j, __first2 + (__i - __first1), __first2 + (__j - __first1),
+                                                  __p, _IsVector{});
+            });
+    });
+}
+
+
 //------------------------------------------------------------------------
 // equal version for sequences with equal length
 //------------------------------------------------------------------------
@@ -1032,6 +1064,16 @@ __pattern_equal(_ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator
     return __internal::__brick_equal(__first1, __last1, __first2, __p, __is_vector);
 }
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _BinaryPredicate>
+bool
+__pattern_equal(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
+                _ForwardIterator2 __first2, _BinaryPredicate __p) noexcept
+{
+    static_assert(__is_backend_tag_v<_Tag>);
+
+    return __internal::__brick_equal(__first1, __last1, __first2, __p, typename _Tag::__is_vector{});
+}
+
 template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _BinaryPredicate,
           class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, bool>
@@ -1048,6 +1090,21 @@ __pattern_equal(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1, _Ran
     });
 }
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
+          class _BinaryPredicate>
+bool
+__pattern_equal(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1,
+                _RandomAccessIterator1 __last1, _RandomAccessIterator2 __first2, _BinaryPredicate __p)
+{
+    return __internal::__except_handler([&]() {
+        return !__internal::__parallel_or(
+            ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1,
+            [__first1, __first2, __p](_RandomAccessIterator1 __i, _RandomAccessIterator1 __j) {
+                return !__internal::__brick_equal(__i, __j, __first2 + (__i - __first1), __p, _IsVector{});
+            });
+    });
+}
+
 //------------------------------------------------------------------------
 // find_if
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index a414fffbb35..2f9ded980ae 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -999,6 +999,30 @@ __pattern_equal(_ExecutionPolicy&& __exec, _Iterator1 __first1, _Iterator1 __las
         oneapi::dpl::__ranges::make_zip_view(__buf1.all_view(), __buf2.all_view()));
 }
 
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2, typename _Pred>
+bool
+__pattern_equal(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator1 __first1, _Iterator1 __last1,
+                _Iterator2 __first2, _Iterator2 __last2, _Pred __pred)
+{
+    if (__last1 == __first1 || __last2 == __first2 || __last1 - __first1 != __last2 - __first2)
+        return false;
+
+    using _Predicate = oneapi::dpl::unseq_backend::single_match_pred<_ExecutionPolicy, equal_predicate<_Pred>>;
+
+    auto __keep1 = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _Iterator1>();
+    auto __buf1 = __keep1(__first1, __last1);
+    auto __keep2 = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _Iterator2>();
+    auto __buf2 = __keep2(__first2, __last2);
+
+    // TODO: in case of confilicting names
+    // __par_backend_hetero::make_wrapped_policy<__par_backend_hetero::__or_policy_wrapper>()
+    return !__par_backend_hetero::__parallel_find_or(
+        ::std::forward<_ExecutionPolicy>(__exec), _Predicate{equal_predicate<_Pred>{__pred}},
+        __par_backend_hetero::__parallel_or_tag{},
+        oneapi::dpl::__ranges::make_zip_view(__buf1.all_view(), __buf2.all_view()));
+}
+
+
 //------------------------------------------------------------------------
 // equal version for sequences with equal length
 //------------------------------------------------------------------------
@@ -1013,6 +1037,15 @@ __pattern_equal(_ExecutionPolicy&& __exec, _Iterator1 __first1, _Iterator1 __las
                                                     /*vector=*/::std::true_type{}, /*parallel=*/::std::true_type{});
 }
 
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2, typename _Pred>
+bool
+__pattern_equal(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator1 __first1, _Iterator1 __last1,
+                _Iterator2 __first2, _Pred __pred)
+{
+    return oneapi::dpl::__internal::__pattern_equal(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1,
+                                                    __first2, __first2 + (__last1 - __first1), __pred);
+}
+
 //------------------------------------------------------------------------
 // find_if
 //------------------------------------------------------------------------

From 0508fc612612ce007b9b83d7aabba8c7bc8854c6 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Mon, 5 Feb 2024 17:24:16 +0100
Subject: [PATCH 077/566] __pattern_equal + tag calls

---
 include/oneapi/dpl/pstl/algorithm_impl.h      | 18 ++++++++++----
 include/oneapi/dpl/pstl/glue_algorithm_impl.h | 20 +++++++++-------
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 24 ++++++++++++++-----
 3 files changed, 44 insertions(+), 18 deletions(-)

diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 1d9003e41bb..ad7cfd9b399 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -1291,8 +1291,13 @@ __pattern_find_end(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first, _R
 {
     if (__last - __first == __s_last - __s_first)
     {
-        const bool __res = __internal::__pattern_equal(::std::forward<_ExecutionPolicy>(__exec), __first, __last,
-                                                       __s_first, __pred, __is_vector, ::std::true_type());
+        // TODO is it correct that we check _RandomAccessIterator2 in __select_backend ?
+        constexpr auto __dispatch_tag =
+            oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1,
+                                                      _RandomAccessIterator2>();
+
+        const bool __res = __internal::__pattern_equal(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
+                                                       __first, __last, __s_first, __pred);
         return __res ? __first : __last;
     }
     else
@@ -1395,8 +1400,13 @@ __pattern_search(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first, _Ran
 {
     if (__last - __first == __s_last - __s_first)
     {
-        const bool __res = __internal::__pattern_equal(::std::forward<_ExecutionPolicy>(__exec), __first, __last,
-                                                       __s_first, __pred, __is_vector, ::std::true_type());
+        // TODO is it correct that we check _RandomAccessIterator2 in __select_backend ?
+        constexpr auto __dispatch_tag =
+            oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1,
+                                                      _RandomAccessIterator2>();
+
+        const bool __res = __internal::__pattern_equal(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
+                                                       __first, __last, __s_first, __pred);
         return __res ? __first : __last;
     }
     else
diff --git a/include/oneapi/dpl/pstl/glue_algorithm_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
index 7a3582ac073..5af0b880aee 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
@@ -776,10 +776,12 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, bool>
 equal(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 __last1, _ForwardIterator2 __first2,
       _BinaryPredicate __p)
 {
-    return oneapi::dpl::__internal::__pattern_equal(
-        ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __p,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1>());
+   // TODO is it correct that we check _ForwardIterator2 in __select_backend ?
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
+
+    return oneapi::dpl::__internal::__pattern_equal(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first1,
+                                                    __last1, __first2, __p);
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2>
@@ -795,10 +797,12 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, bool>
 equal(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 __last1, _ForwardIterator2 __first2,
       _ForwardIterator2 __last2, _BinaryPredicate __p)
 {
-    return oneapi::dpl::__internal::__pattern_equal(
-        ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2, __p,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1>());
+    // TODO is it correct that we check _ForwardIterator2 in __select_backend ?
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
+
+    return oneapi::dpl::__internal::__pattern_equal(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first1,
+                                                    __last1, __first2, __last2, __p);
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2>
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 2f9ded980ae..3fee7daff48 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -1032,9 +1032,12 @@ oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, b
 __pattern_equal(_ExecutionPolicy&& __exec, _Iterator1 __first1, _Iterator1 __last1, _Iterator2 __first2, _Pred __pred,
                 /*vector=*/::std::true_type, /*parallel=*/::std::true_type)
 {
-    return oneapi::dpl::__internal::__pattern_equal(::std::forward<_ExecutionPolicy>(__exec), __first1, __last1,
-                                                    __first2, __first2 + (__last1 - __first1), __pred,
-                                                    /*vector=*/::std::true_type{}, /*parallel=*/::std::true_type{});
+    // TODO is it correct that we check _Iterator2 in __select_backend ?
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator1, _Iterator2>();
+
+    return oneapi::dpl::__internal::__pattern_equal(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first1,
+                                                    __last1, __first2, __first2 + (__last1 - __first1), __pred);
 }
 
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2, typename _Pred>
@@ -1081,8 +1084,12 @@ __pattern_find_end(_ExecutionPolicy&& __exec, _Iterator1 __first, _Iterator1 __l
 
     if (__last - __first == __s_last - __s_first)
     {
-        const bool __res = __pattern_equal(::std::forward<_ExecutionPolicy>(__exec), __first, __last, __s_first, __pred,
-                                           ::std::true_type(), ::std::true_type());
+        // TODO is it correct that we check _Iterator2 in __select_backend ?
+        constexpr auto __dispatch_tag =
+            oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator1, _Iterator2>();
+
+        const bool __res = __pattern_equal(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+                                           __s_first, __pred);
         return __res ? __first : __last;
     }
     else
@@ -1146,9 +1153,14 @@ __pattern_search(_ExecutionPolicy&& __exec, _Iterator1 __first, _Iterator1 __las
 
     if (__last - __first == __s_last - __s_first)
     {
+        // TODO is it correct that we check _Iterator2 in __select_backend ?
+        constexpr auto __dispatch_tag =
+            oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator1, _Iterator2>();
+
         const bool __res = __pattern_equal(
+            __dispatch_tag,
             __par_backend_hetero::make_wrapped_policy<equal_wrapper>(::std::forward<_ExecutionPolicy>(__exec)), __first,
-            __last, __s_first, __pred, ::std::true_type(), ::std::true_type());
+            __last, __s_first, __pred);
         return __res ? __first : __last;
     }
 

From a550ebf6d6a42bf2e3e50f95b9beb84b63e32bb1 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Mon, 5 Feb 2024 17:36:04 +0100
Subject: [PATCH 078/566] __pattern_find_end + tag impls

---
 include/oneapi/dpl/pstl/algorithm_fwd.h       | 11 +++++
 include/oneapi/dpl/pstl/algorithm_impl.h      | 42 +++++++++++++++++++
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 28 +++++++++++++
 3 files changed, 81 insertions(+)

diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index 73bb1bc2f79..1631a73aada 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -501,6 +501,11 @@ __pattern_find_end(_ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _Fo
                    _BinaryPredicate, _IsVector,
                    /*is_parallel=*/::std::false_type) noexcept;
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _BinaryPredicate>
+_ForwardIterator1
+__pattern_find_end(_Tag, _ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2, _ForwardIterator2,
+                   _BinaryPredicate) noexcept;
+
 template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _BinaryPredicate,
           class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator1>
@@ -508,6 +513,12 @@ __pattern_find_end(_ExecutionPolicy&&, _RandomAccessIterator1, _RandomAccessIter
                    _RandomAccessIterator2, _BinaryPredicate, _IsVector,
                    /*is_parallel=*/::std::true_type);
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
+          class _BinaryPredicate>
+_RandomAccessIterator1
+__pattern_find_end(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator1, _RandomAccessIterator1,
+                   _RandomAccessIterator2, _RandomAccessIterator2, _BinaryPredicate);
+
 //------------------------------------------------------------------------
 // find_first_of
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index ad7cfd9b399..ad91b847c06 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -1282,6 +1282,17 @@ __pattern_find_end(_ExecutionPolicy&&, _ForwardIterator1 __first, _ForwardIterat
     return __internal::__brick_find_end(__first, __last, __s_first, __s_last, __pred, __is_vector);
 }
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _BinaryPredicate>
+_ForwardIterator1
+__pattern_find_end(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first, _ForwardIterator1 __last, _ForwardIterator2 __s_first,
+                   _ForwardIterator2 __s_last, _BinaryPredicate __pred) noexcept
+{
+    static_assert(__is_backend_tag_v<_Tag>);
+
+    return __internal::__brick_find_end(__first, __last, __s_first, __s_last, __pred, typename _Tag::__is_vector{});
+}
+
+
 template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _BinaryPredicate,
           class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator1>
@@ -1315,6 +1326,37 @@ __pattern_find_end(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first, _R
     }
 }
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
+          class _BinaryPredicate>
+_RandomAccessIterator1
+__pattern_find_end(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _RandomAccessIterator1 __first,
+                   _RandomAccessIterator1 __last,
+                   _RandomAccessIterator2 __s_first, _RandomAccessIterator2 __s_last, _BinaryPredicate __pred)
+{
+    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
+
+    if (__last - __first == __s_last - __s_first)
+    {
+        const bool __res = __internal::__pattern_equal(__tag, ::std::forward<_ExecutionPolicy>(__exec),
+                                                       __first, __last, __s_first, __pred);
+        return __res ? __first : __last;
+    }
+    else
+    {
+        return __internal::__except_handler([&]() {
+            return __internal::__parallel_find(
+                __backend_tag{},
+                ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+                [__last, __s_first, __s_last, __pred](_RandomAccessIterator1 __i, _RandomAccessIterator1 __j) {
+                    return __internal::__find_subrange(__i, __j, __last, __s_first, __s_last, __pred, false,
+                                                       _IsVector{});
+                },
+                ::std::false_type{});
+        });
+    }
+}
+
+
 //------------------------------------------------------------------------
 // find_first_of
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 3fee7daff48..15ef508753d 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -1106,6 +1106,34 @@ __pattern_find_end(_ExecutionPolicy&& __exec, _Iterator1 __first, _Iterator1 __l
     }
 }
 
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2, typename _Pred>
+_Iterator1
+__pattern_find_end(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator1 __first, _Iterator1 __last,
+                   _Iterator2 __s_first, _Iterator2 __s_last, _Pred __pred)
+{
+    if (__first == __last || __s_last == __s_first || __last - __first < __s_last - __s_first)
+        return __last;
+
+    if (__last - __first == __s_last - __s_first)
+    {
+        const bool __res =
+            __pattern_equal(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __s_first, __pred);
+        return __res ? __first : __last;
+    }
+    else
+    {
+        using _Predicate = unseq_backend::multiple_match_pred<_ExecutionPolicy, _Pred>;
+
+        return __par_backend_hetero::__parallel_find(
+            ::std::forward<_ExecutionPolicy>(__exec),
+            __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__first),
+            __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__last),
+            __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__s_first),
+            __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__s_last), _Predicate{__pred},
+            ::std::false_type{});
+    }
+}
+
 //------------------------------------------------------------------------
 // find_first_of
 //------------------------------------------------------------------------

From a3742b16c102b4c5690651f38e453df51629223d Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Mon, 5 Feb 2024 17:36:24 +0100
Subject: [PATCH 079/566] __pattern_find_end + tag calls

---
 include/oneapi/dpl/pstl/glue_algorithm_impl.h | 9 +++++----
 1 file changed, 5 insertions(+), 4 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_algorithm_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
index 5af0b880aee..3197a4901b4 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
@@ -129,10 +129,11 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _Forward
 find_end(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _ForwardIterator1 __last, _ForwardIterator2 __s_first,
          _ForwardIterator2 __s_last, _BinaryPredicate __pred)
 {
-    return oneapi::dpl::__internal::__pattern_find_end(
-        ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __s_first, __s_last, __pred,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>());
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
+
+    return oneapi::dpl::__internal::__pattern_find_end(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
+                                                       __first, __last, __s_first, __s_last, __pred);
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2>

From 5d8ec601e5be7d45a37bccda4f3bac83468b0af6 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Mon, 5 Feb 2024 17:52:28 +0100
Subject: [PATCH 080/566] __pattern_find_first_of + tag impls

---
 include/oneapi/dpl/pstl/algorithm_fwd.h       | 11 +++++++
 include/oneapi/dpl/pstl/algorithm_impl.h      | 32 ++++++++++++++++++-
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 22 +++++++++++++
 3 files changed, 64 insertions(+), 1 deletion(-)

diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index 1631a73aada..008a150d442 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -539,12 +539,23 @@ oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _Fo
 __pattern_find_first_of(_ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2, _ForwardIterator2,
                         _BinaryPredicate, _IsVector, /*is_parallel=*/::std::false_type) noexcept;
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _BinaryPredicate>
+_ForwardIterator1
+__pattern_find_first_of(_Tag, _ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2,
+                        _ForwardIterator2, _BinaryPredicate) noexcept;
+
 template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _BinaryPredicate,
           class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator1>
 __pattern_find_first_of(_ExecutionPolicy&&, _RandomAccessIterator1, _RandomAccessIterator1, _RandomAccessIterator2,
                         _RandomAccessIterator2, _BinaryPredicate, _IsVector, /*is_parallel=*/::std::true_type);
 
+template <class _IsVector, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2,
+          class _BinaryPredicate>
+_ForwardIterator1
+__pattern_find_first_of(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1,
+                        _ForwardIterator2, _ForwardIterator2, _BinaryPredicate);
+
 //------------------------------------------------------------------------
 // search
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index ad91b847c06..bad6eb6df66 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -1378,7 +1378,7 @@ __brick_find_first_of(_ForwardIterator1 __first, _ForwardIterator1 __last, _Forw
 
 template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _BinaryPredicate,
           class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _ForwardIterator1>
+_ForwardIterator1
 __pattern_find_first_of(_ExecutionPolicy&&, _ForwardIterator1 __first, _ForwardIterator1 __last,
                         _ForwardIterator2 __s_first, _ForwardIterator2 __s_last, _BinaryPredicate __pred,
                         _IsVector __is_vector, /*is_parallel=*/::std::false_type) noexcept
@@ -1386,6 +1386,17 @@ __pattern_find_first_of(_ExecutionPolicy&&, _ForwardIterator1 __first, _ForwardI
     return __internal::__brick_find_first_of(__first, __last, __s_first, __s_last, __pred, __is_vector);
 }
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _BinaryPredicate>
+_ForwardIterator1
+__pattern_find_first_of(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first, _ForwardIterator1 __last,
+                        _ForwardIterator2 __s_first, _ForwardIterator2 __s_last, _BinaryPredicate __pred) noexcept
+{
+    static_assert(__is_backend_tag_v<_Tag>);
+
+    return __internal::__brick_find_first_of(__first, __last, __s_first, __s_last, __pred,
+                                             typename _Tag::__is_vector{});
+}
+
 template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _BinaryPredicate,
           class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _ForwardIterator1>
@@ -1403,6 +1414,25 @@ __pattern_find_first_of(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _F
     });
 }
 
+template <class _IsVector, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2,
+          class _BinaryPredicate>
+_ForwardIterator1
+__pattern_find_first_of(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _ForwardIterator1 __first,
+                        _ForwardIterator1 __last, _ForwardIterator2 __s_first, _ForwardIterator2 __s_last,
+                        _BinaryPredicate __pred)
+{
+    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
+
+    return __internal::__except_handler([&]() {
+        return __internal::__parallel_find(
+            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+            [__s_first, __s_last, &__pred](_ForwardIterator1 __i, _ForwardIterator1 __j) {
+                return __internal::__brick_find_first_of(__i, __j, __s_first, __s_last, __pred, _IsVector{});
+            },
+            ::std::true_type{});
+    });
+}
+
 //------------------------------------------------------------------------
 // search
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 15ef508753d..0b2b439eca2 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -1159,6 +1159,28 @@ __pattern_find_first_of(_ExecutionPolicy&& __exec, _Iterator1 __first, _Iterator
         ::std::true_type{});
 }
 
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2, typename _Pred>
+_Iterator1
+__pattern_find_first_of(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator1 __first,
+                        _Iterator1 __last, _Iterator2 __s_first, _Iterator2 __s_last, _Pred __pred)
+{
+    if (__first == __last || __s_last == __s_first)
+        return __last;
+
+    using _Predicate = unseq_backend::first_match_pred<_ExecutionPolicy, _Pred>;
+
+    // TODO: To check whether it makes sense to iterate over the second sequence in case of
+    // distance(__first, __last) < distance(__s_first, __s_last).
+    return __par_backend_hetero::__parallel_find(
+        ::std::forward<_ExecutionPolicy>(__exec),
+        __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__first),
+        __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__last),
+        __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__s_first),
+        __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__s_last), _Predicate{__pred},
+        ::std::true_type{});
+}
+
+
 //------------------------------------------------------------------------
 // search
 //------------------------------------------------------------------------

From 5a52a0b8d82e6e0809626de4b802b6e4266392e2 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Mon, 5 Feb 2024 17:52:42 +0100
Subject: [PATCH 081/566] __pattern_find_first_of + tag calls

---
 include/oneapi/dpl/pstl/glue_algorithm_impl.h | 9 +++++----
 1 file changed, 5 insertions(+), 4 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_algorithm_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
index 3197a4901b4..576fb10630d 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
@@ -151,10 +151,11 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _Forward
 find_first_of(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _ForwardIterator1 __last,
               _ForwardIterator2 __s_first, _ForwardIterator2 __s_last, _BinaryPredicate __pred)
 {
-    return oneapi::dpl::__internal::__pattern_find_first_of(
-        ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __s_first, __s_last, __pred,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>());
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
+
+    return oneapi::dpl::__internal::__pattern_find_first_of(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
+                                                            __first, __last, __s_first, __s_last, __pred);
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2>

From e95371ad1ae9d6cf88c24f4b4a64d6fb7b388dc7 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Mon, 5 Feb 2024 18:06:03 +0100
Subject: [PATCH 082/566] __pattern_generate + tag impls

---
 include/oneapi/dpl/pstl/algorithm_fwd.h       |  9 +++++++
 include/oneapi/dpl/pstl/algorithm_impl.h      | 25 +++++++++++++++++++
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 12 +++++++++
 3 files changed, 46 insertions(+)

diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index 008a150d442..935c26da900 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -1183,11 +1183,20 @@ oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>
 __pattern_generate(_ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _Generator,
                    /*is_parallel=*/::std::false_type, _IsVector) noexcept;
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Generator>
+void
+__pattern_generate(_Tag, _ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _Generator) noexcept;
+
 template <class _ExecutionPolicy, class _RandomAccessIterator, class _Generator, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator>
 __pattern_generate(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _Generator,
                    /*is_parallel=*/::std::true_type, _IsVector);
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Generator>
+_RandomAccessIterator
+__pattern_generate(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator,
+                   _Generator);
+
 template <class _RandomAccessIterator, class Size, class _Generator>
 _RandomAccessIterator __brick_generate_n(_RandomAccessIterator, Size, _Generator,
                                          /* is_vector = */ ::std::true_type) noexcept;
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index bad6eb6df66..fbc3cb75126 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -3506,6 +3506,15 @@ __pattern_generate(_ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterato
     __internal::__brick_generate(__first, __last, __g, __is_vector);
 }
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Generator>
+void
+__pattern_generate(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _Generator __g) noexcept
+{
+    static_assert(__is_backend_tag_v<_Tag>);
+
+    __internal::__brick_generate(__first, __last, __g, typename _Tag::__is_vector{});
+}
+
 template <class _ExecutionPolicy, class _RandomAccessIterator, class _Generator, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator>
 __pattern_generate(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIterator __last,
@@ -3521,6 +3530,22 @@ __pattern_generate(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _Ra
     });
 }
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Generator>
+_RandomAccessIterator
+__pattern_generate(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
+                   _RandomAccessIterator __last, _Generator __g)
+{
+    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
+
+    return __internal::__except_handler([&]() {
+        __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+                                      [__g](_RandomAccessIterator __begin, _RandomAccessIterator __end) {
+                                          __internal::__brick_generate(__begin, __end, __g, _IsVector{});
+                                      });
+        return __last;
+    });
+}
+
 template <class _RandomAccessIterator, class Size, class _Generator>
 _RandomAccessIterator
 __brick_generate_n(_RandomAccessIterator __first, Size __count, _Generator __g,
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 0b2b439eca2..6995bcb492d 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -556,6 +556,18 @@ __pattern_generate(_ExecutionPolicy&& __exec, _ForwardIterator __first, _Forward
     return __last;
 }
 
+template <typename _BackendTag, typename _ExecutionPolicy, typename _ForwardIterator, typename _Generator>
+_ForwardIterator
+__pattern_generate(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _ForwardIterator __first,
+                   _ForwardIterator __last, _Generator __g)
+{
+    __pattern_walk1(__tag, ::std::forward<_ExecutionPolicy>(__exec),
+                    __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::write>(__first),
+                    __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::write>(__last),
+                    generate_functor<_Generator>{__g});
+    return __last;
+}
+
 //------------------------------------------------------------------------
 // brick_copy, brick_move
 //------------------------------------------------------------------------

From 183ea6192c29daa4a236f532c66346d30ae13c95 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Mon, 5 Feb 2024 18:06:40 +0100
Subject: [PATCH 083/566] __pattern_generate + tag calls

---
 include/oneapi/dpl/pstl/glue_algorithm_impl.h | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_algorithm_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
index 576fb10630d..1fd1ff3232e 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
@@ -467,10 +467,10 @@ template <class _ExecutionPolicy, class _ForwardIterator, class _Generator>
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy>
 generate(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last, _Generator __g)
 {
-    oneapi::dpl::__internal::__pattern_generate(
-        ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __g,
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(),
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>());
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+
+    oneapi::dpl::__internal::__pattern_generate(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first,
+                                                __last, __g);
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator, class _Size, class _Generator>

From 233854b58931309022bbdb907799f45900f9d663 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Mon, 5 Feb 2024 18:17:08 +0100
Subject: [PATCH 084/566] __pattern_generate_n + tag impls

---
 include/oneapi/dpl/pstl/algorithm_fwd.h  |  8 ++++++++
 include/oneapi/dpl/pstl/algorithm_impl.h | 20 ++++++++++++++++++++
 2 files changed, 28 insertions(+)

diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index 935c26da900..f5f7d58dc72 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -1210,11 +1210,19 @@ OutputIterator
 __pattern_generate_n(_ExecutionPolicy&&, OutputIterator, Size, _Generator,
                      /*is_parallel=*/::std::false_type, _IsVector) noexcept;
 
+template <class _Tag, class _ExecutionPolicy, class _OutputIterator, class _Size, class _Generator>
+_OutputIterator
+__pattern_generate_n(_Tag, _ExecutionPolicy&&, _OutputIterator, _Size, _Generator) noexcept;
+
 template <class _ExecutionPolicy, class _RandomAccessIterator, class Size, class _Generator, class _IsVector>
 _RandomAccessIterator
 __pattern_generate_n(_ExecutionPolicy&&, _RandomAccessIterator, Size, _Generator,
                      /*is_parallel=*/::std::true_type, _IsVector);
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Size, class _Generator>
+_RandomAccessIterator
+__pattern_generate_n(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator, _Size, _Generator);
+
 //------------------------------------------------------------------------
 // remove
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index fbc3cb75126..8840e63943b 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -3569,6 +3569,15 @@ __pattern_generate_n(_ExecutionPolicy&&, _OutputIterator __first, _Size __count,
     return __internal::__brick_generate_n(__first, __count, __g, __is_vector);
 }
 
+template <class _Tag, class _ExecutionPolicy, class _OutputIterator, class _Size, class _Generator>
+_OutputIterator
+__pattern_generate_n(_Tag, _ExecutionPolicy&&, _OutputIterator __first, _Size __count, _Generator __g) noexcept
+{
+    static_assert(__is_backend_tag_v<_Tag>);
+
+    return __internal::__brick_generate_n(__first, __count, __g, typename _Tag::__is_vector{});
+}
+
 template <class _ExecutionPolicy, class _RandomAccessIterator, class _Size, class _Generator, class _IsVector>
 _RandomAccessIterator
 __pattern_generate_n(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _Size __count, _Generator __g,
@@ -3580,6 +3589,17 @@ __pattern_generate_n(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _
                                           ::std::true_type(), __is_vector);
 }
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Size, class _Generator>
+_RandomAccessIterator
+__pattern_generate_n(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
+                     _Size __count, _Generator __g)
+{
+    static_assert(__is_random_access_iterator_v<_RandomAccessIterator>,
+                  "Pattern-brick error. Should be a random access iterator.");
+    return __internal::__pattern_generate(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __first + __count,
+                                          __g);
+}
+
 //------------------------------------------------------------------------
 // remove
 //------------------------------------------------------------------------

From c8a61eef2d008df8a06108c79e4f05b803611922 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Mon, 5 Feb 2024 18:17:19 +0100
Subject: [PATCH 085/566] __pattern_generate_n + tag calls

---
 include/oneapi/dpl/pstl/glue_algorithm_impl.h | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_algorithm_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
index 1fd1ff3232e..d1580305c5a 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
@@ -480,10 +480,10 @@ generate_n(_ExecutionPolicy&& __exec, _ForwardIterator __first, _Size __count, _
     if (__count <= 0)
         return __first;
 
-    return oneapi::dpl::__internal::__pattern_generate_n(
-        ::std::forward<_ExecutionPolicy>(__exec), __first, __count, __g,
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>(),
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>());
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+
+    return oneapi::dpl::__internal::__pattern_generate_n(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
+                                                         __first, __count, __g);
 }
 
 // [alg.remove]

From cdc46496cc051cce0eabd78a02c8e504fb33f624 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Mon, 5 Feb 2024 18:27:40 +0100
Subject: [PATCH 086/566] 
 include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h - fix compiler error:
 implement previously absent impl of __pattern_generate_n for hetero policy

---
 .../oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h  | 13 +++++++++++++
 1 file changed, 13 insertions(+)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 6995bcb492d..04e920be109 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -568,6 +568,19 @@ __pattern_generate(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _
     return __last;
 }
 
+template <typename _BackendTag, typename _ExecutionPolicy, typename _ForwardIterator, class _Size, typename _Generator>
+_ForwardIterator
+__pattern_generate_n(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _ForwardIterator __first,
+                     _Size __count, _Generator __g)
+{
+    // TODO: is this new implementation are correct?
+    // Previously we hadn't hetero impl for __pattern_generate_n
+
+    return __pattern_walk1_n(__tag, ::std::forward<_ExecutionPolicy>(__exec),
+                             __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::write>(__first),
+                             __count, generate_functor<_Generator>{__g});
+}
+
 //------------------------------------------------------------------------
 // brick_copy, brick_move
 //------------------------------------------------------------------------

From 4f3b82ed3cbfc0c3138e056312716c7e2a59569c Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Mon, 5 Feb 2024 18:40:40 +0100
Subject: [PATCH 087/566] __pattern_includes + tag impls

---
 include/oneapi/dpl/pstl/algorithm_fwd.h       | 11 +++
 include/oneapi/dpl/pstl/algorithm_impl.h      | 71 +++++++++++++++++++
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 28 ++++++++
 3 files changed, 110 insertions(+)

diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index f5f7d58dc72..202c503a287 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -1305,6 +1305,11 @@ __pattern_includes(_ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _Fo
                    _Compare, _IsVector,
                    /*is_parallel=*/::std::false_type) noexcept;
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Compare>
+bool
+__pattern_includes(_Tag, _ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2, _ForwardIterator2,
+                   _Compare) noexcept;
+
 template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _Compare,
           class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, bool>
@@ -1312,6 +1317,12 @@ __pattern_includes(_ExecutionPolicy&&, _RandomAccessIterator1, _RandomAccessIter
                    _RandomAccessIterator2, _Compare, _IsVector,
                    /*is_parallel=*/::std::true_type);
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
+          class _Compare>
+bool
+__pattern_includes(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator1, _RandomAccessIterator1,
+                   _RandomAccessIterator2, _RandomAccessIterator2, _Compare);
+
 //------------------------------------------------------------------------
 // set_union
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 8840e63943b..299e2b1b78f 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -3787,6 +3787,16 @@ __pattern_includes(_ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardItera
     return ::std::includes(__first1, __last1, __first2, __last2, __comp);
 }
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Compare>
+bool
+__pattern_includes(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
+                   _ForwardIterator2 __first2, _ForwardIterator2 __last2, _Compare __comp) noexcept
+{
+    static_assert(__is_backend_tag_v<_Tag>);
+
+    return ::std::includes(__first1, __last1, __first2, __last2, __comp);
+}
+
 template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _Compare,
           class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, bool>
@@ -3849,6 +3859,67 @@ __pattern_includes(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1, _
     });
 }
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _Compare>
+bool
+__pattern_includes(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1,
+                   _RandomAccessIterator1 __last1, _RandomAccessIterator2 __first2, _RandomAccessIterator2 __last2,
+                   _Compare __comp)
+{
+    if (__first2 == __last2)
+        return true;
+
+    //optimization; {1} - the first sequence, {2} - the second sequence
+    //{1} is empty or size_of{2} > size_of{1}
+    if (__first1 == __last1 || __last2 - __first2 > __last1 - __first1 ||
+        // {1}:     [**********]     or   [**********]
+        // {2}: [***********]                   [***********]
+        __comp(*__first2, *__first1) || __comp(*(__last1 - 1), *(__last2 - 1)))
+        return false;
+
+    __first1 = ::std::lower_bound(__first1, __last1, *__first2, __comp);
+    if (__first1 == __last1)
+        return false;
+
+    if (__last2 - __first2 == 1)
+        return !__comp(*__first1, *__first2) && !__comp(*__first2, *__first1);
+
+    return __internal::__except_handler([&]() {
+        return !__internal::__parallel_or(
+            ::std::forward<_ExecutionPolicy>(__exec), __first2, __last2,
+            [__first1, __last1, __first2, __last2, &__comp](_RandomAccessIterator2 __i, _RandomAccessIterator2 __j) {
+                assert(__j > __i);
+                //assert(__j - __i > 1);
+
+                //1. moving boundaries to "consume" subsequence of equal elements
+                auto __is_equal_sorted = [&__comp](_RandomAccessIterator2 __a, _RandomAccessIterator2 __b) -> bool {
+                    //enough one call of __comp due to compared couple belongs to one sorted sequence
+                    return !__comp(*__a, *__b);
+                };
+
+                //1.1 left bound, case "aaa[aaaxyz...]" - searching "x"
+                if (__i > __first2 && __is_equal_sorted(__i - 1, __i))
+                {
+                    //whole subrange continues to have equal elements - return "no op"
+                    if (__is_equal_sorted(__i, __j - 1))
+                        return false;
+
+                    __i = ::std::upper_bound(__i, __last2, *__i, __comp);
+                }
+
+                //1.2 right bound, case "[...aaa]aaaxyz" - searching "x"
+                if (__j < __last2 && __is_equal_sorted(__j - 1, __j))
+                    __j = ::std::upper_bound(__j, __last2, *__j, __comp);
+
+                //2. testing is __a subsequence of the second range included into the first range
+                auto __b = ::std::lower_bound(__first1, __last1, *__i, __comp);
+
+                assert(!__comp(*(__last1 - 1), *__b));
+                assert(!__comp(*(__j - 1), *__i));
+                return !::std::includes(__b, __last1, __i, __j, __comp);
+            });
+    });
+}
+
 inline constexpr auto __set_algo_cut_off = 1000;
 
 template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _OutputIterator,
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 04e920be109..6a135eb5429 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -1887,6 +1887,34 @@ __pattern_includes(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _Forwa
         __brick_include_type(__comp, __last1 - __first1, __last2 - __first2));
 }
 
+template <typename _BackendTag, typename _ExecutionPolicy, typename _ForwardIterator1, typename _ForwardIterator2,
+          typename _Compare>
+bool
+__pattern_includes(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _ForwardIterator1 __first1,
+                   _ForwardIterator1 __last1, _ForwardIterator2 __first2, _ForwardIterator2 __last2, _Compare __comp)
+{
+    //according to the spec
+    if (__first2 == __last2)
+        return true;
+
+    //optimization; {1} - the first sequence, {2} - the second sequence
+    //{1} is empty or size_of{2} > size_of{1}
+    if (__first1 == __last1 || __last2 - __first2 > __last1 - __first1)
+        return false;
+
+    typedef typename ::std::iterator_traits<_ForwardIterator1>::difference_type _Size1;
+    typedef typename ::std::iterator_traits<_ForwardIterator2>::difference_type _Size2;
+
+    using __brick_include_type = unseq_backend::__brick_includes<_ExecutionPolicy, _Compare, _Size1, _Size2>;
+    return !__par_backend_hetero::__parallel_or(
+        ::std::forward<_ExecutionPolicy>(__exec),
+        __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__first2),
+        __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__last2),
+        __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__first1),
+        __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__last1),
+        __brick_include_type(__comp, __last1 - __first1, __last2 - __first2));
+}
+
 //------------------------------------------------------------------------
 // partial_sort
 //------------------------------------------------------------------------

From 5e7336277f071480af0d111af6d7f6172c666fba Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Mon, 5 Feb 2024 18:40:59 +0100
Subject: [PATCH 088/566] __pattern_includes + tag calls

---
 include/oneapi/dpl/pstl/glue_algorithm_impl.h | 9 +++++----
 1 file changed, 5 insertions(+), 4 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_algorithm_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
index d1580305c5a..e5015e47064 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
@@ -964,10 +964,11 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, bool>
 includes(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 __last1, _ForwardIterator2 __first2,
          _ForwardIterator2 __last2, _Compare __comp)
 {
-    return oneapi::dpl::__internal::__pattern_includes(
-        ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2, __comp,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>());
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
+
+    return oneapi::dpl::__internal::__pattern_includes(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
+                                                       __first1, __last1, __first2, __last2, __comp);
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2>

From 81b5469db7fbdacec7444e1b721c6765503f9d6a Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Mon, 5 Feb 2024 18:56:09 +0100
Subject: [PATCH 089/566] __pattern_inplace_merge + tag impls

---
 include/oneapi/dpl/pstl/algorithm_fwd.h       | 10 ++++
 include/oneapi/dpl/pstl/algorithm_impl.h      | 54 +++++++++++++++++++
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 30 +++++++++++
 3 files changed, 94 insertions(+)

diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index 202c503a287..2132e715aa7 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -1289,12 +1289,22 @@ __pattern_inplace_merge(_ExecutionPolicy&&, _BidirectionalIterator, _Bidirection
                         _Compare, _IsVector,
                         /* is_parallel = */ ::std::false_type) noexcept;
 
+template <class _Tag, class _ExecutionPolicy, class _BidirectionalIterator, class _Compare>
+void
+__pattern_inplace_merge(_Tag, _ExecutionPolicy&&, _BidirectionalIterator, _BidirectionalIterator,
+                        _BidirectionalIterator, _Compare) noexcept;
+
 template <class _ExecutionPolicy, class _RandomAccessIterator, class _Compare, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>
 __pattern_inplace_merge(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _RandomAccessIterator,
                         _Compare, _IsVector,
                         /*is_parallel=*/::std::true_type);
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Compare>
+void
+__pattern_inplace_merge(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator,
+                        _RandomAccessIterator, _Compare);
+
 //------------------------------------------------------------------------
 // includes
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 299e2b1b78f..af5ca222661 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -3732,6 +3732,16 @@ __pattern_inplace_merge(_ExecutionPolicy&&, _BidirectionalIterator __first, _Bid
     __internal::__brick_inplace_merge(__first, __middle, __last, __comp, __is_vector);
 }
 
+template <class _Tag, class _ExecutionPolicy, class _BidirectionalIterator, class _Compare>
+void
+__pattern_inplace_merge(_Tag, _ExecutionPolicy&&, _BidirectionalIterator __first, _BidirectionalIterator __middle,
+                        _BidirectionalIterator __last, _Compare __comp) noexcept
+{
+    static_assert(__is_backend_tag_v<_Tag>);
+
+    __internal::__brick_inplace_merge(__first, __middle, __last, __comp, typename _Tag::__is_vector{});
+}
+
 template <class _ExecutionPolicy, class _RandomAccessIterator, class _Compare, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>
 __pattern_inplace_merge(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIterator __middle,
@@ -3774,6 +3784,50 @@ __pattern_inplace_merge(_ExecutionPolicy&& __exec, _RandomAccessIterator __first
     });
 }
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Compare>
+void
+__pattern_inplace_merge(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
+                        _RandomAccessIterator __middle, _RandomAccessIterator __last, _Compare __comp)
+{
+    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
+
+    if (__first == __last || __first == __middle || __middle == __last)
+    {
+        return;
+    }
+
+    typedef typename ::std::iterator_traits<_RandomAccessIterator>::value_type _Tp;
+    auto __n = __last - __first;
+    __par_backend::__buffer<_ExecutionPolicy, _Tp> __buf(__n);
+    _Tp* __r = __buf.get();
+    __internal::__except_handler([&]() {
+        auto __move_values = [](_RandomAccessIterator __x, _Tp* __z) {
+            if constexpr (::std::is_trivial_v<_Tp>)
+                *__z = ::std::move(*__x);
+            else
+                ::new (::std::addressof(*__z)) _Tp(::std::move(*__x));
+        };
+
+        auto __move_sequences = [](_RandomAccessIterator __first1, _RandomAccessIterator __last1, _Tp* __first2) {
+            return __internal::__brick_uninitialized_move(__first1, __last1, __first2, _IsVector{});
+        };
+
+        __par_backend::__parallel_merge(
+            ::std::forward<_ExecutionPolicy>(__exec), __first, __middle, __middle, __last, __r, __comp,
+            [__n, __move_values, __move_sequences](_RandomAccessIterator __f1, _RandomAccessIterator __l1,
+                                                   _RandomAccessIterator __f2, _RandomAccessIterator __l2, _Tp* __f3,
+                                                   _Compare __comp) {
+                (__utils::__serial_move_merge(__n))(__f1, __l1, __f2, __l2, __f3, __comp, __move_values, __move_values,
+                                                    __move_sequences, __move_sequences);
+                return __f3 + (__l1 - __f1) + (__l2 - __f2);
+            });
+        __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __r, __r + __n,
+                                      [__r, __first](_Tp* __i, _Tp* __j) {
+                                          __brick_move_destroy<_ExecutionPolicy>{}(__i, __j, __first + (__i - __r), _IsVector{});
+                                      });
+    });
+}
+
 //------------------------------------------------------------------------
 // includes
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 6a135eb5429..f6eb9fd44b5 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -1691,6 +1691,36 @@ __pattern_inplace_merge(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator
         __copy_first, __copy_last, __first, __brick_move<_ExecutionPolicy>{}, ::std::true_type{}, ::std::true_type{});
 }
 
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator, typename _Compare>
+void
+__pattern_inplace_merge(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator __first,
+                        _Iterator __middle, _Iterator __last, _Compare __comp)
+{
+    using _ValueType = typename ::std::iterator_traits<_Iterator>::value_type;
+
+    if (__first == __middle || __middle == __last || __first == __last)
+        return;
+
+    assert(__first < __middle && __middle < __last);
+
+    auto __n = __last - __first;
+    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __buf(__exec, __n);
+    auto __copy_first = __buf.get();
+    auto __copy_last = __copy_first + __n;
+
+    __pattern_merge(__exec, __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__first),
+                    __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__middle),
+                    __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__middle),
+                    __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__last),
+                    __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::write>(__copy_first),
+                    __comp, ::std::true_type{}, ::std::true_type{});
+
+    //TODO: optimize copy back depending on Iterator, i.e. set_final_data for host iterator/pointer
+    __pattern_walk2(
+        __tag, __par_backend_hetero::make_wrapped_policy<copy_back_wrapper>(::std::forward<_ExecutionPolicy>(__exec)),
+        __copy_first, __copy_last, __first, __brick_move<_ExecutionPolicy>{});
+}
+
 //------------------------------------------------------------------------
 // sort
 //------------------------------------------------------------------------

From b8e043fe5145420e90fe3fffbb35ade9cc55db3f Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Mon, 5 Feb 2024 18:56:22 +0100
Subject: [PATCH 090/566] __pattern_inplace_merge + tag calls

---
 include/oneapi/dpl/pstl/glue_algorithm_impl.h | 9 +++++----
 1 file changed, 5 insertions(+), 4 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_algorithm_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
index e5015e47064..413ce31fe00 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
@@ -942,10 +942,11 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy>
 inplace_merge(_ExecutionPolicy&& __exec, _BidirectionalIterator __first, _BidirectionalIterator __middle,
               _BidirectionalIterator __last, _Compare __comp)
 {
-    oneapi::dpl::__internal::__pattern_inplace_merge(
-        ::std::forward<_ExecutionPolicy>(__exec), __first, __middle, __last, __comp,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _BidirectionalIterator>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _BidirectionalIterator>());
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _BidirectionalIterator>();
+
+    oneapi::dpl::__internal::__pattern_inplace_merge(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first,
+                                                     __middle, __last, __comp);
 }
 
 template <class _ExecutionPolicy, class _BidirectionalIterator>

From 526ae3010aca144f8a204ca736492825381a1676 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 6 Feb 2024 09:41:14 +0100
Subject: [PATCH 091/566] __pattern_is_heap + tag impls

---
 include/oneapi/dpl/pstl/algorithm_fwd.h       |  9 ++++++++
 include/oneapi/dpl/pstl/algorithm_impl.h      | 23 +++++++++++++++++++
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 17 ++++++++++++++
 3 files changed, 49 insertions(+)

diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index 2132e715aa7..bde93113e9b 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -1481,11 +1481,20 @@ oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, boo
 __pattern_is_heap(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _Compare, _IsVector,
                   /* is_parallel = */ ::std::false_type) noexcept;
 
+template <class _Tag, class _ExecutionPolicy, class _RandomAccessIterator, class _Compare>
+bool
+__pattern_is_heap(_Tag, _ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _Compare) noexcept;
+
 template <class _ExecutionPolicy, class _RandomAccessIterator, class _Compare, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, bool>
 __pattern_is_heap(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _Compare, _IsVector,
                   /* is_parallel = */ ::std::true_type);
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Compare>
+bool
+__pattern_is_heap(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator,
+                  _Compare);
+
 //------------------------------------------------------------------------
 // min_element
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index af5ca222661..973c2d7353b 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -4615,6 +4615,15 @@ __pattern_is_heap(_ExecutionPolicy&&, _RandomAccessIterator __first, _RandomAcce
     return __internal::__brick_is_heap(__first, __last, __comp, __is_vector);
 }
 
+template <class _Tag, class _ExecutionPolicy, class _RandomAccessIterator, class _Compare>
+bool
+__pattern_is_heap(_Tag, _ExecutionPolicy&&, _RandomAccessIterator __first, _RandomAccessIterator __last, _Compare __comp) noexcept
+{
+    static_assert(__is_backend_tag_v<_Tag>);
+
+    return __internal::__brick_is_heap(__first, __last, __comp, typename _Tag::__is_vector{});
+}
+
 template <class _ExecutionPolicy, class _RandomAccessIterator, class _Compare, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, bool>
 __pattern_is_heap(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIterator __last,
@@ -4629,6 +4638,20 @@ __pattern_is_heap(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _Ran
     });
 }
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Compare>
+bool
+__pattern_is_heap(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
+                  _RandomAccessIterator __last, _Compare __comp)
+{
+    return __internal::__except_handler([&]() {
+        return !__parallel_or(::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+                              [__first, __comp](_RandomAccessIterator __i, _RandomAccessIterator __j) {
+                                  return !__internal::__is_heap_local(__first, __i - __first, __j - __first, __comp,
+                                                                      _IsVector{});
+                              });
+    });
+}
+
 //------------------------------------------------------------------------
 // min_element
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index f6eb9fd44b5..23c88d4b769 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -1612,6 +1612,23 @@ __pattern_is_heap(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _Ran
         __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__last), _Predicate{__comp});
 }
 
+template <typename _BackendTag, typename _ExecutionPolicy, typename _RandomAccessIterator, typename _Compare>
+bool
+__pattern_is_heap(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
+                  _RandomAccessIterator __last, _Compare __comp)
+{
+    if (__last - __first < 2)
+        return true;
+
+    using _Predicate =
+        oneapi::dpl::unseq_backend::single_match_pred_by_idx<_ExecutionPolicy, __is_heap_check<_Compare>>;
+
+    return !__par_backend_hetero::__parallel_or(
+        ::std::forward<_ExecutionPolicy>(__exec),
+        __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__first),
+        __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__last), _Predicate{__comp});
+}
+
 //------------------------------------------------------------------------
 // merge
 //------------------------------------------------------------------------

From f61dbc9c69a4815260dc0361af69ce772784317f Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 6 Feb 2024 09:41:25 +0100
Subject: [PATCH 092/566] __pattern_is_heap + tag calls

---
 include/oneapi/dpl/pstl/glue_algorithm_impl.h | 9 +++++----
 1 file changed, 5 insertions(+), 4 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_algorithm_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
index 413ce31fe00..f4a6cf48516 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
@@ -1097,10 +1097,11 @@ template <class _ExecutionPolicy, class _RandomAccessIterator, class _Compare>
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, bool>
 is_heap(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIterator __last, _Compare __comp)
 {
-    return oneapi::dpl::__internal::__pattern_is_heap(
-        ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __comp,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _RandomAccessIterator>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _RandomAccessIterator>());
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
+
+    return oneapi::dpl::__internal::__pattern_is_heap(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first,
+                                                      __last, __comp);
 }
 
 template <class _ExecutionPolicy, class _RandomAccessIterator>

From 42236624b6084b9012c8a6d26b29f0fd35ab8d28 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 6 Feb 2024 10:02:36 +0100
Subject: [PATCH 093/566] __pattern_is_heap_until + tag impls

---
 include/oneapi/dpl/pstl/algorithm_fwd.h       |  9 +++++++
 include/oneapi/dpl/pstl/algorithm_impl.h      | 26 +++++++++++++++++++
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 18 +++++++++++++
 3 files changed, 53 insertions(+)

diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index bde93113e9b..a2d586f4397 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -1459,11 +1459,20 @@ oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _Ra
 __pattern_is_heap_until(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _Compare, _IsVector,
                         /* is_parallel = */ ::std::false_type) noexcept;
 
+template <class _Tag, class _ExecutionPolicy, class _RandomAccessIterator, class _Compare>
+_RandomAccessIterator
+__pattern_is_heap_until(_Tag, _ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _Compare) noexcept;
+
 template <class _ExecutionPolicy, class _RandomAccessIterator, class _Compare, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator>
 __pattern_is_heap_until(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _Compare, _IsVector,
                         /* is_parallel = */ ::std::true_type);
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Compare>
+_RandomAccessIterator
+__pattern_is_heap_until(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator,
+                        _Compare);
+
 //------------------------------------------------------------------------
 // is_heap
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 973c2d7353b..742524bbc3f 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -4529,6 +4529,16 @@ __pattern_is_heap_until(_ExecutionPolicy&&, _RandomAccessIterator __first, _Rand
     return __internal::__brick_is_heap_until(__first, __last, __comp, __is_vector);
 }
 
+template <class _Tag, class _ExecutionPolicy, class _RandomAccessIterator, class _Compare>
+_RandomAccessIterator
+__pattern_is_heap_until(_Tag, _ExecutionPolicy&&, _RandomAccessIterator __first, _RandomAccessIterator __last,
+                        _Compare __comp) noexcept
+{
+    static_assert(__is_backend_tag_v<_Tag>);
+
+    return __internal::__brick_is_heap_until(__first, __last, __comp, typename _Tag::__is_vector{});
+}
+
 template <class _RandomAccessIterator, class _DifferenceType, class _Compare>
 _RandomAccessIterator
 __is_heap_until_local(_RandomAccessIterator __first, _DifferenceType __begin, _DifferenceType __end, _Compare __comp,
@@ -4566,6 +4576,22 @@ __pattern_is_heap_until(_ExecutionPolicy&& __exec, _RandomAccessIterator __first
     });
 }
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Compare>
+_RandomAccessIterator
+__pattern_is_heap_until(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
+                        _RandomAccessIterator __last, _Compare __comp)
+{
+    return __internal::__except_handler([&]() {
+        return __parallel_find(
+            ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+            [__first, __comp](_RandomAccessIterator __i, _RandomAccessIterator __j) {
+                return __internal::__is_heap_until_local(__first, __i - __first, __j - __first, __comp, _IsVector{});
+            },
+            ::std::true_type{});
+    });
+}
+
+
 //------------------------------------------------------------------------
 // is_heap
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 23c88d4b769..115b645e628 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -1595,6 +1595,24 @@ __pattern_is_heap_until(_ExecutionPolicy&& __exec, _RandomAccessIterator __first
         ::std::true_type{});
 }
 
+template <typename _BackendTag, typename _ExecutionPolicy, typename _RandomAccessIterator, typename _Compare>
+_RandomAccessIterator
+__pattern_is_heap_until(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
+                        _RandomAccessIterator __last, _Compare __comp)
+{
+    if (__last - __first < 2)
+        return __last;
+
+    using _Predicate =
+        oneapi::dpl::unseq_backend::single_match_pred_by_idx<_ExecutionPolicy, __is_heap_check<_Compare>>;
+
+    return __par_backend_hetero::__parallel_find(
+        ::std::forward<_ExecutionPolicy>(__exec),
+        __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__first),
+        __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__last), _Predicate{__comp},
+        ::std::true_type{});
+}
+
 template <typename _ExecutionPolicy, typename _RandomAccessIterator, typename _Compare>
 oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, bool>
 __pattern_is_heap(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIterator __last,

From 0e0c0bfbffa2211772e9d858c71c11522975d901 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 6 Feb 2024 10:02:53 +0100
Subject: [PATCH 094/566] __pattern_is_heap_until + tag calls

---
 include/oneapi/dpl/pstl/glue_algorithm_impl.h | 9 +++++----
 1 file changed, 5 insertions(+), 4 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_algorithm_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
index f4a6cf48516..83dd2dfa00f 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
@@ -1079,10 +1079,11 @@ template <class _ExecutionPolicy, class _RandomAccessIterator, class _Compare>
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _RandomAccessIterator>
 is_heap_until(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIterator __last, _Compare __comp)
 {
-    return oneapi::dpl::__internal::__pattern_is_heap_until(
-        ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __comp,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _RandomAccessIterator>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _RandomAccessIterator>());
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
+
+    return oneapi::dpl::__internal::__pattern_is_heap_until(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
+                                                            __first, __last, __comp);
 }
 
 template <class _ExecutionPolicy, class _RandomAccessIterator>

From f5c35e2920fedeebbbbdd5d27aacccb242a336c2 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 6 Feb 2024 10:12:42 +0100
Subject: [PATCH 095/566] __pattern_is_partitioned + tag impls

---
 include/oneapi/dpl/pstl/algorithm_fwd.h       |   9 ++
 include/oneapi/dpl/pstl/algorithm_impl.h      | 106 ++++++++++++++++++
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   |  32 ++++++
 3 files changed, 147 insertions(+)

diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index a2d586f4397..dc540ab2769 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -908,11 +908,20 @@ oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, boo
 __pattern_is_partitioned(_ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _UnaryPredicate, _IsVector,
                          /*is_parallel=*/::std::false_type) noexcept;
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _UnaryPredicate>
+bool
+__pattern_is_partitioned(_Tag, _ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _UnaryPredicate) noexcept;
+
 template <class _ExecutionPolicy, class _RandomAccessIterator, class _UnaryPredicate, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, bool>
 __pattern_is_partitioned(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _UnaryPredicate, _IsVector,
                          /*is_parallel=*/::std::true_type);
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _UnaryPredicate>
+bool
+__pattern_is_partitioned(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator,
+                         _UnaryPredicate);
+
 //------------------------------------------------------------------------
 // partition
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 742524bbc3f..ccc3cec86dc 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -2586,6 +2586,15 @@ __pattern_is_partitioned(_ExecutionPolicy&&, _ForwardIterator __first, _ForwardI
     return __internal::__brick_is_partitioned(__first, __last, __pred, __is_vector);
 }
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _UnaryPredicate>
+bool
+__pattern_is_partitioned(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _UnaryPredicate __pred) noexcept
+{
+    static_assert(__is_backend_tag_v<_Tag>);
+
+    return __internal::__brick_is_partitioned(__first, __last, __pred, typename _Tag::__is_vector{});
+}
+
 template <class _ExecutionPolicy, class _RandomAccessIterator, class _UnaryPredicate, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, bool>
 __pattern_is_partitioned(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIterator __last,
@@ -2683,6 +2692,103 @@ __pattern_is_partitioned(_ExecutionPolicy&& __exec, _RandomAccessIterator __firs
     });
 }
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _UnaryPredicate>
+bool
+__pattern_is_partitioned(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
+                         _RandomAccessIterator __last, _UnaryPredicate __pred)
+{
+    //trivial pre-checks
+    if (__first == __last)
+        return true;
+
+    return __internal::__except_handler([&]() {
+        // State of current range:
+        // broken     - current range is not partitioned by pred
+        // all_true   - all elements in current range satisfy pred
+        // all_false  - all elements in current range don't satisfy pred
+        // true_false - elements satisfy pred are placed before elements that don't satisfy pred
+        enum _ReduceRes
+        {
+            __not_init = -1,
+            __broken,
+            __all_true,
+            __all_false,
+            __true_false
+        };
+        // Array with states that we'll have when state from the left branch is merged with state from the right branch.
+        // State is calculated by formula: new_state = table[left_state * 4 + right_state]
+        const _ReduceRes __table[] = {__broken,     __broken,     __broken,     __broken, __broken,    __all_true,
+                                      __true_false, __true_false, __broken,     __broken, __all_false, __broken,
+                                      __broken,     __broken,     __true_false, __broken};
+        struct _ReduceType
+        {
+            _ReduceRes __val;
+            _RandomAccessIterator __pos;
+        };
+        //a commutative combiner
+        auto __combine = [&__table](_ReduceType __x, _ReduceType __y) {
+            return __x.__pos > __y.__pos ? _ReduceType{__table[__y.__val * 4 + __x.__val], __y.__pos}
+                                         : _ReduceType{__table[__x.__val * 4 + __y.__val], __x.__pos};
+        };
+
+        const _ReduceType __identity{__not_init, __last};
+
+        _ReduceType __result = __par_backend::__parallel_reduce(
+            ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __identity,
+            [&__pred, __combine](_RandomAccessIterator __i, _RandomAccessIterator __j,
+                                              _ReduceType __value) -> _ReduceType {
+                if (__value.__val == __broken)
+                    return _ReduceType{__broken, __i};
+
+                _ReduceType __res{__not_init, __i};
+                // if first element satisfy pred
+                if (__pred(*__i))
+                {
+                    // find first element that don't satisfy pred
+                    _RandomAccessIterator __x =
+                        __internal::__brick_find_if(__i + 1, __j, __not_pred<_UnaryPredicate&>(__pred), _IsVector{});
+                    if (__x != __j)
+                    {
+                        // find first element after "x" that satisfy pred
+                        _RandomAccessIterator __y = __internal::__brick_find_if(__x + 1, __j, __pred, _IsVector{});
+                        // if it was found then range isn't partitioned by pred
+                        if (__y != __j)
+                            return _ReduceType{__broken, __i};
+
+                        __res = _ReduceType{__true_false, __i};
+                    }
+                    else
+                        __res = _ReduceType{__all_true, __i};
+                }
+                else
+                { // if first element doesn't satisfy pred
+                    // then we should find the first element that satisfy pred.
+                    // If we found it then range isn't partitioned by pred
+                    if (__internal::__brick_find_if(__i + 1, __j, __pred, _IsVector{}) != __j)
+                        return _ReduceType{__broken, __i};
+
+                    __res = _ReduceType{__all_false, __i};
+                }
+                // if we have value from left range then we should calculate the result
+                return (__value.__val == __not_init) ? __res : __combine(__value, __res);
+            },
+
+            [__combine](_ReduceType __val1, _ReduceType __val2) -> _ReduceType {
+                if (__val1.__val == __not_init)
+                    return __val2;
+                if (__val2.__val == __not_init)
+                    return __val1;
+                assert(__val1.__val != __not_init && __val2.__val != __not_init);
+
+                if (__val1.__val == __broken || __val2.__val == __broken)
+                    return _ReduceType{__broken, __val1.__pos};
+                // calculate the result for new big range
+                return __combine(__val1, __val2);
+            });
+        return __result.__val != __broken;
+    });
+}
+
 //------------------------------------------------------------------------
 // partition
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 115b645e628..e5428eb37a0 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -1558,6 +1558,38 @@ __pattern_is_partitioned(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator
     return __broken != __reduce_fn(_ReduceValueType{__all_true}, __res);
 }
 
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator, typename _Predicate>
+bool
+__pattern_is_partitioned(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator __first,
+                         _Iterator __last, _Predicate __predicate)
+{
+    if (__last - __first < 2)
+        return true;
+
+    using _ReduceValueType = _IsPartitionedReduceType;
+    auto __reduce_fn = [](_ReduceValueType __a, _ReduceValueType __b) {
+        _ReduceValueType __table[] = {__broken,     __broken,     __broken,     __broken, __broken,    __all_true,
+                                      __true_false, __true_false, __broken,     __broken, __all_false, __broken,
+                                      __broken,     __broken,     __true_false, __broken};
+        return __table[__a * 4 + __b];
+    };
+    auto __transform_fn = [__predicate](auto __gidx, auto __acc) {
+        return (__predicate(__acc[__gidx]) ? __all_true : __all_false);
+    };
+
+    auto __keep = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _Iterator>();
+    auto __buf = __keep(__first, __last);
+
+    auto __res = oneapi::dpl::__par_backend_hetero::__parallel_transform_reduce<_ReduceValueType,
+                                                                                ::std::false_type /*is_commutative*/>(
+                     ::std::forward<_ExecutionPolicy>(__exec), __reduce_fn, __transform_fn,
+                     unseq_backend::__no_init_value{}, // no initial value
+                     __buf.all_view())
+                     .get();
+
+    return __broken != __reduce_fn(_ReduceValueType{__all_true}, __res);
+}
+
 //------------------------------------------------------------------------
 // is_heap / is_heap_until
 //------------------------------------------------------------------------

From b099ad246ffc2a24fd5bc44452f9e6d2683096c5 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 6 Feb 2024 10:12:51 +0100
Subject: [PATCH 096/566] __pattern_is_partitioned + tag calls

---
 include/oneapi/dpl/pstl/glue_algorithm_impl.h | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_algorithm_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
index 83dd2dfa00f..a909b7f9ca0 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
@@ -621,10 +621,10 @@ template <class _ExecutionPolicy, class _ForwardIterator, class _UnaryPredicate>
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, bool>
 is_partitioned(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last, _UnaryPredicate __pred)
 {
-    return oneapi::dpl::__internal::__pattern_is_partitioned(
-        ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __pred,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>());
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+
+    return oneapi::dpl::__internal::__pattern_is_partitioned(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
+                                                             __first, __last, __pred);
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator, class _UnaryPredicate>

From 98ca54c960e1238df607ce58bd8f2cc19e047643 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 6 Feb 2024 10:18:02 +0100
Subject: [PATCH 097/566] __pattern_lexicographical_compare + tag impls

---
 include/oneapi/dpl/pstl/algorithm_fwd.h       | 10 ++++
 include/oneapi/dpl/pstl/algorithm_impl.h      | 20 ++++++++
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 48 +++++++++++++++++++
 3 files changed, 78 insertions(+)

diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index dc540ab2769..fd3d03debfa 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -1609,6 +1609,11 @@ __pattern_lexicographical_compare(_ExecutionPolicy&&, _ForwardIterator1, _Forwar
                                   _ForwardIterator2, _Compare, _IsVector,
                                   /* is_parallel = */ ::std::false_type) noexcept;
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Compare>
+bool
+__pattern_lexicographical_compare(_Tag, _ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2,
+                                  _ForwardIterator2, _Compare) noexcept;
+
 template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _Compare,
           class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, bool>
@@ -1616,6 +1621,11 @@ __pattern_lexicographical_compare(_ExecutionPolicy&&, _RandomAccessIterator1, _R
                                   _RandomAccessIterator2, _RandomAccessIterator2, _Compare, _IsVector,
                                   /* is_parallel = */ ::std::true_type);
 
+template <class _IsVector, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Compare>
+bool
+__pattern_lexicographical_compare(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1,
+                                  _ForwardIterator2, _ForwardIterator2, _Compare) noexcept;
+
 template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Function, class _IsVector,
           class _IsParallel>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _ForwardIterator2>
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index ccc3cec86dc..6e6a8525a69 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -5032,6 +5032,17 @@ __brick_lexicographical_compare(_RandomAccessIterator1 __first1, _RandomAccessIt
     }
 }
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Compare>
+bool
+__pattern_lexicographical_compare(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
+                                  _ForwardIterator2 __first2, _ForwardIterator2 __last2, _Compare __comp) noexcept
+{
+    static_assert(__is_backend_tag_v<_Tag>);
+
+    return __internal::__brick_lexicographical_compare(__first1, __last1, __first2, __last2, __comp,
+                                                       typename _Tag::__is_vector{});
+}
+
 template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Compare, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, bool>
 __pattern_lexicographical_compare(_ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
@@ -5041,6 +5052,15 @@ __pattern_lexicographical_compare(_ExecutionPolicy&&, _ForwardIterator1 __first1
     return __internal::__brick_lexicographical_compare(__first1, __last1, __first2, __last2, __comp, __is_vector);
 }
 
+template <class _IsVector, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Compare>
+bool
+__pattern_lexicographical_compare(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _ForwardIterator1 __first1,
+                                  _ForwardIterator1 __last1, _ForwardIterator2 __first2, _ForwardIterator2 __last2,
+                                  _Compare __comp) noexcept
+{
+    return __internal::__brick_lexicographical_compare(__first1, __last1, __first2, __last2, __comp, _IsVector{});
+}
+
 template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _Compare,
           class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, bool>
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index e5428eb37a0..ce70301da0d 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -1956,6 +1956,54 @@ __pattern_lexicographical_compare(_ExecutionPolicy&& __exec, _Iterator1 __first1
     return __ret_idx ? __ret_idx == 1 : (__last1 - __first1) < (__last2 - __first2);
 }
 
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2, typename _Compare>
+bool
+__pattern_lexicographical_compare(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator1 __first1,
+                                  _Iterator1 __last1, _Iterator2 __first2, _Iterator2 __last2, _Compare __comp)
+{
+    //trivial pre-checks
+    if (__first2 == __last2)
+        return false;
+    if (__first1 == __last1)
+        return true;
+
+    using _Iterator1DifferenceType = typename ::std::iterator_traits<_Iterator1>::difference_type;
+    using _ReduceValueType = int32_t;
+
+    auto __reduce_fn = [](_ReduceValueType __a, _ReduceValueType __b) {
+        bool __is_mismatched = __a != 0;
+        return __a * __is_mismatched + __b * !__is_mismatched;
+    };
+    auto __transform_fn = [__comp](auto __gidx, auto __acc1, auto __acc2) {
+        auto const& __s1_val = __acc1[__gidx];
+        auto const& __s2_val = __acc2[__gidx];
+
+        ::std::int32_t __is_s1_val_less = __comp(__s1_val, __s2_val);
+        ::std::int32_t __is_s1_val_greater = __comp(__s2_val, __s1_val);
+
+        // 1 if __s1_val <  __s2_val, -1 if __s1_val <  __s2_val, 0 if __s1_val == __s2_val
+        return _ReduceValueType{1 * __is_s1_val_less - 1 * __is_s1_val_greater};
+    };
+
+    auto __shared_size = ::std::min(__last1 - __first1, (_Iterator1DifferenceType)(__last2 - __first2));
+
+    auto __keep1 = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _Iterator1>();
+    auto __buf1 = __keep1(__first1, __first1 + __shared_size);
+
+    auto __keep2 = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _Iterator2>();
+    auto __buf2 = __keep2(__first2, __first2 + __shared_size);
+
+    auto __ret_idx =
+        oneapi::dpl::__par_backend_hetero::__parallel_transform_reduce<_ReduceValueType,
+                                                                       ::std::false_type /*is_commutative*/>(
+            ::std::forward<_ExecutionPolicy>(__exec), __reduce_fn, __transform_fn,
+            unseq_backend::__no_init_value{}, // no initial value
+            __buf1.all_view(), __buf2.all_view())
+            .get();
+
+    return __ret_idx ? __ret_idx == 1 : (__last1 - __first1) < (__last2 - __first2);
+}
+
 template <typename _ExecutionPolicy, typename _ForwardIterator1, typename _ForwardIterator2, typename _Compare>
 oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, bool>
 __pattern_includes(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 __last1,

From 0279692e72c5f1e5e9cf4166ff673b6814b0eeb2 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 6 Feb 2024 10:18:11 +0100
Subject: [PATCH 098/566] __pattern_lexicographical_compare + tag calls

---
 include/oneapi/dpl/pstl/glue_algorithm_impl.h | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_algorithm_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
index a909b7f9ca0..8ca191f2c4c 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
@@ -1197,10 +1197,11 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, bool>
 lexicographical_compare(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
                         _ForwardIterator2 __first2, _ForwardIterator2 __last2, _Compare __comp)
 {
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
+
     return oneapi::dpl::__internal::__pattern_lexicographical_compare(
-        ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2, __comp,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>());
+        __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2, __comp);
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2>

From 047a9cb96029b83db7b85767016ebb5b8a63cddb Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 6 Feb 2024 10:35:44 +0100
Subject: [PATCH 099/566] __pattern_merge + tag impls

---
 include/oneapi/dpl/pstl/algorithm_fwd.h       | 12 ++++++
 include/oneapi/dpl/pstl/algorithm_impl.h      | 27 ++++++++++++
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 43 +++++++++++++++++++
 3 files changed, 82 insertions(+)

diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index fd3d03debfa..1a9a270c288 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -1273,6 +1273,12 @@ oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _Ou
 __pattern_merge(_ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2, _ForwardIterator2,
                 _OutputIterator, _Compare, _IsVector, /* is_parallel = */ ::std::false_type) noexcept;
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _OutputIterator,
+          class _Compare>
+_OutputIterator
+__pattern_merge(_Tag, _ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2, _ForwardIterator2,
+                _OutputIterator, _Compare) noexcept;
+
 template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _OutputIterator,
           class _Compare, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _OutputIterator>
@@ -1280,6 +1286,12 @@ __pattern_merge(_ExecutionPolicy&&, _RandomAccessIterator1, _RandomAccessIterato
                 _RandomAccessIterator2, _OutputIterator, _Compare, _IsVector,
                 /* is_parallel = */ ::std::true_type);
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
+          class _RandomAccessIterator3, class _Compare>
+_RandomAccessIterator3
+__pattern_merge(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator1, _RandomAccessIterator1,
+                _RandomAccessIterator2, _RandomAccessIterator2, _RandomAccessIterator3, _Compare);
+
 //------------------------------------------------------------------------
 // inplace_merge
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 6e6a8525a69..04b238611ec 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -3793,6 +3793,18 @@ __pattern_merge(_ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator
     return __internal::__brick_merge(__first1, __last1, __first2, __last2, __d_first, __comp, __is_vector);
 }
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _OutputIterator,
+          class _Compare>
+_OutputIterator
+__pattern_merge(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1 __last1, _ForwardIterator2 __first2,
+                _ForwardIterator2 __last2, _OutputIterator __d_first, _Compare __comp) noexcept
+{
+    static_assert(__is_backend_tag_v<_Tag>);
+
+    return __internal::__brick_merge(__first1, __last1, __first2, __last2, __d_first, __comp,
+                                     typename _Tag::__is_vector{});
+}
+
 template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
           class _RandomAccessIterator3, class _Compare, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator3>
@@ -3809,6 +3821,21 @@ __pattern_merge(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1, _Ran
     return __d_first + (__last1 - __first1) + (__last2 - __first2);
 }
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
+          class _RandomAccessIterator3, class _Compare>
+_RandomAccessIterator3
+__pattern_merge(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1,
+                _RandomAccessIterator1 __last1, _RandomAccessIterator2 __first2, _RandomAccessIterator2 __last2,
+                _RandomAccessIterator3 __d_first, _Compare __comp)
+{
+    __par_backend::__parallel_merge(
+        ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2, __d_first, __comp,
+        [](_RandomAccessIterator1 __f1, _RandomAccessIterator1 __l1, _RandomAccessIterator2 __f2,
+           _RandomAccessIterator2 __l2, _RandomAccessIterator3 __f3,
+           _Compare __comp) { return __internal::__brick_merge(__f1, __l1, __f2, __l2, __f3, __comp, _IsVector{}); });
+    return __d_first + (__last1 - __first1) + (__last2 - __first2);
+}
+
 //------------------------------------------------------------------------
 // inplace_merge
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index ce70301da0d..3fbd8d8757f 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -1725,6 +1725,49 @@ __pattern_merge(_ExecutionPolicy&& __exec, _Iterator1 __first1, _Iterator1 __las
     }
     return __d_first + __n;
 }
+
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2,
+          typename _Iterator3, typename _Compare>
+_Iterator3
+__pattern_merge(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator1 __first1, _Iterator1 __last1,
+                _Iterator2 __first2, _Iterator2 __last2, _Iterator3 __d_first, _Compare __comp)
+{
+    auto __n1 = __last1 - __first1;
+    auto __n2 = __last2 - __first2;
+    auto __n = __n1 + __n2;
+    if (__n == 0)
+        return __d_first;
+
+    //To consider the direct copying pattern call in case just one of sequences is empty.
+    if (__n1 == 0)
+        oneapi::dpl::__internal::__pattern_walk2_brick(
+            __tag,
+            oneapi::dpl::__par_backend_hetero::make_wrapped_policy<copy_back_wrapper>(
+                ::std::forward<_ExecutionPolicy>(__exec)),
+            __first2, __last2, __d_first, oneapi::dpl::__internal::__brick_copy<_ExecutionPolicy>{});
+    else if (__n2 == 0)
+        oneapi::dpl::__internal::__pattern_walk2_brick(
+            __tag,
+            oneapi::dpl::__par_backend_hetero::make_wrapped_policy<copy_back_wrapper2>(
+                ::std::forward<_ExecutionPolicy>(__exec)),
+            __first1, __last1, __d_first, oneapi::dpl::__internal::__brick_copy<_ExecutionPolicy>{});
+    else
+    {
+        auto __keep1 = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _Iterator1>();
+        auto __buf1 = __keep1(__first1, __last1);
+        auto __keep2 = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _Iterator2>();
+        auto __buf2 = __keep2(__first2, __last2);
+
+        auto __keep3 = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::write, _Iterator3>();
+        auto __buf3 = __keep3(__d_first, __d_first + __n);
+
+        __par_backend_hetero::__parallel_merge(::std::forward<_ExecutionPolicy>(__exec), __buf1.all_view(),
+                                               __buf2.all_view(), __buf3.all_view(), __comp)
+            .wait();
+    }
+    return __d_first + __n;
+}
+
 //------------------------------------------------------------------------
 // inplace_merge
 //------------------------------------------------------------------------

From 3f94595d567d9ad45a842cf12941377cf01b9c6b Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 6 Feb 2024 10:36:21 +0100
Subject: [PATCH 100/566] __pattern_merge + tag calls

---
 include/oneapi/dpl/pstl/glue_algorithm_impl.h          |  9 +++++----
 include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h | 10 +++++++---
 2 files changed, 12 insertions(+), 7 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_algorithm_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
index 8ca191f2c4c..165bb4e9ab2 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
@@ -922,10 +922,11 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _Forward
 merge(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 __last1, _ForwardIterator2 __first2,
       _ForwardIterator2 __last2, _ForwardIterator __d_first, _Compare __comp)
 {
-    return oneapi::dpl::__internal::__pattern_merge(
-        ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2, __d_first, __comp,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2, _ForwardIterator>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2, _ForwardIterator>());
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1,
+                                                                              _ForwardIterator2, _ForwardIterator>();
+
+    return oneapi::dpl::__internal::__pattern_merge(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first1,
+                                                    __last1, __first2, __last2, __d_first, __comp);
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _ForwardIterator>
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 3fbd8d8757f..05a80007649 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -2526,11 +2526,13 @@ __pattern_set_union(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _Forw
                                                                            __buf,__comp, unseq_backend::_DifferenceTag()
                                                                           ) - __buf;
     //2. Merge {1} and the difference
+    const auto __dispatch_tag1 = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, decltype(__first1),
+                                                                           decltype(__buf), decltype(__result)>();
     return oneapi::dpl::__internal::__pattern_merge(
+        __dispatch_tag1,
         oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__set_union_copy_case_2>(
             ::std::forward<_ExecutionPolicy>(__exec)),
-        __first1, __last1, __buf, __buf + __n_diff, __result, __comp,
-        /*vector=*/::std::true_type(), /*parallel=*/::std::true_type());
+        __first1, __last1, __buf, __buf + __n_diff, __result, __comp);
 }
 
 //Dummy names to avoid kernel problems
@@ -2619,7 +2621,9 @@ __pattern_set_symmetric_difference(_ExecutionPolicy&& __exec, _ForwardIterator1
         __buf_2;
 
     //3. Merge the differences
-    return oneapi::dpl::__internal::__pattern_merge(::std::forward<_ExecutionPolicy>(__exec), __buf_1,
+    constexpr auto __dispatch_tag1 = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, decltype(__buf_1),
+                                                                               decltype(__buf_2), decltype(__result)>();
+    return oneapi::dpl::__internal::__pattern_merge(__dispatch_tag1, ::std::forward<_ExecutionPolicy>(__exec), __buf_1,
                                                     __buf_1 + __n_diff_1, __buf_2, __buf_2 + __n_diff_2, __result,
                                                     __comp, ::std::true_type(), ::std::true_type());
 }

From f0b93f2a4f037b4d7e73300d02d7ebde7cdc3933 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 6 Feb 2024 10:48:35 +0100
Subject: [PATCH 101/566] __pattern_min_element + tag impls

---
 include/oneapi/dpl/pstl/algorithm_fwd.h       |  9 +++
 include/oneapi/dpl/pstl/algorithm_impl.h      | 40 +++++++++++++
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 58 +++++++++++++++++++
 3 files changed, 107 insertions(+)

diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index 1a9a270c288..49d4a9460e8 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -1542,11 +1542,20 @@ oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _Fo
 __pattern_min_element(_ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _Compare, _IsVector,
                       /* is_parallel = */ ::std::false_type) noexcept;
 
+template <class _Tag, typename _ExecutionPolicy, typename _ForwardIterator, typename _Compare>
+_ForwardIterator
+__pattern_min_element(_Tag, _ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _Compare) noexcept;
+
 template <typename _ExecutionPolicy, typename _RandomAccessIterator, typename _Compare, typename _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator>
 __pattern_min_element(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _Compare, _IsVector,
                       /* is_parallel = */ ::std::true_type);
 
+template <typename _IsVector, typename _ExecutionPolicy, typename _RandomAccessIterator, typename _Compare>
+_RandomAccessIterator
+__pattern_min_element(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator,
+                      _Compare);
+
 //------------------------------------------------------------------------
 // minmax_element
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 04b238611ec..ad1f1937b5f 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -4843,6 +4843,15 @@ __pattern_min_element(_ExecutionPolicy&&, _ForwardIterator __first, _ForwardIter
     return __internal::__brick_min_element(__first, __last, __comp, __is_vector);
 }
 
+template <class _Tag, typename _ExecutionPolicy, typename _ForwardIterator, typename _Compare>
+_ForwardIterator
+__pattern_min_element(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _Compare __comp) noexcept
+{
+    static_assert(__is_backend_tag_v<_Tag>);
+
+    return __internal::__brick_min_element(__first, __last, __comp, typename _Tag::__is_vector{});
+}
+
 template <typename _ExecutionPolicy, typename _RandomAccessIterator, typename _Compare, typename _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator>
 __pattern_min_element(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIterator __last,
@@ -4874,6 +4883,37 @@ __pattern_min_element(_ExecutionPolicy&& __exec, _RandomAccessIterator __first,
     });
 }
 
+template <typename _IsVector, typename _ExecutionPolicy, typename _RandomAccessIterator, typename _Compare>
+_RandomAccessIterator
+__pattern_min_element(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
+                      _RandomAccessIterator __last, _Compare __comp)
+{
+    // a trivial case pre-check
+    if (__last - __first < 2)
+        return __first;
+
+    return __internal::__except_handler([&]() {
+        return __par_backend::__parallel_reduce(
+            ::std::forward<_ExecutionPolicy>(__exec), __first, __last, /*identity*/ __last,
+            [=](_RandomAccessIterator __begin, _RandomAccessIterator __end,
+                _RandomAccessIterator __init) -> _RandomAccessIterator {
+                const _RandomAccessIterator __subresult =
+                    __internal::__brick_min_element(__begin, __end, __comp, _IsVector{});
+                return __init == __last ? __subresult
+                                        : __internal::__cmp_iterators_by_values(__init, __subresult, __comp,
+                                                                                oneapi::dpl::__internal::__pstl_less());
+            },
+            [=](_RandomAccessIterator __it1, _RandomAccessIterator __it2) -> _RandomAccessIterator {
+                if (__it1 == __last)
+                    return __it2;
+                if (__it2 == __last)
+                    return __it1;
+                return __internal::__cmp_iterators_by_values(__it1, __it2, __comp,
+                                                             oneapi::dpl::__internal::__pstl_less());
+            });
+    });
+}
+
 //------------------------------------------------------------------------
 // minmax_element
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 05a80007649..0472c13e8ed 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -706,6 +706,64 @@ __pattern_min_element(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __
     return __first + ::std::get<0>(__ret_idx);
 }
 
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator, typename _Compare>
+_Iterator
+__pattern_min_element(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last,
+                      _Compare __comp)
+{
+    if (__first == __last)
+        return __last;
+
+    using _IteratorValueType = typename ::std::iterator_traits<_Iterator>::value_type;
+    using _IndexValueType = ::std::make_unsigned_t<typename ::std::iterator_traits<_Iterator>::difference_type>;
+    using _ReduceValueType = tuple<_IndexValueType, _IteratorValueType>;
+    // Commutativity of the reduction operator depends on the compilation target (see __reduce_fn below);
+    // __spirv_target_conditional postpones deciding on commutativity to the device code where the
+    // target can be correctly tested.
+    using _Commutative = oneapi::dpl::__internal::__spirv_target_conditional</*_SpirvT*/ ::std::false_type,
+                                                                             /*_NonSpirvT*/ ::std::true_type>;
+    auto __reduce_fn = [__comp](_ReduceValueType __a, _ReduceValueType __b) {
+        using ::std::get;
+        // TODO: Consider removing the non-commutative operator for SPIR-V targets when we see improved performance with the
+        // non-sequential load path in transform_reduce.
+        if constexpr (oneapi::dpl::__internal::__is_spirv_target_v)
+        {
+            // This operator doesn't track the lowest found index in case of equal min. or max. values. Thus, this operator is
+            // not commutative.
+            if (__comp(get<1>(__b), get<1>(__a)))
+            {
+                return __b;
+            }
+            return __a;
+        }
+        else
+        {
+            // This operator keeps track of the lowest found index in case of equal min. or max. values. Thus, this operator is
+            // commutative.
+            bool _is_a_lt_b = __comp(get<1>(__a), get<1>(__b));
+            bool _is_b_lt_a = __comp(get<1>(__b), get<1>(__a));
+
+            if (_is_b_lt_a || (!_is_a_lt_b && get<0>(__b) < get<0>(__a)))
+            {
+                return __b;
+            }
+            return __a;
+        }
+    };
+    auto __transform_fn = [](auto __gidx, auto __acc) { return _ReduceValueType{__gidx, __acc[__gidx]}; };
+
+    auto __keep = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _Iterator>();
+    auto __buf = __keep(__first, __last);
+
+    auto __ret_idx = oneapi::dpl::__par_backend_hetero::__parallel_transform_reduce<_ReduceValueType, _Commutative>(
+                         ::std::forward<_ExecutionPolicy>(__exec), __reduce_fn, __transform_fn,
+                         unseq_backend::__no_init_value{}, // no initial value
+                         __buf.all_view())
+                         .get();
+
+    return __first + ::std::get<0>(__ret_idx);
+}
+
 // TODO:
 //   The following minmax_element implementation
 //   has at worst 2N applications of the predicate

From d7cc061ad64b43c61af8a61cdd2e31310370910d Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 6 Feb 2024 10:48:45 +0100
Subject: [PATCH 102/566] __pattern_min_element + tag calls

---
 include/oneapi/dpl/pstl/glue_algorithm_impl.h | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_algorithm_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
index 165bb4e9ab2..0ccdda7065f 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
@@ -1120,10 +1120,10 @@ template <class _ExecutionPolicy, class _ForwardIterator, class _Compare>
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _ForwardIterator>
 min_element(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last, _Compare __comp)
 {
-    return oneapi::dpl::__internal::__pattern_min_element(
-        ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __comp,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>());
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+
+    return oneapi::dpl::__internal::__pattern_min_element(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
+                                                          __first, __last, __comp);
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator>

From 7ed0def8d66632be23ef80ecda98dc4a98d4fbe4 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 6 Feb 2024 11:01:23 +0100
Subject: [PATCH 103/566] __pattern_minmax_element + tag impls

---
 include/oneapi/dpl/pstl/algorithm_fwd.h       |  9 ++++
 include/oneapi/dpl/pstl/algorithm_impl.h      | 51 +++++++++++++++++++
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 50 ++++++++++++++++++
 3 files changed, 110 insertions(+)

diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index 49d4a9460e8..9af3f6b2176 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -1575,12 +1575,21 @@ oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy,
 __pattern_minmax_element(_ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _Compare, _IsVector,
                          /* is_parallel = */ ::std::false_type) noexcept;
 
+template <class _Tag, typename _ExecutionPolicy, typename _ForwardIterator, typename _Compare>
+::std::pair<_ForwardIterator, _ForwardIterator>
+__pattern_minmax_element(_Tag, _ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _Compare) noexcept;
+
 template <typename _ExecutionPolicy, typename _RandomAccessIterator, typename _Compare, typename _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy,
                                                            ::std::pair<_RandomAccessIterator, _RandomAccessIterator>>
 __pattern_minmax_element(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _Compare, _IsVector,
                          /* is_parallel = */ ::std::true_type);
 
+template <typename _IsVector, typename _ExecutionPolicy, typename _RandomAccessIterator, typename _Compare>
+::std::pair<_RandomAccessIterator, _RandomAccessIterator>
+__pattern_minmax_element(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator,
+                         _Compare);
+
 //------------------------------------------------------------------------
 // mismatch
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index ad1f1937b5f..83efd6ffdc8 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -4947,6 +4947,16 @@ __pattern_minmax_element(_ExecutionPolicy&&, _ForwardIterator __first, _ForwardI
     return __internal::__brick_minmax_element(__first, __last, __comp, __is_vector);
 }
 
+template <class _Tag, typename _ExecutionPolicy, typename _ForwardIterator, typename _Compare>
+::std::pair<_ForwardIterator, _ForwardIterator>
+__pattern_minmax_element(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
+                         _Compare __comp) noexcept
+{
+    static_assert(__is_backend_tag_v<_Tag>);
+
+    return __internal::__brick_minmax_element(__first, __last, __comp, typename _Tag::__is_vector{});
+}
+
 template <typename _ExecutionPolicy, typename _RandomAccessIterator, typename _Compare, typename _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy,
                                                            ::std::pair<_RandomAccessIterator, _RandomAccessIterator>>
@@ -4989,6 +4999,47 @@ __pattern_minmax_element(_ExecutionPolicy&& __exec, _RandomAccessIterator __firs
     });
 }
 
+template <typename _IsVector, typename _ExecutionPolicy, typename _RandomAccessIterator, typename _Compare>
+::std::pair<_RandomAccessIterator, _RandomAccessIterator>
+__pattern_minmax_element(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
+                         _RandomAccessIterator __last, _Compare __comp)
+{
+    // a trivial case pre-check
+    if (__last - __first < 2)
+        return ::std::make_pair(__first, __first);
+
+    return __internal::__except_handler([&]() {
+        typedef ::std::pair<_RandomAccessIterator, _RandomAccessIterator> _Result;
+
+        return __par_backend::__parallel_reduce(
+            ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+            /*identity*/ ::std::make_pair(__last, __last),
+            [=, &__comp](_RandomAccessIterator __begin, _RandomAccessIterator __end, _Result __init) -> _Result {
+                const _Result __subresult = __internal::__brick_minmax_element(__begin, __end, __comp, _IsVector{});
+                if (__init.first == __last) // = identity
+                    return __subresult;
+                return ::std::make_pair(
+                    __internal::__cmp_iterators_by_values(__init.first, __subresult.first, __comp,
+                                                          oneapi::dpl::__internal::__pstl_less()),
+                    __internal::__cmp_iterators_by_values(__init.second, __subresult.second,
+                                                          oneapi::dpl::__internal::__reorder_pred<_Compare>(__comp),
+                                                          oneapi::dpl::__internal::__pstl_greater()));
+            },
+            [=, &__comp](_Result __p1, _Result __p2) -> _Result {
+                if (__p1.first == __last)
+                    return __p2;
+                if (__p2.first == __last)
+                    return __p1;
+                return ::std::make_pair(
+                    __internal::__cmp_iterators_by_values(__p1.first, __p2.first, __comp,
+                                                          oneapi::dpl::__internal::__pstl_less()),
+                    __internal::__cmp_iterators_by_values(__p1.second, __p2.second,
+                                                          oneapi::dpl::__internal::__reorder_pred<_Compare>(__comp),
+                                                          oneapi::dpl::__internal::__pstl_greater()));
+            });
+    });
+}
+
 //------------------------------------------------------------------------
 // mismatch
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 0472c13e8ed..23c9ef38e93 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -832,6 +832,56 @@ __pattern_minmax_element(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator
     return ::std::make_pair<_Iterator, _Iterator>(__first + ::std::get<0>(__ret), __first + ::std::get<1>(__ret));
 }
 
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator, typename _Compare>
+::std::pair<_Iterator, _Iterator>
+__pattern_minmax_element(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator __first,
+                         _Iterator __last, _Compare __comp)
+{
+    if (__first == __last)
+        return ::std::make_pair(__first, __first);
+
+    using _IteratorValueType = typename ::std::iterator_traits<_Iterator>::value_type;
+    using _IndexValueType = ::std::make_unsigned_t<typename ::std::iterator_traits<_Iterator>::difference_type>;
+    using _ReduceValueType = ::std::tuple<_IndexValueType, _IndexValueType, _IteratorValueType, _IteratorValueType>;
+
+    // This operator doesn't track the lowest found index in case of equal min. values and the highest found index in
+    // case of equal max. values. Thus, this operator is not commutative.
+    auto __reduce_fn = [__comp](_ReduceValueType __a, _ReduceValueType __b) {
+        using ::std::get;
+        auto __chosen_for_min = __a;
+        auto __chosen_for_max = __b;
+
+        assert(get<0>(__a) < get<0>(__b));
+        assert(get<1>(__a) < get<1>(__b));
+
+        if (__comp(get<2>(__b), get<2>(__a)))
+            __chosen_for_min = ::std::move(__b);
+        if (__comp(get<3>(__b), get<3>(__a)))
+            __chosen_for_max = ::std::move(__a);
+        return _ReduceValueType{get<0>(__chosen_for_min), get<1>(__chosen_for_max), get<2>(__chosen_for_min),
+                                get<3>(__chosen_for_max)};
+    };
+
+    // TODO: Doesn't work with `zip_iterator`.
+    //       In that case the first and the second arguments of `_ReduceValueType` will be
+    //       a `tuple` of `difference_type`, not the `difference_type` itself.
+    auto __transform_fn = [](auto __gidx, auto __acc) {
+        return _ReduceValueType{__gidx, __gidx, __acc[__gidx], __acc[__gidx]};
+    };
+
+    auto __keep = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _Iterator>();
+    auto __buf = __keep(__first, __last);
+
+    auto __ret = oneapi::dpl::__par_backend_hetero::__parallel_transform_reduce<_ReduceValueType,
+                                                                                ::std::false_type /*is_commutative*/>(
+                     ::std::forward<_ExecutionPolicy>(__exec), __reduce_fn, __transform_fn,
+                     unseq_backend::__no_init_value{}, // no initial value
+                     __buf.all_view())
+                     .get();
+
+    return ::std::make_pair<_Iterator, _Iterator>(__first + ::std::get<0>(__ret), __first + ::std::get<1>(__ret));
+}
+
 //------------------------------------------------------------------------
 // adjacent_find
 //------------------------------------------------------------------------

From 854b42cdb0239aa42efea7749d6791103981a3ad Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 6 Feb 2024 11:01:34 +0100
Subject: [PATCH 104/566] __pattern_minmax_element + tag calls

---
 include/oneapi/dpl/pstl/glue_algorithm_impl.h | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_algorithm_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
index 0ccdda7065f..1ed979c2e74 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
@@ -1155,10 +1155,10 @@ template <class _ExecutionPolicy, class _ForwardIterator, class _Compare>
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, ::std::pair<_ForwardIterator, _ForwardIterator>>
 minmax_element(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last, _Compare __comp)
 {
-    return oneapi::dpl::__internal::__pattern_minmax_element(
-        ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __comp,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>());
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+
+    return oneapi::dpl::__internal::__pattern_minmax_element(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
+                                                             __first, __last, __comp);
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator>

From 3543d130975fc5d00285549f3e4d6f57c3a9bac9 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 6 Feb 2024 11:15:16 +0100
Subject: [PATCH 105/566] __pattern_mismatch + tag impls

---
 include/oneapi/dpl/pstl/algorithm_fwd.h       | 11 +++++++
 include/oneapi/dpl/pstl/algorithm_impl.h      | 33 +++++++++++++++++++
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 21 ++++++++++++
 3 files changed, 65 insertions(+)

diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index 9af3f6b2176..54f4d61341e 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -1612,6 +1612,11 @@ __pattern_mismatch(_ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _Fo
                    _Predicate, _IsVector,
                    /* is_parallel = */ ::std::false_type) noexcept;
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Predicate>
+::std::pair<_ForwardIterator1, _ForwardIterator2>
+__pattern_mismatch(_Tag, _ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2, _ForwardIterator2,
+                   _Predicate) noexcept;
+
 template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _Predicate,
           class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy,
@@ -1619,6 +1624,12 @@ oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy,
 __pattern_mismatch(_ExecutionPolicy&&, _RandomAccessIterator1, _RandomAccessIterator1, _RandomAccessIterator2,
                    _RandomAccessIterator2, _Predicate, _IsVector, /* is_parallel = */ ::std::true_type);
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
+          class _Predicate>
+::std::pair<_RandomAccessIterator1, _RandomAccessIterator2>
+__pattern_mismatch(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator1, _RandomAccessIterator1,
+                   _RandomAccessIterator2, _RandomAccessIterator2, _Predicate);
+
 //------------------------------------------------------------------------
 // lexicographical_compare
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 83efd6ffdc8..bfe8f6ca0af 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -5078,6 +5078,16 @@ __pattern_mismatch(_ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardItera
     return __internal::__brick_mismatch(__first1, __last1, __first2, __last2, __pred, __is_vector);
 }
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Predicate>
+::std::pair<_ForwardIterator1, _ForwardIterator2>
+__pattern_mismatch(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
+                   _ForwardIterator2 __first2, _ForwardIterator2 __last2, _Predicate __pred) noexcept
+{
+    static_assert(__is_backend_tag_v<_Tag>);
+
+    return __internal::__brick_mismatch(__first1, __last1, __first2, __last2, __pred, typename _Tag::__is_vector{});
+}
+
 template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _Predicate,
           class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy,
@@ -5100,6 +5110,29 @@ __pattern_mismatch(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1, _
     });
 }
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
+          class _Predicate>
+::std::pair<_RandomAccessIterator1, _RandomAccessIterator2>
+__pattern_mismatch(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1,
+                   _RandomAccessIterator1 __last1, _RandomAccessIterator2 __first2, _RandomAccessIterator2 __last2,
+                   _Predicate __pred)
+{
+    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
+
+    return __internal::__except_handler([&]() {
+        auto __n = ::std::min(__last1 - __first1, __last2 - __first2);
+        auto __result = __internal::__parallel_find(
+            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first1, __first1 + __n,
+            [__first1, __first2, __pred](_RandomAccessIterator1 __i, _RandomAccessIterator1 __j) {
+                return __internal::__brick_mismatch(__i, __j, __first2 + (__i - __first1), __first2 + (__j - __first1),
+                                                    __pred, _IsVector{})
+                    .first;
+            },
+            ::std::true_type{});
+        return ::std::make_pair(__result, __first2 + (__result - __first1));
+    });
+}
+
 //------------------------------------------------------------------------
 // lexicographical_compare
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 23c9ef38e93..4cd97a19d20 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -1429,6 +1429,27 @@ __pattern_mismatch(_ExecutionPolicy&& __exec, _Iterator1 __first1, _Iterator1 __
     return ::std::make_pair(__first1 + __n, __first2 + __n);
 }
 
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2, typename _Pred>
+::std::pair<_Iterator1, _Iterator2>
+__pattern_mismatch(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator1 __first1, _Iterator1 __last1,
+                   _Iterator2 __first2, _Iterator2 __last2, _Pred __pred)
+{
+    auto __n = ::std::min(__last1 - __first1, __last2 - __first2);
+    if (__n <= 0)
+        return ::std::make_pair(__first1, __first2);
+
+    using _Predicate = oneapi::dpl::unseq_backend::single_match_pred<_ExecutionPolicy, equal_predicate<_Pred>>;
+
+    auto __first_zip = __par_backend_hetero::zip(
+        __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__first1),
+        __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__first2));
+    auto __result =
+        __par_backend_hetero::__parallel_find(::std::forward<_ExecutionPolicy>(__exec), __first_zip, __first_zip + __n,
+                                              _Predicate{equal_predicate<_Pred>{__pred}}, ::std::true_type{});
+    __n = __result - __first_zip;
+    return ::std::make_pair(__first1 + __n, __first2 + __n);
+}
+
 //------------------------------------------------------------------------
 // copy_if
 //------------------------------------------------------------------------

From 4bf5073ade730a3e90a59a3a04982cd411d01e34 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 6 Feb 2024 11:15:28 +0100
Subject: [PATCH 106/566] __pattern_mismatch + tag calls

---
 include/oneapi/dpl/pstl/glue_algorithm_impl.h | 9 +++++----
 1 file changed, 5 insertions(+), 4 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_algorithm_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
index 1ed979c2e74..2173b90d8c6 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
@@ -734,10 +734,11 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy,
 mismatch(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 __last1, _ForwardIterator2 __first2,
          _ForwardIterator2 __last2, _BinaryPredicate __pred)
 {
-    return oneapi::dpl::__internal::__pattern_mismatch(
-        ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2, __pred,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>());
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
+
+    return oneapi::dpl::__internal::__pattern_mismatch(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
+                                                       __first1, __last1, __first2, __last2, __pred);
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _BinaryPredicate>

From 70a4146d59ae3c13206da3a66a1a1e9a5ecd9f52 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 6 Feb 2024 12:20:38 +0100
Subject: [PATCH 107/566] __pattern_nth_element + tag impls

---
 include/oneapi/dpl/pstl/algorithm_fwd.h       | 10 ++++
 include/oneapi/dpl/pstl/algorithm_impl.h      | 53 +++++++++++++++++++
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 15 ++++++
 3 files changed, 78 insertions(+)

diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index 54f4d61341e..f45f54585d4 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -1124,12 +1124,22 @@ __pattern_nth_element(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIt
                       _IsVector,
                       /*is_parallel=*/::std::false_type) noexcept;
 
+template <class _Tag, class _ExecutionPolicy, class _RandomAccessIterator, class _Compare>
+void
+__pattern_nth_element(_Tag, _ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _RandomAccessIterator,
+                      _Compare) noexcept;
+
 template <class _ExecutionPolicy, class _RandomAccessIterator, class _Compare, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>
 __pattern_nth_element(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _RandomAccessIterator, _Compare,
                       _IsVector,
                       /*is_parallel=*/::std::true_type);
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Compare>
+void
+__pattern_nth_element(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator,
+                      _RandomAccessIterator, _Compare);
+
 //------------------------------------------------------------------------
 // fill, fill_n
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index bfe8f6ca0af..b96eb4da63a 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -3406,6 +3406,16 @@ __pattern_nth_element(_ExecutionPolicy&&, _RandomAccessIterator __first, _Random
     ::std::nth_element(__first, __nth, __last, __comp);
 }
 
+template <class _Tag, class _ExecutionPolicy, class _RandomAccessIterator, class _Compare>
+void
+__pattern_nth_element(_Tag, _ExecutionPolicy&&, _RandomAccessIterator __first, _RandomAccessIterator __nth,
+                      _RandomAccessIterator __last, _Compare __comp) noexcept
+{
+    static_assert(__is_backend_tag_v<_Tag>);
+
+    ::std::nth_element(__first, __nth, __last, __comp);
+}
+
 template <class _ExecutionPolicy, class _RandomAccessIterator, class _Compare, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>
 __pattern_nth_element(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIterator __nth,
@@ -3450,6 +3460,49 @@ __pattern_nth_element(_ExecutionPolicy&& __exec, _RandomAccessIterator __first,
     } while (__x != __nth);
 }
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Compare>
+void
+__pattern_nth_element(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
+                      _RandomAccessIterator __nth, _RandomAccessIterator __last, _Compare __comp)
+{
+    if (__first == __last || __nth == __last)
+    {
+        return;
+    }
+
+    using ::std::iter_swap;
+    typedef typename ::std::iterator_traits<_RandomAccessIterator>::value_type _Tp;
+    _RandomAccessIterator __x;
+    do
+    {
+        __x = __internal::__pattern_partition(
+            ::std::forward<_ExecutionPolicy>(__exec), __first + 1, __last,
+            [&__comp, __first](const _Tp& __x) { return __comp(__x, *__first); }, _IsVector{},
+            /*is_parallel=*/::std::true_type());
+        --__x;
+        if (__x != __first)
+        {
+            iter_swap(__first, __x);
+        }
+        // if x > nth then our new range for partition is [first, x)
+        if (__x - __nth > 0)
+        {
+            __last = __x;
+        }
+        // if x < nth then our new range for partition is [x, last)
+        else if (__x - __nth < 0)
+        {
+            // if *x == *nth then we start the new partition at the next index where *x != *nth
+            while (!__comp(*__nth, *__x) && !__comp(*__x, *__nth) && __x - __nth < 0)
+            {
+                ++__x;
+            }
+            iter_swap(__nth, __x);
+            __first = __x;
+        }
+    } while (__x != __nth);
+}
+
 //------------------------------------------------------------------------
 // fill, fill_n
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 4cd97a19d20..10908ebd883 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -2359,6 +2359,21 @@ __pattern_nth_element(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __
                            /*vector*/ ::std::true_type{}, /*parallel*/ ::std::true_type{});
 }
 
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator, typename _Compare>
+void
+__pattern_nth_element(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __nth,
+                      _Iterator __last, _Compare __comp)
+{
+    if (__first == __last || __nth == __last)
+        return;
+
+    // TODO: check partition-based implementation
+    // - try to avoid host dereference issue
+    // - measure performance of the issue-free implementation
+    __pattern_partial_sort(::std::forward<_ExecutionPolicy>(__exec), __first, __nth + 1, __last, __comp,
+                           /*vector*/ ::std::true_type{}, /*parallel*/ ::std::true_type{});
+}
+
 //------------------------------------------------------------------------
 // reverse
 //------------------------------------------------------------------------

From 442a9b91479cf2441c06a77c746dc47d86f89c3e Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 6 Feb 2024 12:20:52 +0100
Subject: [PATCH 108/566] __pattern_nth_element + tag calls

---
 include/oneapi/dpl/pstl/glue_algorithm_impl.h | 9 +++++----
 1 file changed, 5 insertions(+), 4 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_algorithm_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
index 2173b90d8c6..3051269d7f0 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
@@ -1177,10 +1177,11 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy>
 nth_element(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIterator __nth,
             _RandomAccessIterator __last, _Compare __comp)
 {
-    oneapi::dpl::__internal::__pattern_nth_element(
-        ::std::forward<_ExecutionPolicy>(__exec), __first, __nth, __last, __comp,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _RandomAccessIterator>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _RandomAccessIterator>());
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
+
+    oneapi::dpl::__internal::__pattern_nth_element(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first,
+                                                   __nth, __last, __comp);
 }
 
 template <class _ExecutionPolicy, class _RandomAccessIterator>

From 370bf3ee7d22e6deb97e8c85462485bbb27c21f3 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 6 Feb 2024 12:35:03 +0100
Subject: [PATCH 109/566] __pattern_partial_sort + tag impls

---
 include/oneapi/dpl/pstl/algorithm_fwd.h       | 10 ++++++
 include/oneapi/dpl/pstl/algorithm_impl.h      | 32 +++++++++++++++++++
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 16 ++++++++++
 3 files changed, 58 insertions(+)

diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index f45f54585d4..762fa0b813d 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -1057,12 +1057,22 @@ __pattern_partial_sort(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessI
                        _Compare, _IsVector,
                        /*is_parallel=*/::std::false_type) noexcept;
 
+template <class _Tag, class _ExecutionPolicy, class _RandomAccessIterator, class _Compare>
+void
+__pattern_partial_sort(_Tag, _ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _RandomAccessIterator,
+                       _Compare) noexcept;
+
 template <class _ExecutionPolicy, class _RandomAccessIterator, class _Compare, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>
 __pattern_partial_sort(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _RandomAccessIterator,
                        _Compare, _IsVector,
                        /*is_parallel=*/::std::true_type);
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Compare>
+void
+__pattern_partial_sort(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator,
+                       _RandomAccessIterator, _Compare);
+
 //------------------------------------------------------------------------
 // partial_sort_copy
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index b96eb4da63a..9a274d7f3f3 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -3158,6 +3158,16 @@ __pattern_partial_sort(_ExecutionPolicy&&, _RandomAccessIterator __first, _Rando
     ::std::partial_sort(__first, __middle, __last, __comp);
 }
 
+template <class _Tag, class _ExecutionPolicy, class _RandomAccessIterator, class _Compare>
+void
+__pattern_partial_sort(_Tag, _ExecutionPolicy&&, _RandomAccessIterator __first, _RandomAccessIterator __middle,
+                       _RandomAccessIterator __last, _Compare __comp) noexcept
+{
+    static_assert(__is_backend_tag_v<_Tag>);
+
+    ::std::partial_sort(__first, __middle, __last, __comp);
+}
+
 template <class _ExecutionPolicy, class _RandomAccessIterator, class _Compare, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>
 __pattern_partial_sort(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIterator __middle,
@@ -3180,6 +3190,28 @@ __pattern_partial_sort(_ExecutionPolicy&& __exec, _RandomAccessIterator __first,
     });
 }
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Compare>
+void
+__pattern_partial_sort(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
+                       _RandomAccessIterator __middle, _RandomAccessIterator __last, _Compare __comp)
+{
+    const auto __n = __middle - __first;
+    if (__n == 0)
+        return;
+
+    __except_handler([&]() {
+        __par_backend::__parallel_stable_sort(
+            ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __comp,
+            [__n](_RandomAccessIterator __begin, _RandomAccessIterator __end, _Compare __comp) {
+                if (__n < __end - __begin)
+                    ::std::partial_sort(__begin, __begin + __n, __end, __comp);
+                else
+                    ::std::sort(__begin, __end, __comp);
+            },
+            __n);
+    });
+}
+
 //------------------------------------------------------------------------
 // partial_sort_copy
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 10908ebd883..5ad1eda0ece 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -2251,6 +2251,22 @@ __pattern_partial_sort(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator _
         .wait();
 }
 
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator, typename _Compare>
+oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy>
+__pattern_partial_sort(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __mid,
+                       _Iterator __last, _Compare __comp)
+{
+    if (__last - __first < 2)
+        return;
+
+    __par_backend_hetero::__parallel_partial_sort(
+        ::std::forward<_ExecutionPolicy>(__exec),
+        __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read_write>(__first),
+        __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read_write>(__mid),
+        __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read_write>(__last), __comp)
+        .wait();
+}
+
 //------------------------------------------------------------------------
 // partial_sort_copy
 //------------------------------------------------------------------------

From 2500d1e690eccca36b8d39e1dde7e28c390c29de Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 6 Feb 2024 12:35:16 +0100
Subject: [PATCH 110/566] __pattern_partial_sort + tag calls

---
 include/oneapi/dpl/pstl/glue_algorithm_impl.h          | 9 +++++----
 include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h | 8 ++++----
 2 files changed, 9 insertions(+), 8 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_algorithm_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
index 3051269d7f0..20e38400587 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
@@ -838,10 +838,11 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy>
 partial_sort(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIterator __middle,
              _RandomAccessIterator __last, _Compare __comp)
 {
-    oneapi::dpl::__internal::__pattern_partial_sort(
-        ::std::forward<_ExecutionPolicy>(__exec), __first, __middle, __last, __comp,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _RandomAccessIterator>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _RandomAccessIterator>());
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
+
+    oneapi::dpl::__internal::__pattern_partial_sort(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first,
+                                                    __middle, __last, __comp);
 }
 
 template <class _ExecutionPolicy, class _RandomAccessIterator>
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 5ad1eda0ece..56df7e4a06c 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -2368,11 +2368,12 @@ __pattern_nth_element(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __
     if (__first == __last || __nth == __last)
         return;
 
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator>();
+
     // TODO: check partition-based implementation
     // - try to avoid host dereference issue
     // - measure performance of the issue-free implementation
-    __pattern_partial_sort(::std::forward<_ExecutionPolicy>(__exec), __first, __nth + 1, __last, __comp,
-                           /*vector*/ ::std::true_type{}, /*parallel*/ ::std::true_type{});
+    __pattern_partial_sort(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __nth + 1, __last, __comp);
 }
 
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator, typename _Compare>
@@ -2386,8 +2387,7 @@ __pattern_nth_element(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec
     // TODO: check partition-based implementation
     // - try to avoid host dereference issue
     // - measure performance of the issue-free implementation
-    __pattern_partial_sort(::std::forward<_ExecutionPolicy>(__exec), __first, __nth + 1, __last, __comp,
-                           /*vector*/ ::std::true_type{}, /*parallel*/ ::std::true_type{});
+    __pattern_partial_sort(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __nth + 1, __last, __comp);
 }
 
 //------------------------------------------------------------------------

From b415521fe0c6c806a5e8535e65e04f27708ae8f8 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 6 Feb 2024 13:26:18 +0100
Subject: [PATCH 111/566] __pattern_partial_sort_copy + tag impls

---
 include/oneapi/dpl/pstl/algorithm_fwd.h       |  10 ++
 include/oneapi/dpl/pstl/algorithm_impl.h      |  85 ++++++++++++++
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 106 ++++++++++++++++--
 3 files changed, 193 insertions(+), 8 deletions(-)

diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index 762fa0b813d..97e8a1b3de3 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -1084,6 +1084,11 @@ __pattern_partial_sort_copy(_ExecutionPolicy&&, _RandomAccessIterator1, _RandomA
                             _RandomAccessIterator2, _Compare, _IsVector,
                             /*is_parallel=*/::std::false_type) noexcept;
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _RandomAccessIterator, class _Compare>
+_RandomAccessIterator
+__pattern_partial_sort_copy(_Tag, _ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _RandomAccessIterator,
+                            _RandomAccessIterator, _Compare) noexcept;
+
 template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _Compare,
           class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator2>
@@ -1091,6 +1096,11 @@ __pattern_partial_sort_copy(_ExecutionPolicy&&, _RandomAccessIterator1, _RandomA
                             _RandomAccessIterator2, _Compare, _IsVector,
                             /*is_parallel=*/::std::true_type);
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _RandomAccessIterator, class _Compare>
+_RandomAccessIterator
+__pattern_partial_sort_copy(_Tag, _ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _RandomAccessIterator,
+                            _RandomAccessIterator, _Compare) noexcept;
+
 //------------------------------------------------------------------------
 // adjacent_find
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 9a274d7f3f3..20793e4f2c4 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -3225,6 +3225,16 @@ __pattern_partial_sort_copy(_ExecutionPolicy&&, _ForwardIterator __first, _Forwa
     return ::std::partial_sort_copy(__first, __last, __d_first, __d_last, __comp);
 }
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _RandomAccessIterator, class _Compare>
+_RandomAccessIterator
+__pattern_partial_sort_copy(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
+                            _RandomAccessIterator __d_first, _RandomAccessIterator __d_last, _Compare __comp) noexcept
+{
+    static_assert(__is_backend_tag_v<_Tag>);
+
+    return ::std::partial_sort_copy(__first, __last, __d_first, __d_last, __comp);
+}
+
 template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _Compare,
           class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator2>
@@ -3299,6 +3309,81 @@ __pattern_partial_sort_copy(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __
     });
 }
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
+          class _Compare>
+_RandomAccessIterator2
+__pattern_partial_sort_copy(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator1 __first,
+                            _RandomAccessIterator1 __last, _RandomAccessIterator2 __d_first,
+                            _RandomAccessIterator2 __d_last, _Compare __comp)
+{
+    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
+
+    if (__last == __first || __d_last == __d_first)
+    {
+        return __d_first;
+    }
+    auto __n1 = __last - __first;
+    auto __n2 = __d_last - __d_first;
+    return __internal::__except_handler([&]() {
+        if (__n2 >= __n1)
+        {
+            __par_backend::__parallel_stable_sort(
+                ::std::forward<_ExecutionPolicy>(__exec), __d_first, __d_first + __n1, __comp,
+                [__first, __d_first](_RandomAccessIterator2 __i, _RandomAccessIterator2 __j, _Compare __comp) {
+                    _RandomAccessIterator1 __i1 = __first + (__i - __d_first);
+                    _RandomAccessIterator1 __j1 = __first + (__j - __d_first);
+
+                    // 1. Copy elements from input to output
+                    __brick_copy<_ExecutionPolicy>{}(__i1, __j1, __i, _IsVector{});
+                    // 2. Sort elements in output sequence
+                    ::std::sort(__i, __j, __comp);
+                },
+                __n1);
+            return __d_first + __n1;
+        }
+        else
+        {
+            typedef typename ::std::iterator_traits<_RandomAccessIterator1>::value_type _T1;
+            typedef typename ::std::iterator_traits<_RandomAccessIterator2>::value_type _T2;
+            __par_backend::__buffer<_ExecutionPolicy, _T1> __buf(__n1);
+            _T1* __r = __buf.get();
+
+            __par_backend::__parallel_stable_sort(
+                ::std::forward<_ExecutionPolicy>(__exec), __r, __r + __n1, __comp,
+                [__n2, __first, __r](_T1* __i, _T1* __j, _Compare __comp) {
+                    _RandomAccessIterator1 __it = __first + (__i - __r);
+
+                    // 1. Copy elements from input to raw memory
+                    for (_T1* __k = __i; __k != __j; ++__k, ++__it)
+                    {
+                        ::new (__k) _T2(*__it);
+                    }
+
+                    // 2. Sort elements in temporary buffer
+                    if (__n2 < __j - __i)
+                        ::std::partial_sort(__i, __i + __n2, __j, __comp);
+                    else
+                        ::std::sort(__i, __j, __comp);
+                },
+                __n2);
+
+            // 3. Move elements from temporary buffer to output
+            __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __r, __r + __n2,
+                                          [__r, __d_first](_T1* __i, _T1* __j) {
+                                              __brick_move_destroy<_ExecutionPolicy>{}(
+                                                  __i, __j, __d_first + (__i - __r), _IsVector{});
+                                          });
+
+            if constexpr (!::std::is_trivially_destructible_v<_T1>)
+                __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __r + __n2,
+                                              __r + __n1,
+                                              [](_T1* __i, _T1* __j) { __brick_destroy(__i, __j, _IsVector{}); });
+
+            return __d_first + __n2;
+        }
+    });
+}
+
 //------------------------------------------------------------------------
 // adjacent_find
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 56df7e4a06c..b4631412785 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -2315,16 +2315,26 @@ __pattern_partial_sort_copy(_ExecutionPolicy&& __exec, _InIterator __first, _InI
     // as it uses a similar mechanism.
     if (__in_size <= __out_size)
     {
+        constexpr auto __dispatch_tag1 =
+            oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, decltype(__first), decltype(__last),
+                                                      decltype(__out_first)>();
+        using __backend_tag1 = typename decltype(__dispatch_tag1)::__backend_tag;
+
         // If our output buffer is larger than the input buffer, simply copy elements to the output and use
         // full sort on them.
-        auto __out_end = __pattern_walk2</*_IsSync=*/::std::false_type>(
-            __par_backend_hetero::make_wrapped_policy<__initial_copy_1>(__exec), __first, __last, __out_first,
-            __brick_copy<_ExecutionPolicy>{}, ::std::true_type{}, ::std::true_type{});
+        auto __out_end = __pattern_walk2<__backend_tag1, /*_IsSync=*/::std::false_type>(
+            __dispatch_tag1, __par_backend_hetero::make_wrapped_policy<__initial_copy_1>(__exec), __first, __last,
+            __out_first, __brick_copy<_ExecutionPolicy>{});
+
+        constexpr auto __dispatch_tag2 =
+            oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, decltype(__out_first), decltype(__out_end),
+                                                      decltype(__out_first)>();
 
         // Use reqular sort as partial_sort isn't required to be stable
         __pattern_sort(
+            __dispatch_tag2,
             __par_backend_hetero::make_wrapped_policy<__partial_sort_1>(::std::forward<_ExecutionPolicy>(__exec)),
-            __out_first, __out_end, __comp, ::std::true_type{}, ::std::true_type{}, ::std::true_type{});
+            __out_first, __out_end, __comp, ::std::true_type{});
 
         return __out_end;
     }
@@ -2337,9 +2347,85 @@ __pattern_partial_sort_copy(_ExecutionPolicy&& __exec, _InIterator __first, _InI
         oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __buf(__exec, __in_size);
 
         auto __buf_first = __buf.get();
-        auto __buf_last = __pattern_walk2</*_IsSync=*/::std::false_type>(
+
+        constexpr auto __dispatch_tag1 =
+            oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, decltype(__first), decltype(__last),
+                                                      decltype(__buf_first)>();
+        using __backend_tag1 = typename decltype(__dispatch_tag1)::__backend_tag;
+
+        auto __buf_last = __pattern_walk2<__backend_tag1, /*_IsSync=*/::std::false_type>(
             __par_backend_hetero::make_wrapped_policy<__initial_copy_2>(__exec), __first, __last, __buf_first,
-            __brick_copy<_ExecutionPolicy>{}, ::std::true_type{}, ::std::true_type{});
+            __brick_copy<_ExecutionPolicy>{});
+
+        auto __buf_mid = __buf_first + __out_size;
+
+        __par_backend_hetero::__parallel_partial_sort(
+            __par_backend_hetero::make_wrapped_policy<__partial_sort_2>(__exec),
+            __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read_write>(__buf_first),
+            __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read_write>(__buf_mid),
+            __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read_write>(__buf_last), __comp);
+
+        constexpr auto __dispatch_tag2 =
+            oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, decltype(__buf_first), decltype(__buf_mid),
+                                                      decltype(__out_first)>();
+
+        return __pattern_walk2(
+            __dispatch_tag2,
+            __par_backend_hetero::make_wrapped_policy<__copy_back>(::std::forward<_ExecutionPolicy>(__exec)),
+            __buf_first, __buf_mid, __out_first, __brick_copy<_ExecutionPolicy>{});
+    }
+}
+
+template <typename _BackendTag, typename _ExecutionPolicy, typename _InIterator, typename _OutIterator,
+          typename _Compare>
+_OutIterator
+__pattern_partial_sort_copy(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _InIterator __first,
+                            _InIterator __last, _OutIterator __out_first, _OutIterator __out_last, _Compare __comp)
+{
+    using _ValueType = typename ::std::iterator_traits<_InIterator>::value_type;
+
+    auto __in_size = __last - __first;
+    auto __out_size = __out_last - __out_first;
+
+    if (__in_size == 0 || __out_size == 0)
+        return __out_first;
+
+    // TODO: we can avoid a separate __pattern_walk2 for initial copy: it can be done during sort itself
+    // like it's done for CPU version, but it's better to be done together with merge cutoff implementation
+    // as it uses a similar mechanism.
+    if (__in_size <= __out_size)
+    {
+        // If our output buffer is larger than the input buffer, simply copy elements to the output and use
+        // full sort on them.
+        auto __out_end = __pattern_walk2<_BackendTag, /*_IsSync=*/::std::false_type>(
+            __tag, __par_backend_hetero::make_wrapped_policy<__initial_copy_1>(__exec), __first, __last, __out_first,
+            __brick_copy<_ExecutionPolicy>{});
+
+        // Use regular sort as partial_sort isn't required to be stable
+        __pattern_sort(
+            __par_backend_hetero::make_wrapped_policy<__partial_sort_1>(::std::forward<_ExecutionPolicy>(__exec)),
+            __out_first, __out_end, __comp, ::std::true_type{}, ::std::true_type{}, ::std::true_type{});
+
+        return __out_end;
+    }
+    else
+    {
+        // If our input buffer is smaller than the input buffer do the following:
+        // - create a temporary buffer and copy all the elements from the input buffer there
+        // - run partial sort on the temporary buffer
+        // - copy k elements from the temporary buffer to the output buffer.
+        oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __buf(__exec, __in_size);
+
+        auto __buf_first = __buf.get();
+
+        constexpr auto __dispatch_tag1 =
+            oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, decltype(__first), decltype(__last),
+                                                      decltype(__buf_first)>();
+        using __backend_tag1 = typename decltype(__dispatch_tag1)::__backend_tag;
+
+        auto __buf_last = __pattern_walk2<__backend_tag1, /*_IsSync=*/::std::false_type>(
+            __dispatch_tag1, __par_backend_hetero::make_wrapped_policy<__initial_copy_2>(__exec), __first, __last,
+            __buf_first, __brick_copy<_ExecutionPolicy>{});
 
         auto __buf_mid = __buf_first + __out_size;
 
@@ -2349,10 +2435,14 @@ __pattern_partial_sort_copy(_ExecutionPolicy&& __exec, _InIterator __first, _InI
             __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read_write>(__buf_mid),
             __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read_write>(__buf_last), __comp);
 
+        constexpr auto __dispatch_tag2 =
+            oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, decltype(__buf_first), decltype(__buf_mid),
+                                                      decltype(__out_first)>();
+
         return __pattern_walk2(
+            __dispatch_tag2,
             __par_backend_hetero::make_wrapped_policy<__copy_back>(::std::forward<_ExecutionPolicy>(__exec)),
-            __buf_first, __buf_mid, __out_first, __brick_copy<_ExecutionPolicy>{}, ::std::true_type{},
-            ::std::true_type{});
+            __buf_first, __buf_mid, __out_first, __brick_copy<_ExecutionPolicy>{});
     }
 }
 

From 3c14631c6dfed1a8341dddeaa7ef5a7dd0507abc Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 6 Feb 2024 13:37:15 +0100
Subject: [PATCH 112/566] __pattern_partial_sort_copy + tag calls

---
 include/oneapi/dpl/pstl/glue_algorithm_impl.h | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_algorithm_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
index 20e38400587..5a96ceb30fe 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
@@ -861,10 +861,11 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _RandomA
 partial_sort_copy(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last,
                   _RandomAccessIterator __d_first, _RandomAccessIterator __d_last, _Compare __comp)
 {
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator, _RandomAccessIterator>();
+
     return oneapi::dpl::__internal::__pattern_partial_sort_copy(
-        ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __d_first, __d_last, __comp,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator, _RandomAccessIterator>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator, _RandomAccessIterator>());
+        __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __d_first, __d_last, __comp);
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator, class _RandomAccessIterator>

From 04df16cc91e7ebd6e15caafc05922294bd5b6e6c Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 6 Feb 2024 15:23:27 +0100
Subject: [PATCH 113/566] __pattern_partition + tag impls

---
 include/oneapi/dpl/pstl/algorithm_fwd.h       |  9 ++
 include/oneapi/dpl/pstl/algorithm_impl.h      | 85 +++++++++++++++++++
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 10 +++
 3 files changed, 104 insertions(+)

diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index 97e8a1b3de3..2049400a571 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -939,11 +939,20 @@ oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _Fo
 __pattern_partition(_ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _UnaryPredicate, _IsVector,
                     /*is_parallel=*/::std::false_type) noexcept;
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _UnaryPredicate>
+_ForwardIterator
+__pattern_partition(_Tag, _ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _UnaryPredicate) noexcept;
+
 template <class _ExecutionPolicy, class _RandomAccessIterator, class _UnaryPredicate, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator>
 __pattern_partition(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _UnaryPredicate, _IsVector,
                     /*is_parallel=*/::std::true_type);
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _UnaryPredicate>
+_RandomAccessIterator
+__pattern_partition(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator,
+                    _UnaryPredicate);
+
 //------------------------------------------------------------------------
 // stable_partition
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 20793e4f2c4..d8531ae1570 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -2818,6 +2818,16 @@ __pattern_partition(_ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterat
     return __internal::__brick_partition(__first, __last, __pred, __is_vector);
 }
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _UnaryPredicate>
+_ForwardIterator
+__pattern_partition(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
+                    _UnaryPredicate __pred) noexcept
+{
+    static_assert(__is_backend_tag_v<_Tag>);
+
+    return __internal::__brick_partition(__first, __last, __pred, typename _Tag::__is_vector{});
+}
+
 template <class _ExecutionPolicy, class _RandomAccessIterator, class _UnaryPredicate, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator>
 __pattern_partition(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIterator __last,
@@ -2886,6 +2896,81 @@ __pattern_partition(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _R
     });
 }
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _UnaryPredicate>
+_RandomAccessIterator
+__pattern_partition(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
+                    _RandomAccessIterator __last, _UnaryPredicate __pred)
+{
+    // partitioned range: elements before pivot satisfy pred (true part),
+    //                    elements after pivot don't satisfy pred (false part)
+    struct _PartitionRange
+    {
+        _RandomAccessIterator __begin;
+        _RandomAccessIterator __pivot;
+        _RandomAccessIterator __end;
+    };
+
+    return __internal::__except_handler([&]() {
+        _PartitionRange __init{__last, __last, __last};
+
+        // lambda for merging two partitioned ranges to one partitioned range
+        auto __reductor = [&__exec](_PartitionRange __val1, _PartitionRange __val2) -> _PartitionRange {
+            auto __size1 = __val1.__end - __val1.__pivot;
+            auto __size2 = __val2.__pivot - __val2.__begin;
+            auto __new_begin = __val2.__begin - (__val1.__end - __val1.__begin);
+
+            // if all elements in left range satisfy pred then we can move new pivot to pivot of right range
+            if (__val1.__end == __val1.__pivot)
+            {
+                return {__new_begin, __val2.__pivot, __val2.__end};
+            }
+            // if true part of right range greater than false part of left range
+            // then we should swap the false part of left range and last part of true part of right range
+            else if (__size2 > __size1)
+            {
+                constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, decltype(__val1.__pivot), decltype(__val1.__pivot + __size1)>();
+                using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
+                __par_backend::__parallel_for(
+                    __backend_tag{},
+                    ::std::forward<_ExecutionPolicy>(__exec), __val1.__pivot, __val1.__pivot + __size1,
+                    [__val1, __val2, __size1](_RandomAccessIterator __i, _RandomAccessIterator __j) {
+                        __internal::__brick_swap_ranges(__i, __j, (__val2.__pivot - __size1) + (__i - __val1.__pivot),
+                                                        _IsVector{});
+                    });
+                return {__new_begin, __val2.__pivot - __size1, __val2.__end};
+            }
+            // else we should swap the first part of false part of left range and true part of right range
+            else
+            {
+                constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, decltype(__val1.__pivot), decltype(__val1.__pivot + __size2)>();
+                using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
+                __par_backend::__parallel_for(
+                    __backend_tag{},
+                    ::std::forward<_ExecutionPolicy>(__exec), __val1.__pivot, __val1.__pivot + __size2,
+                    [__val1, __val2](_RandomAccessIterator __i, _RandomAccessIterator __j) {
+                        __internal::__brick_swap_ranges(__i, __j, __val2.__begin + (__i - __val1.__pivot), _IsVector{});
+                    });
+                return {__new_begin, __val1.__pivot + __size2, __val2.__end};
+            }
+        };
+
+        _PartitionRange __result = __par_backend::__parallel_reduce(
+            ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __init,
+            [__pred, __reductor](_RandomAccessIterator __i, _RandomAccessIterator __j,
+                                 _PartitionRange __value) -> _PartitionRange {
+                //1. serial partition
+                _RandomAccessIterator __pivot = __internal::__brick_partition(__i, __j, __pred, _IsVector{});
+
+                // 2. merging of two ranges (left and right respectively)
+                return __reductor(__value, {__i, __pivot, __j});
+            },
+            __reductor);
+        return __result.__pivot;
+    });
+}
+
 //------------------------------------------------------------------------
 // stable_partition
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index b4631412785..07855e3e796 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -2075,6 +2075,16 @@ __pattern_partition(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __la
                                       ::std::true_type(), ::std::true_type());
 }
 
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator, typename _UnaryPredicate>
+_Iterator
+__pattern_partition(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last,
+                    _UnaryPredicate __pred)
+{
+    //TODO: consider nonstable approaches
+    return __pattern_stable_partition(::std::forward<_ExecutionPolicy>(__exec), __first, __last, __pred,
+                                      ::std::true_type(), ::std::true_type());
+}
+
 //------------------------------------------------------------------------
 // lexicographical_compare
 //------------------------------------------------------------------------

From 973e89c52e19c58121c1cca7615be23562af2c4a Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 6 Feb 2024 15:24:44 +0100
Subject: [PATCH 114/566] __pattern_partition + tag calls

---
 include/oneapi/dpl/pstl/algorithm_impl.h      | 15 +++++++--------
 include/oneapi/dpl/pstl/glue_algorithm_impl.h |  8 ++++----
 2 files changed, 11 insertions(+), 12 deletions(-)

diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index d8531ae1570..629029f0a77 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -3629,15 +3629,16 @@ __pattern_nth_element(_ExecutionPolicy&& __exec, _RandomAccessIterator __first,
         return;
     }
 
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
+
     using ::std::iter_swap;
     typedef typename ::std::iterator_traits<_RandomAccessIterator>::value_type _Tp;
     _RandomAccessIterator __x;
     do
     {
-        __x = __internal::__pattern_partition(
-            ::std::forward<_ExecutionPolicy>(__exec), __first + 1, __last,
-            [&__comp, __first](const _Tp& __x) { return __comp(__x, *__first); }, __is_vector,
-            /*is_parallel=*/::std::true_type());
+        __x = __internal::__pattern_partition(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first + 1,
+                                              __last,
+                                              [&__comp, __first](const _Tp& __x) { return __comp(__x, *__first); });
         --__x;
         if (__x != __first)
         {
@@ -3677,10 +3678,8 @@ __pattern_nth_element(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec
     _RandomAccessIterator __x;
     do
     {
-        __x = __internal::__pattern_partition(
-            ::std::forward<_ExecutionPolicy>(__exec), __first + 1, __last,
-            [&__comp, __first](const _Tp& __x) { return __comp(__x, *__first); }, _IsVector{},
-            /*is_parallel=*/::std::true_type());
+        __x = __internal::__pattern_partition(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first + 1, __last,
+                                              [&__comp, __first](const _Tp& __x) { return __comp(__x, *__first); });
         --__x;
         if (__x != __first)
         {
diff --git a/include/oneapi/dpl/pstl/glue_algorithm_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
index 5a96ceb30fe..1b7b0c8317e 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
@@ -631,10 +631,10 @@ template <class _ExecutionPolicy, class _ForwardIterator, class _UnaryPredicate>
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _ForwardIterator>
 partition(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last, _UnaryPredicate __pred)
 {
-    return oneapi::dpl::__internal::__pattern_partition(
-        ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __pred,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>());
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+
+    return oneapi::dpl::__internal::__pattern_partition(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
+                                                        __first, __last, __pred);
 }
 
 template <class _ExecutionPolicy, class _BidirectionalIterator, class _UnaryPredicate>

From eaddf2111b33fbd65c4ad5cb32e90a24aae112ce Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 6 Feb 2024 15:53:15 +0100
Subject: [PATCH 115/566] __pattern_partition_copy + tag impls

---
 include/oneapi/dpl/pstl/algorithm_fwd.h       | 12 +++++
 include/oneapi/dpl/pstl/algorithm_impl.h      | 50 +++++++++++++++++++
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 25 ++++++++++
 3 files changed, 87 insertions(+)

diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index 2049400a571..9b380d672b8 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -999,6 +999,12 @@ __pattern_partition_copy(_ExecutionPolicy&&, _ForwardIterator, _ForwardIterator,
                          _UnaryPredicate, _IsVector,
                          /*is_parallelization=*/::std::false_type) noexcept;
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _OutputIterator1, class _OutputIterator2,
+          class _UnaryPredicate>
+::std::pair<_OutputIterator1, _OutputIterator2>
+__pattern_partition_copy(_Tag, _ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _OutputIterator1,
+                         _OutputIterator2, _UnaryPredicate) noexcept;
+
 template <class _ExecutionPolicy, class _RandomAccessIterator, class _OutputIterator1, class _OutputIterator2,
           class _UnaryPredicate, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy,
@@ -1007,6 +1013,12 @@ __pattern_partition_copy(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAcces
                          _OutputIterator2, _UnaryPredicate, _IsVector,
                          /*is_parallelization=*/::std::true_type);
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
+          class _RandomAccessIterator3, class _UnaryPredicate>
+::std::pair<_RandomAccessIterator2, _RandomAccessIterator3>
+__pattern_partition_copy(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator1, _RandomAccessIterator1,
+                         _RandomAccessIterator2, _RandomAccessIterator3, _UnaryPredicate);
+
 //------------------------------------------------------------------------
 // sort
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 629029f0a77..8f910352492 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -3090,6 +3090,18 @@ __pattern_partition_copy(_ExecutionPolicy&&, _ForwardIterator __first, _ForwardI
     return __internal::__brick_partition_copy(__first, __last, __out_true, __out_false, __pred, __is_vector);
 }
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _OutputIterator1, class _OutputIterator2,
+          class _UnaryPredicate>
+::std::pair<_OutputIterator1, _OutputIterator2>
+__pattern_partition_copy(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
+                         _OutputIterator1 __out_true, _OutputIterator2 __out_false, _UnaryPredicate __pred) noexcept
+{
+    static_assert(__is_backend_tag_v<_Tag>);
+
+    return __internal::__brick_partition_copy(__first, __last, __out_true, __out_false, __pred,
+                                              typename _Tag::__is_vector{});
+}
+
 template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
           class _RandomAccessIterator3, class _UnaryPredicate, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy,
@@ -3130,6 +3142,44 @@ __pattern_partition_copy(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __fir
     return __internal::__brick_partition_copy(__first, __last, __out_true, __out_false, __pred, __is_vector);
 }
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
+          class _RandomAccessIterator3, class _UnaryPredicate>
+::std::pair<_RandomAccessIterator2, _RandomAccessIterator3>
+__pattern_partition_copy(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator1 __first,
+                         _RandomAccessIterator1 __last, _RandomAccessIterator2 __out_true,
+                         _RandomAccessIterator3 __out_false, _UnaryPredicate __pred)
+{
+    typedef typename ::std::iterator_traits<_RandomAccessIterator1>::difference_type _DifferenceType;
+    typedef ::std::pair<_DifferenceType, _DifferenceType> _ReturnType;
+    const _DifferenceType __n = __last - __first;
+    if (_DifferenceType(1) < __n)
+    {
+        __par_backend::__buffer<_ExecutionPolicy, bool> __mask_buf(__n);
+        return __internal::__except_handler([&__exec, __n, __first, __out_true, __out_false, __pred, &__mask_buf]() {
+            bool* __mask = __mask_buf.get();
+            _ReturnType __m{};
+            __par_backend::__parallel_strict_scan(
+                ::std::forward<_ExecutionPolicy>(__exec), __n, ::std::make_pair(_DifferenceType(0), _DifferenceType(0)),
+                [=](_DifferenceType __i, _DifferenceType __len) { // Reduce
+                    return __internal::__brick_calc_mask_1<_DifferenceType>(__first + __i, __first + (__i + __len),
+                                                                            __mask + __i, __pred, _IsVector{});
+                },
+                [](const _ReturnType& __x, const _ReturnType& __y) -> _ReturnType {
+                    return ::std::make_pair(__x.first + __y.first, __x.second + __y.second);
+                },                                                                       // Combine
+                [=](_DifferenceType __i, _DifferenceType __len, _ReturnType __initial) { // Scan
+                    __internal::__brick_partition_by_mask(__first + __i, __first + (__i + __len),
+                                                          __out_true + __initial.first, __out_false + __initial.second,
+                                                          __mask + __i, _IsVector{});
+                },
+                [&__m](_ReturnType __total) { __m = __total; });
+            return ::std::make_pair(__out_true + __m.first, __out_false + __m.second);
+        });
+    }
+    // trivial sequence - use serial algorithm
+    return __internal::__brick_partition_copy(__first, __last, __out_true, __out_false, __pred, _IsVector{});
+}
+
 //------------------------------------------------------------------------
 // sort
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 07855e3e796..10097adeb11 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -1561,6 +1561,31 @@ __pattern_partition_copy(_ExecutionPolicy&& __exec, _Iterator1 __first, _Iterato
     return ::std::make_pair(__result1 + __result.second, __result2 + (__last - __first - __result.second));
 }
 
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2,
+          typename _Iterator3, typename _UnaryPredicate>
+::std::pair<_Iterator2, _Iterator3>
+__pattern_partition_copy(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator1 __first,
+                         _Iterator1 __last, _Iterator2 __result1, _Iterator3 __result2, _UnaryPredicate __pred)
+{
+    if (__first == __last)
+        return ::std::make_pair(__result1, __result2);
+
+    using _It1DifferenceType = typename ::std::iterator_traits<_Iterator1>::difference_type;
+    using _ReduceOp = ::std::plus<_It1DifferenceType>;
+
+    unseq_backend::__create_mask<_UnaryPredicate, _It1DifferenceType> __create_mask_op{__pred};
+    unseq_backend::__partition_by_mask<_ReduceOp, /*inclusive*/ ::std::true_type> __copy_by_mask_op{_ReduceOp{}};
+
+    auto __result = __pattern_scan_copy(
+        ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+        __par_backend_hetero::zip(
+            __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::write>(__result1),
+            __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::write>(__result2)),
+        __create_mask_op, __copy_by_mask_op);
+
+    return ::std::make_pair(__result1 + __result.second, __result2 + (__last - __first - __result.second));
+}
+
 //------------------------------------------------------------------------
 // unique_copy
 //------------------------------------------------------------------------

From 7398bf4e44b1f9726b477eef30ef1a5f57750278 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 6 Feb 2024 15:54:49 +0100
Subject: [PATCH 116/566] __pattern_partition_copy + tag calls

---
 include/oneapi/dpl/pstl/glue_algorithm_impl.h          | 9 +++++----
 include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h | 6 ++++--
 2 files changed, 9 insertions(+), 6 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_algorithm_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
index 1b7b0c8317e..21cd4dec70c 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
@@ -655,10 +655,11 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy,
 partition_copy(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last,
                _ForwardIterator1 __out_true, _ForwardIterator2 __out_false, _UnaryPredicate __pred)
 {
-    return oneapi::dpl::__internal::__pattern_partition_copy(
-        ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __out_true, __out_false, __pred,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator, _ForwardIterator1, _ForwardIterator2>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator, _ForwardIterator1, _ForwardIterator2>());
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator,
+                                                                              _ForwardIterator1, _ForwardIterator2>();
+
+    return oneapi::dpl::__internal::__pattern_partition_copy(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
+                                                             __first, __last, __out_true, __out_false, __pred);
 }
 
 // [alg.sort]
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 10097adeb11..694aa4bf0fc 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -2066,8 +2066,10 @@ __pattern_stable_partition(_ExecutionPolicy&& __exec, _Iterator __first, _Iterat
     auto __true_result = __true_buf.get();
     auto __false_result = __false_buf.get();
 
-    auto copy_result = __pattern_partition_copy(__exec, __first, __last, __true_result, __false_result, __pred,
-                                                /*vector=*/::std::true_type{}, /*parallel*/ ::std::true_type{});
+    constexpr auto __dispatch_tag = __select_backend<_ExecutionPolicy, decltype(__first), decltype(__last),
+                                                     decltype(__true_result), decltype(__false_result)>();
+    auto copy_result =
+        __pattern_partition_copy(__dispatch_tag, __exec, __first, __last, __true_result, __false_result, __pred);
     auto true_count = copy_result.first - __true_result;
 
     //TODO: optimize copy back if possible (inplace, decrease number of submits)

From c51a27f5304f5250c64a0a7b433aa66cd081d0ca Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 6 Feb 2024 16:13:57 +0100
Subject: [PATCH 117/566] __pattern_remove_if + tag impls

---
 include/oneapi/dpl/pstl/algorithm_fwd.h       |  9 ++++++
 include/oneapi/dpl/pstl/algorithm_impl.h      | 32 +++++++++++++++++++
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 31 ++++++++++++++++++
 3 files changed, 72 insertions(+)

diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index 9b380d672b8..8bfc1b911e8 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -1299,11 +1299,20 @@ oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _Fo
 __pattern_remove_if(_ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _UnaryPredicate, _IsVector,
                     /*is_parallel*/ ::std::false_type) noexcept;
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _UnaryPredicate>
+_ForwardIterator
+__pattern_remove_if(_Tag, _ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _UnaryPredicate) noexcept;
+
 template <class _ExecutionPolicy, class _RandomAccessIterator, class _UnaryPredicate, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator>
 __pattern_remove_if(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _UnaryPredicate, _IsVector,
                     /*is_parallel*/ ::std::true_type);
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _UnaryPredicate>
+_RandomAccessIterator
+__pattern_remove_if(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator,
+                    _UnaryPredicate);
+
 //------------------------------------------------------------------------
 // merge
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 8f910352492..e3c6a38d7df 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -4042,6 +4042,16 @@ __pattern_remove_if(_ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterat
     return __internal::__brick_remove_if(__first, __last, __pred, __is_vector);
 }
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _UnaryPredicate>
+_ForwardIterator
+__pattern_remove_if(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
+                    _UnaryPredicate __pred) noexcept
+{
+    static_assert(__is_backend_tag_v<_Tag>);
+
+    return __internal::__brick_remove_if(__first, __last, __pred, typename _Tag::__is_vector{});
+}
+
 template <class _ExecutionPolicy, class _RandomAccessIterator, class _UnaryPredicate, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator>
 __pattern_remove_if(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIterator __last,
@@ -4064,6 +4074,28 @@ __pattern_remove_if(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _R
         __is_vector);
 }
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _UnaryPredicate>
+_RandomAccessIterator
+__pattern_remove_if(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
+                    _RandomAccessIterator __last, _UnaryPredicate __pred)
+{
+    typedef typename ::std::iterator_traits<_RandomAccessIterator>::reference _ReferenceType;
+
+    if (__first == __last || __first + 1 == __last)
+    {
+        // Trivial sequence - use serial algorithm
+        return __internal::__brick_remove_if(__first, __last, __pred, _IsVector{});
+    }
+
+    return __internal::__remove_elements(
+        ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+        [&__pred](bool* __b, bool* __e, _RandomAccessIterator __it) {
+            __internal::__brick_walk2(
+                __b, __e, __it, [&__pred](bool& __x, _ReferenceType __y) { __x = !__pred(__y); }, _IsVector{});
+        },
+        _IsVector{});
+}
+
 //------------------------------------------------------------------------
 // merge
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 694aa4bf0fc..8e89a62eb2b 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -1640,6 +1640,37 @@ __pattern_remove_if(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __la
         __copy_first, __copy_last, __first, __brick_copy<_ExecutionPolicy>{}, ::std::true_type{}, ::std::true_type{});
 }
 
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator, typename _Predicate>
+_Iterator
+__pattern_remove_if(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last,
+                    _Predicate __pred)
+{
+    if (__last == __first)
+        return __last;
+
+    using _ValueType = typename ::std::iterator_traits<_Iterator>::value_type;
+
+    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __buf(__exec, __last - __first);
+    auto __copy_first = __buf.get();
+
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, decltype(__first), decltype(__last),
+                                                  decltype(__copy_first)>();
+
+    auto __copy_last =
+        __pattern_copy_if(__dispatch_tag, __exec, __first, __last, __copy_first, __not_pred<_Predicate>{__pred});
+
+    constexpr auto __dispatch_tag1 =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, decltype(__copy_first), decltype(__copy_last),
+                                                  decltype(__first)>();
+
+    //TODO: optimize copy back depending on Iterator, i.e. set_final_data for host iterator/pointer
+    return __pattern_walk2(
+        __dispatch_tag1,
+        __par_backend_hetero::make_wrapped_policy<copy_back_wrapper>(::std::forward<_ExecutionPolicy>(__exec)),
+        __copy_first, __copy_last, __first, __brick_copy<_ExecutionPolicy>{});
+}
+
 template <typename _ExecutionPolicy, typename _Iterator, typename _BinaryPredicate>
 oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, _Iterator>
 __pattern_unique(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last, _BinaryPredicate __pred,

From b16634793da330d224a9cd93e6475f74d5a903b9 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 6 Feb 2024 16:14:11 +0100
Subject: [PATCH 118/566] __pattern_remove_if + tag calls

---
 include/oneapi/dpl/pstl/glue_algorithm_impl.h | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_algorithm_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
index 21cd4dec70c..a4276ff2f0d 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
@@ -514,10 +514,10 @@ template <class _ExecutionPolicy, class _ForwardIterator, class _UnaryPredicate>
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _ForwardIterator>
 remove_if(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last, _UnaryPredicate __pred)
 {
-    return oneapi::dpl::__internal::__pattern_remove_if(
-        ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __pred,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>());
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+
+    return oneapi::dpl::__internal::__pattern_remove_if(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
+                                                        __first, __last, __pred);
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator, class _Tp>

From fddf7bcf2637caa120bd73911022f1443aca98be Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 6 Feb 2024 16:30:46 +0100
Subject: [PATCH 119/566] __pattern_reverse + tag impls

---
 include/oneapi/dpl/pstl/algorithm_fwd.h       |  8 +++++++
 include/oneapi/dpl/pstl/algorithm_impl.h      | 21 +++++++++++++++++++
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 20 ++++++++++++++++++
 3 files changed, 49 insertions(+)

diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index 8bfc1b911e8..8652b22d920 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -817,11 +817,19 @@ oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>
 __pattern_reverse(_ExecutionPolicy&&, _BidirectionalIterator, _BidirectionalIterator, _IsVector,
                   /*is_parallel=*/::std::false_type) noexcept;
 
+template <class _Tag, class _ExecutionPolicy, class _BidirectionalIterator>
+void
+__pattern_reverse(_Tag, _ExecutionPolicy&&, _BidirectionalIterator, _BidirectionalIterator) noexcept;
+
 template <class _ExecutionPolicy, class _RandomAccessIterator, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>
 __pattern_reverse(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _IsVector,
                   /*is_parallel=*/::std::true_type);
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator>
+void
+__pattern_reverse(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator);
+
 //------------------------------------------------------------------------
 // reverse_copy
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index e3c6a38d7df..6d18e4630eb 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -2294,6 +2294,15 @@ __pattern_reverse(_ExecutionPolicy&&, _BidirectionalIterator __first, _Bidirecti
     __internal::__brick_reverse(__first, __last, _is_vector);
 }
 
+template <class _Tag, class _ExecutionPolicy, class _BidirectionalIterator>
+void
+__pattern_reverse(_Tag, _ExecutionPolicy&&, _BidirectionalIterator __first, _BidirectionalIterator __last) noexcept
+{
+    static_assert(__is_backend_tag_v<_Tag>);
+
+    __internal::__brick_reverse(__first, __last, typename _Tag::__is_vector{});
+}
+
 template <class _ExecutionPolicy, class _RandomAccessIterator, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>
 __pattern_reverse(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIterator __last,
@@ -2306,6 +2315,18 @@ __pattern_reverse(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _Ran
         });
 }
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator>
+void
+__pattern_reverse(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
+                  _RandomAccessIterator __last)
+{
+    __par_backend::__parallel_for(
+        ::std::forward<_ExecutionPolicy>(__exec), __first, __first + (__last - __first) / 2,
+        [__first, __last](_RandomAccessIterator __inner_first, _RandomAccessIterator __inner_last) {
+            __internal::__brick_reverse(__inner_first, __inner_last, __last - (__inner_first - __first), _IsVector{});
+        });
+}
+
 //------------------------------------------------------------------------
 // reverse_copy
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 8e89a62eb2b..19da3a640d6 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -2569,6 +2569,26 @@ __pattern_reverse(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last
         .wait();
 }
 
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator>
+void
+__pattern_reverse(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last)
+{
+    using __backend_tag = typename decltype(__tag)::__backend_tag;
+
+    auto __n = __last - __first;
+    if (__n <= 0)
+        return;
+
+    auto __keep = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read_write, _Iterator>();
+    auto __buf = __keep(__first, __last);
+    oneapi::dpl::__par_backend_hetero::__parallel_for(
+        __backend_tag{},
+        ::std::forward<_ExecutionPolicy>(__exec),
+        unseq_backend::__reverse_functor<typename ::std::iterator_traits<_Iterator>::difference_type>{__n}, __n / 2,
+        __buf.all_view())
+        .wait();
+}
+
 //------------------------------------------------------------------------
 // reverse_copy
 //------------------------------------------------------------------------

From e88097890013db6bfe5388a4b8a6b37953754a57 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 6 Feb 2024 16:30:56 +0100
Subject: [PATCH 120/566] __pattern_reverse + tag calls

---
 include/oneapi/dpl/pstl/glue_algorithm_impl.h | 9 +++++----
 1 file changed, 5 insertions(+), 4 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_algorithm_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
index a4276ff2f0d..6ab9208f939 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
@@ -575,10 +575,11 @@ template <class _ExecutionPolicy, class _BidirectionalIterator>
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy>
 reverse(_ExecutionPolicy&& __exec, _BidirectionalIterator __first, _BidirectionalIterator __last)
 {
-    oneapi::dpl::__internal::__pattern_reverse(
-        ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _BidirectionalIterator>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _BidirectionalIterator>());
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _BidirectionalIterator>();
+
+    oneapi::dpl::__internal::__pattern_reverse(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first,
+                                               __last);
 }
 
 template <class _ExecutionPolicy, class _BidirectionalIterator, class _ForwardIterator>

From a43668a12c3ee17b6945d074cc5b4b160e643931 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 6 Feb 2024 16:40:39 +0100
Subject: [PATCH 121/566] __pattern_reverse_copy + tag impls

---
 include/oneapi/dpl/pstl/algorithm_fwd.h       | 10 +++++++
 include/oneapi/dpl/pstl/algorithm_impl.h      | 27 +++++++++++++++++++
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 24 +++++++++++++++++
 3 files changed, 61 insertions(+)

diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index 8652b22d920..7732b7d3847 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -847,11 +847,21 @@ oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _Ou
 __pattern_reverse_copy(_ExecutionPolicy&&, _BidirectionalIterator, _BidirectionalIterator, _OutputIterator, _IsVector,
                        /*is_parallel=*/::std::false_type) noexcept;
 
+template <class _Tag, class _ExecutionPolicy, class _BidirectionalIterator, class _OutputIterator>
+_OutputIterator
+__pattern_reverse_copy(_Tag, _ExecutionPolicy&&, _BidirectionalIterator, _BidirectionalIterator,
+                       _OutputIterator) noexcept;
+
 template <class _ExecutionPolicy, class _RandomAccessIterator, class _OutputIterator, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _OutputIterator>
 __pattern_reverse_copy(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _OutputIterator, _IsVector,
                        /*is_parallel=*/::std::true_type);
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2>
+_RandomAccessIterator2
+__pattern_reverse_copy(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator1, _RandomAccessIterator1,
+                       _RandomAccessIterator2);
+
 //------------------------------------------------------------------------
 // rotate
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 6d18e4630eb..ced2b512d75 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -2359,6 +2359,16 @@ __pattern_reverse_copy(_ExecutionPolicy&&, _BidirectionalIterator __first, _Bidi
     return __internal::__brick_reverse_copy(__first, __last, __d_first, __is_vector);
 }
 
+template <class _Tag, class _ExecutionPolicy, class _BidirectionalIterator, class _OutputIterator>
+_OutputIterator
+__pattern_reverse_copy(_Tag, _ExecutionPolicy&&, _BidirectionalIterator __first, _BidirectionalIterator __last,
+                       _OutputIterator __d_first) noexcept
+{
+    static_assert(__is_backend_tag_v<_Tag>);
+
+    return __internal::__brick_reverse_copy(__first, __last, __d_first, typename _Tag::__is_vector{});
+}
+
 template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator2>
 __pattern_reverse_copy(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first, _RandomAccessIterator1 __last,
@@ -2375,6 +2385,23 @@ __pattern_reverse_copy(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first
     return __d_first + __len;
 }
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2>
+_RandomAccessIterator2
+__pattern_reverse_copy(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator1 __first,
+                       _RandomAccessIterator1 __last, _RandomAccessIterator2 __d_first)
+{
+    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
+
+    auto __len = __last - __first;
+    __par_backend::__parallel_for(
+        __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+        [__first, __len, __d_first](_RandomAccessIterator1 __inner_first, _RandomAccessIterator1 __inner_last) {
+            __internal::__brick_reverse_copy(__inner_first, __inner_last,
+                                             __d_first + (__len - (__inner_last - __first)), _IsVector{});
+        });
+    return __d_first + __len;
+}
+
 //------------------------------------------------------------------------
 // rotate
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 19da3a640d6..87de9cec1fb 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -2616,6 +2616,30 @@ __pattern_reverse_copy(_ExecutionPolicy&& __exec, _BidirectionalIterator __first
     return __result + __n;
 }
 
+template <typename _BackendTag, typename _ExecutionPolicy, typename _BidirectionalIterator, typename _ForwardIterator>
+_ForwardIterator
+__pattern_reverse_copy(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _BidirectionalIterator __first,
+                       _BidirectionalIterator __last, _ForwardIterator __result)
+{
+    auto __n = __last - __first;
+    if (__n <= 0)
+        return __result;
+
+    auto __keep1 =
+        oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _BidirectionalIterator>();
+    auto __buf1 = __keep1(__first, __last);
+    auto __keep2 =
+        oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::write, _ForwardIterator>();
+    auto __buf2 = __keep2(__result, __result + __n);
+    oneapi::dpl::__par_backend_hetero::__parallel_for(
+        _BackendTag{}, ::std::forward<_ExecutionPolicy>(__exec),
+        unseq_backend::__reverse_copy<typename ::std::iterator_traits<_BidirectionalIterator>::difference_type>{__n},
+        __n, __buf1.all_view(), __buf2.all_view())
+        .wait();
+
+    return __result + __n;
+}
+
 //------------------------------------------------------------------------
 // rotate
 //------------------------------------------------------------------------

From 37dd88e11fc02df2c0e681a74ef00da3963b2c1a Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 6 Feb 2024 16:40:59 +0100
Subject: [PATCH 122/566] __pattern_reverse_copy + tag calls

---
 include/oneapi/dpl/pstl/glue_algorithm_impl.h | 9 +++++----
 1 file changed, 5 insertions(+), 4 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_algorithm_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
index 6ab9208f939..c0b07fc6317 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
@@ -587,10 +587,11 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _Forward
 reverse_copy(_ExecutionPolicy&& __exec, _BidirectionalIterator __first, _BidirectionalIterator __last,
              _ForwardIterator __d_first)
 {
-    return oneapi::dpl::__internal::__pattern_reverse_copy(
-        ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __d_first,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _BidirectionalIterator, _ForwardIterator>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _BidirectionalIterator, _ForwardIterator>());
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _BidirectionalIterator, _ForwardIterator>();
+
+    return oneapi::dpl::__internal::__pattern_reverse_copy(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
+                                                           __first, __last, __d_first);
 }
 
 // [alg.rotate]

From 2a7da4662bc003e04f4e4d6047d8545d69d4030b Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 6 Feb 2024 16:55:27 +0100
Subject: [PATCH 123/566] __pattern_rotate + tag impls

---
 include/oneapi/dpl/pstl/algorithm_fwd.h       |  9 +++
 include/oneapi/dpl/pstl/algorithm_impl.h      | 73 +++++++++++++++++++
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 36 +++++++++
 3 files changed, 118 insertions(+)

diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index 7732b7d3847..d36abc46f8e 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -879,11 +879,20 @@ oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _Fo
 __pattern_rotate(_ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _ForwardIterator, _IsVector,
                  /*is_parallel=*/::std::false_type) noexcept;
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator>
+_ForwardIterator
+__pattern_rotate(_Tag, _ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _ForwardIterator) noexcept;
+
 template <class _ExecutionPolicy, class _RandomAccessIterator, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator>
 __pattern_rotate(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _RandomAccessIterator, _IsVector,
                  /*is_parallel=*/::std::true_type);
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator>
+_RandomAccessIterator
+__pattern_rotate(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator,
+                 _RandomAccessIterator);
+
 //------------------------------------------------------------------------
 // rotate_copy
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index ced2b512d75..2d7e9713541 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -2462,6 +2462,15 @@ __pattern_rotate(_ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator
     return __internal::__brick_rotate(__first, __middle, __last, __is_vector);
 }
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator>
+_ForwardIterator
+__pattern_rotate(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __middle, _ForwardIterator __last) noexcept
+{
+    static_assert(__is_backend_tag_v<_Tag>);
+
+    return __internal::__brick_rotate(__first, __middle, __last, typename _Tag::__is_vector{});
+}
+
 template <class _ExecutionPolicy, class _RandomAccessIterator, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator>
 __pattern_rotate(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIterator __middle,
@@ -2524,6 +2533,70 @@ __pattern_rotate(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _Rand
     }
 }
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator>
+_RandomAccessIterator
+__pattern_rotate(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
+                 _RandomAccessIterator __middle, _RandomAccessIterator __last)
+{
+    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
+
+    typedef typename ::std::iterator_traits<_RandomAccessIterator>::value_type _Tp;
+    auto __n = __last - __first;
+    auto __m = __middle - __first;
+    if (__m <= __n / 2)
+    {
+        __par_backend::__buffer<_ExecutionPolicy, _Tp> __buf(__n - __m);
+        return __internal::__except_handler([&__exec, __n, __m, __first, __middle, __last, &__buf]() {
+            _Tp* __result = __buf.get();
+            __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __middle, __last,
+                                          [__middle, __result](_RandomAccessIterator __b, _RandomAccessIterator __e) {
+                                              __internal::__brick_uninitialized_move(
+                                                  __b, __e, __result + (__b - __middle), _IsVector{});
+                                          });
+
+            __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __middle,
+                                          [__last, __middle](_RandomAccessIterator __b, _RandomAccessIterator __e) {
+                                              __internal::__brick_move<_ExecutionPolicy>{}(
+                                                  __b, __e, __b + (__last - __middle), _IsVector{});
+                                          });
+
+            __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __result,
+                                          __result + (__n - __m), [__first, __result](_Tp* __b, _Tp* __e) {
+                                              __brick_move_destroy<_ExecutionPolicy>{}(
+                                                  __b, __e, __first + (__b - __result), _IsVector{});
+                                          });
+
+            return __first + (__last - __middle);
+        });
+    }
+    else
+    {
+        __par_backend::__buffer<_ExecutionPolicy, _Tp> __buf(__m);
+        return __internal::__except_handler([&__exec, __n, __m, __first, __middle, __last, &__buf]() {
+            _Tp* __result = __buf.get();
+            __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __middle,
+                                          [__first, __result](_RandomAccessIterator __b, _RandomAccessIterator __e) {
+                                              __internal::__brick_uninitialized_move(
+                                                  __b, __e, __result + (__b - __first), _IsVector{});
+                                          });
+
+            __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __middle, __last,
+                                          [__first, __middle](_RandomAccessIterator __b, _RandomAccessIterator __e) {
+                                              __internal::__brick_move<_ExecutionPolicy>{}(
+                                                  __b, __e, __first + (__b - __middle), _IsVector{});
+                                          });
+
+            __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __result,
+                                          __result + __m, [__n, __m, __first, __result](_Tp* __b, _Tp* __e) {
+                                              __brick_move_destroy<_ExecutionPolicy>{}(
+                                                  __b, __e, __first + ((__n - __m) + (__b - __result)), _IsVector{});
+                                          });
+
+            return __first + (__last - __middle);
+        });
+    }
+}
+
 //------------------------------------------------------------------------
 // rotate_copy
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 87de9cec1fb..9c162c26226 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -2687,6 +2687,42 @@ __pattern_rotate(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __new_f
     return __first + (__last - __new_first);
 }
 
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator>
+_Iterator
+__pattern_rotate(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __new_first,
+                 _Iterator __last)
+{
+    using __backend_tag = typename decltype(__tag)::__backend_tag;
+
+    auto __n = __last - __first;
+    if (__n <= 0)
+        return __first;
+
+    using _Tp = typename ::std::iterator_traits<_Iterator>::value_type;
+
+    auto __keep = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read_write, _Iterator>();
+    auto __buf = __keep(__first, __last);
+    auto __temp_buf = oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _Tp>(__exec, __n);
+
+    auto __temp_rng =
+        oneapi::dpl::__ranges::all_view<_Tp, __par_backend_hetero::access_mode::write>(__temp_buf.get_buffer());
+
+    const auto __shift = __new_first - __first;
+    oneapi::dpl::__par_backend_hetero::__parallel_for(
+        __backend_tag{}, oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__rotate_wrapper>(__exec),
+        unseq_backend::__rotate_copy<typename ::std::iterator_traits<_Iterator>::difference_type>{__n, __shift}, __n,
+        __buf.all_view(), __temp_rng);
+
+    using _Function = __brick_move<_ExecutionPolicy>;
+    auto __brick = unseq_backend::walk_n<_ExecutionPolicy, _Function>{_Function{}};
+
+    oneapi::dpl::__par_backend_hetero::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
+                                                      __brick, __n, __temp_rng, __buf.all_view())
+        .wait();
+
+    return __first + (__last - __new_first);
+}
+
 //------------------------------------------------------------------------
 // rotate_copy
 //------------------------------------------------------------------------

From 2e467001283c532408e82547d27a7eb525abe0ef Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 6 Feb 2024 16:55:39 +0100
Subject: [PATCH 124/566] __pattern_rotate + tag calls

---
 include/oneapi/dpl/pstl/glue_algorithm_impl.h | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_algorithm_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
index c0b07fc6317..40ca4ecb2d6 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
@@ -600,10 +600,10 @@ template <class _ExecutionPolicy, class _ForwardIterator>
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _ForwardIterator>
 rotate(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __middle, _ForwardIterator __last)
 {
-    return oneapi::dpl::__internal::__pattern_rotate(
-        ::std::forward<_ExecutionPolicy>(__exec), __first, __middle, __last,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>());
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+
+    return oneapi::dpl::__internal::__pattern_rotate(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first,
+                                                     __middle, __last);
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2>

From ece56b26c1c194f04da7c4b1c046a71166786493 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 6 Feb 2024 17:06:07 +0100
Subject: [PATCH 125/566] __pattern_rotate_copy + tag impls

---
 include/oneapi/dpl/pstl/algorithm_fwd.h       | 10 +++++
 include/oneapi/dpl/pstl/algorithm_impl.h      | 44 +++++++++++++++++++
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 30 +++++++++++++
 3 files changed, 84 insertions(+)

diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index d36abc46f8e..1639f795928 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -912,12 +912,22 @@ __pattern_rotate_copy(_ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _F
                       _IsVector,
                       /*is_parallel=*/::std::false_type) noexcept;
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _OutputIterator>
+_OutputIterator
+__pattern_rotate_copy(_Tag, _ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _ForwardIterator,
+                      _OutputIterator) noexcept;
+
 template <class _ExecutionPolicy, class _RandomAccessIterator, class _OutputIterator, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _OutputIterator>
 __pattern_rotate_copy(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _RandomAccessIterator,
                       _OutputIterator, _IsVector,
                       /*is_parallel=*/::std::true_type);
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2>
+_RandomAccessIterator2
+__pattern_rotate_copy(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator1, _RandomAccessIterator1,
+                      _RandomAccessIterator1, _RandomAccessIterator2);
+
 //------------------------------------------------------------------------
 // is_partitioned
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 2d7e9713541..9fbcf8265dc 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -2629,6 +2629,17 @@ __pattern_rotate_copy(_ExecutionPolicy&& __exec, _ForwardIterator __first, _Forw
                                            __result, __is_vector);
 }
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _OutputIterator>
+_OutputIterator
+__pattern_rotate_copy(_Tag, _ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __middle,
+                      _ForwardIterator __last, _OutputIterator __result) noexcept
+{
+    static_assert(__is_backend_tag_v<_Tag>);
+
+    return __internal::__brick_rotate_copy(::std::forward<_ExecutionPolicy>(__exec), __first, __middle, __last,
+                                           __result, typename _Tag::__is_vector{});
+}
+
 template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator2>
 __pattern_rotate_copy(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first, _RandomAccessIterator1 __middle,
@@ -2660,6 +2671,39 @@ __pattern_rotate_copy(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first,
     return __result + (__last - __first);
 }
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2>
+_RandomAccessIterator2
+__pattern_rotate_copy(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator1 __first,
+                      _RandomAccessIterator1 __middle, _RandomAccessIterator1 __last, _RandomAccessIterator2 __result)
+{
+    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
+
+    __par_backend::__parallel_for(
+        __backend_tag{},
+        ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+        [__first, __last, __middle, __result](_RandomAccessIterator1 __b, _RandomAccessIterator1 __e) {
+            __internal::__brick_copy<_ExecutionPolicy> __copy{};
+            if (__b > __middle)
+            {
+                __copy(__b, __e, __result + (__b - __middle), _IsVector{});
+            }
+            else
+            {
+                _RandomAccessIterator2 __new_result = __result + ((__last - __middle) + (__b - __first));
+                if (__e < __middle)
+                {
+                    __copy(__b, __e, __new_result, _IsVector{});
+                }
+                else
+                {
+                    __copy(__b, __middle, __new_result, _IsVector{});
+                    __copy(__middle, __e, __result, _IsVector{});
+                }
+            }
+        });
+    return __result + (__last - __first);
+}
+
 //------------------------------------------------------------------------
 // is_partitioned
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 9c162c26226..c2fad196ee1 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -2755,6 +2755,36 @@ __pattern_rotate_copy(_ExecutionPolicy&& __exec, _BidirectionalIterator __first,
     return __result + __n;
 }
 
+template <typename _BackendTag, typename _ExecutionPolicy, typename _BidirectionalIterator, typename _ForwardIterator>
+_ForwardIterator
+__pattern_rotate_copy(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _BidirectionalIterator __first,
+                      _BidirectionalIterator __new_first, _BidirectionalIterator __last, _ForwardIterator __result)
+{
+    using __backend_tag = typename decltype(__tag)::__backend_tag;
+
+    auto __n = __last - __first;
+    if (__n <= 0)
+        return __result;
+
+    auto __keep1 =
+        oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _BidirectionalIterator>();
+    auto __buf1 = __keep1(__first, __last);
+    auto __keep2 =
+        oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::write, _ForwardIterator>();
+    auto __buf2 = __keep2(__result, __result + __n);
+
+    const auto __shift = __new_first - __first;
+
+    oneapi::dpl::__par_backend_hetero::__parallel_for(
+        __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
+        unseq_backend::__rotate_copy<typename ::std::iterator_traits<_BidirectionalIterator>::difference_type>{__n,
+                                                                                                               __shift},
+        __n, __buf1.all_view(), __buf2.all_view())
+        .wait();
+
+    return __result + __n;
+}
+
 template <typename _ExecutionPolicy, typename _ForwardIterator1, typename _ForwardIterator2, typename _OutputIterator,
           typename _Compare, typename _IsOpDifference>
 oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, _OutputIterator>

From 668271be4aa6ff24bb1bd34688b76ed5af886cf7 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 6 Feb 2024 17:06:19 +0100
Subject: [PATCH 126/566] __pattern_rotate_copy + tag calls

---
 include/oneapi/dpl/pstl/glue_algorithm_impl.h | 9 +++++----
 1 file changed, 5 insertions(+), 4 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_algorithm_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
index 40ca4ecb2d6..bc97670f8a8 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
@@ -611,10 +611,11 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _Forward
 rotate_copy(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _ForwardIterator1 __middle, _ForwardIterator1 __last,
             _ForwardIterator2 __result)
 {
-    return oneapi::dpl::__internal::__pattern_rotate_copy(
-        ::std::forward<_ExecutionPolicy>(__exec), __first, __middle, __last, __result,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>());
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
+
+    return oneapi::dpl::__internal::__pattern_rotate_copy(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
+                                                          __first, __middle, __last, __result);
 }
 
 // [alg.partitions]

From a7aed3937dcf14e413dea6c62a6476099a6e3b84 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 6 Feb 2024 17:17:27 +0100
Subject: [PATCH 127/566] __pattern_search + tag impls

---
 include/oneapi/dpl/pstl/algorithm_fwd.h       | 11 ++++++
 include/oneapi/dpl/pstl/algorithm_impl.h      | 39 +++++++++++++++++++
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 29 ++++++++++++++
 3 files changed, 79 insertions(+)

diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index 1639f795928..d246538403c 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -577,6 +577,11 @@ __pattern_search(_ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _Forw
                  _BinaryPredicate, _IsVector,
                  /*is_parallel=*/::std::false_type) noexcept;
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _BinaryPredicate>
+_ForwardIterator1
+__pattern_search(_Tag, _ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2, _ForwardIterator2,
+                 _BinaryPredicate) noexcept;
+
 template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _BinaryPredicate,
           class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator1>
@@ -584,6 +589,12 @@ __pattern_search(_ExecutionPolicy&&, _RandomAccessIterator1, _RandomAccessIterat
                  _RandomAccessIterator2, _BinaryPredicate, _IsVector,
                  /*is_parallel=*/::std::true_type);
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
+          class _BinaryPredicate>
+_RandomAccessIterator1
+__pattern_search(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator1, _RandomAccessIterator1,
+                 _RandomAccessIterator2, _RandomAccessIterator2, _BinaryPredicate);
+
 //------------------------------------------------------------------------
 // search_n
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 9fbcf8265dc..2dbbaf6f26d 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -1462,6 +1462,16 @@ __pattern_search(_ExecutionPolicy&&, _ForwardIterator1 __first, _ForwardIterator
     return __internal::__brick_search(__first, __last, __s_first, __s_last, __pred, __is_vector);
 }
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _BinaryPredicate>
+_ForwardIterator1
+__pattern_search(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first, _ForwardIterator1 __last,
+                 _ForwardIterator2 __s_first, _ForwardIterator2 __s_last, _BinaryPredicate __pred) noexcept
+{
+    static_assert(__is_backend_tag_v<_Tag>);
+
+    return __internal::__brick_search(__first, __last, __s_first, __s_last, __pred, typename _Tag::__is_vector{});
+}
+
 template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _BinaryPredicate,
           class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator1>
@@ -1496,6 +1506,35 @@ __pattern_search(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first, _Ran
     }
 }
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
+          class _BinaryPredicate>
+_RandomAccessIterator1
+__pattern_search(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _RandomAccessIterator1 __first,
+                 _RandomAccessIterator1 __last, _RandomAccessIterator2 __s_first, _RandomAccessIterator2 __s_last,
+                 _BinaryPredicate __pred)
+{
+    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
+
+    if (__last - __first == __s_last - __s_first)
+    {
+        const bool __res = __internal::__pattern_equal(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+                                                       __s_first, __pred);
+        return __res ? __first : __last;
+    }
+    else
+    {
+        return __internal::__except_handler([&]() {
+            return __internal::__parallel_find(
+                __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+                [__last, __s_first, __s_last, __pred](_RandomAccessIterator1 __i, _RandomAccessIterator1 __j) {
+                    return __internal::__find_subrange(__i, __j, __last, __s_first, __s_last, __pred, true,
+                                                       _IsVector{});
+                },
+                /*_IsFirst=*/::std::true_type{});
+        });
+    }
+}
+
 //------------------------------------------------------------------------
 // search_n
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index c2fad196ee1..99c46da8f16 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -1357,6 +1357,35 @@ __pattern_search(_ExecutionPolicy&& __exec, _Iterator1 __first, _Iterator1 __las
         ::std::true_type{});
 }
 
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2, typename _Pred>
+_Iterator1
+__pattern_search(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator1 __first, _Iterator1 __last,
+                 _Iterator2 __s_first, _Iterator2 __s_last, _Pred __pred)
+{
+    if (__s_last == __s_first)
+        return __first;
+
+    if (__last - __first < __s_last - __s_first)
+        return __last;
+
+    if (__last - __first == __s_last - __s_first)
+    {
+        const bool __res = __pattern_equal(
+            __tag, __par_backend_hetero::make_wrapped_policy<equal_wrapper>(::std::forward<_ExecutionPolicy>(__exec)),
+            __first, __last, __s_first, __pred);
+        return __res ? __first : __last;
+    }
+
+    using _Predicate = unseq_backend::multiple_match_pred<_ExecutionPolicy, _Pred>;
+    return __par_backend_hetero::__parallel_find(
+        ::std::forward<_ExecutionPolicy>(__exec),
+        __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__first),
+        __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__last),
+        __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__s_first),
+        __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__s_last), _Predicate{__pred},
+        ::std::true_type{});
+}
+
 //------------------------------------------------------------------------
 // search_n
 //------------------------------------------------------------------------

From 6dce79e1f30ae1f00a52bfecb63ebae3f5e6d9bd Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 6 Feb 2024 17:20:01 +0100
Subject: [PATCH 128/566] __pattern_search + tag calls

---
 include/oneapi/dpl/pstl/glue_algorithm_impl.h | 9 +++++----
 1 file changed, 5 insertions(+), 4 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_algorithm_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
index bc97670f8a8..c06b4a9ba4e 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
@@ -229,10 +229,11 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _Forward
 search(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _ForwardIterator1 __last, _ForwardIterator2 __s_first,
        _ForwardIterator2 __s_last, _BinaryPredicate __pred)
 {
-    return oneapi::dpl::__internal::__pattern_search(
-        ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __s_first, __s_last, __pred,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>());
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
+
+    return oneapi::dpl::__internal::__pattern_search(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first,
+                                                     __last, __s_first, __s_last, __pred);
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2>

From cdd519334f1d55934ce8af2c538bf8896cfcf9a3 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 6 Feb 2024 17:45:29 +0100
Subject: [PATCH 129/566] __pattern_search_n + tag impls

---
 include/oneapi/dpl/pstl/algorithm_fwd.h       | 11 ++++++
 include/oneapi/dpl/pstl/algorithm_impl.h      | 39 +++++++++++++++++++
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 28 +++++++++++++
 3 files changed, 78 insertions(+)

diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index d246538403c..d53131d1e0c 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -616,6 +616,11 @@ __pattern_search_n(_ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _Size
                    IsVector,
                    /*is_parallel=*/::std::false_type) noexcept;
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Size, class _Tp, class _BinaryPredicate>
+_ForwardIterator
+__pattern_search_n(_Tag, _ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _Size, const _Tp&,
+                   _BinaryPredicate) noexcept;
+
 template <class _ExecutionPolicy, class _RandomAccessIterator, class _Size, class _Tp, class _BinaryPredicate,
           class IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator>
@@ -623,6 +628,12 @@ __pattern_search_n(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessItera
                    _BinaryPredicate, IsVector,
                    /*is_parallel=*/::std::true_type);
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Size, class _Tp,
+          class _BinaryPredicate>
+_RandomAccessIterator
+__pattern_search_n(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _Size,
+                   const _Tp&, _BinaryPredicate);
+
 //------------------------------------------------------------------------
 // copy_n
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 2dbbaf6f26d..461ef02a777 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -1564,6 +1564,16 @@ __pattern_search_n(_ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterato
     return __internal::__brick_search_n(__first, __last, __count, __value, __pred, __is_vector);
 }
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Size, class _Tp, class _BinaryPredicate>
+_ForwardIterator
+__pattern_search_n(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _Size __count,
+                   const _Tp& __value, _BinaryPredicate __pred) noexcept
+{
+    static_assert(__is_backend_tag_v<_Tag>);
+
+    return __internal::__brick_search_n(__first, __last, __count, __value, __pred, typename _Tag::__is_vector{});
+}
+
 template <class _ExecutionPolicy, class _RandomAccessIterator, class _Size, class _Tp, class _BinaryPredicate,
           class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator>
@@ -1592,6 +1602,35 @@ __pattern_search_n(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _Ra
     }
 }
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Size, class _Tp,
+          class _BinaryPredicate>
+_RandomAccessIterator
+__pattern_search_n(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
+                   _RandomAccessIterator __last, _Size __count, const _Tp& __value, _BinaryPredicate __pred)
+{
+    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
+
+    if (static_cast<_Size>(__last - __first) == __count)
+    {
+        const bool __result =
+            !__internal::__pattern_any_of(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+                                          [&__value, &__pred](const _Tp& __val) { return !__pred(__val, __value); });
+        return __result ? __first : __last;
+    }
+    else
+    {
+        return __internal::__except_handler([&__exec, __first, __last, __count, &__value, __pred]() {
+            return __internal::__parallel_find(
+                __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+                [__last, __count, &__value, __pred](_RandomAccessIterator __i, _RandomAccessIterator __j) {
+                    return __internal::__find_subrange(__i, __j, __last, __count, __value, __pred, _IsVector{});
+                },
+                ::std::true_type{});
+        });
+    }
+}
+
+
 //------------------------------------------------------------------------
 // copy_n
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 99c46da8f16..61b5fc06646 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -1433,6 +1433,34 @@ __pattern_search_n(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __las
         _Predicate{__pred, __value, __count}, ::std::true_type{});
 }
 
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator, typename _Size, typename _Tp,
+          typename _BinaryPredicate>
+_Iterator
+__pattern_search_n(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last,
+                   _Size __count, const _Tp& __value, _BinaryPredicate __pred)
+{
+    if (__count <= 0)
+        return __first;
+
+    if (__last - __first < __count)
+        return __last;
+
+    if (__last - __first == __count)
+    {
+        return (!__internal::__pattern_any_of(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+                                              __search_n_unary_predicate<_Tp, _BinaryPredicate>{__value, __pred}))
+                   ? __first
+                   : __last;
+    }
+
+    using _Predicate = unseq_backend::n_elem_match_pred<_ExecutionPolicy, _BinaryPredicate, _Tp, _Size>;
+    return __par_backend_hetero::__parallel_find(
+        _BackendTag{}, ::std::forward<_ExecutionPolicy>(__exec),
+        __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__first),
+        __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__last),
+        _Predicate{__pred, __value, __count}, ::std::true_type{});
+}
+
 //------------------------------------------------------------------------
 // mismatch
 //------------------------------------------------------------------------

From 302c456f0a64caed54ecab58f864f48339c5566e Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 6 Feb 2024 17:46:38 +0100
Subject: [PATCH 130/566] __pattern_search_n + tag calls

---
 include/oneapi/dpl/pstl/glue_algorithm_impl.h | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_algorithm_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
index c06b4a9ba4e..0eec9137088 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
@@ -250,10 +250,10 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _Forward
 search_n(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last, _Size __count,
          const _Tp& __value, _BinaryPredicate __pred)
 {
-    return oneapi::dpl::__internal::__pattern_search_n(
-        ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __count, __value, __pred,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>());
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+
+    return oneapi::dpl::__internal::__pattern_search_n(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
+                                                       __first, __last, __count, __value, __pred);
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator, class _Size, class _Tp>

From 1871dd6cf6b0a97a56a9a57dd2d84ee54e005775 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 7 Feb 2024 09:59:53 +0100
Subject: [PATCH 131/566] __pattern_set_difference + tag impls

---
 include/oneapi/dpl/pstl/algorithm_fwd.h       | 12 ++++
 include/oneapi/dpl/pstl/algorithm_impl.h      | 65 +++++++++++++++++++
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 25 +++++++
 3 files changed, 102 insertions(+)

diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index d53131d1e0c..0d3afa5a85c 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -1547,6 +1547,12 @@ oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _Ou
 __pattern_set_difference(_ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2, _ForwardIterator2,
                          _OutputIterator, _Compare, _IsVector, /*is_parallel=*/::std::false_type) noexcept;
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _OutputIterator,
+          class _Compare>
+_OutputIterator
+__pattern_set_difference(_Tag, _ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2,
+                         _ForwardIterator2, _OutputIterator, _Compare) noexcept;
+
 template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _OutputIterator,
           class _Compare, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _OutputIterator>
@@ -1554,6 +1560,12 @@ __pattern_set_difference(_ExecutionPolicy&&, _RandomAccessIterator1, _RandomAcce
                          _RandomAccessIterator2, _OutputIterator, _Compare, _IsVector,
                          /*is_parallel=*/::std::true_type);
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
+          class _RandomAccessIterator3, class _Compare>
+_RandomAccessIterator3
+__pattern_set_difference(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator1, _RandomAccessIterator1,
+                         _RandomAccessIterator2, _RandomAccessIterator2, _RandomAccessIterator3, _Compare);
+
 //------------------------------------------------------------------------
 // set_symmetric_difference
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 461ef02a777..366561640a3 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -5094,6 +5094,19 @@ __pattern_set_difference(_ExecutionPolicy&&, _ForwardIterator1 __first1, _Forwar
     return __internal::__brick_set_difference(__first1, __last1, __first2, __last2, __result, __comp, __is_vector);
 }
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _OutputIterator,
+          class _Compare>
+_OutputIterator
+__pattern_set_difference(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
+                         _ForwardIterator2 __first2, _ForwardIterator2 __last2, _OutputIterator __result,
+                         _Compare __comp) noexcept
+{
+    static_assert(__is_backend_tag_v<_Tag>);
+
+    return __internal::__brick_set_difference(__first1, __last1, __first2, __last2, __result, __comp,
+                                              typename _Tag::__is_vector{});
+}
+
 template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
           class _RandomAccessIterator3, class _Compare, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator3>
@@ -5148,6 +5161,58 @@ __pattern_set_difference(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __fir
     return ::std::set_difference(__first1, __last1, __first2, __last2, __result, __comp);
 }
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
+          class _RandomAccessIterator3, class _Compare>
+_RandomAccessIterator3
+__pattern_set_difference(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1,
+                         _RandomAccessIterator1 __last1,
+                         _RandomAccessIterator2 __first2, _RandomAccessIterator2 __last2,
+                         _RandomAccessIterator3 __result, _Compare __comp)
+{
+    typedef typename ::std::iterator_traits<_RandomAccessIterator3>::value_type _T;
+    typedef typename ::std::iterator_traits<_RandomAccessIterator1>::difference_type _DifferenceType;
+
+    const auto __n1 = __last1 - __first1;
+    const auto __n2 = __last2 - __first2;
+
+    // {} \ {2}: the difference is empty
+    if (__n1 == 0)
+        return __result;
+
+    // {1} \ {}: parallel copying just first sequence
+    if (__n2 == 0)
+        return __pattern_walk2_brick(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __result,
+                                     __internal::__brick_copy<_ExecutionPolicy>{});
+
+    // testing  whether the sequences are intersected
+    _RandomAccessIterator1 __left_bound_seq_1 = ::std::lower_bound(__first1, __last1, *__first2, __comp);
+    //{1} < {2}: seq 2 is wholly greater than seq 1, so, parallel copying just first sequence
+    if (__left_bound_seq_1 == __last1)
+        return __pattern_walk2_brick(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __result,
+                                     __internal::__brick_copy<_ExecutionPolicy>{});
+
+    // testing  whether the sequences are intersected
+    _RandomAccessIterator2 __left_bound_seq_2 = ::std::lower_bound(__first2, __last2, *__first1, __comp);
+    //{2} < {1}: seq 1 is wholly greater than seq 2, so, parallel copying just first sequence
+    if (__left_bound_seq_2 == __last2)
+        return __internal::__pattern_walk2_brick(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1,
+                                                 __result, __brick_copy<_ExecutionPolicy>{});
+
+    if (__n1 + __n2 > __set_algo_cut_off)
+        return __parallel_set_op(
+            ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2, __result, __comp,
+            [](_DifferenceType __n, _DifferenceType) { return __n; },
+            [](_RandomAccessIterator1 __first1, _RandomAccessIterator1 __last1, _RandomAccessIterator2 __first2,
+               _RandomAccessIterator2 __last2, _T* __result, _Compare __comp) {
+                return oneapi::dpl::__utils::__set_difference_construct(__first1, __last1, __first2, __last2, __result,
+                                                                        __comp, __BrickCopyConstruct<_IsVector>());
+            },
+            _IsVector{});
+
+    // use serial algorithm
+    return ::std::set_difference(__first1, __last1, __first2, __last2, __result, __comp);
+}
+
 //------------------------------------------------------------------------
 // set_symmetric_difference
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 61b5fc06646..a28a3da8531 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -2956,6 +2956,31 @@ __pattern_set_difference(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1,
                                    __result, __comp, unseq_backend::_DifferenceTag());
 }
 
+template <typename _BackendTag, typename _ExecutionPolicy, typename _ForwardIterator1, typename _ForwardIterator2,
+          typename _OutputIterator, typename _Compare>
+_OutputIterator
+__pattern_set_difference(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _ForwardIterator1 __first1,
+                         _ForwardIterator1 __last1, _ForwardIterator2 __first2, _ForwardIterator2 __last2,
+                         _OutputIterator __result, _Compare __comp)
+{
+    // {} \ {2}: the difference is empty
+    if (__first1 == __last1)
+        return __result;
+
+    // {1} \ {}: the difference is {1}
+    if (__first2 == __last2)
+    {
+        return oneapi::dpl::__internal::__pattern_walk2_brick(
+            __tag,
+            oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__set_difference_copy_case_1>(
+                ::std::forward<_ExecutionPolicy>(__exec)),
+            __first1, __last1, __result, oneapi::dpl::__internal::__brick_copy<_ExecutionPolicy>{});
+    }
+
+    return __pattern_hetero_set_op(::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2,
+                                   __result, __comp, unseq_backend::_DifferenceTag());
+}
+
 //Dummy names to avoid kernel problems
 template <typename Name>
 class __set_union_copy_case_1

From a4ad88e08cd6228831bfe22eb09db3699e13dbf8 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 7 Feb 2024 10:00:15 +0100
Subject: [PATCH 132/566] __pattern_set_difference + tag calls

---
 include/oneapi/dpl/pstl/glue_algorithm_impl.h | 9 +++++----
 1 file changed, 5 insertions(+), 4 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_algorithm_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
index 0eec9137088..e5a7dc854c9 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
@@ -1044,10 +1044,11 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _Forward
 set_difference(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
                _ForwardIterator2 __first2, _ForwardIterator2 __last2, _ForwardIterator __result, _Compare __comp)
 {
-    return oneapi::dpl::__internal::__pattern_set_difference(
-        ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2, __result, __comp,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2, _ForwardIterator>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2, _ForwardIterator>());
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1,
+                                                                              _ForwardIterator2, _ForwardIterator>();
+
+    return oneapi::dpl::__internal::__pattern_set_difference(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
+                                                             __first1, __last1, __first2, __last2, __result, __comp);
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _ForwardIterator>

From 9575ac3faf11494dd846bbabfd2fd71c584dec00 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 7 Feb 2024 10:12:23 +0100
Subject: [PATCH 133/566] __pattern_set_intersection + tag impls

---
 include/oneapi/dpl/pstl/algorithm_fwd.h       | 13 ++++
 include/oneapi/dpl/pstl/algorithm_impl.h      | 77 +++++++++++++++++++
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 15 ++++
 3 files changed, 105 insertions(+)

diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index 0d3afa5a85c..4774d55213e 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -1520,6 +1520,12 @@ __pattern_set_intersection(_ExecutionPolicy&&, _ForwardIterator1, _ForwardIterat
                            _ForwardIterator2, _OutputIterator, _Compare, _IsVector,
                            /*is_parallel=*/::std::false_type) noexcept;
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _OutputIterator,
+          class _Compare>
+_OutputIterator
+__pattern_set_intersection(_Tag, _ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2,
+                           _ForwardIterator2, _OutputIterator, _Compare) noexcept;
+
 template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _OutputIterator,
           class _Compare, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _OutputIterator>
@@ -1527,6 +1533,13 @@ __pattern_set_intersection(_ExecutionPolicy&&, _RandomAccessIterator1, _RandomAc
                            _RandomAccessIterator2, _OutputIterator, _Compare, _IsVector,
                            /*is_parallel=*/::std::true_type);
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
+          class _RandomAccessIterator3, class _Compare>
+_RandomAccessIterator3
+__pattern_set_intersection(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator1,
+                           _RandomAccessIterator1, _RandomAccessIterator2, _RandomAccessIterator2,
+                           _RandomAccessIterator3, _Compare);
+
 //------------------------------------------------------------------------
 // set_difference
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 366561640a3..34e00c2e251 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -4996,6 +4996,19 @@ __pattern_set_intersection(_ExecutionPolicy&&, _ForwardIterator1 __first1, _Forw
     return __internal::__brick_set_intersection(__first1, __last1, __first2, __last2, __result, __comp, __is_vector);
 }
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _OutputIterator,
+          class _Compare>
+_OutputIterator
+__pattern_set_intersection(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
+                           _ForwardIterator2 __first2, _ForwardIterator2 __last2, _OutputIterator __result,
+                           _Compare __comp) noexcept
+{
+    static_assert(__is_backend_tag_v<_Tag>);
+
+    return __internal::__brick_set_intersection(__first1, __last1, __first2, __last2, __result, __comp,
+                                                typename _Tag::__is_vector{});
+}
+
 template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
           class _RandomAccessIterator3, class _Compare, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator3>
@@ -5061,6 +5074,70 @@ __pattern_set_intersection(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __f
     return ::std::set_intersection(__left_bound_seq_1, __last1, __left_bound_seq_2, __last2, __result, __comp);
 }
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
+          class _RandomAccessIterator3, class _Compare>
+_RandomAccessIterator3
+__pattern_set_intersection(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1,
+                           _RandomAccessIterator1 __last1, _RandomAccessIterator2 __first2,
+                           _RandomAccessIterator2 __last2, _RandomAccessIterator3 __result, _Compare __comp)
+{
+    typedef typename ::std::iterator_traits<_RandomAccessIterator3>::value_type _T;
+    typedef typename ::std::iterator_traits<_RandomAccessIterator1>::difference_type _DifferenceType;
+
+    const auto __n1 = __last1 - __first1;
+    const auto __n2 = __last2 - __first2;
+
+    // intersection is empty
+    if (__n1 == 0 || __n2 == 0)
+        return __result;
+
+    // testing  whether the sequences are intersected
+    _RandomAccessIterator1 __left_bound_seq_1 = ::std::lower_bound(__first1, __last1, *__first2, __comp);
+    //{1} < {2}: seq 2 is wholly greater than seq 1, so, the intersection is empty
+    if (__left_bound_seq_1 == __last1)
+        return __result;
+
+    // testing  whether the sequences are intersected
+    _RandomAccessIterator2 __left_bound_seq_2 = ::std::lower_bound(__first2, __last2, *__first1, __comp);
+    //{2} < {1}: seq 1 is wholly greater than seq 2, so, the intersection is empty
+    if (__left_bound_seq_2 == __last2)
+        return __result;
+
+    const auto __m1 = __last1 - __left_bound_seq_1 + __n2;
+    if (__m1 > __set_algo_cut_off)
+    {
+        //we know proper offset due to [first1; left_bound_seq_1) < [first2; last2)
+        return __internal::__parallel_set_op(
+            ::std::forward<_ExecutionPolicy>(__exec), __left_bound_seq_1, __last1, __first2, __last2, __result, __comp,
+            [](_DifferenceType __n, _DifferenceType __m) { return ::std::min(__n, __m); },
+            [](_RandomAccessIterator1 __first1, _RandomAccessIterator1 __last1, _RandomAccessIterator2 __first2,
+               _RandomAccessIterator2 __last2, _T* __result, _Compare __comp) {
+                return oneapi::dpl::__utils::__set_intersection_construct(__first1, __last1, __first2, __last2,
+                                                                          __result, __comp);
+            },
+            _IsVector{});
+    }
+
+    const auto __m2 = __last2 - __left_bound_seq_2 + __n1;
+    if (__m2 > __set_algo_cut_off)
+    {
+        //we know proper offset due to [first2; left_bound_seq_2) < [first1; last1)
+        __result = __internal::__parallel_set_op(
+            ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __left_bound_seq_2, __last2, __result, __comp,
+            [](_DifferenceType __n, _DifferenceType __m) { return ::std::min(__n, __m); },
+            [](_RandomAccessIterator1 __first1, _RandomAccessIterator1 __last1, _RandomAccessIterator2 __first2,
+               _RandomAccessIterator2 __last2, _T* __result, _Compare __comp) {
+                return oneapi::dpl::__utils::__set_intersection_construct(__first2, __last2, __first1, __last1,
+                                                                          __result, __comp);
+            },
+            _IsVector{});
+        return __result;
+    }
+
+    // [left_bound_seq_1; last1) and [left_bound_seq_2; last2) - use serial algorithm
+    return ::std::set_intersection(__left_bound_seq_1, __last1, __left_bound_seq_2, __last2, __result, __comp);
+}
+
 //------------------------------------------------------------------------
 // set_difference
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index a28a3da8531..479800c16dc 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -2922,6 +2922,21 @@ __pattern_set_intersection(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1
                                    __result, __comp, unseq_backend::_IntersectionTag());
 }
 
+template <typename _BackendTag, typename _ExecutionPolicy, typename _ForwardIterator1, typename _ForwardIterator2,
+          typename _OutputIterator, typename _Compare>
+_OutputIterator
+__pattern_set_intersection(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _ForwardIterator1 __first1,
+                           _ForwardIterator1 __last1, _ForwardIterator2 __first2, _ForwardIterator2 __last2,
+                           _OutputIterator __result, _Compare __comp)
+{
+    // intersection is empty
+    if (__first1 == __last1 || __first2 == __last2)
+        return __result;
+
+    return __pattern_hetero_set_op(::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2,
+                                   __result, __comp, unseq_backend::_IntersectionTag());
+}
+
 //Dummy names to avoid kernel problems
 template <typename Name>
 class __set_difference_copy_case_1

From 22e757aac474dca124ee8976a0e1b50c5f7385ee Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 7 Feb 2024 10:12:33 +0100
Subject: [PATCH 134/566] __pattern_set_intersection + tag calls

---
 include/oneapi/dpl/pstl/glue_algorithm_impl.h | 9 +++++----
 1 file changed, 5 insertions(+), 4 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_algorithm_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
index e5a7dc854c9..0ef89eae00c 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
@@ -1021,10 +1021,11 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _Forward
 set_intersection(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
                  _ForwardIterator2 __first2, _ForwardIterator2 __last2, _ForwardIterator __result, _Compare __comp)
 {
-    return oneapi::dpl::__internal::__pattern_set_intersection(
-        ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2, __result, __comp,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2, _ForwardIterator>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2, _ForwardIterator>());
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1,
+                                                                              _ForwardIterator2, _ForwardIterator>();
+
+    return oneapi::dpl::__internal::__pattern_set_intersection(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
+                                                               __first1, __last1, __first2, __last2, __result, __comp);
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _ForwardIterator>

From bb130a376fb2fdc995a9d14a3a0de52a235716d5 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 7 Feb 2024 10:25:56 +0100
Subject: [PATCH 135/566] __pattern_set_symmetric_difference + tag impls

---
 include/oneapi/dpl/pstl/algorithm_fwd.h       | 13 ++++
 include/oneapi/dpl/pstl/algorithm_impl.h      | 39 +++++++++++
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 64 ++++++++++++++++++-
 3 files changed, 115 insertions(+), 1 deletion(-)

diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index 4774d55213e..f81eeea1f32 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -1600,6 +1600,12 @@ __pattern_set_symmetric_difference(_ExecutionPolicy&&, _ForwardIterator1, _Forwa
                                    _ForwardIterator2, _OutputIterator, _Compare, _IsVector,
                                    /*is_parallel=*/::std::false_type) noexcept;
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _OutputIterator,
+          class _Compare>
+_OutputIterator
+__pattern_set_symmetric_difference(_Tag, _ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2,
+                                   _ForwardIterator2, _OutputIterator, _Compare) noexcept;
+
 template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _OutputIterator,
           class _Compare, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _OutputIterator>
@@ -1607,6 +1613,13 @@ __pattern_set_symmetric_difference(_ExecutionPolicy&&, _RandomAccessIterator1, _
                                    _RandomAccessIterator2, _RandomAccessIterator2, _OutputIterator, _Compare, _IsVector,
                                    /*is_parallel=*/::std::true_type);
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
+          class _RandomAccessIterator3, class _Compare>
+_RandomAccessIterator3
+__pattern_set_symmetric_difference(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator1,
+                                   _RandomAccessIterator1, _RandomAccessIterator2, _RandomAccessIterator2,
+                                   _RandomAccessIterator3, _Compare);
+
 //------------------------------------------------------------------------
 // is_heap_until
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 34e00c2e251..2d27bc0fd14 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -5325,6 +5325,19 @@ __pattern_set_symmetric_difference(_ExecutionPolicy&&, _ForwardIterator1 __first
                                                         __is_vector);
 }
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _OutputIterator,
+          class _Compare>
+_OutputIterator
+__pattern_set_symmetric_difference(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
+                                   _ForwardIterator2 __first2, _ForwardIterator2 __last2, _OutputIterator __result,
+                                   _Compare __comp) noexcept
+{
+    static_assert(__is_backend_tag_v<_Tag>);
+
+    return __internal::__brick_set_symmetric_difference(__first1, __last1, __first2, __last2, __result, __comp,
+                                                        typename _Tag::__is_vector{});
+}
+
 template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
           class _RandomAccessIterator3, class _Compare, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator3>
@@ -5352,6 +5365,32 @@ __pattern_set_symmetric_difference(_ExecutionPolicy&& __exec, _RandomAccessItera
         __is_vector);
 }
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
+          class _RandomAccessIterator3, class _Compare>
+_RandomAccessIterator3
+__pattern_set_symmetric_difference(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec,
+                                   _RandomAccessIterator1 __first1, _RandomAccessIterator1 __last1,
+                                   _RandomAccessIterator2 __first2, _RandomAccessIterator2 __last2,
+                                   _RandomAccessIterator3 __result, _Compare __comp)
+{
+    const auto __n1 = __last1 - __first1;
+    const auto __n2 = __last2 - __first2;
+
+    // use serial algorithm
+    if (__n1 + __n2 <= __set_algo_cut_off)
+        return ::std::set_symmetric_difference(__first1, __last1, __first2, __last2, __result, __comp);
+
+    typedef typename ::std::iterator_traits<_RandomAccessIterator3>::value_type _T;
+    return __internal::__parallel_set_union_op(
+        ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2, __result, __comp,
+        [](_RandomAccessIterator1 __first1, _RandomAccessIterator1 __last1, _RandomAccessIterator2 __first2,
+           _RandomAccessIterator2 __last2, _T* __result, _Compare __comp) {
+            return oneapi::dpl::__utils::__set_symmetric_difference_construct(
+                __first1, __last1, __first2, __last2, __result, __comp, __BrickCopyConstruct<_IsVector>());
+        },
+        _IsVector{});
+}
+
 //------------------------------------------------------------------------
 // is_heap_until
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 479800c16dc..45560162836 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -3150,7 +3150,69 @@ __pattern_set_symmetric_difference(_ExecutionPolicy&& __exec, _ForwardIterator1
                                                                                decltype(__buf_2), decltype(__result)>();
     return oneapi::dpl::__internal::__pattern_merge(__dispatch_tag1, ::std::forward<_ExecutionPolicy>(__exec), __buf_1,
                                                     __buf_1 + __n_diff_1, __buf_2, __buf_2 + __n_diff_2, __result,
-                                                    __comp, ::std::true_type(), ::std::true_type());
+                                                    __comp);
+}
+
+template <typename _BackendTag, typename _ExecutionPolicy, typename _ForwardIterator1, typename _ForwardIterator2,
+          typename _OutputIterator, typename _Compare>
+_OutputIterator
+__pattern_set_symmetric_difference(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec,
+                                   _ForwardIterator1 __first1, _ForwardIterator1 __last1, _ForwardIterator2 __first2,
+                                   _ForwardIterator2 __last2, _OutputIterator __result, _Compare __comp)
+{
+    if (__first1 == __last1 && __first2 == __last2)
+        return __result;
+
+    //{1} is empty
+    if (__first1 == __last1)
+    {
+        return oneapi::dpl::__internal::__pattern_walk2_brick(
+            __tag,
+            oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__set_symmetric_difference_copy_case_1>(
+                ::std::forward<_ExecutionPolicy>(__exec)),
+            __first2, __last2, __result, oneapi::dpl::__internal::__brick_copy<_ExecutionPolicy>{});
+    }
+
+    //{2} is empty
+    if (__first2 == __last2)
+    {
+        return oneapi::dpl::__internal::__pattern_walk2_brick(
+            __tag,
+            oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__set_symmetric_difference_copy_case_2>(
+                ::std::forward<_ExecutionPolicy>(__exec)),
+            __first1, __last1, __result, oneapi::dpl::__internal::__brick_copy<_ExecutionPolicy>{});
+    }
+
+    typedef typename ::std::iterator_traits<_OutputIterator>::value_type _ValueType;
+
+    // temporary buffers to store intermediate result
+    const auto __n1 = __last1 - __first1;
+    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __diff_1(__exec, __n1);
+    auto __buf_1 = __diff_1.get();
+    const auto __n2 = __last2 - __first2;
+    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __diff_2(__exec, __n2);
+    auto __buf_2 = __diff_2.get();
+
+    //1. Calc difference {1} \ {2}
+    const auto __n_diff_1 =
+        oneapi::dpl::__internal::__pattern_hetero_set_op(
+            oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__set_symmetric_difference_phase_1>(__exec),
+            __first1, __last1, __first2, __last2, __buf_1, __comp, unseq_backend::_DifferenceTag()) -
+        __buf_1;
+
+    //2. Calc difference {2} \ {1}
+    const auto __n_diff_2 =
+        oneapi::dpl::__internal::__pattern_hetero_set_op(
+            oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__set_symmetric_difference_phase_2>(__exec),
+            __first2, __last2, __first1, __last1, __buf_2, __comp, unseq_backend::_DifferenceTag()) -
+        __buf_2;
+
+    //3. Merge the differences
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, decltype(__buf_1),
+                                                                              decltype(__buf_2), decltype(__result)>();
+    return oneapi::dpl::__internal::__pattern_merge(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __buf_1,
+                                                    __buf_1 + __n_diff_1, __buf_2, __buf_2 + __n_diff_2, __result,
+                                                    __comp);
 }
 
 template <typename _Name>

From 7bc65aa5b92106b68685b8d074e361cada8b1ec2 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 7 Feb 2024 10:26:18 +0100
Subject: [PATCH 136/566] __pattern_set_symmetric_difference + tag calls

---
 include/oneapi/dpl/pstl/glue_algorithm_impl.h | 8 +++++---
 1 file changed, 5 insertions(+), 3 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_algorithm_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
index 0ef89eae00c..063349837ff 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
@@ -1070,10 +1070,12 @@ set_symmetric_difference(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1,
                          _ForwardIterator2 __first2, _ForwardIterator2 __last2, _ForwardIterator __result,
                          _Compare __comp)
 {
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1,
+                                                                              _ForwardIterator2, _ForwardIterator>();
+
     return oneapi::dpl::__internal::__pattern_set_symmetric_difference(
-        ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2, __result, __comp,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2, _ForwardIterator>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2, _ForwardIterator>());
+        __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2, __result,
+        __comp);
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _ForwardIterator>

From d14cbf6536974cff7ce8fcfc1e2b5f059eed6d5e Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 7 Feb 2024 10:41:18 +0100
Subject: [PATCH 137/566] __pattern_set_union + tag impls

---
 include/oneapi/dpl/pstl/algorithm_fwd.h       | 12 +++++
 include/oneapi/dpl/pstl/algorithm_impl.h      | 38 +++++++++++++
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 53 +++++++++++++++++++
 3 files changed, 103 insertions(+)

diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index f81eeea1f32..e60f8c82b10 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -1493,12 +1493,24 @@ oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _Ou
 __pattern_set_union(_ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2, _ForwardIterator2,
                     _OutputIterator, _Compare, _IsVector, /*is_parallel=*/::std::false_type) noexcept;
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _OutputIterator,
+          class _Compare>
+_OutputIterator
+__pattern_set_union(_Tag, _ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2,
+                    _ForwardIterator2, _OutputIterator, _Compare) noexcept;
+
 template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _OutputIterator,
           class _Compare, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _OutputIterator>
 __pattern_set_union(_ExecutionPolicy&&, _RandomAccessIterator1, _RandomAccessIterator1, _RandomAccessIterator2,
                     _RandomAccessIterator2, _OutputIterator, _Compare, _IsVector, /*is_parallel=*/::std::true_type);
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
+          class _OutputIterator, class _Compare>
+_OutputIterator
+__pattern_set_union(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator1, _RandomAccessIterator1,
+                    _RandomAccessIterator2, _RandomAccessIterator2, _OutputIterator, _Compare);
+
 //------------------------------------------------------------------------
 // set_intersection
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 2d27bc0fd14..b6117538af8 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -4936,6 +4936,19 @@ __pattern_set_union(_ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIter
     return __internal::__brick_set_union(__first1, __last1, __first2, __last2, __result, __comp, __is_vector);
 }
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _OutputIterator,
+          class _Compare>
+_OutputIterator
+__pattern_set_union(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
+                    _ForwardIterator2 __first2, _ForwardIterator2 __last2, _OutputIterator __result,
+                    _Compare __comp) noexcept
+{
+    static_assert(__is_backend_tag_v<_Tag>);
+
+    return __internal::__brick_set_union(__first1, __last1, __first2, __last2, __result, __comp,
+                                         typename _Tag::__is_vector{});
+}
+
 template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _OutputIterator,
           class _Compare, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _OutputIterator>
@@ -4962,6 +4975,31 @@ __pattern_set_union(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1,
         __is_vector);
 }
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
+          class _OutputIterator, class _Compare>
+_OutputIterator
+__pattern_set_union(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1,
+                    _RandomAccessIterator1 __last1, _RandomAccessIterator2 __first2, _RandomAccessIterator2 __last2,
+                    _OutputIterator __result, _Compare __comp)
+{
+    const auto __n1 = __last1 - __first1;
+    const auto __n2 = __last2 - __first2;
+
+    // use serial algorithm
+    if (__n1 + __n2 <= __set_algo_cut_off)
+        return ::std::set_union(__first1, __last1, __first2, __last2, __result, __comp);
+
+    typedef typename ::std::iterator_traits<_OutputIterator>::value_type _Tp;
+    return __parallel_set_union_op(
+        ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2, __result, __comp,
+        [](_RandomAccessIterator1 __first1, _RandomAccessIterator1 __last1, _RandomAccessIterator2 __first2,
+           _RandomAccessIterator2 __last2, _Tp* __result, _Compare __comp) {
+            return oneapi::dpl::__utils::__set_union_construct(__first1, __last1, __first2, __last2, __result, __comp,
+                                                               __BrickCopyConstruct<_IsVector>());
+        },
+        _IsVector{});
+}
+
 //------------------------------------------------------------------------
 // set_intersection
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 45560162836..4c5475ac68a 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -3060,6 +3060,59 @@ __pattern_set_union(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _Forw
         __first1, __last1, __buf, __buf + __n_diff, __result, __comp);
 }
 
+template <typename _BackendTag, typename _ExecutionPolicy, typename _ForwardIterator1, typename _ForwardIterator2,
+          typename _OutputIterator, typename _Compare>
+_OutputIterator
+__pattern_set_union(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _ForwardIterator1 __first1,
+                    _ForwardIterator1 __last1, _ForwardIterator2 __first2, _ForwardIterator2 __last2,
+                    _OutputIterator __result, _Compare __comp)
+{
+    if (__first1 == __last1 && __first2 == __last2)
+        return __result;
+
+    //{1} is empty
+    if (__first1 == __last1)
+    {
+        return oneapi::dpl::__internal::__pattern_walk2_brick(
+            __tag,
+            oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__set_union_copy_case_1>(
+                ::std::forward<_ExecutionPolicy>(__exec)),
+            __first2, __last2, __result, oneapi::dpl::__internal::__brick_copy<_ExecutionPolicy>{});
+    }
+
+    //{2} is empty
+    if (__first2 == __last2)
+    {
+        return oneapi::dpl::__internal::__pattern_walk2_brick(
+            __tag,
+            oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__set_union_copy_case_2>(
+                ::std::forward<_ExecutionPolicy>(__exec)),
+            __first1, __last1, __result, oneapi::dpl::__internal::__brick_copy<_ExecutionPolicy>{});
+    }
+
+    typedef typename ::std::iterator_traits<_OutputIterator>::value_type _ValueType;
+
+    // temporary buffer to store intermediate result
+    const auto __n2 = __last2 - __first2;
+    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __diff(__exec, __n2);
+    auto __buf = __diff.get();
+
+    //1. Calc difference {2} \ {1}
+    const auto __n_diff =
+        oneapi::dpl::__internal::__pattern_hetero_set_op(__exec, __first2, __last2, __first1, __last1, __buf, __comp,
+                                                         unseq_backend::_DifferenceTag()) -
+        __buf;
+
+    //2. Merge {1} and the difference
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, decltype(__first1),
+                                                                          decltype(__buf), decltype(__result)>();
+    return oneapi::dpl::__internal::__pattern_merge(
+        __dispatch_tag,
+        oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__set_union_copy_case_2>(
+            ::std::forward<_ExecutionPolicy>(__exec)),
+        __first1, __last1, __buf, __buf + __n_diff, __result, __comp);
+}
+
 //Dummy names to avoid kernel problems
 template <typename Name>
 class __set_symmetric_difference_copy_case_1

From 12f9c2821890bd136fcdd96c2fabdcc7d4d1b2c8 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 7 Feb 2024 10:42:16 +0100
Subject: [PATCH 138/566] __pattern_set_union + tag calls

---
 include/oneapi/dpl/pstl/glue_algorithm_impl.h | 9 +++++----
 1 file changed, 5 insertions(+), 4 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_algorithm_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
index 063349837ff..1e0a7ef92ed 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
@@ -998,10 +998,11 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _Forward
 set_union(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 __last1, _ForwardIterator2 __first2,
           _ForwardIterator2 __last2, _ForwardIterator __result, _Compare __comp)
 {
-    return oneapi::dpl::__internal::__pattern_set_union(
-        ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2, __result, __comp,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2, _ForwardIterator>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2, _ForwardIterator>());
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1,
+                                                                              _ForwardIterator2, _ForwardIterator>();
+
+    return oneapi::dpl::__internal::__pattern_set_union(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
+                                                        __first1, __last1, __first2, __last2, __result, __comp);
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _ForwardIterator>

From c33acee85fb5d2b2fdc49ddb99b4e336a431675e Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 7 Feb 2024 10:56:35 +0100
Subject: [PATCH 139/566] __pattern_shift_left + tag impls

---
 include/oneapi/dpl/pstl/algorithm_fwd.h       | 10 ++++
 include/oneapi/dpl/pstl/algorithm_impl.h      | 55 +++++++++++++++++++
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 20 +++++++
 3 files changed, 85 insertions(+)

diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index e60f8c82b10..2946cbd7f5c 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -1862,12 +1862,22 @@ __pattern_shift_left(_ExecutionPolicy&&, _ForwardIterator, _ForwardIterator,
                      typename ::std::iterator_traits<_ForwardIterator>::difference_type, _IsVector,
                      /*is_parallel=*/::std::false_type) noexcept;
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator>
+_ForwardIterator
+__pattern_shift_left(_Tag, _ExecutionPolicy&&, _ForwardIterator, _ForwardIterator,
+                     typename ::std::iterator_traits<_ForwardIterator>::difference_type) noexcept;
+
 template <class _ExecutionPolicy, class _ForwardIterator, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _ForwardIterator>
 __pattern_shift_left(_ExecutionPolicy&&, _ForwardIterator, _ForwardIterator,
                      typename ::std::iterator_traits<_ForwardIterator>::difference_type, _IsVector,
                      /*is_parallel=*/::std::true_type);
 
+template <class _IsVector, class _ExecutionPolicy, class _ForwardIterator>
+_ForwardIterator
+__pattern_shift_left(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _ForwardIterator, _ForwardIterator,
+                     typename ::std::iterator_traits<_ForwardIterator>::difference_type);
+
 template <class _ExecutionPolicy, class _BidirectionalIterator, class _IsVector, class _IsParallel>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _BidirectionalIterator>
 __pattern_shift_right(_ExecutionPolicy&&, _BidirectionalIterator, _BidirectionalIterator,
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index b6117538af8..4ac553ab79d 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -6148,6 +6148,16 @@ __pattern_shift_left(_ExecutionPolicy&&, _ForwardIterator __first, _ForwardItera
     return __brick_shift_left(__first, __last, __n, __is_vector);
 }
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator>
+_ForwardIterator
+__pattern_shift_left(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
+                     typename ::std::iterator_traits<_ForwardIterator>::difference_type __n) noexcept
+{
+    static_assert(__is_backend_tag_v<_Tag>);
+
+    return __brick_shift_left(__first, __last, __n, typename _Tag::__is_vector{});
+}
+
 template <class _ExecutionPolicy, class _ForwardIterator, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _ForwardIterator>
 __pattern_shift_left(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last,
@@ -6192,6 +6202,51 @@ __pattern_shift_left(_ExecutionPolicy&& __exec, _ForwardIterator __first, _Forwa
     return __first + __size_res;
 }
 
+template <class _IsVector, class _ExecutionPolicy, class _ForwardIterator>
+_ForwardIterator
+__pattern_shift_left(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _ForwardIterator __first,
+                     _ForwardIterator __last, typename ::std::iterator_traits<_ForwardIterator>::difference_type __n)
+{
+    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
+
+    //If (n > 0 && n < m), returns first + (m - n). Otherwise, if n  > 0, returns first. Otherwise, returns last.
+    if (__n <= 0)
+        return __last;
+    auto __size = __last - __first;
+    if (__n >= __size)
+        return __first;
+
+    using _DiffType = typename ::std::iterator_traits<_ForwardIterator>::difference_type;
+
+    _DiffType __mid = __size / 2 + __size % 2;
+    _DiffType __size_res = __size - __n;
+
+    //1. n >= size/2; there is enough memory to 'total' parallel copying
+    if (__n >= __mid)
+    {
+        __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __n, __size,
+                                      [__first, __n](_DiffType __i, _DiffType __j) {
+                                          __brick_move<_ExecutionPolicy>{}(__first + __i, __first + __j,
+                                                                           __first + __i - __n, _IsVector{});
+                                      });
+    }
+    else //2. n < size/2; there is not enough memory to parallel copying; doing parallel copying by n elements
+    {
+        //TODO: to consider parallel processing by the 'internal' loop (but we may probably get cache locality issues)
+        for (auto __k = __n; __k < __size; __k += __n)
+        {
+            auto __end = ::std::min(__k + __n, __size);
+            __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __k, __end,
+                                          [__first, __n](_DiffType __i, _DiffType __j) {
+                                              __brick_move<_ExecutionPolicy>{}(__first + __i, __first + __j,
+                                                                               __first + __i - __n, _IsVector{});
+                                          });
+        }
+    }
+
+    return __first + __size_res;
+}
+
 template <class _ExecutionPolicy, class _BidirectionalIterator, class _IsVector, class _IsParallel>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _BidirectionalIterator>
 __pattern_shift_right(_ExecutionPolicy&& __exec, _BidirectionalIterator __first, _BidirectionalIterator __last,
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 4c5475ac68a..2dcf2ccc8b2 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -3358,6 +3358,26 @@ __pattern_shift_right(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __
     return __last - __res;
 }
 
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator>
+_Iterator
+__pattern_shift_left(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last,
+                     typename ::std::iterator_traits<_Iterator>::difference_type __n)
+{
+    //If (n > 0 && n < m), returns first + (m - n). Otherwise, if n  > 0, returns first. Otherwise, returns last.
+    auto __size = __last - __first;
+    if (__n <= 0)
+        return __last;
+    if (__n >= __size)
+        return __first;
+
+    auto __keep = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read_write, _Iterator>();
+    auto __buf = __keep(__first, __last);
+
+    auto __res =
+        oneapi::dpl::__internal::__pattern_shift_left(::std::forward<_ExecutionPolicy>(__exec), __buf.all_view(), __n);
+    return __first + __res;
+}
+
 } // namespace __internal
 } // namespace dpl
 } // namespace oneapi

From 7660a96c5f140c6c505ed563f9ebbb73dbbef08a Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 7 Feb 2024 10:56:46 +0100
Subject: [PATCH 140/566] __pattern_shift_left + tag calls

---
 include/oneapi/dpl/pstl/glue_algorithm_impl.h | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_algorithm_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
index 1e0a7ef92ed..376c39fef56 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
@@ -1235,10 +1235,10 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _Forward
 shift_left(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last,
            typename ::std::iterator_traits<_ForwardIterator>::difference_type __n)
 {
-    return oneapi::dpl::__internal::__pattern_shift_left(
-        ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __n,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>());
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+
+    return oneapi::dpl::__internal::__pattern_shift_left(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
+                                                         __first, __last, __n);
 }
 
 // [shift.right]

From 2e12b116291fe90c4ccfa3e711ca03f89d44ee77 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 7 Feb 2024 11:20:01 +0100
Subject: [PATCH 141/566] __pattern_shift_right + tag impls

---
 include/oneapi/dpl/pstl/algorithm_impl.h      | 20 +++++++++++++++++
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 22 +++++++++++++++++++
 2 files changed, 42 insertions(+)

diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 4ac553ab79d..46771e8cf72 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -6261,6 +6261,26 @@ __pattern_shift_right(_ExecutionPolicy&& __exec, _BidirectionalIterator __first,
     return __res.base();
 }
 
+template <class _Tag, class _ExecutionPolicy, class _BidirectionalIterator>
+_BidirectionalIterator
+__pattern_shift_right(_Tag, _ExecutionPolicy&& __exec, _BidirectionalIterator __first, _BidirectionalIterator __last,
+                      typename ::std::iterator_traits<_BidirectionalIterator>::difference_type __n)
+{
+    static_assert(__is_backend_tag_v<_Tag>);
+
+    using _ReverseIterator = typename ::std::reverse_iterator<_BidirectionalIterator>;
+
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, decltype(_ReverseIterator(__last)),
+                                                  decltype(_ReverseIterator(__first))>();
+
+    auto __res =
+        oneapi::dpl::__internal::__pattern_shift_left(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
+                                                      _ReverseIterator(__last), _ReverseIterator(__first), __n);
+
+    return __res.base();
+}
+
 } // namespace __internal
 } // namespace dpl
 } // namespace oneapi
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 2dcf2ccc8b2..9dffa37b99f 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -3358,6 +3358,28 @@ __pattern_shift_right(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __
     return __last - __res;
 }
 
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator>
+_Iterator
+__pattern_shift_right(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last,
+                      typename ::std::iterator_traits<_Iterator>::difference_type __n)
+{
+    //If (n > 0 && n < m), returns first + n. Otherwise, if n  > 0, returns last. Otherwise, returns first.
+    auto __size = __last - __first;
+    if (__n <= 0)
+        return __first;
+    if (__n >= __size)
+        return __last;
+
+    auto __keep = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read_write, _Iterator>();
+    auto __buf = __keep(__first, __last);
+
+    //A shift right is the shift left with a reverse logic.
+    auto __rng = oneapi::dpl::__ranges::reverse_view_simple<decltype(__buf.all_view())>{__buf.all_view()};
+    auto __res = oneapi::dpl::__internal::__pattern_shift_left(::std::forward<_ExecutionPolicy>(__exec), __rng, __n);
+
+    return __last - __res;
+}
+
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator>
 _Iterator
 __pattern_shift_left(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last,

From 6caf1f20b0d938a786450f1bc5d6d7a92e69ef8d Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 7 Feb 2024 11:20:15 +0100
Subject: [PATCH 142/566] __pattern_shift_right + tag calls

---
 include/oneapi/dpl/pstl/glue_algorithm_impl.h | 9 +++++----
 1 file changed, 5 insertions(+), 4 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_algorithm_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
index 376c39fef56..9bd59d079a7 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
@@ -1248,10 +1248,11 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _Bidirec
 shift_right(_ExecutionPolicy&& __exec, _BidirectionalIterator __first, _BidirectionalIterator __last,
             typename ::std::iterator_traits<_BidirectionalIterator>::difference_type __n)
 {
-    return oneapi::dpl::__internal::__pattern_shift_right(
-        ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __n,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _BidirectionalIterator>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _BidirectionalIterator>());
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _BidirectionalIterator>();
+
+    return oneapi::dpl::__internal::__pattern_shift_right(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
+                                                          __first, __last, __n);
 }
 
 } // namespace dpl

From 6c4634c165472c68d173fa7e8853b6d467ea901e Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 7 Feb 2024 11:33:35 +0100
Subject: [PATCH 143/566] __pattern_sort + tag impls

---
 include/oneapi/dpl/pstl/algorithm_fwd.h       | 10 +++++++
 include/oneapi/dpl/pstl/algorithm_impl.h      | 26 +++++++++++++++++++
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   |  9 +++++++
 3 files changed, 45 insertions(+)

diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index 2946cbd7f5c..26c0c17a261 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -1088,12 +1088,22 @@ oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>
 __pattern_sort(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _Compare, _IsVector /*is_vector*/,
                /*is_parallel=*/::std::false_type, _IsMoveConstructible) noexcept;
 
+template <class _Tag, class _ExecutionPolicy, class _RandomAccessIterator, class _Compare, class _IsMoveConstructible>
+void
+__pattern_sort(_Tag, _ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _Compare,
+               _IsMoveConstructible) noexcept;
+
 template <class _ExecutionPolicy, class _RandomAccessIterator, class _Compare, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>
 __pattern_sort(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _Compare, _IsVector /*is_vector*/,
                /*is_parallel=*/::std::true_type,
                /*is_move_constructible=*/::std::true_type);
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Compare>
+void
+__pattern_sort(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _Compare,
+               /*is_move_constructible=*/::std::true_type);
+
 //------------------------------------------------------------------------
 // stable_sort
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 46771e8cf72..12e119ce397 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -3436,6 +3436,16 @@ __pattern_sort(_ExecutionPolicy&&, _RandomAccessIterator __first, _RandomAccessI
     ::std::sort(__first, __last, __comp);
 }
 
+template <class _Tag, class _ExecutionPolicy, class _RandomAccessIterator, class _Compare, class _IsMoveConstructible>
+void
+__pattern_sort(_Tag, _ExecutionPolicy&&, _RandomAccessIterator __first, _RandomAccessIterator __last, _Compare __comp,
+               _IsMoveConstructible) noexcept
+{
+    static_assert(__is_backend_tag_v<_Tag>);
+
+    ::std::sort(__first, __last, __comp);
+}
+
 template <class _ExecutionPolicy, class _RandomAccessIterator, class _Compare, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>
 __pattern_sort(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIterator __last, _Compare __comp,
@@ -3451,6 +3461,22 @@ __pattern_sort(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _Random
     });
 }
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Compare>
+void
+__pattern_sort(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
+               _RandomAccessIterator __last, _Compare __comp,
+               /*is_move_constructible=*/::std::true_type)
+{
+    __internal::__except_handler([&]() {
+        __par_backend::__parallel_stable_sort(
+            ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __comp,
+            [](_RandomAccessIterator __first, _RandomAccessIterator __last, _Compare __comp) {
+                ::std::sort(__first, __last, __comp);
+            },
+            __last - __first);
+    });
+}
+
 //------------------------------------------------------------------------
 // stable_sort
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 9dffa37b99f..f76b3f9a660 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -2100,6 +2100,15 @@ __pattern_sort(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last, _
                                   oneapi::dpl::identity{});
 }
 
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator, typename _Compare>
+void
+__pattern_sort(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last,
+               _Compare __comp, /*is_move_constructible=*/::std::true_type)
+{
+    __stable_sort_with_projection(::std::forward<_ExecutionPolicy>(__exec), __first, __last, __comp,
+                                  oneapi::dpl::identity{});
+}
+
 //------------------------------------------------------------------------
 // stable_sort
 //------------------------------------------------------------------------

From 08bd8ba9dc63e6bbfc67e0f9320833de9ea3b2fa Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 7 Feb 2024 11:34:25 +0100
Subject: [PATCH 144/566] __pattern_sort + tag calls

---
 include/oneapi/dpl/pstl/glue_algorithm_impl.h          | 7 +++++--
 include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h | 7 +++++--
 2 files changed, 10 insertions(+), 4 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_algorithm_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
index 9bd59d079a7..d187b575446 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
@@ -672,11 +672,14 @@ template <class _ExecutionPolicy, class _RandomAccessIterator, class _Compare>
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy>
 sort(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIterator __last, _Compare __comp)
 {
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
+
     typedef typename ::std::iterator_traits<_RandomAccessIterator>::value_type _InputType;
+
     oneapi::dpl::__internal::__pattern_sort(
+        __dispatch_tag,
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __comp,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _RandomAccessIterator>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _RandomAccessIterator>(),
         typename ::std::is_move_constructible<_InputType>::type());
 }
 
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index f76b3f9a660..39a8a039bd6 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -2536,9 +2536,12 @@ __pattern_partial_sort_copy(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&&
             __brick_copy<_ExecutionPolicy>{});
 
         // Use regular sort as partial_sort isn't required to be stable
-        __pattern_sort(
+        constexpr auto __dispatch_tag =
+            oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, decltype(__out_first), decltype(__out_end)>();
+
+        __pattern_sort(__dispatch_tag,
             __par_backend_hetero::make_wrapped_policy<__partial_sort_1>(::std::forward<_ExecutionPolicy>(__exec)),
-            __out_first, __out_end, __comp, ::std::true_type{}, ::std::true_type{}, ::std::true_type{});
+            __out_first, __out_end, __comp, ::std::true_type{});
 
         return __out_end;
     }

From 8c29d9245b155f94b7ba9a6c9fe243c27a4414fc Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 7 Feb 2024 11:45:24 +0100
Subject: [PATCH 145/566] __pattern_sort_by_key + tag impls

---
 include/oneapi/dpl/pstl/algorithm_fwd.h       | 12 ++++++
 include/oneapi/dpl/pstl/algorithm_impl.h      | 42 +++++++++++++++++++
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 14 +++++++
 3 files changed, 68 insertions(+)

diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index 26c0c17a261..0be83b0cd80 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -1130,6 +1130,12 @@ __pattern_sort_by_key(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __keys_f
                       _RandomAccessIterator1 __keys_last, _RandomAccessIterator2 __values_first, _Compare __comp,
                       _IsVector /*vector=*/, /*is_parallel=*/::std::false_type) noexcept;
 
+template <class _Tag, typename _ExecutionPolicy, typename _RandomAccessIterator1, typename _RandomAccessIterator2,
+          typename _Compare>
+void
+__pattern_sort_by_key(_Tag, _ExecutionPolicy&&, _RandomAccessIterator1, _RandomAccessIterator1, _RandomAccessIterator2,
+                      _Compare) noexcept;
+
 template <typename _ExecutionPolicy, typename _RandomAccessIterator1, typename _RandomAccessIterator2,
           typename _Compare, typename _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>
@@ -1137,6 +1143,12 @@ __pattern_sort_by_key(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __keys_f
                       _RandomAccessIterator1 __keys_last, _RandomAccessIterator2 __values_first, _Compare __comp,
                       _IsVector /*vector=*/, /*is_parallel=*/::std::true_type);
 
+template <typename _IsVector, typename _ExecutionPolicy, typename _RandomAccessIterator1,
+          typename _RandomAccessIterator2, typename _Compare>
+void
+__pattern_sort_by_key(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator1, _RandomAccessIterator1,
+                      _RandomAccessIterator2, _Compare);
+
 //------------------------------------------------------------------------
 // partial_sort
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 12e119ce397..025e442cbb1 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -3524,6 +3524,24 @@ __pattern_sort_by_key(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __keys_f
     ::std::sort(__beg, __end, __cmp_f);
 }
 
+template <class _Tag, typename _ExecutionPolicy, typename _RandomAccessIterator1, typename _RandomAccessIterator2,
+          typename _Compare>
+void
+__pattern_sort_by_key(_Tag, _ExecutionPolicy&& __exec, _RandomAccessIterator1 __keys_first,
+                      _RandomAccessIterator1 __keys_last, _RandomAccessIterator2 __values_first,
+                      _Compare __comp) noexcept
+{
+    static_assert(__is_backend_tag_v<_Tag>);
+
+    auto __beg = oneapi::dpl::make_zip_iterator(__keys_first, __values_first);
+    auto __end = __beg + (__keys_last - __keys_first);
+    auto __cmp_f = [__comp](const auto& __a, const auto& __b) {
+        return __comp(::std::get<0>(__a), ::std::get<0>(__b));
+    };
+
+    ::std::sort(__beg, __end, __cmp_f);
+}
+
 template <typename _ExecutionPolicy, typename _RandomAccessIterator1, typename _RandomAccessIterator2,
           typename _Compare, typename _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>
@@ -3549,6 +3567,30 @@ __pattern_sort_by_key(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __keys_f
     });
 }
 
+template <typename _IsVector, typename _ExecutionPolicy, typename _RandomAccessIterator1,
+          typename _RandomAccessIterator2, typename _Compare>
+void
+__pattern_sort_by_key(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator1 __keys_first,
+                      _RandomAccessIterator1 __keys_last, _RandomAccessIterator2 __values_first, _Compare __comp)
+{
+    static_assert(
+        ::std::is_move_constructible_v<typename ::std::iterator_traits<_RandomAccessIterator1>::value_type> &&
+            ::std::is_move_constructible_v<typename ::std::iterator_traits<_RandomAccessIterator2>::value_type>,
+        "The keys and values should be move constructible in case of parallel execution.");
+
+    auto __beg = oneapi::dpl::make_zip_iterator(__keys_first, __values_first);
+    auto __end = __beg + (__keys_last - __keys_first);
+    auto __cmp_f = [__comp](const auto& __a, const auto& __b) {
+        return __comp(::std::get<0>(__a), ::std::get<0>(__b));
+    };
+
+    __internal::__except_handler([&]() {
+        __par_backend::__parallel_stable_sort(
+            ::std::forward<_ExecutionPolicy>(__exec), __beg, __end, __cmp_f,
+            [](auto __first, auto __last, auto __cmp) { ::std::sort(__first, __last, __cmp); }, __end - __beg);
+    });
+}
+
 //------------------------------------------------------------------------
 // partial_sort
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 39a8a039bd6..141b989fc91 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -2137,6 +2137,20 @@ __pattern_sort_by_key(_ExecutionPolicy&& __exec, _Iterator1 __keys_first, _Itera
                                   [](const auto& __a) { return ::std::get<0>(__a); });
 }
 
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2, typename _Compare>
+void
+__pattern_sort_by_key(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator1 __keys_first,
+                      _Iterator1 __keys_last, _Iterator2 __values_first, _Compare __comp)
+{
+    static_assert(::std::is_move_constructible_v<typename ::std::iterator_traits<_Iterator1>::value_type> &&
+                      ::std::is_move_constructible_v<typename ::std::iterator_traits<_Iterator2>::value_type>,
+                  "The keys and values should be move constructible in case of parallel execution.");
+
+    auto __beg = oneapi::dpl::make_zip_iterator(__keys_first, __values_first);
+    auto __end = __beg + (__keys_last - __keys_first);
+    __stable_sort_with_projection(::std::forward<_ExecutionPolicy>(__exec), __beg, __end, __comp,
+                                  [](const auto& __a) { return ::std::get<0>(__a); });
+}
 
 template <typename _ExecutionPolicy, typename _Iterator, typename _UnaryPredicate>
 oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, _Iterator>

From 22cc335b5813ef4cc63f31cb2f9184a174764a5b Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 7 Feb 2024 11:45:52 +0100
Subject: [PATCH 146/566] __pattern_sort_by_key + tag calls

---
 include/oneapi/dpl/pstl/glue_algorithm_impl.h | 9 +++++----
 1 file changed, 5 insertions(+), 4 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_algorithm_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
index d187b575446..aba552030e0 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
@@ -719,10 +719,11 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy>
 sort_by_key(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __keys_first, _RandomAccessIterator1 __keys_last,
             _RandomAccessIterator2 __values_first, _Compare __comp)
 {
-    oneapi::dpl::__internal::__pattern_sort_by_key(
-        ::std::forward<_ExecutionPolicy>(__exec), __keys_first, __keys_last, __values_first, __comp,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2>());
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2>();
+
+    oneapi::dpl::__internal::__pattern_sort_by_key(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
+                                                   __keys_first, __keys_last, __values_first, __comp);
 }
 
 template <typename _ExecutionPolicy, typename _RandomAccessIterator1, typename _RandomAccessIterator2>

From 38ea27e2d335b3b1744926529f9a0e18509913c6 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 7 Feb 2024 12:00:53 +0100
Subject: [PATCH 147/566] __pattern_stable_partition + tag impls

---
 include/oneapi/dpl/pstl/algorithm_fwd.h       | 10 +++
 include/oneapi/dpl/pstl/algorithm_impl.h      | 61 +++++++++++++++++++
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 46 ++++++++++++++
 3 files changed, 117 insertions(+)

diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index 0be83b0cd80..60b02a29dc8 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -1030,11 +1030,21 @@ __pattern_stable_partition(_ExecutionPolicy&&, _BidirectionalIterator, _Bidirect
                            _IsVector,
                            /*is_parallelization=*/::std::false_type) noexcept;
 
+template <class _Tag, class _ExecutionPolicy, class _BidirectionalIterator, class _UnaryPredicate>
+_BidirectionalIterator
+__pattern_stable_partition(_Tag, _ExecutionPolicy&&, _BidirectionalIterator, _BidirectionalIterator,
+                           _UnaryPredicate) noexcept;
+
 template <class _ExecutionPolicy, class _RandomAccessIterator, class _UnaryPredicate, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator>
 __pattern_stable_partition(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _UnaryPredicate, _IsVector,
                            /*is_parallelization=*/::std::true_type);
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _UnaryPredicate>
+_RandomAccessIterator
+__pattern_stable_partition(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator,
+                           _UnaryPredicate);
+
 //------------------------------------------------------------------------
 // partition_copy
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 025e442cbb1..0744dbe33d0 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -3244,6 +3244,16 @@ __pattern_stable_partition(_ExecutionPolicy&&, _BidirectionalIterator __first, _
     return __internal::__brick_stable_partition(__first, __last, __pred, __is_vector);
 }
 
+template <class _Tag, class _ExecutionPolicy, class _BidirectionalIterator, class _UnaryPredicate>
+_BidirectionalIterator
+__pattern_stable_partition(_Tag, _ExecutionPolicy&&, _BidirectionalIterator __first, _BidirectionalIterator __last,
+                           _UnaryPredicate __pred) noexcept
+{
+    static_assert(__is_backend_tag_v<_Tag>);
+
+    return __internal::__brick_stable_partition(__first, __last, __pred, typename _Tag::__is_vector{});
+}
+
 template <class _ExecutionPolicy, class _RandomAccessIterator, class _UnaryPredicate, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator>
 __pattern_stable_partition(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIterator __last,
@@ -3296,6 +3306,57 @@ __pattern_stable_partition(_ExecutionPolicy&& __exec, _RandomAccessIterator __fi
     });
 }
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _UnaryPredicate>
+_RandomAccessIterator
+__pattern_stable_partition(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
+                           _RandomAccessIterator __last, _UnaryPredicate __pred)
+{
+    // partitioned range: elements before pivot satisfy pred (true part),
+    //                    elements after pivot don't satisfy pred (false part)
+    struct _PartitionRange
+    {
+        _RandomAccessIterator __begin;
+        _RandomAccessIterator __pivot;
+        _RandomAccessIterator __end;
+    };
+
+    return __internal::__except_handler([&]() {
+        _PartitionRange __init{__last, __last, __last};
+
+        // lambda for merging two partitioned ranges to one partitioned range
+        auto __reductor = [](_PartitionRange __val1, _PartitionRange __val2) -> _PartitionRange {
+            auto __size1 = __val1.__end - __val1.__pivot;
+            auto __new_begin = __val2.__begin - (__val1.__end - __val1.__begin);
+
+            // if all elements in left range satisfy pred then we can move new pivot to pivot of right range
+            if (__val1.__end == __val1.__pivot)
+            {
+                return {__new_begin, __val2.__pivot, __val2.__end};
+            }
+            // if true part of right range greater than false part of left range
+            // then we should swap the false part of left range and last part of true part of right range
+            else
+            {
+                __internal::__brick_rotate(__val1.__pivot, __val2.__begin, __val2.__pivot, _IsVector{});
+                return {__new_begin, __val2.__pivot - __size1, __val2.__end};
+            }
+        };
+
+        _PartitionRange __result = __par_backend::__parallel_reduce(
+            ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __init,
+            [&__pred, __reductor](_RandomAccessIterator __i, _RandomAccessIterator __j,
+                                  _PartitionRange __value) -> _PartitionRange {
+                //1. serial stable_partition
+                _RandomAccessIterator __pivot = __internal::__brick_stable_partition(__i, __j, __pred, _IsVector{});
+
+                // 2. merging of two ranges (left and right respectively)
+                return __reductor(__value, {__i, __pivot, __j});
+            },
+            __reductor);
+        return __result.__pivot;
+    });
+}
+
 //------------------------------------------------------------------------
 // partition_copy
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 141b989fc91..03a96684233 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -2203,6 +2203,52 @@ __pattern_stable_partition(_ExecutionPolicy&& __exec, _Iterator __first, _Iterat
     return __first + true_count;
 }
 
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator, typename _UnaryPredicate>
+_Iterator
+__pattern_stable_partition(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator __first,
+                           _Iterator __last, _UnaryPredicate __pred)
+{
+    if (__last == __first)
+        return __last;
+    else if (__last - __first < 2)
+        return __pattern_any_of(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __pred) ? __last
+                                                                                                          : __first;
+
+    using _ValueType = typename ::std::iterator_traits<_Iterator>::value_type;
+
+    auto __n = __last - __first;
+
+    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __true_buf(__exec, __n);
+    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __false_buf(__exec, __n);
+    auto __true_result = __true_buf.get();
+    auto __false_result = __false_buf.get();
+
+    constexpr auto __dispatch_tag = __select_backend<_ExecutionPolicy, decltype(__first), decltype(__last),
+                                                     decltype(__true_result), decltype(__false_result)>();
+    auto copy_result =
+        __pattern_partition_copy(__dispatch_tag, __exec, __first, __last, __true_result, __false_result, __pred);
+    auto true_count = copy_result.first - __true_result;
+
+    //TODO: optimize copy back if possible (inplace, decrease number of submits)
+    constexpr auto __dispatch_tag1 =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, decltype(__true_result),
+                                                  decltype(copy_result.first), decltype(__first)>();
+    using __backend_tag1 = typename decltype(__dispatch_tag1)::__backend_tag;
+    __pattern_walk2<__backend_tag1, /*_IsSync=*/::std::false_type>(
+        __dispatch_tag1, __par_backend_hetero::make_wrapped_policy<copy_back_wrapper>(__exec), __true_result,
+        copy_result.first, __first, __brick_move<_ExecutionPolicy>{});
+
+    constexpr auto __dispatch_tag2 =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, decltype(__false_result),
+                                                  decltype(copy_result.second), decltype(__first + true_count)>();
+    __pattern_walk2(
+        __dispatch_tag2,
+        __par_backend_hetero::make_wrapped_policy<copy_back_wrapper2>(::std::forward<_ExecutionPolicy>(__exec)),
+        __false_result, copy_result.second, __first + true_count, __brick_move<_ExecutionPolicy>{});
+
+    return __first + true_count;
+}
+
 template <typename _ExecutionPolicy, typename _Iterator, typename _UnaryPredicate>
 oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, _Iterator>
 __pattern_partition(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last, _UnaryPredicate __pred,

From a481bf5e739f6dd0d2fbb7c6d1b4b06324765914 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 7 Feb 2024 12:01:44 +0100
Subject: [PATCH 148/566] __pattern_stable_partition + tag calls

---
 include/oneapi/dpl/pstl/glue_algorithm_impl.h          |  9 +++++----
 include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h | 10 ++++++----
 2 files changed, 11 insertions(+), 8 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_algorithm_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
index aba552030e0..27d8afeaa30 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
@@ -646,10 +646,11 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _Bidirec
 stable_partition(_ExecutionPolicy&& __exec, _BidirectionalIterator __first, _BidirectionalIterator __last,
                  _UnaryPredicate __pred)
 {
-    return oneapi::dpl::__internal::__pattern_stable_partition(
-        ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __pred,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _BidirectionalIterator>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _BidirectionalIterator>());
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _BidirectionalIterator>();
+
+    return oneapi::dpl::__internal::__pattern_stable_partition(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
+                                                               __first, __last, __pred);
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator, class _ForwardIterator1, class _ForwardIterator2,
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 03a96684233..38f734668dc 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -2254,9 +2254,10 @@ oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, _
 __pattern_partition(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last, _UnaryPredicate __pred,
                     /*vector*/ ::std::true_type, /*parallel*/ ::std::true_type)
 {
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator>();
+
     //TODO: consider nonstable approaches
-    return __pattern_stable_partition(::std::forward<_ExecutionPolicy>(__exec), __first, __last, __pred,
-                                      ::std::true_type(), ::std::true_type());
+    return __pattern_stable_partition(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __pred);
 }
 
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator, typename _UnaryPredicate>
@@ -2264,9 +2265,10 @@ _Iterator
 __pattern_partition(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last,
                     _UnaryPredicate __pred)
 {
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator>();
+
     //TODO: consider nonstable approaches
-    return __pattern_stable_partition(::std::forward<_ExecutionPolicy>(__exec), __first, __last, __pred,
-                                      ::std::true_type(), ::std::true_type());
+    return __pattern_stable_partition(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __pred);
 }
 
 //------------------------------------------------------------------------

From ed0cab97743318cf619ec8561a9d2758ba3fe28a Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 7 Feb 2024 12:15:20 +0100
Subject: [PATCH 149/566] __pattern_stable_sort + tag impls

---
 include/oneapi/dpl/pstl/algorithm_fwd.h       |  9 +++++++
 include/oneapi/dpl/pstl/algorithm_impl.h      | 24 +++++++++++++++++++
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   |  9 +++++++
 3 files changed, 42 insertions(+)

diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index 60b02a29dc8..73220412b0c 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -1124,12 +1124,21 @@ __pattern_stable_sort(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIt
                       _IsVector /*is_vector*/,
                       /*is_parallel=*/::std::false_type) noexcept;
 
+template <class _Tag, class _ExecutionPolicy, class _RandomAccessIterator, class _Compare>
+void
+__pattern_stable_sort(_Tag, _ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _Compare) noexcept;
+
 template <class _ExecutionPolicy, class _RandomAccessIterator, class _Compare, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>
 __pattern_stable_sort(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _Compare,
                       _IsVector /*is_vector*/,
                       /*is_parallel=*/::std::true_type);
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Compare>
+void
+__pattern_stable_sort(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator,
+                      _Compare);
+
 //------------------------------------------------------------------------
 // sort_by_key
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 0744dbe33d0..9ca164bbdb2 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -3550,6 +3550,15 @@ __pattern_stable_sort(_ExecutionPolicy&&, _RandomAccessIterator __first, _Random
     ::std::stable_sort(__first, __last, __comp);
 }
 
+template <class _Tag, class _ExecutionPolicy, class _RandomAccessIterator, class _Compare>
+void
+__pattern_stable_sort(_Tag, _ExecutionPolicy&&, _RandomAccessIterator __first, _RandomAccessIterator __last, _Compare __comp) noexcept
+{
+    static_assert(__is_backend_tag_v<_Tag>);
+
+    ::std::stable_sort(__first, __last, __comp);
+}
+
 template <class _ExecutionPolicy, class _RandomAccessIterator, class _Compare, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>
 __pattern_stable_sort(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIterator __last,
@@ -3565,6 +3574,21 @@ __pattern_stable_sort(_ExecutionPolicy&& __exec, _RandomAccessIterator __first,
     });
 }
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Compare>
+void
+__pattern_stable_sort(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
+                      _RandomAccessIterator __last, _Compare __comp)
+{
+    __internal::__except_handler([&]() {
+        __par_backend::__parallel_stable_sort(
+            ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __comp,
+            [](_RandomAccessIterator __first, _RandomAccessIterator __last, _Compare __comp) {
+                ::std::stable_sort(__first, __last, __comp);
+            },
+            __last - __first);
+    });
+}
+
 //------------------------------------------------------------------------
 // sort_by_key
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 38f734668dc..cf284bc7c3f 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -2121,6 +2121,15 @@ __pattern_stable_sort(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __
                                   oneapi::dpl::identity{});
 }
 
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator, typename _Compare>
+void
+__pattern_stable_sort(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last,
+                      _Compare __comp)
+{
+    __stable_sort_with_projection(::std::forward<_ExecutionPolicy>(__exec), __first, __last, __comp,
+                                  oneapi::dpl::identity{});
+}
+
 template <typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2, typename _Compare>
 oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy>
 __pattern_sort_by_key(_ExecutionPolicy&& __exec, _Iterator1 __keys_first, _Iterator1 __keys_last,

From 8f6de3844d9adf2a3c8e72268fbbcf075a7ac2e0 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 7 Feb 2024 12:15:30 +0100
Subject: [PATCH 150/566] __pattern_stable_sort + tag calls

---
 include/oneapi/dpl/pstl/glue_algorithm_impl.h | 9 +++++----
 1 file changed, 5 insertions(+), 4 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_algorithm_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
index 27d8afeaa30..abbff064991 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
@@ -698,10 +698,11 @@ template <class _ExecutionPolicy, class _RandomAccessIterator, class _Compare>
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy>
 stable_sort(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIterator __last, _Compare __comp)
 {
-    oneapi::dpl::__internal::__pattern_stable_sort(
-        ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __comp,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _RandomAccessIterator>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _RandomAccessIterator>());
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
+
+    oneapi::dpl::__internal::__pattern_stable_sort(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first,
+                                                   __last, __comp);
 }
 
 template <class _ExecutionPolicy, class _RandomAccessIterator>

From 832e93d701ffc418e25082f618494d3fe98e16f1 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 7 Feb 2024 12:25:47 +0100
Subject: [PATCH 151/566] __pattern_swap + tag impls

---
 include/oneapi/dpl/pstl/algorithm_impl.h              | 10 ++++++++++
 .../oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h    | 11 +++++++++++
 2 files changed, 21 insertions(+)

diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 9ca164bbdb2..eee3f82927a 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -6224,6 +6224,16 @@ __pattern_swap(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIt
                            __is_parallel);
 }
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Function>
+_ForwardIterator2
+__pattern_swap(_Tag __tag, _ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
+               _ForwardIterator2 __first2, _Function __f)
+{
+    static_assert(__is_backend_tag_v<_Tag>);
+
+    return __pattern_walk2(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __f);
+}
+
 //------------------------------------------------------------------------
 // shift_left
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index cf284bc7c3f..05087fe6bf1 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -204,6 +204,17 @@ __pattern_swap(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIt
         __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __f);
 }
 
+template <typename _BackendTag, typename _ExecutionPolicy, typename _ForwardIterator1, typename _ForwardIterator2,
+          typename _Function>
+_ForwardIterator2
+__pattern_swap(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _ForwardIterator1 __first1,
+               _ForwardIterator1 __last1, _ForwardIterator2 __first2, _Function __f)
+{
+    return __pattern_walk2<_BackendTag, /*_IsSync=*/::std::true_type, __par_backend_hetero::access_mode::read_write,
+                           __par_backend_hetero::access_mode::read_write>(
+        __tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __f);
+}
+
 //------------------------------------------------------------------------
 // walk3
 //------------------------------------------------------------------------

From 68f6ad1f08a830c1a7661726f439bf241e145c3d Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 7 Feb 2024 12:26:12 +0100
Subject: [PATCH 152/566] __pattern_swap + tag calls

---
 include/oneapi/dpl/pstl/glue_algorithm_impl.h | 17 +++++++++--------
 1 file changed, 9 insertions(+), 8 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_algorithm_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
index abbff064991..a332a039c1c 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
@@ -313,14 +313,15 @@ swap_ranges(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardItera
 {
     typedef typename ::std::iterator_traits<_ForwardIterator1>::reference _ReferenceType1;
     typedef typename ::std::iterator_traits<_ForwardIterator2>::reference _ReferenceType2;
-    return oneapi::dpl::__internal::__pattern_swap(
-        ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2,
-        [](_ReferenceType1 __x, _ReferenceType2 __y) {
-            using ::std::swap;
-            swap(__x, __y);
-        },
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>());
+
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
+
+    return oneapi::dpl::__internal::__pattern_swap(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first1,
+                                                   __last1, __first2, [](_ReferenceType1 __x, _ReferenceType2 __y) {
+                                                       using ::std::swap;
+                                                       swap(__x, __y);
+                                                   });
 }
 
 // [alg.transform]

From d94d02b47ca269e76515fca6e721d3dc142dd4ee Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 7 Feb 2024 12:41:10 +0100
Subject: [PATCH 153/566] __pattern_unique + tag impls

---
 include/oneapi/dpl/pstl/algorithm_fwd.h       |  9 +++++
 include/oneapi/dpl/pstl/algorithm_impl.h      | 35 +++++++++++++++++++
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 28 +++++++++++++++
 3 files changed, 72 insertions(+)

diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index 73220412b0c..d9210c80d9e 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -775,11 +775,20 @@ oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _Fo
 __pattern_unique(_ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _BinaryPredicate, _IsVector,
                  /*is_parallel=*/::std::false_type) noexcept;
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _BinaryPredicate>
+_ForwardIterator
+__pattern_unique(_Tag, _ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _BinaryPredicate) noexcept;
+
 template <class _ExecutionPolicy, class _RandomAccessIterator, class _BinaryPredicate, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator>
 __pattern_unique(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _BinaryPredicate, _IsVector,
                  /*is_parallel=*/::std::true_type);
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _BinaryPredicate>
+_RandomAccessIterator
+__pattern_unique(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator,
+                 _BinaryPredicate);
+
 //------------------------------------------------------------------------
 // unique_copy
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index eee3f82927a..bf6d352e6a6 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -2096,6 +2096,15 @@ __pattern_unique(_ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator
     return __internal::__brick_unique(__first, __last, __pred, __is_vector);
 }
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _BinaryPredicate>
+_ForwardIterator
+__pattern_unique(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _BinaryPredicate __pred) noexcept
+{
+    static_assert(__is_backend_tag_v<_Tag>);
+
+    return __internal::__brick_unique(__first, __last, __pred, typename _Tag::__is_vector{});
+}
+
 // That function is shared between two algorithms - remove_if (__pattern_remove_if) and unique (pattern unique). But a mask calculation is different.
 // So, a caller passes _CalcMask brick into remove_elements.
 template <class _ExecutionPolicy, class _ForwardIterator, class _CalcMask, class _IsVector>
@@ -2204,6 +2213,32 @@ __pattern_unique(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _Rand
         __is_vector);
 }
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _BinaryPredicate>
+_RandomAccessIterator
+__pattern_unique(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
+                 _RandomAccessIterator __last, _BinaryPredicate __pred)
+{
+    typedef typename ::std::iterator_traits<_RandomAccessIterator>::reference _ReferenceType;
+
+    if (__first == __last)
+    {
+        return __last;
+    }
+    if (__first + 1 == __last || __first + 2 == __last)
+    {
+        // Trivial sequence - use serial algorithm
+        return __internal::__brick_unique(__first, __last, __pred, _IsVector{});
+    }
+    return __internal::__remove_elements(
+        ::std::forward<_ExecutionPolicy>(__exec), ++__first, __last,
+        [&__pred](bool* __b, bool* __e, _RandomAccessIterator __it) {
+            __internal::__brick_walk3(
+                __b, __e, __it - 1, __it,
+                [&__pred](bool& __x, _ReferenceType __y, _ReferenceType __z) { __x = !__pred(__y, __z); }, _IsVector{});
+        },
+        _IsVector{});
+}
+
 //------------------------------------------------------------------------
 // unique_copy
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 05087fe6bf1..747f0d59380 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -1767,6 +1767,34 @@ __pattern_unique(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last,
         __copy_first, __copy_last, __first, __brick_copy<_ExecutionPolicy>{});
 }
 
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator, typename _BinaryPredicate>
+_Iterator
+__pattern_unique(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last,
+                 _BinaryPredicate __pred)
+{
+    if (__last - __first < 2)
+        return __last;
+
+    using _ValueType = typename ::std::iterator_traits<_Iterator>::value_type;
+
+    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __buf(__exec, __last - __first);
+    auto __copy_first = __buf.get();
+    auto __copy_last = __pattern_unique_copy(__exec, __first, __last, __copy_first, __pred,
+                                             /*vector=*/::std::true_type{}, /*parallel*/ ::std::true_type{});
+
+    constexpr auto __dispatch_tag1 =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, decltype(__copy_first), decltype(__copy_last),
+                                                  decltype(__first)>();
+    using __backend_tag1 = typename decltype(__dispatch_tag1)::__backend_tag;
+
+    //TODO: optimize copy back depending on Iterator, i.e. set_final_data for host iterator/pointer
+    return __pattern_walk2<__backend_tag1, /*_IsSync=*/::std::true_type, __par_backend_hetero::access_mode::read_write,
+                           __par_backend_hetero::access_mode::read_write>(
+        __dispatch_tag1,
+        __par_backend_hetero::make_wrapped_policy<copy_back_wrapper>(::std::forward<_ExecutionPolicy>(__exec)),
+        __copy_first, __copy_last, __first, __brick_copy<_ExecutionPolicy>{});
+}
+
 //------------------------------------------------------------------------
 // is_partitioned
 //------------------------------------------------------------------------

From 8f9bc07177659d93498eb217499fd14334de2bfb Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 7 Feb 2024 12:41:37 +0100
Subject: [PATCH 154/566] __pattern_unique + tag calls

---
 include/oneapi/dpl/pstl/glue_algorithm_impl.h | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_algorithm_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
index a332a039c1c..adc9f6f4702 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
@@ -538,10 +538,10 @@ template <class _ExecutionPolicy, class _ForwardIterator, class _BinaryPredicate
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _ForwardIterator>
 unique(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last, _BinaryPredicate __pred)
 {
-    return oneapi::dpl::__internal::__pattern_unique(
-        ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __pred,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>());
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+
+    return oneapi::dpl::__internal::__pattern_unique(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first,
+                                                     __last, __pred);
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator>

From ff9bb256839399d363c09137c67847c7ce11141a Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 7 Feb 2024 12:56:10 +0100
Subject: [PATCH 155/566] __pattern_unique_copy + tag impls

---
 include/oneapi/dpl/pstl/algorithm_fwd.h       | 11 ++++
 include/oneapi/dpl/pstl/algorithm_impl.h      | 59 +++++++++++++++++++
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 19 ++++++
 3 files changed, 89 insertions(+)

diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index d9210c80d9e..b9893fa63df 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -807,6 +807,11 @@ oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _Ou
 __pattern_unique_copy(_ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _OutputIterator, _BinaryPredicate,
                       _IsVector, /*parallel=*/::std::false_type) noexcept;
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _OutputIterator, class _BinaryPredicate>
+_OutputIterator
+__pattern_unique_copy(_Tag, _ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _OutputIterator,
+                      _BinaryPredicate) noexcept;
+
 template <class _ExecutionPolicy, class _DifferenceType, class _RandomAccessIterator, class _BinaryPredicate>
 _DifferenceType
 __brick_calc_mask_2(_RandomAccessIterator, _RandomAccessIterator, bool* __restrict, _BinaryPredicate,
@@ -823,6 +828,12 @@ oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _Ou
 __pattern_unique_copy(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _OutputIterator,
                       _BinaryPredicate, _IsVector, /*parallel=*/::std::true_type);
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
+          class _BinaryPredicate>
+_RandomAccessIterator2
+__pattern_unique_copy(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator1, _RandomAccessIterator1,
+                      _RandomAccessIterator2, _BinaryPredicate);
+
 //------------------------------------------------------------------------
 // reverse
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index bf6d352e6a6..d49e89bf990 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -2272,6 +2272,16 @@ __pattern_unique_copy(_ExecutionPolicy&&, _ForwardIterator __first, _ForwardIter
     return __internal::__brick_unique_copy(__first, __last, __result, __pred, __is_vector);
 }
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _OutputIterator, class _BinaryPredicate>
+_OutputIterator
+__pattern_unique_copy(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
+                      _OutputIterator __result, _BinaryPredicate __pred) noexcept
+{
+    static_assert(__is_backend_tag_v<_Tag>);
+
+    return __internal::__brick_unique_copy(__first, __last, __result, __pred, typename _Tag::__is_vector{});
+}
+
 template <class _DifferenceType, class _RandomAccessIterator, class _BinaryPredicate>
 _DifferenceType
 __brick_calc_mask_2(_RandomAccessIterator __first, _RandomAccessIterator __last, bool* __restrict __mask,
@@ -2344,6 +2354,55 @@ __pattern_unique_copy(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first,
     return __internal::__brick_unique_copy(__first, __last, __result, __pred, __is_vector);
 }
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
+          class _BinaryPredicate>
+_RandomAccessIterator2
+__pattern_unique_copy(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator1 __first,
+                      _RandomAccessIterator1 __last, _RandomAccessIterator2 __result, _BinaryPredicate __pred)
+{
+    typedef typename ::std::iterator_traits<_RandomAccessIterator1>::difference_type _DifferenceType;
+    const _DifferenceType __n = __last - __first;
+    if (_DifferenceType(2) < __n)
+    {
+        __par_backend::__buffer<_ExecutionPolicy, bool> __mask_buf(__n);
+        if (_DifferenceType(2) < __n)
+        {
+            return __internal::__except_handler([&__exec, __n, __first, __result, __pred, &__mask_buf]() {
+                bool* __mask = __mask_buf.get();
+                _DifferenceType __m{};
+                __par_backend::__parallel_strict_scan(
+                    ::std::forward<_ExecutionPolicy>(__exec), __n, _DifferenceType(0),
+                    [=](_DifferenceType __i, _DifferenceType __len) -> _DifferenceType { // Reduce
+                        _DifferenceType __extra = 0;
+                        if (__i == 0)
+                        {
+                            // Special boundary case
+                            __mask[__i] = true;
+                            if (--__len == 0)
+                                return 1;
+                            ++__i;
+                            ++__extra;
+                        }
+                        return __internal::__brick_calc_mask_2<_DifferenceType>(__first + __i, __first + (__i + __len),
+                                                                                __mask + __i, __pred, _IsVector{}) +
+                               __extra;
+                    },
+                    ::std::plus<_DifferenceType>(),                                              // Combine
+                    [=](_DifferenceType __i, _DifferenceType __len, _DifferenceType __initial) { // Scan
+                        // Phase 2 is same as for __pattern_copy_if
+                        __internal::__brick_copy_by_mask(
+                            __first + __i, __first + (__i + __len), __result + __initial, __mask + __i,
+                            [](_RandomAccessIterator1 __x, _RandomAccessIterator2 __z) { *__z = *__x; }, _IsVector{});
+                    },
+                    [&__m](_DifferenceType __total) { __m = __total; });
+                return __result + __m;
+            });
+        }
+    }
+    // trivial sequence - use serial algorithm
+    return __internal::__brick_unique_copy(__first, __last, __result, __pred, _IsVector{});
+}
+
 //------------------------------------------------------------------------
 // reverse
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 747f0d59380..32654de3adc 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -1676,6 +1676,25 @@ __pattern_unique_copy(_ExecutionPolicy&& __exec, _Iterator1 __first, _Iterator1
     return __result_first + __result.second;
 }
 
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2,
+          typename _BinaryPredicate>
+_Iterator2
+__pattern_unique_copy(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator1 __first, _Iterator1 __last,
+                      _Iterator2 __result_first, _BinaryPredicate __pred)
+{
+    using _It1DifferenceType = typename ::std::iterator_traits<_Iterator1>::difference_type;
+    unseq_backend::__copy_by_mask<::std::plus<_It1DifferenceType>, oneapi::dpl::__internal::__pstl_assign,
+                                  /*inclusive*/ ::std::true_type, 1>
+        __copy_by_mask_op;
+    __create_mask_unique_copy<__not_pred<_BinaryPredicate>, _It1DifferenceType> __create_mask_op{
+        __not_pred<_BinaryPredicate>{__pred}};
+
+    auto __result = __pattern_scan_copy(::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result_first,
+                                        __create_mask_op, __copy_by_mask_op);
+
+    return __result_first + __result.second;
+}
+
 template <typename _Name>
 class copy_back_wrapper
 {

From f93e8675bd1dc3a508a2216c7c88ebaedfeee7b5 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 7 Feb 2024 12:57:07 +0100
Subject: [PATCH 156/566] __pattern_unique_copy + tag calls

---
 include/oneapi/dpl/pstl/glue_algorithm_impl.h        |  9 +++++----
 .../oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h   | 12 ++++++++----
 2 files changed, 13 insertions(+), 8 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_algorithm_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
index adc9f6f4702..9236c9f22ff 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
@@ -557,10 +557,11 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _Forward
 unique_copy(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _ForwardIterator1 __last, _ForwardIterator2 __result,
             _BinaryPredicate __pred)
 {
-    return oneapi::dpl::__internal::__pattern_unique_copy(
-        ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result, __pred,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>());
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
+
+    return oneapi::dpl::__internal::__pattern_unique_copy(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
+                                                          __first, __last, __result, __pred);
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2>
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 32654de3adc..ed2ff5efc7a 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -1770,8 +1770,10 @@ __pattern_unique(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last,
 
     oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __buf(__exec, __last - __first);
     auto __copy_first = __buf.get();
-    auto __copy_last = __pattern_unique_copy(__exec, __first, __last, __copy_first, __pred,
-                                             /*vector=*/::std::true_type{}, /*parallel*/ ::std::true_type{});
+    auto __copy_last =
+        __pattern_unique_copy(oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, decltype(__first),
+                                                                        decltype(__last), decltype(__copy_first)>(),
+                              __exec, __first, __last, __copy_first, __pred);
 
     constexpr auto __dispatch_tag1 =
         oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, decltype(__copy_first), decltype(__copy_last),
@@ -1798,8 +1800,10 @@ __pattern_unique(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _It
 
     oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __buf(__exec, __last - __first);
     auto __copy_first = __buf.get();
-    auto __copy_last = __pattern_unique_copy(__exec, __first, __last, __copy_first, __pred,
-                                             /*vector=*/::std::true_type{}, /*parallel*/ ::std::true_type{});
+    auto __copy_last =
+        __pattern_unique_copy(oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, decltype(__first),
+                                                                        decltype(__last), decltype(__copy_first)>(),
+                              __exec, __first, __last, __copy_first, __pred);
 
     constexpr auto __dispatch_tag1 =
         oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, decltype(__copy_first), decltype(__copy_last),

From 072914fa3f5d92a08e2f6ee09b5e0b782a1dcae2 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 7 Feb 2024 13:21:08 +0100
Subject: [PATCH 157/566] __pattern_walk3_transform_if + tag impls

---
 include/oneapi/dpl/pstl/algorithm_fwd.h         |  6 ++++++
 include/oneapi/dpl/pstl/algorithm_impl.h        | 13 +++++++++++++
 .../dpl/pstl/hetero/algorithm_impl_hetero.h     | 17 +++++++++++++++++
 3 files changed, 36 insertions(+)

diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index b9893fa63df..ec2cd79a135 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -394,6 +394,12 @@ __pattern_walk3_transform_if(_ExecutionPolicy&& __exec, _ForwardIterator1 __firs
                              _ForwardIterator2 __first2, _ForwardIterator3 __first3, _Function __func,
                              _IsVector __is_vector, _IsParallel __is_parallel) noexcept;
 
+template <class _Tag, typename _ExecutionPolicy, typename _ForwardIterator1, typename _ForwardIterator2,
+          typename _ForwardIterator3, typename _Function>
+_ForwardIterator3
+__pattern_walk3_transform_if(_Tag, _ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2,
+                             _ForwardIterator3, _Function) noexcept;
+
 //------------------------------------------------------------------------
 // equal
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index d49e89bf990..705b8d278f6 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -948,6 +948,19 @@ __pattern_walk3_transform_if(_ExecutionPolicy&& __exec, _ForwardIterator1 __firs
                            __is_vector, __is_parallel);
 }
 
+template <class _Tag, typename _ExecutionPolicy, typename _ForwardIterator1, typename _ForwardIterator2,
+          typename _ForwardIterator3, typename _Function>
+_ForwardIterator3
+__pattern_walk3_transform_if(_Tag __tag, _ExecutionPolicy&& __exec, _ForwardIterator1 __first1,
+                             _ForwardIterator1 __last1, _ForwardIterator2 __first2, _ForwardIterator3 __first3,
+                             _Function __func) noexcept
+{
+    static_assert(__is_backend_tag_v<_Tag>);
+
+    return __pattern_walk3(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __first3,
+                           __func);
+}
+
 //------------------------------------------------------------------------
 // equal
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index ed2ff5efc7a..4e72e3af667 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -480,6 +480,23 @@ __pattern_walk3_transform_if(_ExecutionPolicy&& __exec, _ForwardIterator1 __firs
         /*vector=*/::std::true_type{}, /*parallel*/ ::std::true_type{});
 }
 
+template <typename _BackendTag, typename _ExecutionPolicy, typename _ForwardIterator1, typename _ForwardIterator2,
+          typename _ForwardIterator3, typename _Function>
+_ForwardIterator3
+__pattern_walk3_transform_if(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _ForwardIterator1 __first1,
+                             _ForwardIterator1 __last1, _ForwardIterator2 __first2, _ForwardIterator3 __first3,
+                             _Function __func)
+{
+    // Require `read_write` access mode for output sequence to force a copy in for host iterators to capture incoming
+    // values of the output sequence for elements where the predicate is false.
+    return __pattern_walk3<_BackendTag, __par_backend_hetero::access_mode::read,
+                           __par_backend_hetero::access_mode::read, __par_backend_hetero::access_mode::read_write>(
+        __tag,
+        __par_backend_hetero::make_wrapped_policy<__walk3_transform_if_wrapper>(
+            ::std::forward<_ExecutionPolicy>(__exec)),
+        __first1, __last1, __first2, __first3, __func);
+}
+
 //------------------------------------------------------------------------
 // fill
 //------------------------------------------------------------------------

From 81b8a5749ede28dc4117505b28cd0aac13fa03e5 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 7 Feb 2024 13:21:28 +0100
Subject: [PATCH 158/566] __pattern_walk3_transform_if + tag calls

---
 include/oneapi/dpl/pstl/glue_algorithm_impl.h | 11 ++++++-----
 1 file changed, 6 insertions(+), 5 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_algorithm_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
index 9236c9f22ff..19b0208acca 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
@@ -377,12 +377,13 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _Forward
 transform_if(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
              _ForwardIterator2 __first2, _ForwardIterator3 __result, _BinaryOperation __op, _BinaryPredicate __pred)
 {
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1,
+                                                                              _ForwardIterator2, _ForwardIterator3>();
+
     return oneapi::dpl::__internal::__pattern_walk3_transform_if(
-        ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __result,
-        oneapi::dpl::__internal::__transform_if_binary_functor<_BinaryOperation, _BinaryPredicate>(::std::move(__op),
-                                                                                                   ::std::move(__pred)),
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2, _ForwardIterator3>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2, _ForwardIterator3>());
+        __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __result,
+        oneapi::dpl::__internal::__transform_if_binary_functor<_BinaryOperation, _BinaryPredicate>(
+            ::std::move(__op), ::std::move(__pred)));
 }
 
 // [alg.replace]

From 8bf565ce406d47a839fb3905c07ec9aaceaab67d Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 7 Feb 2024 14:49:57 +0100
Subject: [PATCH 159/566] __pattern_transform_reduce + tag impls

---
 .../dpl/pstl/hetero/numeric_impl_hetero.h     | 53 +++++++++++++++
 include/oneapi/dpl/pstl/numeric_fwd.h         | 36 ++++++++++
 include/oneapi/dpl/pstl/numeric_impl.h        | 65 +++++++++++++++++++
 3 files changed, 154 insertions(+)

diff --git a/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
index 90cf7764f4d..e581a66546c 100644
--- a/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
@@ -66,6 +66,35 @@ __pattern_transform_reduce(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __f
         .get();
 }
 
+template <typename _BackendTag, typename _ExecutionPolicy, typename _RandomAccessIterator1,
+          typename _RandomAccessIterator2, typename _Tp, typename _BinaryOperation1, typename _BinaryOperation2>
+_Tp
+__pattern_transform_reduce(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1,
+                           _RandomAccessIterator1 __last1, _RandomAccessIterator2 __first2, _Tp __init,
+                           _BinaryOperation1 __binary_op1, _BinaryOperation2 __binary_op2)
+{
+    if (__first1 == __last1)
+        return __init;
+
+    using _Functor = unseq_backend::walk_n<_ExecutionPolicy, _BinaryOperation2>;
+    using _RepackedTp = __par_backend_hetero::__repacked_tuple_t<_Tp>;
+
+    auto __n = __last1 - __first1;
+    auto __keep1 =
+        oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _RandomAccessIterator1>();
+    auto __buf1 = __keep1(__first1, __last1);
+    auto __keep2 =
+        oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _RandomAccessIterator2>();
+    auto __buf2 = __keep2(__first2, __first2 + __n);
+
+    return oneapi::dpl::__par_backend_hetero::__parallel_transform_reduce<_RepackedTp,
+                                                                          ::std::true_type /*is_commutative*/>(
+               ::std::forward<_ExecutionPolicy>(__exec), __binary_op1, _Functor{__binary_op2},
+               unseq_backend::__init_value<_RepackedTp>{__init}, // initial value
+               __buf1.all_view(), __buf2.all_view())
+        .get();
+}
+
 //------------------------------------------------------------------------
 // transform_reduce (with unary and binary functions)
 //------------------------------------------------------------------------
@@ -94,6 +123,30 @@ __pattern_transform_reduce(_ExecutionPolicy&& __exec, _ForwardIterator __first,
         .get();
 }
 
+template <typename _BackendTag, typename _ExecutionPolicy, typename _ForwardIterator, typename _Tp,
+          typename _BinaryOperation, typename _UnaryOperation>
+_Tp
+__pattern_transform_reduce(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _ForwardIterator __first,
+                           _ForwardIterator __last, _Tp __init, _BinaryOperation __binary_op,
+                           _UnaryOperation __unary_op)
+{
+    if (__first == __last)
+        return __init;
+
+    using _Functor = unseq_backend::walk_n<_ExecutionPolicy, _UnaryOperation>;
+    using _RepackedTp = __par_backend_hetero::__repacked_tuple_t<_Tp>;
+
+    auto __keep = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _ForwardIterator>();
+    auto __buf = __keep(__first, __last);
+
+    return oneapi::dpl::__par_backend_hetero::__parallel_transform_reduce<_RepackedTp,
+                                                                          ::std::true_type /*is_commutative*/>(
+               ::std::forward<_ExecutionPolicy>(__exec), __binary_op, _Functor{__unary_op},
+               unseq_backend::__init_value<_RepackedTp>{__init}, // initial value
+               __buf.all_view())
+        .get();
+}
+
 //------------------------------------------------------------------------
 // transform_scan
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/numeric_fwd.h b/include/oneapi/dpl/pstl/numeric_fwd.h
index aab7edc0150..e060327043d 100644
--- a/include/oneapi/dpl/pstl/numeric_fwd.h
+++ b/include/oneapi/dpl/pstl/numeric_fwd.h
@@ -50,6 +50,12 @@ __pattern_transform_reduce(_ExecutionPolicy&&, _ForwardIterator1, _ForwardIterat
                            _BinaryOperation1, _BinaryOperation2, _IsVector,
                            /*is_parallel=*/::std::false_type) noexcept;
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Tp,
+          class _BinaryOperation1, class _BinaryOperation2>
+_Tp
+__pattern_transform_reduce(_Tag, _ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2, _Tp,
+                           _BinaryOperation1, _BinaryOperation2) noexcept;
+
 template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _Tp,
           class _BinaryOperation1, class _BinaryOperation2, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _Tp>
@@ -57,6 +63,12 @@ __pattern_transform_reduce(_ExecutionPolicy&&, _RandomAccessIterator1, _RandomAc
                            _Tp, _BinaryOperation1, _BinaryOperation2, _IsVector __is_vector,
                            /*is_parallel=*/::std::true_type);
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
+          class _Tp, class _BinaryOperation1, class _BinaryOperation2>
+_Tp
+__pattern_transform_reduce(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator1,
+                           _RandomAccessIterator1, _RandomAccessIterator2, _Tp, _BinaryOperation1, _BinaryOperation2);
+
 //------------------------------------------------------------------------
 // transform_reduce (version with unary and binary functions)
 //------------------------------------------------------------------------
@@ -76,6 +88,12 @@ __pattern_transform_reduce(_ExecutionPolicy&&, _ForwardIterator, _ForwardIterato
                            _UnaryOperation, _IsVector,
                            /*is_parallel=*/::std::false_type) noexcept;
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Tp,
+          class _BinaryOperation1, class _BinaryOperation2>
+_Tp
+__pattern_transform_reduce(_Tag, _ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2, _Tp,
+                           _BinaryOperation1, _BinaryOperation2 __bnary_op2) noexcept;
+
 template <class _ExecutionPolicy, class _RandomAccessIterator, class _Tp, class _BinaryOperation, class _UnaryOperation,
           class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _Tp>
@@ -83,6 +101,24 @@ __pattern_transform_reduce(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAcc
                            _UnaryOperation, _IsVector,
                            /*is_parallel=*/::std::true_type);
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
+          class _Tp, class _BinaryOperation1, class _BinaryOperation2>
+_Tp
+__pattern_transform_reduce(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator1,
+                           _RandomAccessIterator1, _RandomAccessIterator2, _Tp, _BinaryOperation1, _BinaryOperation2);
+
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Tp, class _BinaryOperation,
+          class _UnaryOperation>
+_Tp
+__pattern_transform_reduce(_Tag, _ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _Tp, _BinaryOperation,
+                           _UnaryOperation) noexcept;
+
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Tp, class _BinaryOperation,
+          class _UnaryOperation>
+_Tp
+__pattern_transform_reduce(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator,
+                           _Tp, _BinaryOperation, _UnaryOperation);
+
 //------------------------------------------------------------------------
 // transform_exclusive_scan
 //
diff --git a/include/oneapi/dpl/pstl/numeric_impl.h b/include/oneapi/dpl/pstl/numeric_impl.h
index 8e65580918c..8253106a7c5 100644
--- a/include/oneapi/dpl/pstl/numeric_impl.h
+++ b/include/oneapi/dpl/pstl/numeric_impl.h
@@ -72,6 +72,19 @@ __pattern_transform_reduce(_ExecutionPolicy&&, _ForwardIterator1 __first1, _Forw
     return __brick_transform_reduce(__first1, __last1, __first2, __init, __binary_op1, __binary_op2, __is_vector);
 }
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Tp,
+          class _BinaryOperation1, class _BinaryOperation2>
+_Tp
+__pattern_transform_reduce(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
+                           _ForwardIterator2 __first2, _Tp __init, _BinaryOperation1 __binary_op1,
+                           _BinaryOperation2 __binary_op2) noexcept
+{
+    static_assert(__is_backend_tag_v<_Tag>);
+
+    return __brick_transform_reduce(__first1, __last1, __first2, __init, __binary_op1, __binary_op2,
+                                    typename _Tag::__is_vector{});
+}
+
 template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _Tp,
           class _BinaryOperation1, class _BinaryOperation2, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _Tp>
@@ -95,6 +108,29 @@ __pattern_transform_reduce(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __f
     });
 }
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
+          class _Tp, class _BinaryOperation1, class _BinaryOperation2>
+_Tp
+__pattern_transform_reduce(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1,
+                           _RandomAccessIterator1 __last1, _RandomAccessIterator2 __first2, _Tp __init,
+                           _BinaryOperation1 __binary_op1, _BinaryOperation2 __binary_op2)
+{
+    return __internal::__except_handler([&]() {
+        return __par_backend::__parallel_transform_reduce(
+            ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1,
+            [__first1, __first2, __binary_op2](_RandomAccessIterator1 __i) mutable {
+                return __binary_op2(*__i, *(__first2 + (__i - __first1)));
+            },
+            __init,
+            __binary_op1, // Combine
+            [__first1, __first2, __binary_op1, __binary_op2](_RandomAccessIterator1 __i, _RandomAccessIterator1 __j,
+                                                             _Tp __init) -> _Tp {
+                return __internal::__brick_transform_reduce(__i, __j, __first2 + (__i - __first1), __init, __binary_op1,
+                                                            __binary_op2, _IsVector{});
+            });
+    });
+}
+
 //------------------------------------------------------------------------
 // transform_reduce (version with unary and binary functions)
 //------------------------------------------------------------------------
@@ -133,6 +169,18 @@ __pattern_transform_reduce(_ExecutionPolicy&&, _ForwardIterator __first, _Forwar
     return __internal::__brick_transform_reduce(__first, __last, __init, __binary_op, __unary_op, __is_vector);
 }
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Tp, class _BinaryOperation,
+          class _UnaryOperation>
+_Tp
+__pattern_transform_reduce(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _Tp __init,
+                           _BinaryOperation __binary_op, _UnaryOperation __unary_op) noexcept
+{
+    static_assert(__is_backend_tag_v<_Tag>);
+
+    return __internal::__brick_transform_reduce(__first, __last, __init, __binary_op, __unary_op,
+                                                typename _Tag::__is_vector{});
+}
+
 template <class _ExecutionPolicy, class _RandomAccessIterator, class _Tp, class _BinaryOperation, class _UnaryOperation,
           class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _Tp>
@@ -150,6 +198,23 @@ __pattern_transform_reduce(_ExecutionPolicy&& __exec, _RandomAccessIterator __fi
     });
 }
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Tp, class _BinaryOperation,
+          class _UnaryOperation>
+_Tp
+__pattern_transform_reduce(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
+                           _RandomAccessIterator __last, _Tp __init, _BinaryOperation __binary_op,
+                           _UnaryOperation __unary_op)
+{
+    return __internal::__except_handler([&]() {
+        return __par_backend::__parallel_transform_reduce(
+            ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+            [__unary_op](_RandomAccessIterator __i) mutable { return __unary_op(*__i); }, __init, __binary_op,
+            [__unary_op, __binary_op](_RandomAccessIterator __i, _RandomAccessIterator __j, _Tp __init) {
+                return __internal::__brick_transform_reduce(__i, __j, __init, __binary_op, __unary_op, _IsVector{});
+            });
+    });
+}
+
 //------------------------------------------------------------------------
 // transform_exclusive_scan
 //

From a32df2280222e8af0be85fba0ff9350a1bf59125 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 7 Feb 2024 14:50:12 +0100
Subject: [PATCH 160/566] __pattern_transform_reduce + tag calls

---
 include/oneapi/dpl/pstl/glue_numeric_impl.h | 28 ++++++++++++---------
 1 file changed, 16 insertions(+), 12 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_numeric_impl.h b/include/oneapi/dpl/pstl/glue_numeric_impl.h
index 9c450f79a2b..82d5fe9e8a9 100644
--- a/include/oneapi/dpl/pstl/glue_numeric_impl.h
+++ b/include/oneapi/dpl/pstl/glue_numeric_impl.h
@@ -70,11 +70,13 @@ transform_reduce(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _Forward
                  _ForwardIterator2 __first2, _Tp __init)
 {
     typedef typename ::std::iterator_traits<_ForwardIterator1>::value_type _InputType;
+
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
+
     return oneapi::dpl::__internal::__pattern_transform_reduce(
-        ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __init, ::std::plus<_InputType>(),
-        ::std::multiplies<_InputType>(),
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>());
+        __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __init,
+        ::std::plus<_InputType>(), ::std::multiplies<_InputType>());
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Tp, class _BinaryOperation1,
@@ -83,10 +85,12 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _Tp>
 transform_reduce(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
                  _ForwardIterator2 __first2, _Tp __init, _BinaryOperation1 __binary_op1, _BinaryOperation2 __binary_op2)
 {
-    return oneapi::dpl::__internal::__pattern_transform_reduce(
-        ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __init, __binary_op1, __binary_op2,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>());
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
+
+    return oneapi::dpl::__internal::__pattern_transform_reduce(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
+                                                               __first1, __last1, __first2, __init, __binary_op1,
+                                                               __binary_op2);
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator, class _Tp, class _BinaryOperation, class _UnaryOperation>
@@ -94,10 +98,10 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _Tp>
 transform_reduce(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last, _Tp __init,
                  _BinaryOperation __binary_op, _UnaryOperation __unary_op)
 {
-    return oneapi::dpl::__internal::__pattern_transform_reduce(
-        ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __init, __binary_op, __unary_op,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator>());
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+
+    return oneapi::dpl::__internal::__pattern_transform_reduce(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
+                                                               __first, __last, __init, __binary_op, __unary_op);
 }
 
 // [exclusive.scan]

From db44b15df58833e9c246168026a5d3925689572e Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 7 Feb 2024 15:49:59 +0100
Subject: [PATCH 161/566] __pattern_transform_scan + tag impls

---
 .../dpl/pstl/hetero/numeric_impl_hetero.h     |  30 ++++++
 include/oneapi/dpl/pstl/numeric_fwd.h         |  24 +++++
 include/oneapi/dpl/pstl/numeric_impl.h        | 102 ++++++++++++++++++
 3 files changed, 156 insertions(+)

diff --git a/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
index e581a66546c..cdaf9c8a606 100644
--- a/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
@@ -252,6 +252,21 @@ __pattern_transform_scan(_ExecutionPolicy&& __exec, _Iterator1 __first, _Iterato
                                          __unary_op, _InitType{__init}, __binary_op, _Inclusive{});
 }
 
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2,
+          typename _UnaryOperation, typename _Type,
+          typename _BinaryOperation, typename _Inclusive>
+_Iterator2
+__pattern_transform_scan(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator1 __first,
+                         _Iterator1 __last, _Iterator2 __result,
+                         _UnaryOperation __unary_op, _Type __init, _BinaryOperation __binary_op, _Inclusive)
+{
+    using _RepackedType = __par_backend_hetero::__repacked_tuple_t<_Type>;
+    using _InitType = unseq_backend::__init_value<_RepackedType>;
+
+    return __pattern_transform_scan_base(::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result,
+                                         __unary_op, _InitType{__init}, __binary_op, _Inclusive{});
+}
+
 // scan without initial element
 template <typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2, typename _UnaryOperation,
           typename _BinaryOperation, typename _Inclusive>
@@ -268,6 +283,21 @@ __pattern_transform_scan(_ExecutionPolicy&& __exec, _Iterator1 __first, _Iterato
                                          __unary_op, _InitType{}, __binary_op, _Inclusive{});
 }
 
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2,
+          typename _UnaryOperation, typename _BinaryOperation, typename _Inclusive>
+_Iterator2
+__pattern_transform_scan(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator1 __first,
+                         _Iterator1 __last, _Iterator2 __result, _UnaryOperation __unary_op,
+                         _BinaryOperation __binary_op, _Inclusive)
+{
+    using _Type = typename ::std::iterator_traits<_Iterator1>::value_type;
+    using _RepackedType = __par_backend_hetero::__repacked_tuple_t<_Type>;
+    using _InitType = unseq_backend::__no_init_value<_RepackedType>;
+
+    return __pattern_transform_scan_base(::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result,
+                                         __unary_op, _InitType{}, __binary_op, _Inclusive{});
+}
+
 //------------------------------------------------------------------------
 // adjacent_difference
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/numeric_fwd.h b/include/oneapi/dpl/pstl/numeric_fwd.h
index e060327043d..28e95d0175c 100644
--- a/include/oneapi/dpl/pstl/numeric_fwd.h
+++ b/include/oneapi/dpl/pstl/numeric_fwd.h
@@ -142,6 +142,12 @@ __pattern_transform_scan(_ExecutionPolicy&&, _ForwardIterator, _ForwardIterator,
                          _BinaryOperation, _Inclusive, _IsVector,
                          /*is_parallel=*/::std::false_type) noexcept;
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _OutputIterator, class _UnaryOperation,
+          class _Tp, class _BinaryOperation, class _Inclusive>
+_OutputIterator
+__pattern_transform_scan(_Tag, _ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _OutputIterator, _UnaryOperation,
+                         _Tp, _BinaryOperation, _Inclusive) noexcept;
+
 template <class _ExecutionPolicy, class _RandomAccessIterator, class _OutputIterator, class _UnaryOperation, class _Tp,
           class _BinaryOperation, class _Inclusive, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy_conditional<
@@ -150,6 +156,12 @@ __pattern_transform_scan(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAcces
                          _UnaryOperation, _Tp, _BinaryOperation, _Inclusive, _IsVector,
                          /*is_parallel=*/::std::true_type);
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _OutputIterator,
+          class _UnaryOperation, class _Tp, class _BinaryOperation, class _Inclusive>
+::std::enable_if_t<!::std::is_floating_point_v<_Tp>, _OutputIterator>
+__pattern_transform_scan(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator,
+                         _OutputIterator, _UnaryOperation, _Tp, _BinaryOperation, _Inclusive);
+
 template <class _ExecutionPolicy, class _RandomAccessIterator, class _OutputIterator, class _UnaryOperation, class _Tp,
           class _BinaryOperation, class _Inclusive, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy_conditional<_ExecutionPolicy,
@@ -158,6 +170,12 @@ __pattern_transform_scan(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAcces
                          _UnaryOperation, _Tp, _BinaryOperation, _Inclusive, _IsVector,
                          /*is_parallel=*/::std::true_type);
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _OutputIterator,
+          class _UnaryOperation, class _Tp, class _BinaryOperation, class _Inclusive>
+::std::enable_if_t<::std::is_floating_point_v<_Tp>, _OutputIterator>
+__pattern_transform_scan(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator,
+                         _OutputIterator, _UnaryOperation, _Tp, _BinaryOperation, _Inclusive);
+
 // transform_scan without initial element
 template <class _ExecutionPolicy, class _ForwardIterator, class _OutputIterator, class _UnaryOperation,
           class _BinaryOperation, class _Inclusive, class _IsVector, class _IsParallel>
@@ -166,6 +184,12 @@ __pattern_transform_scan(_ExecutionPolicy&& __exec, _ForwardIterator __first, _F
                          _OutputIterator __result, _UnaryOperation __unary_op, _BinaryOperation __binary_op, _Inclusive,
                          _IsVector __is_vector, _IsParallel __is_parallel);
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _OutputIterator, class _UnaryOperation,
+          class _BinaryOperation, class _Inclusive>
+_OutputIterator
+__pattern_transform_scan(_Tag, _ExecutionPolicy&& __exec, _ForwardIterator, _ForwardIterator, _OutputIterator,
+                         _UnaryOperation, _BinaryOperation, _Inclusive);
+
 //------------------------------------------------------------------------
 // adjacent_difference
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/numeric_impl.h b/include/oneapi/dpl/pstl/numeric_impl.h
index 8253106a7c5..28fcec58ab3 100644
--- a/include/oneapi/dpl/pstl/numeric_impl.h
+++ b/include/oneapi/dpl/pstl/numeric_impl.h
@@ -305,6 +305,20 @@ __pattern_transform_scan(_ExecutionPolicy&&, _ForwardIterator __first, _ForwardI
         .first;
 }
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _OutputIterator, class _UnaryOperation,
+          class _Tp, class _BinaryOperation, class _Inclusive>
+_OutputIterator
+__pattern_transform_scan(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
+                         _OutputIterator __result, _UnaryOperation __unary_op, _Tp __init, _BinaryOperation __binary_op,
+                         _Inclusive) noexcept
+{
+    static_assert(__is_backend_tag_v<_Tag>);
+
+    return __internal::__brick_transform_scan(__first, __last, __result, __unary_op, __init, __binary_op, _Inclusive(),
+                                              typename _Tag::__is_vector{})
+        .first;
+}
+
 template <class _ExecutionPolicy, class _RandomAccessIterator, class _OutputIterator, class _UnaryOperation, class _Tp,
           class _BinaryOperation, class _Inclusive, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy_conditional<
@@ -336,6 +350,36 @@ __pattern_transform_scan(_ExecutionPolicy&& __exec, _RandomAccessIterator __firs
     });
 }
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _OutputIterator,
+          class _UnaryOperation, class _Tp, class _BinaryOperation, class _Inclusive>
+::std::enable_if_t<!::std::is_floating_point_v<_Tp>, _OutputIterator>
+__pattern_transform_scan(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
+                         _RandomAccessIterator __last, _OutputIterator __result, _UnaryOperation __unary_op, _Tp __init,
+                         _BinaryOperation __binary_op, _Inclusive)
+{
+    typedef typename ::std::iterator_traits<_RandomAccessIterator>::difference_type _DifferenceType;
+
+    return __internal::__except_handler([&]() {
+        __par_backend::__parallel_transform_scan(
+            ::std::forward<_ExecutionPolicy>(__exec), __last - __first,
+            [__first, __unary_op](_DifferenceType __i) mutable { return __unary_op(__first[__i]); }, __init,
+            __binary_op,
+            [__first, __unary_op, __binary_op](_DifferenceType __i, _DifferenceType __j, _Tp __init) {
+                // Execute serial __brick_transform_reduce, due to the explicit SIMD vectorization (reduction) requires a commutative operation for the guarantee of correct scan.
+                return __internal::__brick_transform_reduce(__first + __i, __first + __j, __init, __binary_op,
+                                                            __unary_op,
+                                                            /*__is_vector*/ ::std::false_type());
+            },
+            [__first, __unary_op, __binary_op, __result](_DifferenceType __i, _DifferenceType __j,
+                                                                      _Tp __init) {
+                return __internal::__brick_transform_scan(__first + __i, __first + __j, __result + __i, __unary_op,
+                                                          __init, __binary_op, _Inclusive(), _IsVector{})
+                    .second;
+            });
+        return __result + (__last - __first);
+    });
+}
+
 template <class _ExecutionPolicy, class _RandomAccessIterator, class _OutputIterator, class _UnaryOperation, class _Tp,
           class _BinaryOperation, class _Inclusive, class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy_conditional<_ExecutionPolicy,
@@ -373,6 +417,42 @@ __pattern_transform_scan(_ExecutionPolicy&& __exec, _RandomAccessIterator __firs
     });
 }
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _OutputIterator,
+          class _UnaryOperation, class _Tp, class _BinaryOperation, class _Inclusive>
+::std::enable_if_t<::std::is_floating_point_v<_Tp>, _OutputIterator>
+__pattern_transform_scan(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
+                         _RandomAccessIterator __last, _OutputIterator __result, _UnaryOperation __unary_op, _Tp __init,
+                         _BinaryOperation __binary_op, _Inclusive)
+{
+    typedef typename ::std::iterator_traits<_RandomAccessIterator>::difference_type _DifferenceType;
+    _DifferenceType __n = __last - __first;
+
+    if (__n <= 0)
+    {
+        return __result;
+    }
+    return __internal::__except_handler([&]() {
+        __par_backend::__parallel_strict_scan(
+            ::std::forward<_ExecutionPolicy>(__exec), __n, __init,
+            [__first, __unary_op, __binary_op, __result](_DifferenceType __i, _DifferenceType __len) {
+                return __internal::__brick_transform_scan(__first + __i, __first + (__i + __len), __result + __i,
+                                                          __unary_op, _Tp{}, __binary_op, _Inclusive(), _IsVector{})
+                    .second;
+            },
+            __binary_op,
+            [__result, &__binary_op](_DifferenceType __i, _DifferenceType __len, _Tp __initial) {
+                return *(::std::transform(__result + __i, __result + __i + __len, __result + __i,
+                                          [&__initial, &__binary_op](const _Tp& __x) {
+                                              _ONEDPL_PRAGMA_FORCEINLINE
+                                              return __binary_op(__initial, __x);
+                                          }) -
+                         1);
+            },
+            [](_Tp) {});
+        return __result + (__last - __first);
+    });
+}
+
 // transform_scan without initial element
 template <class _ExecutionPolicy, class _ForwardIterator, class _OutputIterator, class _UnaryOperation,
           class _BinaryOperation, class _Inclusive, class _IsVector, class _IsParallel>
@@ -395,6 +475,28 @@ __pattern_transform_scan(_ExecutionPolicy&& __exec, _ForwardIterator __first, _F
     }
 }
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _OutputIterator, class _UnaryOperation,
+          class _BinaryOperation, class _Inclusive>
+_OutputIterator
+__pattern_transform_scan(_Tag __tag, _ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last,
+                         _OutputIterator __result, _UnaryOperation __unary_op, _BinaryOperation __binary_op, _Inclusive)
+{
+    static_assert(__is_backend_tag_v<_Tag>);
+
+    typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _ValueType;
+    if (__first != __last)
+    {
+        _ValueType __tmp = __unary_op(*__first);
+        *__result = __tmp;
+        return __pattern_transform_scan(::std::forward<_ExecutionPolicy>(__exec), ++__first, __last, ++__result,
+                                        __unary_op, __tmp, __binary_op, _Inclusive(), __is_vector, __is_parallel);
+    }
+    else
+    {
+        return __result;
+    }
+}
+
 //------------------------------------------------------------------------
 // adjacent_difference
 //------------------------------------------------------------------------

From 8af57fb21dead97d159bac7c971cac88ca0d2a75 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 7 Feb 2024 15:15:49 +0100
Subject: [PATCH 162/566] __pattern_transform_scan + tag calls

---
 include/oneapi/dpl/pstl/glue_numeric_impl.h | 33 +++++++++++----------
 include/oneapi/dpl/pstl/numeric_impl.h      | 14 ++++++---
 2 files changed, 28 insertions(+), 19 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_numeric_impl.h b/include/oneapi/dpl/pstl/glue_numeric_impl.h
index 82d5fe9e8a9..91c8b7a7c7b 100644
--- a/include/oneapi/dpl/pstl/glue_numeric_impl.h
+++ b/include/oneapi/dpl/pstl/glue_numeric_impl.h
@@ -225,11 +225,12 @@ transform_exclusive_scan(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _
                          _ForwardIterator2 __result, _Tp __init, _BinaryOperation __binary_op,
                          _UnaryOperation __unary_op)
 {
-    return oneapi::dpl::__internal::__pattern_transform_scan(
-        ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result, __unary_op, __init, __binary_op,
-        /*inclusive=*/::std::false_type(),
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>());
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
+
+    return oneapi::dpl::__internal::__pattern_transform_scan(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
+                                                             __first, __last, __result, __unary_op, __init, __binary_op,
+                                                             /*inclusive=*/::std::false_type());
 }
 
 // [transform.inclusive.scan]
@@ -241,11 +242,12 @@ transform_inclusive_scan(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _
                          _ForwardIterator2 __result, _BinaryOperation __binary_op, _UnaryOperation __unary_op,
                          _Tp __init)
 {
-    return oneapi::dpl::__internal::__pattern_transform_scan(
-        ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result, __unary_op, __init, __binary_op,
-        /*inclusive=*/::std::true_type(),
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>());
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
+
+    return oneapi::dpl::__internal::__pattern_transform_scan(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
+                                                             __first, __last, __result, __unary_op, __init, __binary_op,
+                                                             /*inclusive=*/::std::true_type());
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _UnaryOperation,
@@ -254,11 +256,12 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _Forward
 transform_inclusive_scan(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _ForwardIterator1 __last,
                          _ForwardIterator2 __result, _BinaryOperation __binary_op, _UnaryOperation __unary_op)
 {
-    return oneapi::dpl::__internal::__pattern_transform_scan(
-        ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result, __unary_op, __binary_op,
-        /*inclusive=*/::std::true_type(),
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>());
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
+
+    return oneapi::dpl::__internal::__pattern_transform_scan(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
+                                                             __first, __last, __result, __unary_op, __binary_op,
+                                                             /*inclusive=*/::std::true_type());
 }
 
 // [adjacent.difference]
diff --git a/include/oneapi/dpl/pstl/numeric_impl.h b/include/oneapi/dpl/pstl/numeric_impl.h
index 28fcec58ab3..683531a9844 100644
--- a/include/oneapi/dpl/pstl/numeric_impl.h
+++ b/include/oneapi/dpl/pstl/numeric_impl.h
@@ -466,8 +466,13 @@ __pattern_transform_scan(_ExecutionPolicy&& __exec, _ForwardIterator __first, _F
     {
         _ValueType __tmp = __unary_op(*__first);
         *__result = __tmp;
-        return __pattern_transform_scan(::std::forward<_ExecutionPolicy>(__exec), ++__first, __last, ++__result,
-                                        __unary_op, __tmp, __binary_op, _Inclusive(), __is_vector, __is_parallel);
+
+        constexpr auto __dispatch_tag =
+            oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator, _OutputIterator>();
+
+        return __pattern_transform_scan(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), ++__first, __last,
+                                        ++__result,
+                                        __unary_op, __tmp, __binary_op, _Inclusive());
     }
     else
     {
@@ -488,8 +493,9 @@ __pattern_transform_scan(_Tag __tag, _ExecutionPolicy&& __exec, _ForwardIterator
     {
         _ValueType __tmp = __unary_op(*__first);
         *__result = __tmp;
-        return __pattern_transform_scan(::std::forward<_ExecutionPolicy>(__exec), ++__first, __last, ++__result,
-                                        __unary_op, __tmp, __binary_op, _Inclusive(), __is_vector, __is_parallel);
+
+        return __pattern_transform_scan(__tag, ::std::forward<_ExecutionPolicy>(__exec), ++__first, __last, ++__result,
+                                        __unary_op, __tmp, __binary_op, _Inclusive());
     }
     else
     {

From 57de95f3b04de03297b9c742d98a681547e21289 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 7 Feb 2024 15:34:46 +0100
Subject: [PATCH 163/566] __pattern_adjacent_difference + tag impls

---
 .../dpl/pstl/hetero/numeric_impl_hetero.h     | 55 +++++++++++++++++++
 include/oneapi/dpl/pstl/numeric_fwd.h         | 11 ++++
 include/oneapi/dpl/pstl/numeric_impl.h        | 34 ++++++++++++
 3 files changed, 100 insertions(+)

diff --git a/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
index cdaf9c8a606..a051fccfd28 100644
--- a/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
@@ -365,6 +365,61 @@ __pattern_adjacent_difference(_ExecutionPolicy&& __exec, _ForwardIterator1 __fir
     }
 }
 
+template <typename _BackendTag, typename _ExecutionPolicy, typename _ForwardIterator1, typename _ForwardIterator2,
+          typename _BinaryOperation>
+_ForwardIterator2
+__pattern_adjacent_difference(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _ForwardIterator1 __first,
+                              _ForwardIterator1 __last, _ForwardIterator2 __d_first, _BinaryOperation __op)
+{
+    auto __n = __last - __first;
+    if (__n <= 0)
+        return __d_first;
+
+    using _It1ValueT = typename ::std::iterator_traits<_ForwardIterator1>::value_type;
+    using _It2ValueTRef = typename ::std::iterator_traits<_ForwardIterator2>::reference;
+
+    _ForwardIterator2 __d_last = __d_first + __n;
+
+#if !__SYCL_UNNAMED_LAMBDA__
+    // if we have the only element, just copy it according to the specification
+    if (__n == 1)
+    {
+        return __internal::__except_handler([&__exec, __first, __last, __d_first, __d_last, &__op, __tag]() {
+            auto __wrapped_policy = __par_backend_hetero::make_wrapped_policy<adjacent_difference_wrapper>(
+                ::std::forward<_ExecutionPolicy>(__exec));
+
+            __internal::__pattern_walk2_brick(__tag, __wrapped_policy, __first, __last, __d_first,
+                                              __internal::__brick_copy<decltype(__wrapped_policy)>{});
+
+            return __d_last;
+        });
+    }
+    else
+#endif
+    {
+        return __internal::__except_handler([&__exec, __first, __last, __d_first, __d_last, &__op, __n]() {
+            auto __fn = [__op](_It1ValueT __in1, _It1ValueT __in2, _It2ValueTRef __out1) {
+                __out1 = __op(__in2, __in1); // This move assignment is allowed by the C++ standard draft N4810
+            };
+
+            auto __keep1 =
+                oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _ForwardIterator1>();
+            auto __buf1 = __keep1(__first, __last);
+            auto __keep2 =
+                oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::write, _ForwardIterator2>();
+            auto __buf2 = __keep2(__d_first, __d_last);
+
+            using _Function = unseq_backend::walk_adjacent_difference<_ExecutionPolicy, decltype(__fn)>;
+
+            oneapi::dpl::__par_backend_hetero::__parallel_for(__exec, _Function{__fn}, __n, __buf1.all_view(),
+                                                              __buf2.all_view())
+                .wait();
+
+            return __d_last;
+        });
+    }
+}
+
 } // namespace __internal
 } // namespace dpl
 } // namespace oneapi
diff --git a/include/oneapi/dpl/pstl/numeric_fwd.h b/include/oneapi/dpl/pstl/numeric_fwd.h
index 28e95d0175c..fdd4928c230 100644
--- a/include/oneapi/dpl/pstl/numeric_fwd.h
+++ b/include/oneapi/dpl/pstl/numeric_fwd.h
@@ -209,12 +209,23 @@ oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _Ou
 __pattern_adjacent_difference(_ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _OutputIterator, _BinaryOperation,
                               _IsVector, /*is_parallel*/ ::std::false_type) noexcept;
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _OutputIterator, class _BinaryOperation>
+_OutputIterator
+__pattern_adjacent_difference(_Tag, _ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _OutputIterator,
+                              _BinaryOperation) noexcept;
+
 template <class _ExecutionPolicy, class _RandomAccessIterator, class _OutputIterator, class _BinaryOperation,
           class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _OutputIterator>
 __pattern_adjacent_difference(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _OutputIterator,
                               _BinaryOperation, _IsVector, /*is_parallel*/ ::std::true_type);
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
+          class _BinaryOperation>
+_RandomAccessIterator2
+__pattern_adjacent_difference(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator1,
+                              _RandomAccessIterator1, _RandomAccessIterator2, _BinaryOperation);
+
 } // namespace __internal
 } // namespace dpl
 } // namespace oneapi
diff --git a/include/oneapi/dpl/pstl/numeric_impl.h b/include/oneapi/dpl/pstl/numeric_impl.h
index 683531a9844..39d9e16242e 100644
--- a/include/oneapi/dpl/pstl/numeric_impl.h
+++ b/include/oneapi/dpl/pstl/numeric_impl.h
@@ -543,6 +543,16 @@ __pattern_adjacent_difference(_ExecutionPolicy&&, _ForwardIterator __first, _For
     return __internal::__brick_adjacent_difference(__first, __last, __d_first, __op, __is_vector);
 }
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _OutputIterator, class _BinaryOperation>
+_OutputIterator
+__pattern_adjacent_difference(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
+                              _OutputIterator __d_first, _BinaryOperation __op) noexcept
+{
+    static_assert(__is_backend_tag_v<_Tag>);
+
+    return __internal::__brick_adjacent_difference(__first, __last, __d_first, __op, typename _Tag::__is_vector{});
+}
+
 template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _BinaryOperation,
           class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator2>
@@ -567,6 +577,30 @@ __pattern_adjacent_difference(_ExecutionPolicy&& __exec, _RandomAccessIterator1
     return __d_first + (__last - __first);
 }
 
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
+          class _BinaryOperation>
+_RandomAccessIterator2
+__pattern_adjacent_difference(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec,
+                              _RandomAccessIterator1 __first, _RandomAccessIterator1 __last,
+                              _RandomAccessIterator2 __d_first, _BinaryOperation __op)
+{
+    assert(__first != __last);
+    typedef typename ::std::iterator_traits<_RandomAccessIterator1>::reference _ReferenceType1;
+    typedef typename ::std::iterator_traits<_RandomAccessIterator2>::reference _ReferenceType2;
+
+    *__d_first = *__first;
+    __par_backend::__parallel_for(
+        ::std::forward<_ExecutionPolicy>(__exec), __first, __last - 1,
+        [&__op, __d_first, __first](_RandomAccessIterator1 __b, _RandomAccessIterator1 __e) {
+            _RandomAccessIterator2 __d_b = __d_first + (__b - __first);
+            __internal::__brick_walk3(
+                __b, __e, __b + 1, __d_b + 1,
+                [&__op](_ReferenceType1 __x, _ReferenceType1 __y, _ReferenceType2 __z) { __z = __op(__y, __x); },
+                _IsVector{});
+        });
+    return __d_first + (__last - __first);
+}
+
 } // namespace __internal
 } // namespace dpl
 } // namespace oneapi

From 9894ddb7dd18ae01efd271c1826a08e2bbfbe815 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 7 Feb 2024 15:33:58 +0100
Subject: [PATCH 164/566] __pattern_adjacent_difference + tag calls

---
 include/oneapi/dpl/pstl/glue_numeric_impl.h | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_numeric_impl.h b/include/oneapi/dpl/pstl/glue_numeric_impl.h
index 91c8b7a7c7b..d6a98e6529f 100644
--- a/include/oneapi/dpl/pstl/glue_numeric_impl.h
+++ b/include/oneapi/dpl/pstl/glue_numeric_impl.h
@@ -271,14 +271,14 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _Forward
 adjacent_difference(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _ForwardIterator1 __last,
                     _ForwardIterator2 __d_first, _BinaryOperation __op)
 {
-
     if (__first == __last)
         return __d_first;
 
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
+
     return oneapi::dpl::__internal::__pattern_adjacent_difference(
-        ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __d_first, __op,
-        oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>(),
-        oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>());
+        __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __d_first, __op);
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2>

From a795d1013a3e0aaa67ef67fbc95e563c51791c41 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 8 Feb 2024 10:54:48 +0100
Subject: [PATCH 165/566] __pattern_histogram + tag impls

---
 .../oneapi/dpl/pstl/hetero/histogram_impl_hetero.h    | 11 ++++++-----
 include/oneapi/dpl/pstl/histogram_impl.h              |  8 +++++---
 2 files changed, 11 insertions(+), 8 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/histogram_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/histogram_impl_hetero.h
index 27179622b6b..fbdeb161c29 100644
--- a/include/oneapi/dpl/pstl/hetero/histogram_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/histogram_impl_hetero.h
@@ -117,11 +117,12 @@ struct __hist_fill_zeros_wrapper
 {
 };
 
-template <typename _ExecutionPolicy, typename _RandomAccessIterator1, typename _Size, typename _BinHash,
-          typename _RandomAccessIterator2>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy>
-__pattern_histogram(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first, _RandomAccessIterator1 __last,
-                    _Size __num_bins, _BinHash&& __func, _RandomAccessIterator2 __histogram_first)
+template <typename _BackendTag, typename _ExecutionPolicy, typename _RandomAccessIterator1, typename _Size,
+          typename _BinHash, typename _RandomAccessIterator2>
+void
+__pattern_histogram(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _RandomAccessIterator1 __first,
+                    _RandomAccessIterator1 __last, _Size __num_bins, _BinHash&& __func,
+                    _RandomAccessIterator2 __histogram_first)
 {
     //If there are no histogram bins there is nothing to do
     if (__num_bins > 0)
diff --git a/include/oneapi/dpl/pstl/histogram_impl.h b/include/oneapi/dpl/pstl/histogram_impl.h
index 4c8f5204793..04893469ce5 100644
--- a/include/oneapi/dpl/pstl/histogram_impl.h
+++ b/include/oneapi/dpl/pstl/histogram_impl.h
@@ -32,12 +32,14 @@ namespace dpl
 namespace __internal
 {
 
-template <typename _ExecutionPolicy, typename _RandomAccessIterator1, typename _Size, typename _IdxHashFunc,
+template <class _Tag, typename _ExecutionPolicy, typename _RandomAccessIterator1, typename _Size, typename _IdxHashFunc,
           typename _RandomAccessIterator2>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>
-__pattern_histogram(_ExecutionPolicy&& exec, _RandomAccessIterator1 __first, _RandomAccessIterator1 __last,
+void
+__pattern_histogram(_Tag, _ExecutionPolicy&& exec, _RandomAccessIterator1 __first, _RandomAccessIterator1 __last,
                     _Size __num_bins, _IdxHashFunc __func, _RandomAccessIterator2 __histogram_first)
 {
+    static_assert(__is_backend_tag_v<_Tag>);
+
     static_assert(sizeof(_Size) == 0 /*false*/,
                   "Histogram API is currently unsupported for policies other than device execution policies");
 }

From f4dd2947461e6fe41e1078242b3443d815102796 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 8 Feb 2024 10:55:04 +0100
Subject: [PATCH 166/566] __pattern_histogram + tag calls

---
 include/oneapi/dpl/pstl/histogram_impl.h | 10 +++++++++-
 1 file changed, 9 insertions(+), 1 deletion(-)

diff --git a/include/oneapi/dpl/pstl/histogram_impl.h b/include/oneapi/dpl/pstl/histogram_impl.h
index 04893469ce5..a67455df13e 100644
--- a/include/oneapi/dpl/pstl/histogram_impl.h
+++ b/include/oneapi/dpl/pstl/histogram_impl.h
@@ -52,7 +52,11 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _RandomA
 histogram(_ExecutionPolicy&& exec, _RandomAccessIterator1 first, _RandomAccessIterator1 last, _Size num_bins,
           _ValueType first_bin_min_val, _ValueType last_bin_max_val, _RandomAccessIterator2 histogram_first)
 {
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2>();
+
     oneapi::dpl::__internal::__pattern_histogram(
+        __dispatch_tag,
         ::std::forward<_ExecutionPolicy>(exec), first, last, num_bins,
         oneapi::dpl::__internal::__evenly_divided_binhash<_ValueType>(first_bin_min_val, last_bin_max_val, num_bins),
         histogram_first);
@@ -66,9 +70,13 @@ histogram(_ExecutionPolicy&& exec, _RandomAccessIterator1 first, _RandomAccessIt
           _RandomAccessIterator2 boundary_first, _RandomAccessIterator2 boundary_last,
           _RandomAccessIterator3 histogram_first)
 {
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2,
+                                                  _RandomAccessIterator3>();
+
     ::std::ptrdiff_t num_bins = boundary_last - boundary_first - 1;
     oneapi::dpl::__internal::__pattern_histogram(
-        ::std::forward<_ExecutionPolicy>(exec), first, last, num_bins,
+        __dispatch_tag, ::std::forward<_ExecutionPolicy>(exec), first, last, num_bins,
         oneapi::dpl::__internal::__custom_boundary_binhash{boundary_first, boundary_last}, histogram_first);
     return histogram_first + num_bins;
 }

From 89fe1ed5b553e79c247dab6ed37e05f9fdfb3a0b Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 8 Feb 2024 12:57:22 +0100
Subject: [PATCH 167/566] __pattern_for_loop_n + tag impls

---
 .../experimental/internal/for_loop_impl.h     | 142 +++++++++---------
 1 file changed, 69 insertions(+), 73 deletions(-)

diff --git a/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h b/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
index 9f4314f20ee..f323c14a225 100644
--- a/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
+++ b/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
@@ -162,37 +162,63 @@ class __reduction_pack
 };
 
 // Sequenced version of for_loop_n
-template <typename _ExecutionPolicy, typename _Ip, typename _Size, typename _Function, typename... _Rest>
+template <class _Tag, typename _ExecutionPolicy, typename _Ip, typename _Size, typename _Function, typename... _Rest>
 void
-__pattern_for_loop_n(_ExecutionPolicy&&, _Ip __first, _Size __n, _Function __f, __single_stride_type,
-                     /*vector=*/::std::false_type, /*parallel=*/::std::false_type, _Rest&&... __rest) noexcept
+__pattern_for_loop_n(_Tag, _ExecutionPolicy&&, _Ip __first, _Size __n, _Function __f, __single_stride_type,
+                     _Rest&&... __rest) noexcept
 {
+    static_assert(__is_backend_tag_serial_v<_Tag>);
+
     __reduction_pack<_Rest...> __pack{__reduction_pack_tag(), ::std::forward<_Rest>(__rest)...};
 
-    for (_Size __i = 0; __i < __n; ++__i, ++__first)
-        __pack.__apply_func(__f, __first, __i);
+    if constexpr (!typename _Tag::__is_vector{})
+    {
+        for (_Size __i = 0; __i < __n; ++__i, ++__first)
+            __pack.__apply_func(__f, __first, __i);
+    }
+    else
+    {
+        oneapi::dpl::__internal::__brick_walk1(
+            __n, [&__pack, __first, __f](_Size __idx) { __pack.__apply_func(__f, __first + __idx, __idx); },
+            ::std::true_type{});
+    }
 
     __pack.__finalize(__n);
 }
 
-template <typename _ExecutionPolicy, typename _Ip, typename _Size, typename _Function, typename _Sp, typename... _Rest>
+template <class _Tag, typename _ExecutionPolicy, typename _Ip, typename _Size, typename _Function, typename _Sp,
+          typename... _Rest>
 void
-__pattern_for_loop_n(_ExecutionPolicy&&, _Ip __first, _Size __n, _Function __f, _Sp __stride,
-                     /*vector=*/::std::false_type, /*parallel=*/::std::false_type, _Rest&&... __rest) noexcept
+__pattern_for_loop_n(_Tag, _ExecutionPolicy&&, _Ip __first, _Size __n, _Function __f, _Sp __stride,
+                     _Rest&&... __rest) noexcept
 {
+    static_assert(__is_backend_tag_serial_v<_Tag>);
+
     __reduction_pack<_Rest...> __pack{__reduction_pack_tag(), ::std::forward<_Rest>(__rest)...};
 
-    // Simple loop from 0 to __n is not suitable here as we need to ensure that __first is always
-    // <= than the end iterator, even if it's not dereferenced. Some implementation might place
-    // validation checks to enforce this invariant.
-    if (__n > 0)
+    if constexpr (!typename _Tag::__is_vector{})
     {
-        for (_Size __i = 0; __i < __n - 1; ++__i, oneapi::dpl::__internal::__advance(__first, __stride))
+        // Simple loop from 0 to __n is not suitable here as we need to ensure that __first is always
+        // <= than the end iterator, even if it's not dereferenced. Some implementation might place
+        // validation checks to enforce this invariant.
+        if (__n > 0)
         {
-            __pack.__apply_func(__f, __first, __i);
-        }
+            for (_Size __i = 0; __i < __n - 1; ++__i, oneapi::dpl::__internal::__advance(__first, __stride))
+            {
+                __pack.__apply_func(__f, __first, __i);
+            }
 
-        __pack.__apply_func(__f, __first, __n - 1);
+            __pack.__apply_func(__f, __first, __n - 1);
+        }
+    }
+    else
+    {
+        oneapi::dpl::__internal::__brick_walk1(
+            __n,
+            [&__pack, __first, __f, __stride](_Size __idx) {
+                __pack.__apply_func(__f, __first + __idx * __stride, __idx);
+            },
+            ::std::true_type{});
     }
 
     __pack.__finalize(__n);
@@ -338,36 +364,6 @@ __pattern_for_loop(_ExecutionPolicy&&, _Ip __first, _Ip __last, _Function __f, _
     __pack.__finalize(__ordinal_position);
 }
 
-// Vectorized version of for_loop_n
-template <typename _ExecutionPolicy, typename _Ip, typename _Size, typename _Function, typename... _Rest>
-void
-__pattern_for_loop_n(_ExecutionPolicy&&, _Ip __first, _Size __n, _Function __f, __single_stride_type,
-                     /*vector=*/::std::true_type, /*parallel=*/::std::false_type, _Rest&&... __rest) noexcept
-{
-    __reduction_pack<_Rest...> __pack{__reduction_pack_tag(), ::std::forward<_Rest>(__rest)...};
-
-    oneapi::dpl::__internal::__brick_walk1(
-        __n, [&__pack, __first, __f](_Size __idx) { __pack.__apply_func(__f, __first + __idx, __idx); },
-        ::std::true_type{});
-
-    __pack.__finalize(__n);
-}
-
-template <typename _ExecutionPolicy, typename _Ip, typename _Size, typename _Function, typename _Sp, typename... _Rest>
-void
-__pattern_for_loop_n(_ExecutionPolicy&&, _Ip __first, _Size __n, _Function __f, _Sp __stride,
-                     /*vector=*/::std::true_type, /*parallel=*/::std::false_type, _Rest&&... __rest) noexcept
-{
-    __reduction_pack<_Rest...> __pack{__reduction_pack_tag(), ::std::forward<_Rest>(__rest)...};
-
-    oneapi::dpl::__internal::__brick_walk1(
-        __n,
-        [&__pack, __first, __f, __stride](_Size __idx) { __pack.__apply_func(__f, __first + __idx * __stride, __idx); },
-        ::std::true_type{});
-
-    __pack.__finalize(__n);
-}
-
 // Vectorized version of for_loop
 template <typename _ExecutionPolicy, typename _Ip, typename _Function, typename _Sp, typename... _Rest>
 void
@@ -386,11 +382,11 @@ __pattern_for_loop(_ExecutionPolicy&& __exec, _Ip __first, _Ip __last, _Function
 // perhaps it's better to check for presence of reduction object and call parallel_for routine instead.
 // TODO: need to add a static_assert for match between rest and f's arguments, currently there is a lot
 // of unclear error in cast of mismatch.
-template <typename _ExecutionPolicy, typename _Ip, typename _Size, typename _Function, typename _IsVector,
+template <typename _IsVector, typename _ExecutionPolicy, typename _Ip, typename _Size, typename _Function,
           typename... _Rest>
 void
-__pattern_for_loop_n(_ExecutionPolicy&& __exec, _Ip __first, _Size __n, _Function __f, __single_stride_type,
-                     _IsVector __is_vector, /*parallel=*/::std::true_type, _Rest&&... __rest)
+__pattern_for_loop_n(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _Ip __first, _Size __n, _Function __f,
+                     __single_stride_type, _Rest&&... __rest)
 {
     using __pack_type = __reduction_pack<_Rest...>;
 
@@ -398,34 +394,34 @@ __pattern_for_loop_n(_ExecutionPolicy&& __exec, _Ip __first, _Size __n, _Functio
     const __pack_type __identity{__reduction_pack_tag(), ::std::forward<_Rest>(__rest)...};
 
     oneapi::dpl::__internal::__except_handler([&]() {
-        return __par_backend::__parallel_reduce(::std::forward<_ExecutionPolicy>(__exec), _Size(0), __n, __identity,
-                                                [__is_vector, __first, __f](_Size __i, _Size __j, __pack_type __value) {
-                                                    const auto __subseq_start = __first + __i;
-                                                    const auto __length = __j - __i;
-
-                                                    oneapi::dpl::__internal::__brick_walk1(
-                                                        __length,
-                                                        [&__value, __f, __i, __subseq_start](_Size __idx) {
-                                                            __value.__apply_func(__f, __subseq_start + __idx,
-                                                                                 __i + __idx);
-                                                        },
-                                                        __is_vector);
-
-                                                    return __value;
-                                                },
-                                                [](__pack_type __lhs, const __pack_type& __rhs) {
-                                                    __lhs.__combine(__rhs);
-                                                    return __lhs;
-                                                })
+        return __par_backend::__parallel_reduce(
+                   ::std::forward<_ExecutionPolicy>(__exec), _Size(0), __n, __identity,
+                   [__first, __f](_Size __i, _Size __j, __pack_type __value) {
+                       const auto __subseq_start = __first + __i;
+                       const auto __length = __j - __i;
+
+                       oneapi::dpl::__internal::__brick_walk1(
+                           __length,
+                           [&__value, __f, __i, __subseq_start](_Size __idx) {
+                               __value.__apply_func(__f, __subseq_start + __idx, __i + __idx);
+                           },
+                           _IsVector{});
+
+                       return __value;
+                   },
+                   [](__pack_type __lhs, const __pack_type& __rhs) {
+                       __lhs.__combine(__rhs);
+                       return __lhs;
+                   })
             .__finalize(__n);
     });
 }
 
-template <typename _ExecutionPolicy, typename _Ip, typename _Size, typename _Function, typename _Sp, typename _IsVector,
+template <typename _IsVector, typename _ExecutionPolicy, typename _Ip, typename _Size, typename _Function, typename _Sp,
           typename... _Rest>
 void
-__pattern_for_loop_n(_ExecutionPolicy&& __exec, _Ip __first, _Size __n, _Function __f, _Sp __stride,
-                     _IsVector __is_vector, /*parallel=*/::std::true_type, _Rest&&... __rest)
+__pattern_for_loop_n(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _Ip __first, _Size __n, _Function __f,
+                     _Sp __stride, _Rest&&... __rest)
 {
     using __pack_type = __reduction_pack<_Rest...>;
 
@@ -435,7 +431,7 @@ __pattern_for_loop_n(_ExecutionPolicy&& __exec, _Ip __first, _Size __n, _Functio
     oneapi::dpl::__internal::__except_handler([&]() {
         return __par_backend::__parallel_reduce(
                    ::std::forward<_ExecutionPolicy>(__exec), _Size(0), __n, __identity,
-                   [__is_vector, __first, __f, __stride](_Size __i, _Size __j, __pack_type __value) {
+                   [__first, __f, __stride](_Size __i, _Size __j, __pack_type __value) {
                        const auto __subseq_start = __first + __i * __stride;
                        const auto __length = __j - __i;
 
@@ -444,7 +440,7 @@ __pattern_for_loop_n(_ExecutionPolicy&& __exec, _Ip __first, _Size __n, _Functio
                            [&__value, __f, __i, __subseq_start, __stride](_Size __idx) {
                                __value.__apply_func(__f, __subseq_start + __idx * __stride, __i + __idx);
                            },
-                           __is_vector);
+                           _IsVector{});
 
                        return __value;
                    },

From 584bae221ebb68513333af202f6489710bf0f55a Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 8 Feb 2024 12:57:54 +0100
Subject: [PATCH 168/566] __pattern_for_loop_n + tag calls

---
 .../experimental/internal/for_loop_impl.h     | 35 ++++++++++++-------
 1 file changed, 22 insertions(+), 13 deletions(-)

diff --git a/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h b/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
index f323c14a225..69e07159b1d 100644
--- a/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
+++ b/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
@@ -252,9 +252,9 @@ __pattern_for_loop(_ExecutionPolicy&& __exec, _Ip __first, _Ip __last, _Function
                    /*vector=*/::std::false_type, /*parallel=*/::std::false_type, _Rest&&... __rest) noexcept
 {
     oneapi::dpl::__internal::__pattern_for_loop_n(
-        ::std::forward<_ExecutionPolicy>(__exec), __first,
+        __serial_tag<::std::false_type>{}, ::std::forward<_ExecutionPolicy>(__exec), __first,
         oneapi::dpl::__internal::__calculate_input_sequence_length(__first, __last, __stride), __f, __stride,
-        ::std::false_type{}, ::std::false_type{}, ::std::forward<_Rest>(__rest)...);
+        ::std::forward<_Rest>(__rest)...);
 }
 
 template <typename _Ip, typename _Function, typename _Sp, typename _Pack, typename _IndexType>
@@ -371,9 +371,9 @@ __pattern_for_loop(_ExecutionPolicy&& __exec, _Ip __first, _Ip __last, _Function
                    /*vector=*/::std::true_type, /*parallel=*/::std::false_type, _Rest&&... __rest) noexcept
 {
     oneapi::dpl::__internal::__pattern_for_loop_n(
-        ::std::forward<_ExecutionPolicy>(__exec), __first,
+        __serial_tag<::std::true_type>{}, ::std::forward<_ExecutionPolicy>(__exec), __first,
         oneapi::dpl::__internal::__calculate_input_sequence_length(__first, __last, __stride), __f, __stride,
-        ::std::true_type{}, ::std::false_type{}, ::std::forward<_Rest>(__rest)...);
+        ::std::forward<_Rest>(__rest)...);
 }
 
 // Parallel version of for_loop_n
@@ -460,9 +460,9 @@ __pattern_for_loop(_ExecutionPolicy&& __exec, _Ip __first, _Ip __last, _Function
                    /*parallel=*/::std::true_type, _Rest&&... __rest)
 {
     oneapi::dpl::__internal::__pattern_for_loop_n(
-        ::std::forward<_ExecutionPolicy>(__exec), __first,
+        __parallel_tag<_IsVector>{}, ::std::forward<_ExecutionPolicy>(__exec), __first,
         oneapi::dpl::__internal::__calculate_input_sequence_length(__first, __last, __stride), __f, __stride,
-        __is_vector, ::std::true_type{}, ::std::forward<_Rest>(__rest)...);
+        ::std::forward<_Rest>(__rest)...);
 }
 
 // Helper structure to split code functions for integral and iterator types so the return
@@ -508,7 +508,7 @@ struct __use_par_vec_helper<_Ip, ::std::enable_if_t<!::std::is_integral_v<_Ip>>>
 
 // Special versions for for_loop: handles both iterators and integral types(treated as random access iterators)
 template <typename _ExecutionPolicy, typename _Ip>
-auto
+constexpr auto
 __use_vectorization()
     -> decltype(__use_par_vec_helper<_Ip>::template __use_vector<_ExecutionPolicy>())
 {
@@ -516,7 +516,7 @@ __use_vectorization()
 }
 
 template <typename _ExecutionPolicy, typename _Ip>
-auto
+constexpr auto
 __use_parallelization()
     -> decltype(__use_par_vec_helper<_Ip>::template __use_parallel<_ExecutionPolicy>())
 {
@@ -542,11 +542,20 @@ void
 __for_loop_n_impl(_ExecutionPolicy&& __exec, _Ip __start, _Size __n, _Fp&& __f, _Sp __stride,
                   ::std::tuple<_Rest...>&& __t, ::std::index_sequence<_Is...>)
 {
-    oneapi::dpl::__internal::__pattern_for_loop_n(
-        ::std::forward<_ExecutionPolicy>(__exec), __start, __n, __f, __stride,
-        oneapi::dpl::__internal::__use_vectorization<_ExecutionPolicy, _Ip>(),
-        oneapi::dpl::__internal::__use_parallelization<_ExecutionPolicy, _Ip>(),
-        ::std::get<_Is>(::std::move(__t))...);
+    using _IsVector = decltype(oneapi::dpl::__internal::__use_vectorization<_ExecutionPolicy, _Ip>());
+
+    if constexpr (oneapi::dpl::__internal::__use_parallelization<_ExecutionPolicy, _Ip>())
+    {
+        oneapi::dpl::__internal::__pattern_for_loop_n(__parallel_tag<_IsVector>{},
+                                                      ::std::forward<_ExecutionPolicy>(__exec), __start, __n, __f,
+                                                      __stride, ::std::get<_Is>(::std::move(__t))...);
+    }
+    else
+    {
+        oneapi::dpl::__internal::__pattern_for_loop_n(__serial_tag<_IsVector>{},
+                                                      ::std::forward<_ExecutionPolicy>(__exec), __start, __n, __f,
+                                                      __stride, ::std::get<_Is>(::std::move(__t))...);
+    }
 }
 
 template <typename _ExecutionPolicy, typename _Ip, typename _Sp, typename... _Rest>

From c0ccfca9739501430a8c493414f8a23d477eb86e Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 8 Feb 2024 13:28:23 +0100
Subject: [PATCH 169/566] __pattern_for_loop + tag impls

---
 .../experimental/internal/for_loop_impl.h     | 114 +++++++++---------
 1 file changed, 58 insertions(+), 56 deletions(-)

diff --git a/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h b/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
index 69e07159b1d..2f5b511acb4 100644
--- a/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
+++ b/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
@@ -242,19 +242,61 @@ struct __is_random_access_or_integral<_Ip,
 {
 };
 
+template <typename _Ip, typename _Function, typename _Sp, typename _Pack, typename _IndexType>
+::std::enable_if_t<
+    ::std::is_same_v<typename ::std::iterator_traits<_Ip>::iterator_category, ::std::bidirectional_iterator_tag>,
+    _IndexType>
+__execute_loop_strided(_Ip __first, _Ip __last, _Function __f, _Sp __stride, _Pack& __pack, _IndexType) noexcept;
+
+template <typename _Ip, typename _Function, typename _Sp, typename _Pack, typename _IndexType>
+::std::enable_if_t<
+    ::std::is_same_v<typename ::std::iterator_traits<_Ip>::iterator_category, ::std::forward_iterator_tag> ||
+        ::std::is_same_v<typename ::std::iterator_traits<_Ip>::iterator_category, ::std::input_iterator_tag>,
+    _IndexType>
+__execute_loop_strided(_Ip __first, _Ip __last, _Function __f, _Sp __stride, _Pack& __pack, _IndexType) noexcept;
+
 template <typename _Ip>
 inline constexpr bool __is_random_access_or_integral_v = __is_random_access_or_integral<_Ip>::value;
 
 // Sequenced version of for_loop for RAI and integral types
-template <typename _ExecutionPolicy, typename _Ip, typename _Function, typename _Sp, typename... _Rest>
-::std::enable_if_t<__is_random_access_or_integral_v<_Ip>>
-__pattern_for_loop(_ExecutionPolicy&& __exec, _Ip __first, _Ip __last, _Function __f, _Sp __stride,
-                   /*vector=*/::std::false_type, /*parallel=*/::std::false_type, _Rest&&... __rest) noexcept
+// Vectorized version of for_loop
+template <class _Tag, typename _ExecutionPolicy, typename _Ip, typename _Function, typename _Sp, typename... _Rest>
+void
+__pattern_for_loop(_Tag __tag, _ExecutionPolicy&& __exec, _Ip __first, _Ip __last, _Function __f, _Sp __stride, _Rest&&... __rest) noexcept
 {
-    oneapi::dpl::__internal::__pattern_for_loop_n(
-        __serial_tag<::std::false_type>{}, ::std::forward<_ExecutionPolicy>(__exec), __first,
-        oneapi::dpl::__internal::__calculate_input_sequence_length(__first, __last, __stride), __f, __stride,
-        ::std::forward<_Rest>(__rest)...);
+    static_assert(__is_backend_tag_serial_v<_Tag>);
+
+    if constexpr (__is_random_access_or_integral_v<_Ip>)
+    {
+        oneapi::dpl::__internal::__pattern_for_loop_n(
+            __tag, ::std::forward<_ExecutionPolicy>(__exec), __first,
+            oneapi::dpl::__internal::__calculate_input_sequence_length(__first, __last, __stride), __f, __stride,
+            ::std::forward<_Rest>(__rest)...);
+    }
+    else
+    {
+        __reduction_pack<_Rest...> __pack{__reduction_pack_tag(), ::std::forward<_Rest>(__rest)...};
+
+        // Make sure that our index type is able to hold all the possible values
+        using __index_type = typename __difference<_Ip>::__type;
+        __index_type __ordinal_position = 0;
+
+        if (__stride == 1)
+        {
+            // Avoid check for i % stride on each iteration for the most common case.
+            for (; __first != __last; ++__first, ++__ordinal_position)
+                __pack.__apply_func(__f, __first, __ordinal_position);
+        }
+        else
+        {
+            __ordinal_position =
+                oneapi::dpl::__internal::__execute_loop_strided(__first, __last, __f, __stride, __pack,
+                                                                // Only passed to deduce the type for internal counter
+                                                                __index_type{});
+        }
+
+        __pack.__finalize(__ordinal_position);
+    }
 }
 
 template <typename _Ip, typename _Function, typename _Sp, typename _Pack, typename _IndexType>
@@ -318,11 +360,12 @@ __execute_loop_strided(_Ip __first, _Ip __last, _Function __f, _Sp __stride, _Pa
 }
 
 // Sequenced version of for_loop for non-RAI and non-integral types
-template <typename _ExecutionPolicy, typename _Ip, typename _Function, typename... _Rest>
+template <class _Tag, typename _ExecutionPolicy, typename _Ip, typename _Function, typename... _Rest>
 ::std::enable_if_t<!__is_random_access_or_integral_v<_Ip>>
-__pattern_for_loop(_ExecutionPolicy&&, _Ip __first, _Ip __last, _Function __f, __single_stride_type,
-                   /*vector=*/::std::false_type, /*parallel=*/::std::false_type, _Rest&&... __rest) noexcept
+__pattern_for_loop(_Tag, _ExecutionPolicy&&, _Ip __first, _Ip __last, _Function __f, __single_stride_type, _Rest&&... __rest) noexcept
 {
+    static_assert(__is_backend_tag_serial_v<_Tag>);
+
     __reduction_pack<_Rest...> __pack{__reduction_pack_tag(), ::std::forward<_Rest>(__rest)...};
 
     // Make sure that our index type is able to hold all the possible values
@@ -336,46 +379,6 @@ __pattern_for_loop(_ExecutionPolicy&&, _Ip __first, _Ip __last, _Function __f, _
     __pack.__finalize(__ordinal_position);
 }
 
-template <typename _ExecutionPolicy, typename _Ip, typename _Function, typename _Sp, typename... _Rest>
-::std::enable_if_t<!__is_random_access_or_integral_v<_Ip>>
-__pattern_for_loop(_ExecutionPolicy&&, _Ip __first, _Ip __last, _Function __f, _Sp __stride,
-                   /*vector=*/::std::false_type, /*parallel=*/::std::false_type, _Rest&&... __rest) noexcept
-{
-    __reduction_pack<_Rest...> __pack{__reduction_pack_tag(), ::std::forward<_Rest>(__rest)...};
-
-    // Make sure that our index type is able to hold all the possible values
-    using __index_type = typename __difference<_Ip>::__type;
-    __index_type __ordinal_position = 0;
-
-    if (__stride == 1)
-    {
-        // Avoid check for i % stride on each iteration for the most common case.
-        for (; __first != __last; ++__first, ++__ordinal_position)
-            __pack.__apply_func(__f, __first, __ordinal_position);
-    }
-    else
-    {
-        __ordinal_position =
-            oneapi::dpl::__internal::__execute_loop_strided(__first, __last, __f, __stride, __pack,
-                                                            // Only passed to deduce the type for internal counter
-                                                            __index_type{});
-    }
-
-    __pack.__finalize(__ordinal_position);
-}
-
-// Vectorized version of for_loop
-template <typename _ExecutionPolicy, typename _Ip, typename _Function, typename _Sp, typename... _Rest>
-void
-__pattern_for_loop(_ExecutionPolicy&& __exec, _Ip __first, _Ip __last, _Function __f, _Sp __stride,
-                   /*vector=*/::std::true_type, /*parallel=*/::std::false_type, _Rest&&... __rest) noexcept
-{
-    oneapi::dpl::__internal::__pattern_for_loop_n(
-        __serial_tag<::std::true_type>{}, ::std::forward<_ExecutionPolicy>(__exec), __first,
-        oneapi::dpl::__internal::__calculate_input_sequence_length(__first, __last, __stride), __f, __stride,
-        ::std::forward<_Rest>(__rest)...);
-}
-
 // Parallel version of for_loop_n
 
 // TODO: Using parallel_reduce when we don't have a reduction object in the pack might be ineffective,
@@ -452,15 +455,14 @@ __pattern_for_loop_n(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec,
     });
 }
 
-template <typename _ExecutionPolicy, typename _Ip, typename _Function, typename _Sp, typename _IsVector,
+template <typename _IsVector, typename _ExecutionPolicy, typename _Ip, typename _Function, typename _Sp,
           typename... _Rest>
 void
-__pattern_for_loop(_ExecutionPolicy&& __exec, _Ip __first, _Ip __last, _Function __f, _Sp __stride,
-                   _IsVector __is_vector,
-                   /*parallel=*/::std::true_type, _Rest&&... __rest)
+__pattern_for_loop(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _Ip __first, _Ip __last, _Function __f,
+                   _Sp __stride, _Rest&&... __rest)
 {
     oneapi::dpl::__internal::__pattern_for_loop_n(
-        __parallel_tag<_IsVector>{}, ::std::forward<_ExecutionPolicy>(__exec), __first,
+        __tag, ::std::forward<_ExecutionPolicy>(__exec), __first,
         oneapi::dpl::__internal::__calculate_input_sequence_length(__first, __last, __stride), __f, __stride,
         ::std::forward<_Rest>(__rest)...);
 }

From 31dc09e2e67843d2276d3230128ffb09f195a6bf Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 8 Feb 2024 13:28:39 +0100
Subject: [PATCH 170/566] __pattern_for_loop + tag calls

---
 .../experimental/internal/for_loop_impl.h     | 19 ++++++++++++++-----
 1 file changed, 14 insertions(+), 5 deletions(-)

diff --git a/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h b/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
index 2f5b511acb4..170cc0f0d62 100644
--- a/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
+++ b/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
@@ -531,11 +531,20 @@ void
 __for_loop_impl(_ExecutionPolicy&& __exec, _Ip __start, _Ip __finish, _Fp&& __f, _Sp __stride,
                 ::std::tuple<_Rest...>&& __t, ::std::index_sequence<_Is...>)
 {
-    oneapi::dpl::__internal::__pattern_for_loop(
-        ::std::forward<_ExecutionPolicy>(__exec), __start, __finish, __f, __stride,
-        oneapi::dpl::__internal::__use_vectorization<_ExecutionPolicy, _Ip>(),
-        oneapi::dpl::__internal::__use_parallelization<_ExecutionPolicy, _Ip>(),
-        ::std::get<_Is>(::std::move(__t))...);
+    using _IsVector = decltype(oneapi::dpl::__internal::__use_vectorization<_ExecutionPolicy, _Ip>());
+
+    if constexpr (oneapi::dpl::__internal::__use_parallelization<_ExecutionPolicy, _Ip>())
+    {
+        oneapi::dpl::__internal::__pattern_for_loop(__parallel_tag<_IsVector>{},
+                                                    ::std::forward<_ExecutionPolicy>(__exec), __start, __finish, __f,
+                                                    __stride, ::std::get<_Is>(::std::move(__t))...);
+    }
+    else
+    {
+        oneapi::dpl::__internal::__pattern_for_loop(__serial_tag<_IsVector>{},
+                                                    ::std::forward<_ExecutionPolicy>(__exec), __start, __finish, __f,
+                                                    __stride, ::std::get<_Is>(::std::move(__t))...);
+    }
 }
 
 template <typename _ExecutionPolicy, typename _Ip, typename _Size, typename _Fp, typename _Sp, typename... _Rest,

From 436c362eb31520f351b42ba7546657e30522bf0e Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 9 Feb 2024 14:09:47 +0100
Subject: [PATCH 171/566] __pattern_scan_copy + tag impls

---
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 27 +++++++++++++++++++
 1 file changed, 27 insertions(+)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 4e72e3af667..88585811c6b 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -1567,6 +1567,33 @@ __pattern_scan_copy(_ExecutionPolicy&& __exec, _Iterator1 __first, _Iterator1 __
     return ::std::make_pair(__output_first + __n, __num_copied);
 }
 
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator1, typename _IteratorOrTuple,
+          typename _CreateMaskOp, typename _CopyByMaskOp>
+::std::pair<_IteratorOrTuple, typename ::std::iterator_traits<_Iterator1>::difference_type>
+__pattern_scan_copy(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Iterator1 __first, _Iterator1 __last,
+                    _IteratorOrTuple __output_first, _CreateMaskOp __create_mask_op, _CopyByMaskOp __copy_by_mask_op)
+{
+    using _It1DifferenceType = typename ::std::iterator_traits<_Iterator1>::difference_type;
+
+    if (__first == __last)
+        return ::std::make_pair(__output_first, _It1DifferenceType{0});
+
+    _It1DifferenceType __n = __last - __first;
+
+    auto __keep1 = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _Iterator1>();
+    auto __buf1 = __keep1(__first, __last);
+    auto __keep2 =
+        oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::write, _IteratorOrTuple>();
+    auto __buf2 = __keep2(__output_first, __output_first + __n);
+
+    auto __res =
+        __par_backend_hetero::__parallel_scan_copy(::std::forward<_ExecutionPolicy>(__exec), __buf1.all_view(),
+                                                   __buf2.all_view(), __n, __create_mask_op, __copy_by_mask_op);
+
+    ::std::size_t __num_copied = __res.get();
+    return ::std::make_pair(__output_first + __n, __num_copied);
+}
+
 template <typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2, typename _Predicate>
 oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, _Iterator2>
 __pattern_copy_if(_ExecutionPolicy&& __exec, _Iterator1 __first, _Iterator1 __last, _Iterator2 __result_first,

From 47e30fc9c11425ba8e6d852ee110a235193d30b3 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 9 Feb 2024 14:10:18 +0100
Subject: [PATCH 172/566] __pattern_scan_copy + tag calls

---
 include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 88585811c6b..0c5cd09cb4d 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -1689,7 +1689,7 @@ __pattern_partition_copy(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __e
     unseq_backend::__partition_by_mask<_ReduceOp, /*inclusive*/ ::std::true_type> __copy_by_mask_op{_ReduceOp{}};
 
     auto __result = __pattern_scan_copy(
-        ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+        __tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
         __par_backend_hetero::zip(
             __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::write>(__result1),
             __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::write>(__result2)),
@@ -1733,8 +1733,8 @@ __pattern_unique_copy(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec
     __create_mask_unique_copy<__not_pred<_BinaryPredicate>, _It1DifferenceType> __create_mask_op{
         __not_pred<_BinaryPredicate>{__pred}};
 
-    auto __result = __pattern_scan_copy(::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result_first,
-                                        __create_mask_op, __copy_by_mask_op);
+    auto __result = __pattern_scan_copy(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+                                        __result_first, __create_mask_op, __copy_by_mask_op);
 
     return __result_first + __result.second;
 }

From d1ca14fa0cf248ace53cda64b7ad87bb1f8e2ccd Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 9 Feb 2024 14:10:40 +0100
Subject: [PATCH 173/566] __pattern_hetero_set_op + tag impls

---
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 65 +++++++++++++++++++
 1 file changed, 65 insertions(+)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 0c5cd09cb4d..0091ef00e46 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -3111,6 +3111,71 @@ __pattern_set_intersection(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1
                                    __result, __comp, unseq_backend::_IntersectionTag());
 }
 
+template <typename _BackendTag, typename _ExecutionPolicy, typename _ForwardIterator1, typename _ForwardIterator2,
+          typename _OutputIterator, typename _Compare, typename _IsOpDifference>
+_OutputIterator
+__pattern_hetero_set_op(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _ForwardIterator1 __first1,
+                        _ForwardIterator1 __last1, _ForwardIterator2 __first2, _ForwardIterator2 __last2,
+                        _OutputIterator __result, _Compare __comp, _IsOpDifference)
+{
+    typedef typename ::std::iterator_traits<_ForwardIterator1>::difference_type _Size1;
+    typedef typename ::std::iterator_traits<_ForwardIterator2>::difference_type _Size2;
+
+    const _Size1 __n1 = __last1 - __first1;
+    const _Size2 __n2 = __last2 - __first2;
+
+    //Algo is based on the recommended approach of set_intersection algo for GPU: binary search + scan (copying by mask).
+    using _ReduceOp = ::std::plus<_Size1>;
+    using _Assigner = unseq_backend::__scan_assigner;
+    using _NoAssign = unseq_backend::__scan_no_assign;
+    using _MaskAssigner = unseq_backend::__mask_assigner<2>;
+    using _InitType = unseq_backend::__no_init_value<_Size1>;
+    using _DataAcc = unseq_backend::walk_n<_ExecutionPolicy, oneapi::dpl::__internal::__no_op>;
+
+    _ReduceOp __reduce_op;
+    _Assigner __assign_op;
+    _DataAcc __get_data_op;
+    unseq_backend::__copy_by_mask<_ReduceOp, oneapi::dpl::__internal::__pstl_assign, /*inclusive*/ ::std::true_type, 2>
+        __copy_by_mask_op;
+    unseq_backend::__brick_set_op<_ExecutionPolicy, _Compare, _Size1, _Size2, _IsOpDifference> __create_mask_op{
+        __comp, __n1, __n2};
+
+    // temporary buffer to store boolean mask
+    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, int32_t> __mask_buf(__exec, __n1);
+
+    auto __keep1 =
+        oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _ForwardIterator1>();
+    auto __buf1 = __keep1(__first1, __last1);
+    auto __keep2 =
+        oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _ForwardIterator2>();
+    auto __buf2 = __keep2(__first2, __last2);
+
+    auto __keep3 = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::write, _OutputIterator>();
+    auto __buf3 = __keep3(__result, __result + __n1);
+
+    auto __result_size =
+        __par_backend_hetero::__parallel_transform_scan_base(
+            ::std::forward<_ExecutionPolicy>(__exec),
+            oneapi::dpl::__ranges::make_zip_view(
+                __buf1.all_view(), __buf2.all_view(),
+                oneapi::dpl::__ranges::all_view<int32_t, __par_backend_hetero::access_mode::read_write>(
+                    __mask_buf.get_buffer())),
+            __buf3.all_view(), __reduce_op, _InitType{},
+            // local scan
+            unseq_backend::__scan</*inclusive*/ ::std::true_type, _ExecutionPolicy, _ReduceOp, _DataAcc, _Assigner,
+                                  _MaskAssigner, decltype(__create_mask_op), _InitType>{
+                __reduce_op, __get_data_op, __assign_op, _MaskAssigner{}, __create_mask_op},
+            // scan between groups
+            unseq_backend::__scan</*inclusive=*/::std::true_type, _ExecutionPolicy, _ReduceOp, _DataAcc, _NoAssign,
+                                  _Assigner, _DataAcc, _InitType>{__reduce_op, __get_data_op, _NoAssign{}, __assign_op,
+                                                                  __get_data_op},
+            // global scan
+            __copy_by_mask_op)
+            .get();
+
+    return __result + __result_size;
+}
+
 template <typename _BackendTag, typename _ExecutionPolicy, typename _ForwardIterator1, typename _ForwardIterator2,
           typename _OutputIterator, typename _Compare>
 _OutputIterator

From ea0f7ec8df149341ae2539c7585eb2011e4c3e55 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 9 Feb 2024 14:10:53 +0100
Subject: [PATCH 174/566] __pattern_hetero_set_op + tag calls

---
 .../dpl/pstl/hetero/algorithm_impl_hetero.h    | 18 ++++++++++++------
 1 file changed, 12 insertions(+), 6 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 0091ef00e46..52b89025106 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -3187,8 +3187,8 @@ __pattern_set_intersection(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& _
     if (__first1 == __last1 || __first2 == __last2)
         return __result;
 
-    return __pattern_hetero_set_op(::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2,
-                                   __result, __comp, unseq_backend::_IntersectionTag());
+    return __pattern_hetero_set_op(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2,
+                                   __last2, __result, __comp, unseq_backend::_IntersectionTag());
 }
 
 //Dummy names to avoid kernel problems
@@ -3246,8 +3246,8 @@ __pattern_set_difference(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __e
             __first1, __last1, __result, oneapi::dpl::__internal::__brick_copy<_ExecutionPolicy>{});
     }
 
-    return __pattern_hetero_set_op(::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2,
-                                   __result, __comp, unseq_backend::_DifferenceTag());
+    return __pattern_hetero_set_op(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2,
+                                   __last2, __result, __comp, unseq_backend::_DifferenceTag());
 }
 
 //Dummy names to avoid kernel problems
@@ -3353,8 +3353,10 @@ __pattern_set_union(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec,
 
     //1. Calc difference {2} \ {1}
     const auto __n_diff =
-        oneapi::dpl::__internal::__pattern_hetero_set_op(__exec, __first2, __last2, __first1, __last1, __buf, __comp,
-                                                         unseq_backend::_DifferenceTag()) -
+        oneapi::dpl::__internal::__pattern_hetero_set_op(
+            oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, decltype(__first2), decltype(__last2),
+                                                      decltype(__first1), decltype(__last1), decltype(__buf)>(),
+            __exec, __first2, __last2, __first1, __last1, __buf, __comp, unseq_backend::_DifferenceTag()) -
         __buf;
 
     //2. Merge {1} and the difference
@@ -3503,6 +3505,8 @@ __pattern_set_symmetric_difference(__hetero_tag<_BackendTag> __tag, _ExecutionPo
     //1. Calc difference {1} \ {2}
     const auto __n_diff_1 =
         oneapi::dpl::__internal::__pattern_hetero_set_op(
+            oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, decltype(__first1), decltype(__last1),
+                                                      decltype(__first2), decltype(__last2), decltype(__buf_1)>(),
             oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__set_symmetric_difference_phase_1>(__exec),
             __first1, __last1, __first2, __last2, __buf_1, __comp, unseq_backend::_DifferenceTag()) -
         __buf_1;
@@ -3510,6 +3514,8 @@ __pattern_set_symmetric_difference(__hetero_tag<_BackendTag> __tag, _ExecutionPo
     //2. Calc difference {2} \ {1}
     const auto __n_diff_2 =
         oneapi::dpl::__internal::__pattern_hetero_set_op(
+            oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, decltype(__first2), decltype(__last2),
+                                                      decltype(__first1), decltype(__last1), decltype(__buf_2)>(),
             oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__set_symmetric_difference_phase_2>(__exec),
             __first2, __last2, __first1, __last1, __buf_2, __comp, unseq_backend::_DifferenceTag()) -
         __buf_2;

From 669d73da1416b5458ba0f8271741bcb1445b34de Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 9 Feb 2024 14:11:14 +0100
Subject: [PATCH 175/566] __pattern_transform_scan_base + tag impls

---
 .../dpl/pstl/hetero/numeric_impl_hetero.h     | 63 +++++++++++++++++++
 1 file changed, 63 insertions(+)

diff --git a/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
index a051fccfd28..89e2782a0c5 100644
--- a/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
@@ -252,6 +252,69 @@ __pattern_transform_scan(_ExecutionPolicy&& __exec, _Iterator1 __first, _Iterato
                                          __unary_op, _InitType{__init}, __binary_op, _Inclusive{});
 }
 
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2,
+          typename _UnaryOperation, typename _InitType, typename _BinaryOperation, typename _Inclusive>
+_Iterator2
+__pattern_transform_scan_base(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Iterator1 __first,
+                              _Iterator1 __last, _Iterator2 __result, _UnaryOperation __unary_op, _InitType __init,
+                              _BinaryOperation __binary_op, _Inclusive)
+{
+    if (__first == __last)
+        return __result;
+
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator1, _Iterator2>();
+
+    const auto __n = __last - __first;
+
+    auto __keep1 = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _Iterator1>();
+    auto __buf1 = __keep1(__first, __last);
+
+    // This is a temporary workaround for an in-place exclusive scan while the SYCL backend scan pattern is not fixed.
+    const bool __is_scan_inplace_exclusive = __n > 1 && !_Inclusive{} && __iterators_possibly_equal(__first, __result);
+    if (!__is_scan_inplace_exclusive)
+    {
+        auto __keep2 = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::write, _Iterator2>();
+        auto __buf2 = __keep2(__result, __result + __n);
+
+        oneapi::dpl::__par_backend_hetero::__parallel_transform_scan(::std::forward<_ExecutionPolicy>(__exec),
+                                                                     __buf1.all_view(), __buf2.all_view(), __n,
+                                                                     __unary_op, __init, __binary_op, _Inclusive{})
+            .wait();
+    }
+    else
+    {
+        assert(__n > 1);
+        assert(!_Inclusive{});
+        assert(__iterators_possibly_equal(__first, __result));
+
+        using _Type = typename _InitType::__value_type;
+
+        auto __policy =
+            __par_backend_hetero::make_wrapped_policy<ExecutionPolicyWrapper>(::std::forward<_ExecutionPolicy>(__exec));
+        using _NewExecutionPolicy = decltype(__policy);
+
+        // Create temporary buffer
+        oneapi::dpl::__par_backend_hetero::__buffer<_NewExecutionPolicy, _Type> __tmp_buf(__policy, __n);
+        auto __first_tmp = __tmp_buf.get();
+        auto __last_tmp = __first_tmp + __n;
+        auto __keep2 = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::write, _Iterator2>();
+        auto __buf2 = __keep2(__first_tmp, __last_tmp);
+
+        // Run main algorithm and save data into temporary buffer
+        oneapi::dpl::__par_backend_hetero::__parallel_transform_scan(__policy, __buf1.all_view(), __buf2.all_view(),
+                                                                     __n, __unary_op, __init, __binary_op, _Inclusive{})
+            .wait();
+
+        // Move data from temporary buffer into results
+        oneapi::dpl::__internal::__pattern_walk2_brick(__dispatch_tag, ::std::move(__policy), __first_tmp, __last_tmp,
+                                                       __result,
+                                                       oneapi::dpl::__internal::__brick_move<_NewExecutionPolicy>{});
+
+        //TODO: optimize copy back depending on Iterator, i.e. set_final_data for host iterator/pointer
+    }
+
+    return __result + __n;
+}
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2,
           typename _UnaryOperation, typename _Type,
           typename _BinaryOperation, typename _Inclusive>

From 6d81e7fa273ad7fb11b9c22da0e1f3541741078a Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 9 Feb 2024 14:11:30 +0100
Subject: [PATCH 176/566] __pattern_transform_scan_base + tag calls

---
 include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
index 89e2782a0c5..e1af0f7a73c 100644
--- a/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
@@ -326,7 +326,7 @@ __pattern_transform_scan(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __e
     using _RepackedType = __par_backend_hetero::__repacked_tuple_t<_Type>;
     using _InitType = unseq_backend::__init_value<_RepackedType>;
 
-    return __pattern_transform_scan_base(::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result,
+    return __pattern_transform_scan_base(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result,
                                          __unary_op, _InitType{__init}, __binary_op, _Inclusive{});
 }
 
@@ -357,7 +357,7 @@ __pattern_transform_scan(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __e
     using _RepackedType = __par_backend_hetero::__repacked_tuple_t<_Type>;
     using _InitType = unseq_backend::__no_init_value<_RepackedType>;
 
-    return __pattern_transform_scan_base(::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result,
+    return __pattern_transform_scan_base(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result,
                                          __unary_op, _InitType{}, __binary_op, _Inclusive{});
 }
 

From 062988d2894f360772661c1a6b0216ae9712045c Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 9 Feb 2024 17:37:02 +0100
Subject: [PATCH 177/566] 
 include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h - fix compile
 error: no matching function for call to '__parallel_find_or'

---
 .../pstl/hetero/dpcpp/parallel_backend_sycl.h | 104 ++++++++++++++++++
 1 file changed, 104 insertions(+)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
index b437154f666..6038150b03b 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
@@ -1205,6 +1205,109 @@ __parallel_find_or(_ExecutionPolicy&& __exec, _Brick __f, _BrickTag __brick_tag,
         return __result != __init_value ? __result : __rng_n;
 }
 
+// Base pattern for __parallel_or and __parallel_find. The execution depends on tag type _BrickTag.
+template <typename _ExecutionPolicy, typename _Brick, typename _BrickTag, typename... _Ranges>
+::std::conditional_t<
+    ::std::is_same_v<_BrickTag, __parallel_or_tag>, bool,
+    oneapi::dpl::__internal::__difference_t<typename oneapi::dpl::__ranges::__get_first_range_type<_Ranges...>::type>>
+__parallel_find_or(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&& __exec, _Brick __f,
+                   _BrickTag __brick_tag, _Ranges&&... __rngs)
+{
+    using _CustomName = oneapi::dpl::__internal::__policy_kernel_name<_ExecutionPolicy>;
+    using _AtomicType = typename _BrickTag::_AtomicType;
+    using _FindOrKernel =
+        oneapi::dpl::__par_backend_hetero::__internal::__kernel_name_generator<__find_or_kernel, _CustomName, _Brick,
+                                                                               _BrickTag, _Ranges...>;
+
+    constexpr bool __or_tag_check = ::std::is_same_v<_BrickTag, __parallel_or_tag>;
+    auto __rng_n = oneapi::dpl::__ranges::__get_first_range_size(__rngs...);
+    assert(__rng_n > 0);
+
+    // TODO: find a way to generalize getting of reliable work-group size
+    auto __wgroup_size = oneapi::dpl::__internal::__max_work_group_size(__exec);
+#if _ONEDPL_COMPILE_KERNEL
+    auto __kernel = __internal::__kernel_compiler<_FindOrKernel>::__compile(__exec);
+    __wgroup_size = ::std::min(__wgroup_size, oneapi::dpl::__internal::__kernel_work_group_size(__exec, __kernel));
+#endif
+    auto __max_cu = oneapi::dpl::__internal::__max_compute_units(__exec);
+
+    auto __n_groups = (__rng_n - 1) / __wgroup_size + 1;
+    // TODO: try to change __n_groups with another formula for more perfect load balancing
+    __n_groups = ::std::min(__n_groups, decltype(__n_groups)(__max_cu));
+
+    auto __n_iter = (__rng_n - 1) / (__n_groups * __wgroup_size) + 1;
+
+    _PRINT_INFO_IN_DEBUG_MODE(__exec, __wgroup_size, __max_cu);
+
+    _AtomicType __init_value = _BrickTag::__init_value(__rng_n);
+    auto __result = __init_value;
+
+    auto __pred = oneapi::dpl::__par_backend_hetero::__early_exit_find_or<_ExecutionPolicy, _Brick>{__f};
+
+    // scope is to copy data back to __result after destruction of temporary sycl:buffer
+    {
+        auto __temp = sycl::buffer<_AtomicType, 1>(&__result, 1); // temporary storage for global atomic
+
+        // main parallel_for
+        __exec.queue().submit([&](sycl::handler& __cgh) {
+            oneapi::dpl::__ranges::__require_access(__cgh, __rngs...);
+            auto __temp_acc = __temp.template get_access<access_mode::read_write>(__cgh);
+
+            // create local accessor to connect atomic with
+            __dpl_sycl::__local_accessor<_AtomicType> __temp_local(1, __cgh);
+#if _ONEDPL_COMPILE_KERNEL && _ONEDPL_KERNEL_BUNDLE_PRESENT
+            __cgh.use_kernel_bundle(__kernel.get_kernel_bundle());
+#endif
+            __cgh.parallel_for<_FindOrKernel>(
+#if _ONEDPL_COMPILE_KERNEL && !_ONEDPL_KERNEL_BUNDLE_PRESENT
+                __kernel,
+#endif
+                sycl::nd_range</*dim=*/1>(sycl::range</*dim=*/1>(__n_groups * __wgroup_size),
+                                          sycl::range</*dim=*/1>(__wgroup_size)),
+                [=](sycl::nd_item</*dim=*/1> __item_id) {
+                    auto __local_idx = __item_id.get_local_id(0);
+
+                    __dpl_sycl::__atomic_ref<_AtomicType, sycl::access::address_space::global_space> __found(
+                        *__dpl_sycl::__get_accessor_ptr(__temp_acc));
+                    __dpl_sycl::__atomic_ref<_AtomicType, sycl::access::address_space::local_space> __found_local(
+                        *__dpl_sycl::__get_accessor_ptr(__temp_local));
+
+                    // 1. Set initial value to local atomic
+                    if (__local_idx == 0)
+                        __found_local.store(__init_value);
+                    __dpl_sycl::__group_barrier(__item_id);
+
+                    // 2. Find any element that satisfies pred and set local atomic value to global atomic
+                    constexpr auto __comp = typename _BrickTag::_Compare{};
+                    __pred(__item_id, __n_iter, __wgroup_size, __comp, __found_local, __brick_tag, __rngs...);
+                    __dpl_sycl::__group_barrier(__item_id);
+
+                    // Set local atomic value to global atomic
+                    if (__local_idx == 0 && __comp(__found_local.load(), __found.load()))
+                    {
+                        if constexpr (__or_tag_check)
+                            __found.store(1);
+                        else
+                        {
+                            for (auto __old = __found.load(); __comp(__found_local.load(), __old);
+                                 __old = __found.load())
+                            {
+                                __found.compare_exchange_strong(__old, __found_local.load());
+                            }
+                        }
+                    }
+                });
+        });
+        //The end of the scope  -  a point of synchronization (on temporary sycl buffer destruction)
+    }
+
+    if constexpr (__or_tag_check)
+        return __result;
+    else
+        return __result != __init_value ? __result : __rng_n;
+}
+
+
 //------------------------------------------------------------------------
 // parallel_or - sync pattern
 //------------------------------------------------------------------------
@@ -1301,6 +1404,7 @@ __parallel_find(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&
         typename ::std::conditional<_IsFirst::value, __parallel_find_forward_tag<decltype(__buf.all_view())>,
                                     __parallel_find_backward_tag<decltype(__buf.all_view())>>::type;
     return __first + oneapi::dpl::__par_backend_hetero::__parallel_find_or(
+                         oneapi::dpl::__internal::__device_backend_tag{},
                          __par_backend_hetero::make_wrapped_policy<__find_policy_wrapper>(
                              ::std::forward<_ExecutionPolicy>(__exec)),
                          __f, _TagType{}, __buf.all_view());

From 112654a2033b5a04757833996d8ff3a7eebb54fd Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 14 Feb 2024 10:51:40 +0100
Subject: [PATCH 178/566] Fix review comment: I guess a backend type should be
 depended on the environment and a logic of backend selection should be like
 that we have currently...

---
 include/oneapi/dpl/pstl/execution_impl.h      | 23 +++++++++++++++++--
 include/oneapi/dpl/pstl/omp/parallel_for.h    |  2 +-
 .../oneapi/dpl/pstl/omp/parallel_for_each.h   |  2 +-
 .../oneapi/dpl/pstl/parallel_backend_serial.h |  5 ++--
 include/oneapi/dpl/pstl/parallel_impl.h       | 10 +++++---
 5 files changed, 33 insertions(+), 9 deletions(-)

diff --git a/include/oneapi/dpl/pstl/execution_impl.h b/include/oneapi/dpl/pstl/execution_impl.h
index ad29a81b0c5..33ada982ff7 100644
--- a/include/oneapi/dpl/pstl/execution_impl.h
+++ b/include/oneapi/dpl/pstl/execution_impl.h
@@ -19,6 +19,7 @@
 #include <iterator>
 #include <type_traits>
 
+#include "parallel_backend.h"
 #include "execution_defs.h"
 #include "iterator_defs.h"
 
@@ -96,10 +97,28 @@ __is_parallelization_preferred()
 // backend selector with tags
 //------------------------------------------------------------------------
 
+struct __serial_backend_tag
+{
+};
+
 struct __tbb_backend_tag
 {
 };
 
+struct __omp_backend_tag
+{
+};
+
+#if _ONEDPL_PAR_BACKEND_TBB
+using __par_backend_tag = __tbb_backend_tag;
+#elif _ONEDPL_PAR_BACKEND_OPENMP
+using __par_backend_tag = __omp_backend_tag;
+#elif _ONEDPL_PAR_BACKEND_SERIAL
+using __par_backend_tag = __serial_backend_tag;
+#else
+#    error "Parallel backend was not specified"
+#endif
+
 template <class _IsVector>
 struct __serial_tag
 {
@@ -112,7 +131,7 @@ struct __parallel_tag
     using __is_vector = _IsVector;
     // backend tag can be change depending on
     // TBB availability in the environment
-    using __backend_tag = __tbb_backend_tag;
+    using __backend_tag = __par_backend_tag;
 };
 
 struct __parallel_forward_tag
@@ -120,7 +139,7 @@ struct __parallel_forward_tag
     using __is_vector = ::std::false_type;
     // backend tag can be change depending on
     // TBB availability in the environment
-    using __backend_tag = __tbb_backend_tag;
+    using __backend_tag = __par_backend_tag;
 };
 
 template <class _IsVector, class... _IteratorTypes>
diff --git a/include/oneapi/dpl/pstl/omp/parallel_for.h b/include/oneapi/dpl/pstl/omp/parallel_for.h
index 296b24e282b..6f870dc4b4a 100644
--- a/include/oneapi/dpl/pstl/omp/parallel_for.h
+++ b/include/oneapi/dpl/pstl/omp/parallel_for.h
@@ -68,7 +68,7 @@ __parallel_for(_ExecutionPolicy&&, _Index __first, _Index __last, _Fp __f)
 
 template <class _ExecutionPolicy, class _Index, class _Fp>
 void
-__parallel_for(oneapi::dpl::__internal::__tbb_backend_tag, _ExecutionPolicy&&, _Index __first, _Index __last, _Fp __f)
+__parallel_for(oneapi::dpl::__internal::__omp_backend_tag, _ExecutionPolicy&&, _Index __first, _Index __last, _Fp __f)
 {
     if (omp_in_parallel())
     {
diff --git a/include/oneapi/dpl/pstl/omp/parallel_for_each.h b/include/oneapi/dpl/pstl/omp/parallel_for_each.h
index 510cd04c352..9cc4f249f80 100644
--- a/include/oneapi/dpl/pstl/omp/parallel_for_each.h
+++ b/include/oneapi/dpl/pstl/omp/parallel_for_each.h
@@ -63,7 +63,7 @@ __parallel_for_each(_ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterat
 
 template <class _ExecutionPolicy, class _ForwardIterator, class _Fp>
 void
-__parallel_for_each(oneapi::dpl::__internal::__tbb_backend_tag, _ExecutionPolicy&&, _ForwardIterator __first,
+__parallel_for_each(oneapi::dpl::__internal::__omp_backend_tag, _ExecutionPolicy&&, _ForwardIterator __first,
                     _ForwardIterator __last, _Fp __f)
 {
     if (omp_in_parallel())
diff --git a/include/oneapi/dpl/pstl/parallel_backend_serial.h b/include/oneapi/dpl/pstl/parallel_backend_serial.h
index 577fcbd36a0..f8bc92c5fc8 100644
--- a/include/oneapi/dpl/pstl/parallel_backend_serial.h
+++ b/include/oneapi/dpl/pstl/parallel_backend_serial.h
@@ -73,7 +73,8 @@ __parallel_for(_ExecutionPolicy&&, _Index __first, _Index __last, _Fp __f)
 
 template <class _ExecutionPolicy, class _Index, class _Fp>
 void
-__parallel_for(oneapi::dpl::__internal::__tbb_backend_tag, _ExecutionPolicy&&, _Index __first, _Index __last, _Fp __f)
+__parallel_for(oneapi::dpl::__internal::__serial_backend_tag, _ExecutionPolicy&&, _Index __first, _Index __last,
+               _Fp __f)
 {
     __f(__first, __last);
 }
@@ -157,7 +158,7 @@ __parallel_for_each(_ExecutionPolicy&&, _ForwardIterator __begin, _ForwardIterat
 
 template <class _ExecutionPolicy, class _ForwardIterator, class _Fp>
 void
-__parallel_for_each(oneapi::dpl::__internal::__tbb_backend_tag, _ExecutionPolicy&&, _ForwardIterator __begin,
+__parallel_for_each(oneapi::dpl::__internal::__serial_backend_tag, _ExecutionPolicy&&, _ForwardIterator __begin,
                     _ForwardIterator __end, _Fp __f)
 {
     for (auto __iter = __begin; __iter != __end; ++__iter)
diff --git a/include/oneapi/dpl/pstl/parallel_impl.h b/include/oneapi/dpl/pstl/parallel_impl.h
index 752db370260..d30fc053cde 100644
--- a/include/oneapi/dpl/pstl/parallel_impl.h
+++ b/include/oneapi/dpl/pstl/parallel_impl.h
@@ -66,10 +66,14 @@ __parallel_find(_ExecutionPolicy&& __exec, _Index __first, _Index __last, _Brick
     return __extremum != __initial_dist ? __first + __extremum : __last;
 }
 
-template <class _ExecutionPolicy, class _Index, class _Brick, class _IsFirst>
+template <class _Tag, class _ExecutionPolicy, class _Index, class _Brick, class _IsFirst>
 _Index
-__parallel_find(__tbb_backend_tag __tag, _ExecutionPolicy&& __exec, _Index __first, _Index __last, _Brick __f, _IsFirst)
+__parallel_find(_Tag __tag, _ExecutionPolicy&& __exec, _Index __first, _Index __last, _Brick __f, _IsFirst)
 {
+    static_assert(__is_backend_tag_v<_Tag>);
+
+    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
+
     typedef typename ::std::iterator_traits<_Index>::difference_type _DifferenceType;
     const _DifferenceType __n = __last - __first;
     _DifferenceType __initial_dist = _IsFirst::value ? __n : -1;
@@ -78,7 +82,7 @@ __parallel_find(__tbb_backend_tag __tag, _ExecutionPolicy&& __exec, _Index __fir
 
     ::std::atomic<_DifferenceType> __extremum(__initial_dist);
     // TODO: find out what is better here: parallel_for or parallel_reduce
-    __par_backend::__parallel_for(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+    __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
                                   [__comp, __f, __first, &__extremum](_Index __i, _Index __j) {
                                       // See "Reducing Contention Through Priority Updates", PPoPP '13, for discussion of
                                       // why using a shared variable scales fairly well in this situation.

From bd5a343d59fd36ef7c7f408a1a9d690a8b2e4a4a Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 14 Feb 2024 11:25:37 +0100
Subject: [PATCH 179/566] __parallel_transform_reduce + tag impls

---
 .../hetero/dpcpp/parallel_backend_sycl_fpga.h |  11 ++
 .../dpcpp/parallel_backend_sycl_reduce.h      | 100 ++++++++++++++++++
 .../dpl/pstl/omp/parallel_transform_reduce.h  |  30 ++++++
 .../oneapi/dpl/pstl/parallel_backend_serial.h |   8 ++
 .../oneapi/dpl/pstl/parallel_backend_tbb.h    |  12 +++
 5 files changed, 161 insertions(+)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
index f2f07b70f64..85f4f0cf528 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
@@ -145,6 +145,17 @@ __parallel_transform_reduce(_ExecutionPolicy&& __exec, _ReduceOp __reduce_op, _T
         __exec.__device_policy(), __reduce_op, __transform_op, __init, ::std::forward<_Ranges>(__rngs)...);
 }
 
+template <typename _Tp, typename _Commutative, typename _ExecutionPolicy, typename _ReduceOp, typename _TransformOp,
+          typename _InitType, typename... _Ranges>
+auto
+__parallel_transform_reduce(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionPolicy&& __exec,
+                            _ReduceOp __reduce_op, _TransformOp __transform_op, _InitType __init, _Ranges&&... __rngs)
+{
+    // workaround until we implement more performant version for patterns
+    return oneapi::dpl::__par_backend_hetero::__parallel_transform_reduce<_Tp, _Commutative>(
+        __exec.__device_policy(), __reduce_op, __transform_op, __init, ::std::forward<_Ranges>(__rngs)...);
+}
+
 //------------------------------------------------------------------------
 // parallel_transform_scan
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h
index 52b036ab4b3..6eab6ac03ec 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h
@@ -511,6 +511,106 @@ __parallel_transform_reduce(_ExecutionPolicy&& __exec, _ReduceOp __reduce_op, _T
         ::std::forward<_Ranges>(__rngs)...);
 }
 
+template <typename _Tp, typename _Commutative, typename _ExecutionPolicy, typename _ReduceOp, typename _TransformOp,
+          typename _InitType, typename... _Ranges>
+auto
+__parallel_transform_reduce(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&& __exec,
+                            _ReduceOp __reduce_op, _TransformOp __transform_op, _InitType __init, _Ranges&&... __rngs)
+{
+    auto __n = oneapi::dpl::__ranges::__get_first_range_size(__rngs...);
+    assert(__n > 0);
+
+    // Get the work group size adjusted to the local memory limit.
+    // Pessimistically double the memory requirement to take into account memory used by compiled kernel.
+    // TODO: find a way to generalize getting of reliable work-group size.
+    ::std::size_t __work_group_size = oneapi::dpl::__internal::__slm_adjusted_work_group_size(__exec, sizeof(_Tp) * 2);
+
+    // Use single work group implementation if array < __work_group_size * __iters_per_work_item.
+    if (__work_group_size >= 256)
+    {
+        if (__n <= 256)
+        {
+            return __parallel_transform_reduce_small_impl<_Tp, 256, 1, _Commutative>(
+                ::std::forward<_ExecutionPolicy>(__exec), __n, __reduce_op, __transform_op, __init,
+                ::std::forward<_Ranges>(__rngs)...);
+        }
+        else if (__n <= 512)
+        {
+            return __parallel_transform_reduce_small_impl<_Tp, 256, 2, _Commutative>(
+                ::std::forward<_ExecutionPolicy>(__exec), __n, __reduce_op, __transform_op, __init,
+                ::std::forward<_Ranges>(__rngs)...);
+        }
+        else if (__n <= 1024)
+        {
+            return __parallel_transform_reduce_small_impl<_Tp, 256, 4, _Commutative>(
+                ::std::forward<_ExecutionPolicy>(__exec), __n, __reduce_op, __transform_op, __init,
+                ::std::forward<_Ranges>(__rngs)...);
+        }
+        else if (__n <= 2048)
+        {
+            return __parallel_transform_reduce_small_impl<_Tp, 256, 8, _Commutative>(
+                ::std::forward<_ExecutionPolicy>(__exec), __n, __reduce_op, __transform_op, __init,
+                ::std::forward<_Ranges>(__rngs)...);
+        }
+        else if (__n <= 4096)
+        {
+            return __parallel_transform_reduce_small_impl<_Tp, 256, 16, _Commutative>(
+                ::std::forward<_ExecutionPolicy>(__exec), __n, __reduce_op, __transform_op, __init,
+                ::std::forward<_Ranges>(__rngs)...);
+        }
+        else if (__n <= 8192)
+        {
+            return __parallel_transform_reduce_small_impl<_Tp, 256, 32, _Commutative>(
+                ::std::forward<_ExecutionPolicy>(__exec), __n, __reduce_op, __transform_op, __init,
+                ::std::forward<_Ranges>(__rngs)...);
+        }
+
+        // Use two-step tree reduction.
+        // First step reduces __work_group_size * __iters_per_work_item_device_kernel elements.
+        // Second step reduces __work_group_size * __iters_per_work_item_work_group_kernel elements.
+        else if (__n <= 2097152)
+        {
+            return __parallel_transform_reduce_mid_impl<_Tp, 256, 32, 1, _Commutative>(
+                ::std::forward<_ExecutionPolicy>(__exec), __n, __reduce_op, __transform_op, __init,
+                ::std::forward<_Ranges>(__rngs)...);
+        }
+        else if (__n <= 4194304)
+        {
+            return __parallel_transform_reduce_mid_impl<_Tp, 256, 32, 2, _Commutative>(
+                ::std::forward<_ExecutionPolicy>(__exec), __n, __reduce_op, __transform_op, __init,
+                ::std::forward<_Ranges>(__rngs)...);
+        }
+        else if (__n <= 8388608)
+        {
+            return __parallel_transform_reduce_mid_impl<_Tp, 256, 32, 4, _Commutative>(
+                ::std::forward<_ExecutionPolicy>(__exec), __n, __reduce_op, __transform_op, __init,
+                ::std::forward<_Ranges>(__rngs)...);
+        }
+        else if (__n <= 16777216)
+        {
+            return __parallel_transform_reduce_mid_impl<_Tp, 256, 32, 8, _Commutative>(
+                ::std::forward<_ExecutionPolicy>(__exec), __n, __reduce_op, __transform_op, __init,
+                ::std::forward<_Ranges>(__rngs)...);
+        }
+        else if (__n <= 33554432)
+        {
+            return __parallel_transform_reduce_mid_impl<_Tp, 256, 32, 16, _Commutative>(
+                ::std::forward<_ExecutionPolicy>(__exec), __n, __reduce_op, __transform_op, __init,
+                ::std::forward<_Ranges>(__rngs)...);
+        }
+        else if (__n <= 67108864)
+        {
+            return __parallel_transform_reduce_mid_impl<_Tp, 256, 32, 32, _Commutative>(
+                ::std::forward<_ExecutionPolicy>(__exec), __n, __reduce_op, __transform_op, __init,
+                ::std::forward<_Ranges>(__rngs)...);
+        }
+    }
+    // Otherwise use a recursive tree reduction.
+    return __parallel_transform_reduce_impl<_Tp, 32, _Commutative>::submit(
+        ::std::forward<_ExecutionPolicy>(__exec), __n, __work_group_size, __reduce_op, __transform_op, __init,
+        ::std::forward<_Ranges>(__rngs)...);
+}
+
 } // namespace __par_backend_hetero
 } // namespace dpl
 } // namespace oneapi
diff --git a/include/oneapi/dpl/pstl/omp/parallel_transform_reduce.h b/include/oneapi/dpl/pstl/omp/parallel_transform_reduce.h
index d94e5fd36e9..aa400692f51 100644
--- a/include/oneapi/dpl/pstl/omp/parallel_transform_reduce.h
+++ b/include/oneapi/dpl/pstl/omp/parallel_transform_reduce.h
@@ -112,6 +112,36 @@ __parallel_transform_reduce(_ExecutionPolicy&&, _RandomAccessIterator __first, _
     return __result;
 }
 
+template <class _ExecutionPolicy, class _RandomAccessIterator, class _UnaryOp, class _Value, class _Combiner,
+          class _Reduction>
+_Value
+__parallel_transform_reduce(oneapi::dpl::__internal::__omp_backend_tag, _ExecutionPolicy&&,
+                            _RandomAccessIterator __first, _RandomAccessIterator __last, _UnaryOp __unary_op,
+                            _Value __init, _Combiner __combiner, _Reduction __reduction)
+{
+    _Value __result = __init;
+    if (omp_in_parallel())
+    {
+        // We don't create a nested parallel region in an existing parallel
+        // region: just create tasks
+        __result = oneapi::dpl::__omp_backend::__transform_reduce_body(__first, __last, __unary_op, __init, __combiner,
+                                                                       __reduction);
+    }
+    else
+    {
+        // Create a parallel region, and a single thread will create tasks
+        // for the region.
+        _PSTL_PRAGMA(omp parallel)
+        _PSTL_PRAGMA(omp single nowait)
+        {
+            __result = oneapi::dpl::__omp_backend::__transform_reduce_body(__first, __last, __unary_op, __init,
+                                                                           __combiner, __reduction);
+        }
+    }
+
+    return __result;
+}
+
 } // namespace __omp_backend
 } // namespace dpl
 } // namespace oneapi
diff --git a/include/oneapi/dpl/pstl/parallel_backend_serial.h b/include/oneapi/dpl/pstl/parallel_backend_serial.h
index f8bc92c5fc8..654859e7c6e 100644
--- a/include/oneapi/dpl/pstl/parallel_backend_serial.h
+++ b/include/oneapi/dpl/pstl/parallel_backend_serial.h
@@ -102,6 +102,14 @@ __parallel_transform_reduce(_ExecutionPolicy&&, _Index __first, _Index __last, _
     return __reduce(__first, __last, __init);
 }
 
+template <class _ExecutionPolicy, class _Index, class _UnaryOp, class _Tp, class _BinaryOp, class _Reduce>
+_Tp
+__parallel_transform_reduce(oneapi::dpl::__internal::__serial_backend_tag, _ExecutionPolicy&&, _Index __first,
+                            _Index __last, _UnaryOp, _Tp __init, _BinaryOp, _Reduce __reduce)
+{
+    return __reduce(__first, __last, __init);
+}
+
 template <class _ExecutionPolicy, typename _Index, typename _Tp, typename _Rp, typename _Cp, typename _Sp, typename _Ap>
 void
 __parallel_strict_scan(_ExecutionPolicy&&, _Index __n, _Tp __initial, _Rp __reduce, _Cp __combine, _Sp __scan,
diff --git a/include/oneapi/dpl/pstl/parallel_backend_tbb.h b/include/oneapi/dpl/pstl/parallel_backend_tbb.h
index 67db87432e5..fb2339808e3 100644
--- a/include/oneapi/dpl/pstl/parallel_backend_tbb.h
+++ b/include/oneapi/dpl/pstl/parallel_backend_tbb.h
@@ -232,6 +232,18 @@ __parallel_transform_reduce(_ExecutionPolicy&&, _Index __first, _Index __last, _
     return __body.sum();
 }
 
+template <class _ExecutionPolicy, class _Index, class _Up, class _Tp, class _Cp, class _Rp>
+_Tp
+__parallel_transform_reduce(oneapi::dpl::__internal::__tbb_backend_tag, _ExecutionPolicy&&, _Index __first,
+                            _Index __last, _Up __u, _Tp __init, _Cp __combine, _Rp __brick_reduce)
+{
+    __tbb_backend::__par_trans_red_body<_Index, _Up, _Tp, _Cp, _Rp> __body(__u, __init, __combine, __brick_reduce);
+    // The grain size of 3 is used in order to provide minimum 2 elements for each body
+    tbb::this_task_arena::isolate(
+        [__first, __last, &__body]() { tbb::parallel_reduce(tbb::blocked_range<_Index>(__first, __last, 3), __body); });
+    return __body.sum();
+}
+
 //------------------------------------------------------------------------
 // parallel_scan
 //------------------------------------------------------------------------

From f6af9aeac36450db976090efe8271cdb77fc9d04 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 14 Feb 2024 12:42:24 +0100
Subject: [PATCH 180/566] __parallel_transform_reduce + tag calls

---
 .../internal/async_impl/async_impl_hetero.h   | 11 ++++++++--
 include/oneapi/dpl/pstl/numeric_impl.h        | 22 ++++++++++++++-----
 2 files changed, 26 insertions(+), 7 deletions(-)

diff --git a/include/oneapi/dpl/internal/async_impl/async_impl_hetero.h b/include/oneapi/dpl/internal/async_impl/async_impl_hetero.h
index cd080b1e6a1..3e91088e92e 100644
--- a/include/oneapi/dpl/internal/async_impl/async_impl_hetero.h
+++ b/include/oneapi/dpl/internal/async_impl/async_impl_hetero.h
@@ -139,9 +139,13 @@ __pattern_transform_reduce_async(_ExecutionPolicy&& __exec, _RandomAccessIterato
         oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _RandomAccessIterator2>();
     auto __buf2 = __keep2(__first2, __first2 + __n);
 
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     return oneapi::dpl::__par_backend_hetero::__parallel_transform_reduce<_RepackedTp,
                                                                           ::std::true_type /*is_commutative*/>(
-        ::std::forward<_ExecutionPolicy>(__exec), __binary_op1, _Functor{__binary_op2},
+        __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __binary_op1, _Functor{__binary_op2},
         unseq_backend::__init_value<_RepackedTp>{__init}, // initial value
         __buf1.all_view(), __buf2.all_view());
 }
@@ -166,9 +170,12 @@ __pattern_transform_reduce_async(_ExecutionPolicy&& __exec, _ForwardIterator __f
     auto __keep = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _ForwardIterator>();
     auto __buf = __keep(__first, __last);
 
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     return oneapi::dpl::__par_backend_hetero::__parallel_transform_reduce<_RepackedTp,
                                                                           ::std::true_type /*is_commutative*/>(
-        ::std::forward<_ExecutionPolicy>(__exec), __binary_op, _Functor{__unary_op},
+        __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __binary_op, _Functor{__unary_op},
         unseq_backend::__init_value<_RepackedTp>{__init}, // initial value
         __buf.all_view());
 }
diff --git a/include/oneapi/dpl/pstl/numeric_impl.h b/include/oneapi/dpl/pstl/numeric_impl.h
index 39d9e16242e..dc3eec4045d 100644
--- a/include/oneapi/dpl/pstl/numeric_impl.h
+++ b/include/oneapi/dpl/pstl/numeric_impl.h
@@ -92,9 +92,13 @@ __pattern_transform_reduce(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __f
                            _RandomAccessIterator2 __first2, _Tp __init, _BinaryOperation1 __binary_op1,
                            _BinaryOperation2 __binary_op2, _IsVector __is_vector, /*is_parallel=*/::std::true_type)
 {
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     return __internal::__except_handler([&]() {
         return __par_backend::__parallel_transform_reduce(
-            ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1,
+            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1,
             [__first1, __first2, __binary_op2](_RandomAccessIterator1 __i) mutable {
                 return __binary_op2(*__i, *(__first2 + (__i - __first1)));
             },
@@ -115,9 +119,11 @@ __pattern_transform_reduce(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& _
                            _RandomAccessIterator1 __last1, _RandomAccessIterator2 __first2, _Tp __init,
                            _BinaryOperation1 __binary_op1, _BinaryOperation2 __binary_op2)
 {
+    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
+
     return __internal::__except_handler([&]() {
         return __par_backend::__parallel_transform_reduce(
-            ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1,
+            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1,
             [__first1, __first2, __binary_op2](_RandomAccessIterator1 __i) mutable {
                 return __binary_op2(*__i, *(__first2 + (__i - __first1)));
             },
@@ -188,9 +194,13 @@ __pattern_transform_reduce(_ExecutionPolicy&& __exec, _RandomAccessIterator __fi
                            _Tp __init, _BinaryOperation __binary_op, _UnaryOperation __unary_op, _IsVector __is_vector,
                            /*is_parallel=*/::std::true_type)
 {
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     return __internal::__except_handler([&]() {
         return __par_backend::__parallel_transform_reduce(
-            ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
             [__unary_op](_RandomAccessIterator __i) mutable { return __unary_op(*__i); }, __init, __binary_op,
             [__unary_op, __binary_op, __is_vector](_RandomAccessIterator __i, _RandomAccessIterator __j, _Tp __init) {
                 return __internal::__brick_transform_reduce(__i, __j, __init, __binary_op, __unary_op, __is_vector);
@@ -201,13 +211,15 @@ __pattern_transform_reduce(_ExecutionPolicy&& __exec, _RandomAccessIterator __fi
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Tp, class _BinaryOperation,
           class _UnaryOperation>
 _Tp
-__pattern_transform_reduce(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
+__pattern_transform_reduce(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
                            _RandomAccessIterator __last, _Tp __init, _BinaryOperation __binary_op,
                            _UnaryOperation __unary_op)
 {
+    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
+
     return __internal::__except_handler([&]() {
         return __par_backend::__parallel_transform_reduce(
-            ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
             [__unary_op](_RandomAccessIterator __i) mutable { return __unary_op(*__i); }, __init, __binary_op,
             [__unary_op, __binary_op](_RandomAccessIterator __i, _RandomAccessIterator __j, _Tp __init) {
                 return __internal::__brick_transform_reduce(__i, __j, __init, __binary_op, __unary_op, _IsVector{});

From 92d385d6e1bba2ea030a1b2dc3d87cac59ecc8f1 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 14 Feb 2024 14:25:02 +0100
Subject: [PATCH 181/566] __parallel_for + tag calls

---
 .../internal/async_impl/async_impl_hetero.h   |  26 +++-
 .../oneapi/dpl/internal/binary_search_impl.h  |  18 ++-
 include/oneapi/dpl/pstl/algorithm_impl.h      | 129 ++++++++++++++----
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   |  73 ++++++----
 .../dpl/pstl/hetero/histogram_impl_hetero.h   |   2 +-
 .../dpl/pstl/hetero/numeric_impl_hetero.h     |   9 +-
 include/oneapi/dpl/pstl/numeric_impl.h        |  12 +-
 include/oneapi/dpl/pstl/parallel_impl.h       |  28 +++-
 8 files changed, 229 insertions(+), 68 deletions(-)

diff --git a/include/oneapi/dpl/internal/async_impl/async_impl_hetero.h b/include/oneapi/dpl/internal/async_impl/async_impl_hetero.h
index 3e91088e92e..232957a9b4a 100644
--- a/include/oneapi/dpl/internal/async_impl/async_impl_hetero.h
+++ b/include/oneapi/dpl/internal/async_impl/async_impl_hetero.h
@@ -42,9 +42,12 @@ __pattern_walk1_async(_ExecutionPolicy&& __exec, _ForwardIterator __first, _Forw
         oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read_write, _ForwardIterator>();
     auto __buf = __keep(__first, __last);
 
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     auto __future_obj = oneapi::dpl::__par_backend_hetero::__parallel_for(
-        ::std::forward<_ExecutionPolicy>(__exec), unseq_backend::walk_n<_ExecutionPolicy, _Function>{__f}, __n,
-        __buf.all_view());
+        __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
+        unseq_backend::walk_n<_ExecutionPolicy, _Function>{__f}, __n, __buf.all_view());
     return __future_obj;
 }
 
@@ -66,9 +69,13 @@ __pattern_walk2_async(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _Fo
     auto __keep2 = oneapi::dpl::__ranges::__get_sycl_range<__acc_mode2, _ForwardIterator2>();
     auto __buf2 = __keep2(__first2, __first2 + __n);
 
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     auto __future = oneapi::dpl::__par_backend_hetero::__parallel_for(
-        ::std::forward<_ExecutionPolicy>(__exec), unseq_backend::walk_n<_ExecutionPolicy, _Function>{__f}, __n,
-        __buf1.all_view(), __buf2.all_view());
+        __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
+        unseq_backend::walk_n<_ExecutionPolicy, _Function>{__f}, __n, __buf1.all_view(), __buf2.all_view());
 
     if constexpr (_IsSync::value)
         __future.wait();
@@ -95,9 +102,14 @@ __pattern_walk3_async(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _Fo
         oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::write, _ForwardIterator3>();
     auto __buf3 = __keep3(__first3, __first3 + __n);
 
-    auto __future = oneapi::dpl::__par_backend_hetero::__parallel_for(
-        ::std::forward<_ExecutionPolicy>(__exec), unseq_backend::walk_n<_ExecutionPolicy, _Function>{__f}, __n,
-        __buf1.all_view(), __buf2.all_view(), __buf3.all_view());
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1,
+                                                                              _ForwardIterator2, _ForwardIterator3>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
+    auto __future =
+        oneapi::dpl::__par_backend_hetero::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
+                                                          unseq_backend::walk_n<_ExecutionPolicy, _Function>{__f}, __n,
+                                                          __buf1.all_view(), __buf2.all_view(), __buf3.all_view());
 
     return __future.__make_future(__first3 + __n);
 }
diff --git a/include/oneapi/dpl/internal/binary_search_impl.h b/include/oneapi/dpl/internal/binary_search_impl.h
index 0c689fe6b8e..3fd155f03a0 100644
--- a/include/oneapi/dpl/internal/binary_search_impl.h
+++ b/include/oneapi/dpl/internal/binary_search_impl.h
@@ -111,6 +111,10 @@ oneapi::dpl::__internal::__enable_if_hetero_execution_policy<Policy, OutputItera
 lower_bound_impl(Policy&& policy, InputIterator1 start, InputIterator1 end, InputIterator2 value_start,
                  InputIterator2 value_end, OutputIterator result, StrictWeakOrdering comp)
 {
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<Policy, InputIterator1, InputIterator2, OutputIterator>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     namespace __bknd = __par_backend_hetero;
     const auto size = ::std::distance(start, end);
 
@@ -128,7 +132,7 @@ lower_bound_impl(Policy&& policy, InputIterator1 start, InputIterator1 end, Inpu
     auto keep_result = oneapi::dpl::__ranges::__get_sycl_range<__bknd::access_mode::read_write, OutputIterator>();
     auto result_buf = keep_result(result, result + value_size);
     auto zip_vw = make_zip_view(input_buf.all_view(), value_buf.all_view(), result_buf.all_view());
-    __bknd::__parallel_for(::std::forward<Policy>(policy),
+    __bknd::__parallel_for(__backend_tag{}, ::std::forward<Policy>(policy),
                            custom_brick<StrictWeakOrdering, decltype(size), lower_bound>{comp, size}, value_size,
                            zip_vw)
         .wait();
@@ -141,6 +145,10 @@ oneapi::dpl::__internal::__enable_if_hetero_execution_policy<Policy, OutputItera
 upper_bound_impl(Policy&& policy, InputIterator1 start, InputIterator1 end, InputIterator2 value_start,
                  InputIterator2 value_end, OutputIterator result, StrictWeakOrdering comp)
 {
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<Policy, InputIterator1, InputIterator2, OutputIterator>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     namespace __bknd = __par_backend_hetero;
     const auto size = ::std::distance(start, end);
 
@@ -158,7 +166,7 @@ upper_bound_impl(Policy&& policy, InputIterator1 start, InputIterator1 end, Inpu
     auto keep_result = oneapi::dpl::__ranges::__get_sycl_range<__bknd::access_mode::read_write, OutputIterator>();
     auto result_buf = keep_result(result, result + value_size);
     auto zip_vw = make_zip_view(input_buf.all_view(), value_buf.all_view(), result_buf.all_view());
-    __bknd::__parallel_for(::std::forward<Policy>(policy),
+    __bknd::__parallel_for(__backend_tag{}, ::std::forward<Policy>(policy),
                            custom_brick<StrictWeakOrdering, decltype(size), upper_bound>{comp, size}, value_size,
                            zip_vw)
         .wait();
@@ -171,6 +179,10 @@ oneapi::dpl::__internal::__enable_if_hetero_execution_policy<Policy, OutputItera
 binary_search_impl(Policy&& policy, InputIterator1 start, InputIterator1 end, InputIterator2 value_start,
                    InputIterator2 value_end, OutputIterator result, StrictWeakOrdering comp)
 {
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<Policy, InputIterator1, InputIterator2, OutputIterator>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     namespace __bknd = __par_backend_hetero;
     const auto size = ::std::distance(start, end);
 
@@ -188,7 +200,7 @@ binary_search_impl(Policy&& policy, InputIterator1 start, InputIterator1 end, In
     auto keep_result = oneapi::dpl::__ranges::__get_sycl_range<__bknd::access_mode::read_write, OutputIterator>();
     auto result_buf = keep_result(result, result + value_size);
     auto zip_vw = make_zip_view(input_buf.all_view(), value_buf.all_view(), result_buf.all_view());
-    __bknd::__parallel_for(::std::forward<Policy>(policy),
+    __bknd::__parallel_for(__backend_tag{}, ::std::forward<Policy>(policy),
                            custom_brick<StrictWeakOrdering, decltype(size), binary_search>{comp, size}, value_size,
                            zip_vw)
         .wait();
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 705b8d278f6..d32664bfcc6 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -226,8 +226,12 @@ __pattern_walk1(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _Rando
                 _IsVector __is_vector,
                 /*parallel=*/::std::true_type)
 {
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     __internal::__except_handler([&]() {
-        __par_backend::__parallel_for(::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+        __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
                                       [__f, __is_vector](_RandomAccessIterator __i, _RandomAccessIterator __j) {
                                           __internal::__brick_walk1(__i, __j, __f, __is_vector);
                                       });
@@ -272,9 +276,13 @@ __pattern_walk_brick(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _
                      _Brick __brick,
                      /*parallel=*/::std::true_type)
 {
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     constexpr auto __is_vector = __internal::__is_vectorization_preferred<_ExecutionPolicy, _RandomAccessIterator>();
     __internal::__except_handler([&]() {
-        __par_backend::__parallel_for(::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+        __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
                                       [__brick, __is_vector](_RandomAccessIterator __i, _RandomAccessIterator __j) {
                                           __brick(__i, __j, __is_vector);
                                       });
@@ -377,9 +385,13 @@ oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _Ra
 __pattern_walk_brick_n(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _Size __n, _Brick __brick,
                        /*is_parallel=*/::std::true_type)
 {
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     constexpr auto __is_vector = __internal::__is_vectorization_preferred<_ExecutionPolicy, _RandomAccessIterator>();
     return __internal::__except_handler([&]() {
-        __par_backend::__parallel_for(::std::forward<_ExecutionPolicy>(__exec), __first, __first + __n,
+        __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __first + __n,
                                       [__brick, __is_vector](_RandomAccessIterator __i, _RandomAccessIterator __j) {
                                           __brick(__i, __j - __i, __is_vector);
                                       });
@@ -470,9 +482,13 @@ oneapi::dpl::__internal::__enable_if_host_execution_policy_conditional<
 __pattern_walk2(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1, _RandomAccessIterator1 __last1,
                 _RandomAccessIterator2 __first2, _Function __f, _IsVector __is_vector, /*parallel=*/::std::true_type)
 {
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     return __internal::__except_handler([&]() {
         __par_backend::__parallel_for(
-            ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1,
+            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1,
             [__f, __first1, __first2, __is_vector](_RandomAccessIterator1 __i, _RandomAccessIterator1 __j) {
                 __internal::__brick_walk2(__i, __j, __first2 + (__i - __first1), __f, __is_vector);
             });
@@ -485,9 +501,11 @@ _RandomAccessIterator2
 __pattern_walk2(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1,
                 _RandomAccessIterator1 __last1, _RandomAccessIterator2 __first2, _Function __f)
 {
+    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
+
     return __internal::__except_handler([&]() {
         __par_backend::__parallel_for(
-            ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1,
+            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1,
             [__f, __first1, __first2](_RandomAccessIterator1 __i, _RandomAccessIterator1 __j) {
                 __internal::__brick_walk2(__i, __j, __first2 + (__i - __first1), __f, _IsVector{});
             });
@@ -619,11 +637,16 @@ oneapi::dpl::__internal::__enable_if_host_execution_policy_conditional<
 __pattern_walk2_brick(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1, _RandomAccessIterator1 __last1,
                       _RandomAccessIterator2 __first2, _Brick __brick, /*parallel=*/::std::true_type)
 {
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     constexpr auto __is_vector =
         __internal::__is_vectorization_preferred<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2>();
 
     return __except_handler([&]() {
         __par_backend::__parallel_for(
+            __backend_tag{},
             ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1,
             [&__is_vector, __first1, __first2, __brick](_RandomAccessIterator1 __i, _RandomAccessIterator1 __j) {
                 __brick(__i, __j, __first2 + (__i - __first1), __is_vector);
@@ -714,11 +737,16 @@ oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _Ra
 __pattern_walk2_brick_n(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1, _Size __n,
                         _RandomAccessIterator2 __first2, _Brick __brick, /*parallel=*/::std::true_type)
 {
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     constexpr auto __is_vector =
         __internal::__is_vectorization_preferred<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2>();
 
     return __except_handler([&]() {
         __par_backend::__parallel_for(
+            __backend_tag{},
             ::std::forward<_ExecutionPolicy>(__exec), __first1, __first1 + __n,
             [&__is_vector, __first1, __first2, __brick](_RandomAccessIterator1 __i, _RandomAccessIterator1 __j) {
                 __brick(__i, __j - __i, __first2 + (__i - __first1), __is_vector);
@@ -820,8 +848,14 @@ __pattern_walk3(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1, _Ran
                 _RandomAccessIterator2 __first2, _RandomAccessIterator3 __first3, _Function __f, _IsVector __is_vector,
                 /*parallel=*/::std::true_type)
 {
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2,
+                                                  _RandomAccessIterator3>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     return __internal::__except_handler([&]() {
         __par_backend::__parallel_for(
+            __backend_tag{},
             ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1,
             [__f, __first1, __first2, __first3, __is_vector](_RandomAccessIterator1 __i, _RandomAccessIterator1 __j) {
                 __internal::__brick_walk3(__i, __j, __first2 + (__i - __first1), __first3 + (__i - __first1), __f,
@@ -837,8 +871,11 @@ _RandomAccessIterator3
 __pattern_walk3(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1, _RandomAccessIterator1 __last1,
                 _RandomAccessIterator2 __first2, _RandomAccessIterator3 __first3, _Function __f)
 {
+    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
+
     return __internal::__except_handler([&]() {
         __par_backend::__parallel_for(
+            __backend_tag{},
             ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1,
             [__f, __first1, __first2, __first3](_RandomAccessIterator1 __i, _RandomAccessIterator1 __j) {
                 __internal::__brick_walk3(__i, __j, __first2 + (__i - __first1), __first3 + (__i - __first1), __f,
@@ -2191,8 +2228,8 @@ __remove_elements(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardI
             [&__m](_DifferenceType __total) { __m = __total; });
 
         // 3. Elements from result are moved to [first, last)
-        __par_backend::__parallel_for(::std::forward<_ExecutionPolicy>(__exec), __result, __result + __m,
-                                      [__result, __first, __is_vector](_Tp* __i, _Tp* __j) {
+        __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __result,
+                                      __result + __m, [__result, __first, __is_vector](_Tp* __i, _Tp* __j) {
                                           __brick_move_destroy<_ExecutionPolicy>{}(__i, __j, __first + (__i - __result),
                                                                                    __is_vector);
                                       });
@@ -2493,8 +2530,12 @@ oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>
 __pattern_reverse(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIterator __last,
                   _IsVector __is_vector, /*is_parallel=*/::std::true_type)
 {
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     __par_backend::__parallel_for(
-        ::std::forward<_ExecutionPolicy>(__exec), __first, __first + (__last - __first) / 2,
+        __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __first + (__last - __first) / 2,
         [__is_vector, __first, __last](_RandomAccessIterator __inner_first, _RandomAccessIterator __inner_last) {
             __internal::__brick_reverse(__inner_first, __inner_last, __last - (__inner_first - __first), __is_vector);
         });
@@ -2505,8 +2546,10 @@ void
 __pattern_reverse(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
                   _RandomAccessIterator __last)
 {
+    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
+
     __par_backend::__parallel_for(
-        ::std::forward<_ExecutionPolicy>(__exec), __first, __first + (__last - __first) / 2,
+        __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __first + (__last - __first) / 2,
         [__first, __last](_RandomAccessIterator __inner_first, _RandomAccessIterator __inner_last) {
             __internal::__brick_reverse(__inner_first, __inner_last, __last - (__inner_first - __first), _IsVector{});
         });
@@ -2559,8 +2602,12 @@ oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _Ra
 __pattern_reverse_copy(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first, _RandomAccessIterator1 __last,
                        _RandomAccessIterator2 __d_first, _IsVector __is_vector, /*is_parallel=*/::std::true_type)
 {
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     auto __len = __last - __first;
-    __par_backend::__parallel_for(::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+    __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
                                   [__is_vector, __first, __len, __d_first](_RandomAccessIterator1 __inner_first,
                                                                            _RandomAccessIterator1 __inner_last) {
                                       __internal::__brick_reverse_copy(__inner_first, __inner_last,
@@ -2661,6 +2708,10 @@ oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _Ra
 __pattern_rotate(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIterator __middle,
                  _RandomAccessIterator __last, _IsVector __is_vector, /*is_parallel=*/::std::true_type)
 {
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     typedef typename ::std::iterator_traits<_RandomAccessIterator>::value_type _Tp;
     auto __n = __last - __first;
     auto __m = __middle - __first;
@@ -2670,19 +2721,19 @@ __pattern_rotate(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _Rand
         return __internal::__except_handler([&__exec, __n, __m, __first, __middle, __last, __is_vector, &__buf]() {
             _Tp* __result = __buf.get();
             __par_backend::__parallel_for(
-                ::std::forward<_ExecutionPolicy>(__exec), __middle, __last,
+                __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __middle, __last,
                 [__middle, __result, __is_vector](_RandomAccessIterator __b, _RandomAccessIterator __e) {
                     __internal::__brick_uninitialized_move(__b, __e, __result + (__b - __middle), __is_vector);
                 });
 
             __par_backend::__parallel_for(
-                ::std::forward<_ExecutionPolicy>(__exec), __first, __middle,
+                __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __middle,
                 [__last, __middle, __is_vector](_RandomAccessIterator __b, _RandomAccessIterator __e) {
                     __internal::__brick_move<_ExecutionPolicy>{}(__b, __e, __b + (__last - __middle), __is_vector);
                 });
 
-            __par_backend::__parallel_for(::std::forward<_ExecutionPolicy>(__exec), __result, __result + (__n - __m),
-                                          [__first, __result, __is_vector](_Tp* __b, _Tp* __e) {
+            __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __result,
+                                          __result + (__n - __m), [__first, __result, __is_vector](_Tp* __b, _Tp* __e) {
                                               __brick_move_destroy<_ExecutionPolicy>{}(
                                                   __b, __e, __first + (__b - __result), __is_vector);
                                           });
@@ -2696,18 +2747,19 @@ __pattern_rotate(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _Rand
         return __internal::__except_handler([&__exec, __n, __m, __first, __middle, __last, __is_vector, &__buf]() {
             _Tp* __result = __buf.get();
             __par_backend::__parallel_for(
-                ::std::forward<_ExecutionPolicy>(__exec), __first, __middle,
+                __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __middle,
                 [__first, __result, __is_vector](_RandomAccessIterator __b, _RandomAccessIterator __e) {
                     __internal::__brick_uninitialized_move(__b, __e, __result + (__b - __first), __is_vector);
                 });
 
             __par_backend::__parallel_for(
-                ::std::forward<_ExecutionPolicy>(__exec), __middle, __last,
+                __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __middle, __last,
                 [__first, __middle, __is_vector](_RandomAccessIterator __b, _RandomAccessIterator __e) {
                     __internal::__brick_move<_ExecutionPolicy>{}(__b, __e, __first + (__b - __middle), __is_vector);
                 });
 
-            __par_backend::__parallel_for(::std::forward<_ExecutionPolicy>(__exec), __result, __result + __m,
+            __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __result,
+                                          __result + __m,
                                           [__n, __m, __first, __result, __is_vector](_Tp* __b, _Tp* __e) {
                                               __brick_move_destroy<_ExecutionPolicy>{}(
                                                   __b, __e, __first + ((__n - __m) + (__b - __result)), __is_vector);
@@ -2831,7 +2883,12 @@ __pattern_rotate_copy(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first,
                       _RandomAccessIterator1 __last, _RandomAccessIterator2 __result, _IsVector __is_vector,
                       /*is_parallel=*/::std::true_type)
 {
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     __par_backend::__parallel_for(
+        __backend_tag{},
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
         [__first, __last, __middle, __result, __is_vector](_RandomAccessIterator1 __b, _RandomAccessIterator1 __e) {
             __internal::__brick_copy<_ExecutionPolicy> __copy{};
@@ -3212,6 +3269,7 @@ __pattern_partition(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _R
             else if (__size2 > __size1)
             {
                 __par_backend::__parallel_for(
+                    __backend_tag{},
                     ::std::forward<_ExecutionPolicy>(__exec), __val1.__pivot, __val1.__pivot + __size1,
                     [__val1, __val2, __size1, __is_vector](_RandomAccessIterator __i, _RandomAccessIterator __j) {
                         __internal::__brick_swap_ranges(__i, __j, (__val2.__pivot - __size1) + (__i - __val1.__pivot),
@@ -3223,6 +3281,7 @@ __pattern_partition(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _R
             else
             {
                 __par_backend::__parallel_for(
+                    __backend_tag{},
                     ::std::forward<_ExecutionPolicy>(__exec), __val1.__pivot, __val1.__pivot + __size2,
                     [__val1, __val2, __is_vector](_RandomAccessIterator __i, _RandomAccessIterator __j) {
                         __internal::__brick_swap_ranges(__i, __j, __val2.__begin + (__i - __val1.__pivot), __is_vector);
@@ -3931,7 +3990,7 @@ __pattern_partial_sort_copy(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __
                 __n2);
 
             // 3. Move elements from temporary buffer to output
-            __par_backend::__parallel_for(::std::forward<_ExecutionPolicy>(__exec), __r, __r + __n2,
+            __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __r, __r + __n2,
                                           [__r, __d_first, __is_vector](_T1* __i, _T1* __j) {
                                               __brick_move_destroy<_ExecutionPolicy>{}(
                                                   __i, __j, __d_first + (__i - __r), __is_vector);
@@ -3939,7 +3998,7 @@ __pattern_partial_sort_copy(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __
 
             if constexpr (!::std::is_trivially_destructible_v<_T1>)
                 __par_backend::__parallel_for(
-                    ::std::forward<_ExecutionPolicy>(__exec), __r + __n2, __r + __n1,
+                    __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __r + __n2, __r + __n1,
                     [__is_vector](_T1* __i, _T1* __j) { __brick_destroy(__i, __j, __is_vector); });
 
             return __d_first + __n2;
@@ -4305,8 +4364,13 @@ __pattern_fill(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _Random
                const _Tp& __value,
                /*is_parallel=*/::std::true_type, _IsVector __is_vector)
 {
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     return __internal::__except_handler([&__exec, __first, __last, &__value, __is_vector]() {
         __par_backend::__parallel_for(
+            __backend_tag{},
             ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
             [&__value, __is_vector](_RandomAccessIterator __begin, _RandomAccessIterator __end) {
                 __internal::__brick_fill<_Tp, _ExecutionPolicy>{__value}(__begin, __end, __is_vector);
@@ -4434,8 +4498,12 @@ __pattern_generate(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _Ra
                    _Generator __g,
                    /*is_parallel=*/::std::true_type, _IsVector __is_vector)
 {
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     return __internal::__except_handler([&]() {
-        __par_backend::__parallel_for(::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+        __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
                                       [__g, __is_vector](_RandomAccessIterator __begin, _RandomAccessIterator __end) {
                                           __internal::__brick_generate(__begin, __end, __g, __is_vector);
                                       });
@@ -4724,6 +4792,11 @@ __pattern_inplace_merge(_ExecutionPolicy&& __exec, _RandomAccessIterator __first
     {
         return;
     }
+
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     typedef typename ::std::iterator_traits<_RandomAccessIterator>::value_type _Tp;
     auto __n = __last - __first;
     __par_backend::__buffer<_ExecutionPolicy, _Tp> __buf(__n);
@@ -4749,10 +4822,11 @@ __pattern_inplace_merge(_ExecutionPolicy&& __exec, _RandomAccessIterator __first
                                                     __move_sequences, __move_sequences);
                 return __f3 + (__l1 - __f1) + (__l2 - __f2);
             });
-        __par_backend::__parallel_for(
-            ::std::forward<_ExecutionPolicy>(__exec), __r, __r + __n, [__r, __first, __is_vector](_Tp* __i, _Tp* __j) {
-                __brick_move_destroy<_ExecutionPolicy>{}(__i, __j, __first + (__i - __r), __is_vector);
-            });
+        __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __r, __r + __n,
+                                      [__r, __first, __is_vector](_Tp* __i, _Tp* __j) {
+                                          __brick_move_destroy<_ExecutionPolicy>{}(__i, __j, __first + (__i - __r),
+                                                                                   __is_vector);
+                                      });
     });
 }
 
@@ -6434,6 +6508,9 @@ __pattern_shift_left(_ExecutionPolicy&& __exec, _ForwardIterator __first, _Forwa
                      typename ::std::iterator_traits<_ForwardIterator>::difference_type __n, _IsVector __is_vector,
                      /*is_parallel=*/::std::true_type)
 {
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     //If (n > 0 && n < m), returns first + (m - n). Otherwise, if n  > 0, returns first. Otherwise, returns last.
     if (__n <= 0)
         return __last;
@@ -6449,7 +6526,7 @@ __pattern_shift_left(_ExecutionPolicy&& __exec, _ForwardIterator __first, _Forwa
     //1. n >= size/2; there is enough memory to 'total' parallel copying
     if (__n >= __mid)
     {
-        __par_backend::__parallel_for(::std::forward<_ExecutionPolicy>(__exec), __n, __size,
+        __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __n, __size,
                                       [__first, __n, __is_vector](_DiffType __i, _DiffType __j) {
                                           __brick_move<_ExecutionPolicy>{}(__first + __i, __first + __j,
                                                                            __first + __i - __n, __is_vector);
@@ -6461,7 +6538,7 @@ __pattern_shift_left(_ExecutionPolicy&& __exec, _ForwardIterator __first, _Forwa
         for (auto __k = __n; __k < __size; __k += __n)
         {
             auto __end = ::std::min(__k + __n, __size);
-            __par_backend::__parallel_for(::std::forward<_ExecutionPolicy>(__exec), __k, __end,
+            __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __k, __end,
                                           [__first, __n, __is_vector](_DiffType __i, _DiffType __j) {
                                               __brick_move<_ExecutionPolicy>{}(__first + __i, __first + __j,
                                                                                __first + __i - __n, __is_vector);
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 52b89025106..7df8d6d7a0d 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -52,7 +52,10 @@ __pattern_walk1(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIte
         oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read_write, _ForwardIterator>();
     auto __buf = __keep(__first, __last);
 
-    oneapi::dpl::__par_backend_hetero::__parallel_for(::std::forward<_ExecutionPolicy>(__exec),
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
+    oneapi::dpl::__par_backend_hetero::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
                                                       unseq_backend::walk_n<_ExecutionPolicy, _Function>{__f}, __n,
                                                       __buf.all_view())
         .wait();
@@ -126,9 +129,13 @@ __pattern_walk2(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardI
     auto __keep2 = oneapi::dpl::__ranges::__get_sycl_range<__acc_mode2, _ForwardIterator2>();
     auto __buf2 = __keep2(__first2, __first2 + __n);
 
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     auto __future_obj = oneapi::dpl::__par_backend_hetero::__parallel_for(
-        ::std::forward<_ExecutionPolicy>(__exec), unseq_backend::walk_n<_ExecutionPolicy, _Function>{__f}, __n,
-        __buf1.all_view(), __buf2.all_view());
+        __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
+        unseq_backend::walk_n<_ExecutionPolicy, _Function>{__f}, __n, __buf1.all_view(), __buf2.all_view());
 
     if constexpr (_IsSync())
         __future_obj.wait();
@@ -155,8 +162,8 @@ __pattern_walk2(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _ForwardIt
     auto __buf2 = __keep2(__first2, __first2 + __n);
 
     auto __future_obj = oneapi::dpl::__par_backend_hetero::__parallel_for(
-        ::std::forward<_ExecutionPolicy>(__exec), unseq_backend::walk_n<_ExecutionPolicy, _Function>{__f}, __n,
-        __buf1.all_view(), __buf2.all_view());
+        _BackendTag{}, ::std::forward<_ExecutionPolicy>(__exec),
+        unseq_backend::walk_n<_ExecutionPolicy, _Function>{__f}, __n, __buf1.all_view(), __buf2.all_view());
 
     if constexpr (_IsSync())
         __future_obj.wait();
@@ -240,7 +247,11 @@ __pattern_walk3(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardI
     auto __keep3 = oneapi::dpl::__ranges::__get_sycl_range<__acc_mode3, _ForwardIterator3>();
     auto __buf3 = __keep3(__first3, __first3 + __n);
 
-    oneapi::dpl::__par_backend_hetero::__parallel_for(::std::forward<_ExecutionPolicy>(__exec),
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1,
+                                                                              _ForwardIterator2, _ForwardIterator3>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
+    oneapi::dpl::__par_backend_hetero::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
                                                       unseq_backend::walk_n<_ExecutionPolicy, _Function>{__f}, __n,
                                                       __buf1.all_view(), __buf2.all_view(), __buf3.all_view())
         .wait();
@@ -269,7 +280,7 @@ __pattern_walk3(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _ForwardIt
     auto __keep3 = oneapi::dpl::__ranges::__get_sycl_range<__acc_mode3, _ForwardIterator3>();
     auto __buf3 = __keep3(__first3, __first3 + __n);
 
-    oneapi::dpl::__par_backend_hetero::__parallel_for(::std::forward<_ExecutionPolicy>(__exec),
+    oneapi::dpl::__par_backend_hetero::__parallel_for(_BackendTag{}, ::std::forward<_ExecutionPolicy>(__exec),
                                                       unseq_backend::walk_n<_ExecutionPolicy, _Function>{__f}, __n,
                                                       __buf1.all_view(), __buf2.all_view(), __buf3.all_view())
         .wait();
@@ -2806,10 +2817,13 @@ __pattern_reverse(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last
     if (__n <= 0)
         return;
 
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     auto __keep = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read_write, _Iterator>();
     auto __buf = __keep(__first, __last);
     oneapi::dpl::__par_backend_hetero::__parallel_for(
-        ::std::forward<_ExecutionPolicy>(__exec),
+        __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
         unseq_backend::__reverse_functor<typename ::std::iterator_traits<_Iterator>::difference_type>{__n}, __n / 2,
         __buf.all_view())
         .wait();
@@ -2828,8 +2842,7 @@ __pattern_reverse(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _I
     auto __keep = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read_write, _Iterator>();
     auto __buf = __keep(__first, __last);
     oneapi::dpl::__par_backend_hetero::__parallel_for(
-        __backend_tag{},
-        ::std::forward<_ExecutionPolicy>(__exec),
+        __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
         unseq_backend::__reverse_functor<typename ::std::iterator_traits<_Iterator>::difference_type>{__n}, __n / 2,
         __buf.all_view())
         .wait();
@@ -2852,9 +2865,14 @@ __pattern_reverse_copy(_ExecutionPolicy&& __exec, _BidirectionalIterator __first
     auto __buf1 = __keep1(__first, __last);
     auto __keep2 =
         oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::write, _ForwardIterator>();
+
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _BidirectionalIterator>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     auto __buf2 = __keep2(__result, __result + __n);
     oneapi::dpl::__par_backend_hetero::__parallel_for(
-        ::std::forward<_ExecutionPolicy>(__exec),
+        __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
         unseq_backend::__reverse_copy<typename ::std::iterator_traits<_BidirectionalIterator>::difference_type>{__n},
         __n, __buf1.all_view(), __buf2.all_view())
         .wait();
@@ -2918,16 +2936,20 @@ __pattern_rotate(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __new_f
         oneapi::dpl::__ranges::all_view<_Tp, __par_backend_hetero::access_mode::write>(__temp_buf.get_buffer());
 
     const auto __shift = __new_first - __first;
+
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     oneapi::dpl::__par_backend_hetero::__parallel_for(
-        oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__rotate_wrapper>(__exec),
+        __backend_tag{}, oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__rotate_wrapper>(__exec),
         unseq_backend::__rotate_copy<typename ::std::iterator_traits<_Iterator>::difference_type>{__n, __shift}, __n,
         __buf.all_view(), __temp_rng);
 
     using _Function = __brick_move<_ExecutionPolicy>;
     auto __brick = unseq_backend::walk_n<_ExecutionPolicy, _Function>{_Function{}};
 
-    oneapi::dpl::__par_backend_hetero::__parallel_for(::std::forward<_ExecutionPolicy>(__exec), __brick, __n,
-                                                      __temp_rng, __buf.all_view())
+    oneapi::dpl::__par_backend_hetero::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
+                                                      __brick, __n, __temp_rng, __buf.all_view())
         .wait();
 
     return __first + (__last - __new_first);
@@ -2938,8 +2960,6 @@ _Iterator
 __pattern_rotate(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __new_first,
                  _Iterator __last)
 {
-    using __backend_tag = typename decltype(__tag)::__backend_tag;
-
     auto __n = __last - __first;
     if (__n <= 0)
         return __first;
@@ -2955,14 +2975,14 @@ __pattern_rotate(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _It
 
     const auto __shift = __new_first - __first;
     oneapi::dpl::__par_backend_hetero::__parallel_for(
-        __backend_tag{}, oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__rotate_wrapper>(__exec),
+        _BackendTag{}, oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__rotate_wrapper>(__exec),
         unseq_backend::__rotate_copy<typename ::std::iterator_traits<_Iterator>::difference_type>{__n, __shift}, __n,
         __buf.all_view(), __temp_rng);
 
     using _Function = __brick_move<_ExecutionPolicy>;
     auto __brick = unseq_backend::walk_n<_ExecutionPolicy, _Function>{_Function{}};
 
-    oneapi::dpl::__par_backend_hetero::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
+    oneapi::dpl::__par_backend_hetero::__parallel_for(_BackendTag{}, ::std::forward<_ExecutionPolicy>(__exec),
                                                       __brick, __n, __temp_rng, __buf.all_view())
         .wait();
 
@@ -2978,6 +2998,10 @@ __pattern_rotate_copy(_ExecutionPolicy&& __exec, _BidirectionalIterator __first,
                       _BidirectionalIterator __last, _ForwardIterator __result, /*vector=*/::std::true_type,
                       /*parallel=*/::std::true_type)
 {
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _BidirectionalIterator, _ForwardIterator>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     auto __n = __last - __first;
     if (__n <= 0)
         return __result;
@@ -2992,6 +3016,7 @@ __pattern_rotate_copy(_ExecutionPolicy&& __exec, _BidirectionalIterator __first,
     const auto __shift = __new_first - __first;
 
     oneapi::dpl::__par_backend_hetero::__parallel_for(
+        __backend_tag{},
         ::std::forward<_ExecutionPolicy>(__exec),
         unseq_backend::__rotate_copy<typename ::std::iterator_traits<_BidirectionalIterator>::difference_type>{__n,
                                                                                                                __shift},
@@ -3006,8 +3031,6 @@ _ForwardIterator
 __pattern_rotate_copy(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _BidirectionalIterator __first,
                       _BidirectionalIterator __new_first, _BidirectionalIterator __last, _ForwardIterator __result)
 {
-    using __backend_tag = typename decltype(__tag)::__backend_tag;
-
     auto __n = __last - __first;
     if (__n <= 0)
         return __result;
@@ -3022,7 +3045,7 @@ __pattern_rotate_copy(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec
     const auto __shift = __new_first - __first;
 
     oneapi::dpl::__par_backend_hetero::__parallel_for(
-        __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
+        _BackendTag{}, ::std::forward<_ExecutionPolicy>(__exec),
         unseq_backend::__rotate_copy<typename ::std::iterator_traits<_BidirectionalIterator>::difference_type>{__n,
                                                                                                                __shift},
         __n, __buf1.all_view(), __buf2.all_view())
@@ -3547,6 +3570,9 @@ __pattern_shift_left(_ExecutionPolicy&& __exec, _Range __rng, oneapi::dpl::__int
     _DiffType __mid = __size / 2 + __size % 2;
     _DiffType __size_res = __size - __n;
 
+    //TODO: required to implement correct tag selection here
+    using __backend_tag = oneapi::dpl::__internal::__device_backend_tag;
+
     //1. n >= size/2; 'size - _n' parallel copying
     if (__n >= __mid)
     {
@@ -3557,14 +3583,15 @@ __pattern_shift_left(_ExecutionPolicy&& __exec, _Range __rng, oneapi::dpl::__int
         auto __src = oneapi::dpl::__ranges::drop_view_simple<_Range, _DiffType>(__rng, __n);
         auto __dst = oneapi::dpl::__ranges::take_view_simple<_Range, _DiffType>(__rng, __size_res);
 
-        oneapi::dpl::__par_backend_hetero::__parallel_for(::std::forward<_ExecutionPolicy>(__exec), __brick, __size_res,
-                                                          __src, __dst)
+        oneapi::dpl::__par_backend_hetero::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
+                                                          __brick, __size_res, __src, __dst)
             .wait();
     }
     else //2. n < size/2; 'n' parallel copying
     {
         auto __brick = unseq_backend::__brick_shift_left<_ExecutionPolicy, _DiffType>{__size, __n};
         oneapi::dpl::__par_backend_hetero::__parallel_for(
+            __backend_tag{},
             oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__shift_left_right>(
                 ::std::forward<_ExecutionPolicy>(__exec)),
             __brick, __n, __rng)
diff --git a/include/oneapi/dpl/pstl/hetero/histogram_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/histogram_impl_hetero.h
index fbdeb161c29..c594e5c25a0 100644
--- a/include/oneapi/dpl/pstl/hetero/histogram_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/histogram_impl_hetero.h
@@ -144,7 +144,7 @@ __pattern_histogram(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec,
         //fill histogram bins with zeros
 
         auto __init_event = oneapi::dpl::__par_backend_hetero::__parallel_for(
-            oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__hist_fill_zeros_wrapper>(__exec),
+            _BackendTag{}, oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__hist_fill_zeros_wrapper>(__exec),
             unseq_backend::walk_n<_ExecutionPolicy, decltype(__fill_func)>{__fill_func}, __num_bins, __bins);
 
         if (__n > 0)
diff --git a/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
index e1af0f7a73c..3edccc031e6 100644
--- a/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
@@ -382,6 +382,7 @@ __pattern_adjacent_difference(_ExecutionPolicy&& __exec, _ForwardIterator1 __fir
         return __d_first;
 
     const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
 
     using _It1ValueT = typename ::std::iterator_traits<_ForwardIterator1>::value_type;
     using _It2ValueTRef = typename ::std::iterator_traits<_ForwardIterator2>::reference;
@@ -419,8 +420,8 @@ __pattern_adjacent_difference(_ExecutionPolicy&& __exec, _ForwardIterator1 __fir
 
             using _Function = unseq_backend::walk_adjacent_difference<_ExecutionPolicy, decltype(__fn)>;
 
-            oneapi::dpl::__par_backend_hetero::__parallel_for(__exec, _Function{__fn}, __n, __buf1.all_view(),
-                                                              __buf2.all_view())
+            oneapi::dpl::__par_backend_hetero::__parallel_for(__backend_tag{}, __exec, _Function{__fn}, __n,
+                                                              __buf1.all_view(), __buf2.all_view())
                 .wait();
 
             return __d_last;
@@ -474,8 +475,8 @@ __pattern_adjacent_difference(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&
 
             using _Function = unseq_backend::walk_adjacent_difference<_ExecutionPolicy, decltype(__fn)>;
 
-            oneapi::dpl::__par_backend_hetero::__parallel_for(__exec, _Function{__fn}, __n, __buf1.all_view(),
-                                                              __buf2.all_view())
+            oneapi::dpl::__par_backend_hetero::__parallel_for(_BackendTag{}, __exec, _Function{__fn}, __n,
+                                                              __buf1.all_view(), __buf2.all_view())
                 .wait();
 
             return __d_last;
diff --git a/include/oneapi/dpl/pstl/numeric_impl.h b/include/oneapi/dpl/pstl/numeric_impl.h
index dc3eec4045d..1bf5ec91e9f 100644
--- a/include/oneapi/dpl/pstl/numeric_impl.h
+++ b/include/oneapi/dpl/pstl/numeric_impl.h
@@ -572,13 +572,17 @@ __pattern_adjacent_difference(_ExecutionPolicy&& __exec, _RandomAccessIterator1
                               _RandomAccessIterator2 __d_first, _BinaryOperation __op, _IsVector __is_vector,
                               /*is_parallel=*/::std::true_type)
 {
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     assert(__first != __last);
     typedef typename ::std::iterator_traits<_RandomAccessIterator1>::reference _ReferenceType1;
     typedef typename ::std::iterator_traits<_RandomAccessIterator2>::reference _ReferenceType2;
 
     *__d_first = *__first;
     __par_backend::__parallel_for(
-        ::std::forward<_ExecutionPolicy>(__exec), __first, __last - 1,
+        __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last - 1,
         [&__op, __is_vector, __d_first, __first](_RandomAccessIterator1 __b, _RandomAccessIterator1 __e) {
             _RandomAccessIterator2 __d_b = __d_first + (__b - __first);
             __internal::__brick_walk3(
@@ -596,13 +600,17 @@ __pattern_adjacent_difference(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&
                               _RandomAccessIterator1 __first, _RandomAccessIterator1 __last,
                               _RandomAccessIterator2 __d_first, _BinaryOperation __op)
 {
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     assert(__first != __last);
     typedef typename ::std::iterator_traits<_RandomAccessIterator1>::reference _ReferenceType1;
     typedef typename ::std::iterator_traits<_RandomAccessIterator2>::reference _ReferenceType2;
 
     *__d_first = *__first;
     __par_backend::__parallel_for(
-        ::std::forward<_ExecutionPolicy>(__exec), __first, __last - 1,
+        __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last - 1,
         [&__op, __d_first, __first](_RandomAccessIterator1 __b, _RandomAccessIterator1 __e) {
             _RandomAccessIterator2 __d_b = __d_first + (__b - __first);
             __internal::__brick_walk3(
diff --git a/include/oneapi/dpl/pstl/parallel_impl.h b/include/oneapi/dpl/pstl/parallel_impl.h
index d30fc053cde..57f3fa737e6 100644
--- a/include/oneapi/dpl/pstl/parallel_impl.h
+++ b/include/oneapi/dpl/pstl/parallel_impl.h
@@ -42,9 +42,12 @@ __parallel_find(_ExecutionPolicy&& __exec, _Index __first, _Index __last, _Brick
 
     constexpr auto __comp = ::std::conditional_t<_IsFirst::value, __pstl_less, __pstl_greater>{};
 
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Index>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     ::std::atomic<_DifferenceType> __extremum(__initial_dist);
     // TODO: find out what is better here: parallel_for or parallel_reduce
-    __par_backend::__parallel_for(::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+    __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
                                   [__comp, __f, __first, &__extremum](_Index __i, _Index __j) {
                                       // See "Reducing Contention Through Priority Updates", PPoPP '13, for discussion of
                                       // why using a shared variable scales fairly well in this situation.
@@ -112,8 +115,29 @@ template <class _ExecutionPolicy, class _Index, class _Brick>
 bool
 __parallel_or(_ExecutionPolicy&& __exec, _Index __first, _Index __last, _Brick __f)
 {
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Index>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
+    ::std::atomic<bool> __found(false);
+    __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+                                  [__f, &__found](_Index __i, _Index __j) {
+                                      if (!__found.load(::std::memory_order_relaxed) && __f(__i, __j))
+                                      {
+                                          __found.store(true, ::std::memory_order_relaxed);
+                                          __par_backend::__cancel_execution();
+                                      }
+                                  });
+    return __found;
+}
+
+template <class _IsVector, class _ExecutionPolicy, class _Index, class _Brick>
+bool
+__parallel_or(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _Index __first, _Index __last, _Brick __f)
+{
+    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
+
     ::std::atomic<bool> __found(false);
-    __par_backend::__parallel_for(::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+    __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
                                   [__f, &__found](_Index __i, _Index __j) {
                                       if (!__found.load(::std::memory_order_relaxed) && __f(__i, __j))
                                       {

From dc34c04364ea352f34477f6764af5f6556365774 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 14 Feb 2024 15:04:52 +0100
Subject: [PATCH 182/566] __parallel_find + tag impls

---
 .../pstl/hetero/dpcpp/parallel_backend_sycl.h | 19 ++++++++++++++++
 .../hetero/dpcpp/parallel_backend_sycl_fpga.h | 22 +++++++++++++++++++
 include/oneapi/dpl/pstl/parallel_impl.h       |  7 +++---
 3 files changed, 44 insertions(+), 4 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
index 6038150b03b..0cc786fa94a 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
@@ -1374,6 +1374,25 @@ __parallel_find(_ExecutionPolicy&& __exec, _Iterator1 __first, _Iterator1 __last
                          __f, _TagType{}, __buf.all_view(), __s_buf.all_view());
 }
 
+template <typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2, typename _Brick, typename _IsFirst>
+_Iterator1
+__parallel_find(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&& __exec, _Iterator1 __first,
+                _Iterator1 __last, _Iterator2 __s_first, _Iterator2 __s_last, _Brick __f, _IsFirst)
+{
+    auto __keep = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _Iterator1>();
+    auto __buf = __keep(__first, __last);
+    auto __s_keep = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _Iterator2>();
+    auto __s_buf = __s_keep(__s_first, __s_last);
+
+    using _TagType = ::std::conditional_t<_IsFirst::value, __parallel_find_forward_tag<decltype(__buf.all_view())>,
+                                          __parallel_find_backward_tag<decltype(__buf.all_view())>>;
+    return __first + oneapi::dpl::__par_backend_hetero::__parallel_find_or(
+                         __par_backend_hetero::make_wrapped_policy<__find_policy_wrapper>(
+                             ::std::forward<_ExecutionPolicy>(__exec)),
+                         __f, _TagType{}, __buf.all_view(), __s_buf.all_view());
+}
+
+
 // Special overload for single sequence cases.
 // TODO: check if similar pattern may apply to other algorithms. If so, these overloads should be moved out of
 // backend code.
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
index 85f4f0cf528..69228bb3601 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
@@ -261,6 +261,18 @@ __parallel_find(_ExecutionPolicy&& __exec, _Iterator1 __first, _Iterator1 __last
                                                               __s_last, __f, __is_first);
 }
 
+template <typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2, typename _Brick, typename _IsFirst>
+_Iterator1
+__parallel_find(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionPolicy&& __exec, _Iterator1 __first,
+                _Iterator1 __last, _Iterator2 __s_first,
+                _Iterator2 __s_last, _Brick __f, _IsFirst __is_first)
+{
+    // workaround until we implement more performant version for patterns
+    return oneapi::dpl::__par_backend_hetero::__parallel_find(oneapi::dpl::__internal::__device_backend_tag{},
+                                                              __exec.__device_policy(), __first, __last, __s_first,
+                                                              __s_last, __f, __is_first);
+}
+
 template <typename _ExecutionPolicy, typename _Iterator, typename _Brick, typename _IsFirst>
 oneapi::dpl::__internal::__enable_if_fpga_execution_policy<_ExecutionPolicy, _Iterator>
 __parallel_find(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last, _Brick __f, _IsFirst __is_first)
@@ -270,6 +282,16 @@ __parallel_find(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last,
                                                               __is_first);
 }
 
+template <typename _ExecutionPolicy, typename _Iterator, typename _Brick, typename _IsFirst>
+_Iterator
+__parallel_find(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionPolicy&& __exec, _Iterator __first,
+                _Iterator __last, _Brick __f, _IsFirst __is_first)
+{
+    // workaround until we implement more performant version for patterns
+    return oneapi::dpl::__par_backend_hetero::__parallel_find(
+        oneapi::dpl::__internal::__device_backend_tag{}, __exec.__device_policy(), __first, __last, __f, __is_first);
+}
+
 //------------------------------------------------------------------------
 // parallel_merge
 //-----------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/parallel_impl.h b/include/oneapi/dpl/pstl/parallel_impl.h
index 57f3fa737e6..2d3f327bf53 100644
--- a/include/oneapi/dpl/pstl/parallel_impl.h
+++ b/include/oneapi/dpl/pstl/parallel_impl.h
@@ -69,12 +69,11 @@ __parallel_find(_ExecutionPolicy&& __exec, _Index __first, _Index __last, _Brick
     return __extremum != __initial_dist ? __first + __extremum : __last;
 }
 
-template <class _Tag, class _ExecutionPolicy, class _Index, class _Brick, class _IsFirst>
+template <class _IsVector, class _ExecutionPolicy, class _Index, class _Brick, class _IsFirst>
 _Index
-__parallel_find(_Tag __tag, _ExecutionPolicy&& __exec, _Index __first, _Index __last, _Brick __f, _IsFirst)
+__parallel_find(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Index __first, _Index __last, _Brick __f,
+                _IsFirst)
 {
-    static_assert(__is_backend_tag_v<_Tag>);
-
     using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
 
     typedef typename ::std::iterator_traits<_Index>::difference_type _DifferenceType;

From 2f948ac73289b7dcd0a6e4c9a74f6abde2c30229 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 14 Feb 2024 15:06:34 +0100
Subject: [PATCH 183/566] __parallel_find + tag calls

---
 include/oneapi/dpl/pstl/algorithm_impl.h      | 60 ++++++++++-----
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 73 ++++++++++++-------
 .../hetero/dpcpp/parallel_backend_sycl_fpga.h |  3 +-
 3 files changed, 89 insertions(+), 47 deletions(-)

diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index d32664bfcc6..607579fcad6 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -1350,13 +1350,13 @@ __pattern_find_end(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first, _R
                    _RandomAccessIterator2 __s_first, _RandomAccessIterator2 __s_last, _BinaryPredicate __pred,
                    _IsVector __is_vector, /*is_parallel=*/::std::true_type)
 {
+    // TODO is it correct that we check _RandomAccessIterator2 in __select_backend ?
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     if (__last - __first == __s_last - __s_first)
     {
-        // TODO is it correct that we check _RandomAccessIterator2 in __select_backend ?
-        constexpr auto __dispatch_tag =
-            oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1,
-                                                      _RandomAccessIterator2>();
-
         const bool __res = __internal::__pattern_equal(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
                                                        __first, __last, __s_first, __pred);
         return __res ? __first : __last;
@@ -1365,7 +1365,7 @@ __pattern_find_end(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first, _R
     {
         return __internal::__except_handler([&]() {
             return __internal::__parallel_find(
-                ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+                __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
                 [__last, __s_first, __s_last, __pred, __is_vector](_RandomAccessIterator1 __i,
                                                                    _RandomAccessIterator1 __j) {
                     return __internal::__find_subrange(__i, __j, __last, __s_first, __s_last, __pred, false,
@@ -1395,8 +1395,7 @@ __pattern_find_end(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _
     {
         return __internal::__except_handler([&]() {
             return __internal::__parallel_find(
-                __backend_tag{},
-                ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+                __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
                 [__last, __s_first, __s_last, __pred](_RandomAccessIterator1 __i, _RandomAccessIterator1 __j) {
                     return __internal::__find_subrange(__i, __j, __last, __s_first, __s_last, __pred, false,
                                                        _IsVector{});
@@ -1454,9 +1453,13 @@ __pattern_find_first_of(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _F
                         _ForwardIterator2 __s_first, _ForwardIterator2 __s_last, _BinaryPredicate __pred,
                         _IsVector __is_vector, /*is_parallel=*/::std::true_type)
 {
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     return __internal::__except_handler([&]() {
         return __internal::__parallel_find(
-            ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
             [__s_first, __s_last, &__pred, __is_vector](_ForwardIterator1 __i, _ForwardIterator1 __j) {
                 return __internal::__brick_find_first_of(__i, __j, __s_first, __s_last, __pred, __is_vector);
             },
@@ -1530,13 +1533,13 @@ __pattern_search(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first, _Ran
                  _IsVector __is_vector,
                  /*is_parallel=*/::std::true_type)
 {
+    // TODO is it correct that we check _RandomAccessIterator2 in __select_backend ?
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     if (__last - __first == __s_last - __s_first)
     {
-        // TODO is it correct that we check _RandomAccessIterator2 in __select_backend ?
-        constexpr auto __dispatch_tag =
-            oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1,
-                                                      _RandomAccessIterator2>();
-
         const bool __res = __internal::__pattern_equal(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
                                                        __first, __last, __s_first, __pred);
         return __res ? __first : __last;
@@ -1545,7 +1548,7 @@ __pattern_search(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first, _Ran
     {
         return __internal::__except_handler([&]() {
             return __internal::__parallel_find(
-                ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+                __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
                 [__last, __s_first, __s_last, __pred, __is_vector](_RandomAccessIterator1 __i,
                                                                    _RandomAccessIterator1 __j) {
                     return __internal::__find_subrange(__i, __j, __last, __s_first, __s_last, __pred, true,
@@ -1631,6 +1634,10 @@ __pattern_search_n(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _Ra
                    _Size __count, const _Tp& __value, _BinaryPredicate __pred, _IsVector __is_vector,
                    /*is_parallel=*/::std::true_type)
 {
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     if (static_cast<_Size>(__last - __first) == __count)
     {
         const bool __result = !__internal::__pattern_any_of(
@@ -1643,7 +1650,7 @@ __pattern_search_n(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _Ra
     {
         return __internal::__except_handler([&__exec, __first, __last, __count, &__value, __pred, __is_vector]() {
             return __internal::__parallel_find(
-                ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+                __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
                 [__last, __count, &__value, __pred, __is_vector](_RandomAccessIterator __i, _RandomAccessIterator __j) {
                     return __internal::__find_subrange(__i, __j, __last, __count, __value, __pred, __is_vector);
                 },
@@ -5831,9 +5838,13 @@ oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _Ra
 __pattern_is_heap_until(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIterator __last,
                         _Compare __comp, _IsVector __is_vector, /* is_parallel = */ ::std::true_type)
 {
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     return __internal::__except_handler([&]() {
         return __parallel_find(
-            ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
             [__first, __comp, __is_vector](_RandomAccessIterator __i, _RandomAccessIterator __j) {
                 return __internal::__is_heap_until_local(__first, __i - __first, __j - __first, __comp, __is_vector);
             },
@@ -5848,7 +5859,7 @@ __pattern_is_heap_until(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _R
 {
     return __internal::__except_handler([&]() {
         return __parallel_find(
-            ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+            __tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
             [__first, __comp](_RandomAccessIterator __i, _RandomAccessIterator __j) {
                 return __internal::__is_heap_until_local(__first, __i - __first, __j - __first, __comp, _IsVector{});
             },
@@ -6228,10 +6239,14 @@ __pattern_mismatch(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1, _
                    _RandomAccessIterator2 __first2, _RandomAccessIterator2 __last2, _Predicate __pred,
                    _IsVector __is_vector, /* is_parallel = */ ::std::true_type)
 {
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     return __internal::__except_handler([&]() {
         auto __n = ::std::min(__last1 - __first1, __last2 - __first2);
         auto __result = __internal::__parallel_find(
-            ::std::forward<_ExecutionPolicy>(__exec), __first1, __first1 + __n,
+            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first1, __first1 + __n,
             [__first1, __first2, __pred, __is_vector](_RandomAccessIterator1 __i, _RandomAccessIterator1 __j) {
                 return __internal::__brick_mismatch(__i, __j, __first2 + (__i - __first1), __first2 + (__j - __first1),
                                                     __pred, __is_vector)
@@ -6362,13 +6377,18 @@ __pattern_lexicographical_compare(_ExecutionPolicy&& __exec, _RandomAccessIterat
     }
     else
     {
+        constexpr auto __dispatch_tag =
+            oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1,
+                                                      _RandomAccessIterator2>();
+        using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
         typedef typename ::std::iterator_traits<_RandomAccessIterator1>::reference _RefType1;
         typedef typename ::std::iterator_traits<_RandomAccessIterator2>::reference _RefType2;
         --__last1;
         --__last2;
         auto __n = ::std::min(__last1 - __first1, __last2 - __first2);
         auto __result = __internal::__parallel_find(
-            ::std::forward<_ExecutionPolicy>(__exec), __first1, __first1 + __n,
+            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first1, __first1 + __n,
             [__first1, __first2, &__comp, __is_vector](_RandomAccessIterator1 __i, _RandomAccessIterator1 __j) {
                 return __internal::__brick_mismatch(
                            __i, __j, __first2 + (__i - __first1), __first2 + (__j - __first1),
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 7df8d6d7a0d..34707777978 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -991,11 +991,14 @@ __pattern_adjacent_find(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator
     if (__last - __first < 2)
         return __last;
 
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     using _Predicate =
         oneapi::dpl::unseq_backend::single_match_pred<_ExecutionPolicy, adjacent_find_fn<_BinaryPredicate>>;
 
     auto __result = __par_backend_hetero::__parallel_find(
-        ::std::forward<_ExecutionPolicy>(__exec),
+        __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
         __par_backend_hetero::zip(
             __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__first),
             __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__first + 1)),
@@ -1024,7 +1027,7 @@ __pattern_adjacent_find(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __ex
         oneapi::dpl::unseq_backend::single_match_pred<_ExecutionPolicy, adjacent_find_fn<_BinaryPredicate>>;
 
     auto __result = __par_backend_hetero::__parallel_find(
-        ::std::forward<_ExecutionPolicy>(__exec),
+        _BackendTag{}, ::std::forward<_ExecutionPolicy>(__exec),
         __par_backend_hetero::zip(
             __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__first),
             __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__first + 1)),
@@ -1254,12 +1257,13 @@ __pattern_find_end(_ExecutionPolicy&& __exec, _Iterator1 __first, _Iterator1 __l
     if (__first == __last || __s_last == __s_first || __last - __first < __s_last - __s_first)
         return __last;
 
+    // TODO is it correct that we check _Iterator2 in __select_backend ?
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator1, _Iterator2>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     if (__last - __first == __s_last - __s_first)
     {
-        // TODO is it correct that we check _Iterator2 in __select_backend ?
-        constexpr auto __dispatch_tag =
-            oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator1, _Iterator2>();
-
         const bool __res = __pattern_equal(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
                                            __s_first, __pred);
         return __res ? __first : __last;
@@ -1269,7 +1273,7 @@ __pattern_find_end(_ExecutionPolicy&& __exec, _Iterator1 __first, _Iterator1 __l
         using _Predicate = unseq_backend::multiple_match_pred<_ExecutionPolicy, _Pred>;
 
         return __par_backend_hetero::__parallel_find(
-            ::std::forward<_ExecutionPolicy>(__exec),
+            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
             __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__first),
             __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__last),
             __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__s_first),
@@ -1297,7 +1301,7 @@ __pattern_find_end(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _
         using _Predicate = unseq_backend::multiple_match_pred<_ExecutionPolicy, _Pred>;
 
         return __par_backend_hetero::__parallel_find(
-            ::std::forward<_ExecutionPolicy>(__exec),
+            _BackendTag{}, ::std::forward<_ExecutionPolicy>(__exec),
             __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__first),
             __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__last),
             __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__s_first),
@@ -1318,12 +1322,16 @@ __pattern_find_first_of(_ExecutionPolicy&& __exec, _Iterator1 __first, _Iterator
     if (__first == __last || __s_last == __s_first)
         return __last;
 
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator1, _Iterator2>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     using _Predicate = unseq_backend::first_match_pred<_ExecutionPolicy, _Pred>;
 
     // TODO: To check whether it makes sense to iterate over the second sequence in case of
     // distance(__first, __last) < distance(__s_first, __s_last).
     return __par_backend_hetero::__parallel_find(
-        ::std::forward<_ExecutionPolicy>(__exec),
+        __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
         __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__first),
         __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__last),
         __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__s_first),
@@ -1344,7 +1352,7 @@ __pattern_find_first_of(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __ex
     // TODO: To check whether it makes sense to iterate over the second sequence in case of
     // distance(__first, __last) < distance(__s_first, __s_last).
     return __par_backend_hetero::__parallel_find(
-        ::std::forward<_ExecutionPolicy>(__exec),
+        _BackendTag{}, ::std::forward<_ExecutionPolicy>(__exec),
         __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__first),
         __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__last),
         __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__s_first),
@@ -1373,12 +1381,13 @@ __pattern_search(_ExecutionPolicy&& __exec, _Iterator1 __first, _Iterator1 __las
     if (__last - __first < __s_last - __s_first)
         return __last;
 
+    // TODO is it correct that we check _Iterator2 in __select_backend ?
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator1, _Iterator2>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     if (__last - __first == __s_last - __s_first)
     {
-        // TODO is it correct that we check _Iterator2 in __select_backend ?
-        constexpr auto __dispatch_tag =
-            oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator1, _Iterator2>();
-
         const bool __res = __pattern_equal(
             __dispatch_tag,
             __par_backend_hetero::make_wrapped_policy<equal_wrapper>(::std::forward<_ExecutionPolicy>(__exec)), __first,
@@ -1388,7 +1397,7 @@ __pattern_search(_ExecutionPolicy&& __exec, _Iterator1 __first, _Iterator1 __las
 
     using _Predicate = unseq_backend::multiple_match_pred<_ExecutionPolicy, _Pred>;
     return __par_backend_hetero::__parallel_find(
-        ::std::forward<_ExecutionPolicy>(__exec),
+        __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
         __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__first),
         __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__last),
         __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__s_first),
@@ -1398,7 +1407,7 @@ __pattern_search(_ExecutionPolicy&& __exec, _Iterator1 __first, _Iterator1 __las
 
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2, typename _Pred>
 _Iterator1
-__pattern_search(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator1 __first, _Iterator1 __last,
+__pattern_search(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Iterator1 __first, _Iterator1 __last,
                  _Iterator2 __s_first, _Iterator2 __s_last, _Pred __pred)
 {
     if (__s_last == __s_first)
@@ -1417,7 +1426,7 @@ __pattern_search(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _It
 
     using _Predicate = unseq_backend::multiple_match_pred<_ExecutionPolicy, _Pred>;
     return __par_backend_hetero::__parallel_find(
-        ::std::forward<_ExecutionPolicy>(__exec),
+        _BackendTag{}, ::std::forward<_ExecutionPolicy>(__exec),
         __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__first),
         __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__last),
         __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__s_first),
@@ -1465,8 +1474,12 @@ __pattern_search_n(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __las
     }
 
     using _Predicate = unseq_backend::n_elem_match_pred<_ExecutionPolicy, _BinaryPredicate, _Tp, _Size>;
+
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     return __par_backend_hetero::__parallel_find(
-        ::std::forward<_ExecutionPolicy>(__exec),
+        __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
         __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__first),
         __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__last),
         _Predicate{__pred, __value, __count}, ::std::true_type{});
@@ -1513,14 +1526,18 @@ __pattern_mismatch(_ExecutionPolicy&& __exec, _Iterator1 __first1, _Iterator1 __
     if (__n <= 0)
         return ::std::make_pair(__first1, __first2);
 
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator1, _Iterator2>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     using _Predicate = oneapi::dpl::unseq_backend::single_match_pred<_ExecutionPolicy, equal_predicate<_Pred>>;
 
     auto __first_zip = __par_backend_hetero::zip(
         __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__first1),
         __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__first2));
-    auto __result =
-        __par_backend_hetero::__parallel_find(::std::forward<_ExecutionPolicy>(__exec), __first_zip, __first_zip + __n,
-                                              _Predicate{equal_predicate<_Pred>{__pred}}, ::std::true_type{});
+    auto __result = __par_backend_hetero::__parallel_find(
+        __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first_zip, __first_zip + __n,
+        _Predicate{equal_predicate<_Pred>{__pred}}, ::std::true_type{});
     __n = __result - __first_zip;
     return ::std::make_pair(__first1 + __n, __first2 + __n);
 }
@@ -1539,9 +1556,9 @@ __pattern_mismatch(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _
     auto __first_zip = __par_backend_hetero::zip(
         __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__first1),
         __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__first2));
-    auto __result =
-        __par_backend_hetero::__parallel_find(::std::forward<_ExecutionPolicy>(__exec), __first_zip, __first_zip + __n,
-                                              _Predicate{equal_predicate<_Pred>{__pred}}, ::std::true_type{});
+    auto __result = __par_backend_hetero::__parallel_find(
+        _BackendTag{}, ::std::forward<_ExecutionPolicy>(__exec), __first_zip, __first_zip + __n,
+        _Predicate{equal_predicate<_Pred>{__pred}}, ::std::true_type{});
     __n = __result - __first_zip;
     return ::std::make_pair(__first1 + __n, __first2 + __n);
 }
@@ -1976,11 +1993,15 @@ __pattern_is_heap_until(_ExecutionPolicy&& __exec, _RandomAccessIterator __first
     if (__last - __first < 2)
         return __last;
 
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     using _Predicate =
         oneapi::dpl::unseq_backend::single_match_pred_by_idx<_ExecutionPolicy, __is_heap_check<_Compare>>;
 
     return __par_backend_hetero::__parallel_find(
-        ::std::forward<_ExecutionPolicy>(__exec),
+        __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
         __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__first),
         __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__last), _Predicate{__comp},
         ::std::true_type{});
@@ -1998,7 +2019,7 @@ __pattern_is_heap_until(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __ex
         oneapi::dpl::unseq_backend::single_match_pred_by_idx<_ExecutionPolicy, __is_heap_check<_Compare>>;
 
     return __par_backend_hetero::__parallel_find(
-        ::std::forward<_ExecutionPolicy>(__exec),
+        _BackendTag{}, ::std::forward<_ExecutionPolicy>(__exec),
         __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__first),
         __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__last), _Predicate{__comp},
         ::std::true_type{});
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
index 69228bb3601..64898115e0c 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
@@ -257,7 +257,8 @@ __parallel_find(_ExecutionPolicy&& __exec, _Iterator1 __first, _Iterator1 __last
 {
 
     // workaround until we implement more performant version for patterns
-    return oneapi::dpl::__par_backend_hetero::__parallel_find(__exec.__device_policy(), __first, __last, __s_first,
+    return oneapi::dpl::__par_backend_hetero::__parallel_find(oneapi::dpl::__internal::__device_backend_tag{},
+                                                              __exec.__device_policy(), __first, __last, __s_first,
                                                               __s_last, __f, __is_first);
 }
 

From 5533a8683c167716e475e9a9438f52d46630849b Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 15 Feb 2024 12:47:48 +0100
Subject: [PATCH 184/566] __pattern_mismatch + tag impls

---
 include/oneapi/dpl/pstl/algorithm_impl.h | 5 ++---
 1 file changed, 2 insertions(+), 3 deletions(-)

diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 607579fcad6..f2a1ddb0458 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -6260,7 +6260,7 @@ __pattern_mismatch(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1, _
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
           class _Predicate>
 ::std::pair<_RandomAccessIterator1, _RandomAccessIterator2>
-__pattern_mismatch(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1,
+__pattern_mismatch(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1,
                    _RandomAccessIterator1 __last1, _RandomAccessIterator2 __first2, _RandomAccessIterator2 __last2,
                    _Predicate __pred)
 {
@@ -6380,7 +6380,6 @@ __pattern_lexicographical_compare(_ExecutionPolicy&& __exec, _RandomAccessIterat
         constexpr auto __dispatch_tag =
             oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1,
                                                       _RandomAccessIterator2>();
-        using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
 
         typedef typename ::std::iterator_traits<_RandomAccessIterator1>::reference _RefType1;
         typedef typename ::std::iterator_traits<_RandomAccessIterator2>::reference _RefType2;
@@ -6388,7 +6387,7 @@ __pattern_lexicographical_compare(_ExecutionPolicy&& __exec, _RandomAccessIterat
         --__last2;
         auto __n = ::std::min(__last1 - __first1, __last2 - __first2);
         auto __result = __internal::__parallel_find(
-            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first1, __first1 + __n,
+            __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __first1 + __n,
             [__first1, __first2, &__comp, __is_vector](_RandomAccessIterator1 __i, _RandomAccessIterator1 __j) {
                 return __internal::__brick_mismatch(
                            __i, __j, __first2 + (__i - __first1), __first2 + (__j - __first1),

From 48e0a2eb8903f80f3fd5e48249fc30c05d88f491 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 15 Feb 2024 11:55:48 +0100
Subject: [PATCH 185/566] __internal::__parallel_find + tag calls

---
 include/oneapi/dpl/pstl/algorithm_impl.h | 43 ++++++++----------------
 1 file changed, 14 insertions(+), 29 deletions(-)

diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index f2a1ddb0458..c759267de54 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -1189,14 +1189,12 @@ __pattern_find_if(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIt
 
 template <class _IsVector, class _ExecutionPolicy, class _ForwardIterator, class _Predicate>
 _ForwardIterator
-__pattern_find_if(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _ForwardIterator __first,
+__pattern_find_if(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _ForwardIterator __first,
                   _ForwardIterator __last, _Predicate __pred)
 {
-    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
-
     return __except_handler([&]() {
         return __parallel_find(
-            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+            __tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
             [__pred](_ForwardIterator __i, _ForwardIterator __j) {
                 return __brick_find_if(__i, __j, __pred, _IsVector{});
             },
@@ -1353,7 +1351,6 @@ __pattern_find_end(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first, _R
     // TODO is it correct that we check _RandomAccessIterator2 in __select_backend ?
     constexpr auto __dispatch_tag =
         oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
 
     if (__last - __first == __s_last - __s_first)
     {
@@ -1365,7 +1362,7 @@ __pattern_find_end(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first, _R
     {
         return __internal::__except_handler([&]() {
             return __internal::__parallel_find(
-                __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+                __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
                 [__last, __s_first, __s_last, __pred, __is_vector](_RandomAccessIterator1 __i,
                                                                    _RandomAccessIterator1 __j) {
                     return __internal::__find_subrange(__i, __j, __last, __s_first, __s_last, __pred, false,
@@ -1383,8 +1380,6 @@ __pattern_find_end(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _
                    _RandomAccessIterator1 __last,
                    _RandomAccessIterator2 __s_first, _RandomAccessIterator2 __s_last, _BinaryPredicate __pred)
 {
-    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
-
     if (__last - __first == __s_last - __s_first)
     {
         const bool __res = __internal::__pattern_equal(__tag, ::std::forward<_ExecutionPolicy>(__exec),
@@ -1395,7 +1390,7 @@ __pattern_find_end(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _
     {
         return __internal::__except_handler([&]() {
             return __internal::__parallel_find(
-                __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+                __tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
                 [__last, __s_first, __s_last, __pred](_RandomAccessIterator1 __i, _RandomAccessIterator1 __j) {
                     return __internal::__find_subrange(__i, __j, __last, __s_first, __s_last, __pred, false,
                                                        _IsVector{});
@@ -1455,11 +1450,10 @@ __pattern_find_first_of(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _F
 {
     constexpr auto __dispatch_tag =
         oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
 
     return __internal::__except_handler([&]() {
         return __internal::__parallel_find(
-            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+            __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
             [__s_first, __s_last, &__pred, __is_vector](_ForwardIterator1 __i, _ForwardIterator1 __j) {
                 return __internal::__brick_find_first_of(__i, __j, __s_first, __s_last, __pred, __is_vector);
             },
@@ -1470,7 +1464,7 @@ __pattern_find_first_of(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _F
 template <class _IsVector, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2,
           class _BinaryPredicate>
 _ForwardIterator1
-__pattern_find_first_of(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _ForwardIterator1 __first,
+__pattern_find_first_of(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _ForwardIterator1 __first,
                         _ForwardIterator1 __last, _ForwardIterator2 __s_first, _ForwardIterator2 __s_last,
                         _BinaryPredicate __pred)
 {
@@ -1478,7 +1472,7 @@ __pattern_find_first_of(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _F
 
     return __internal::__except_handler([&]() {
         return __internal::__parallel_find(
-            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+            __tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
             [__s_first, __s_last, &__pred](_ForwardIterator1 __i, _ForwardIterator1 __j) {
                 return __internal::__brick_find_first_of(__i, __j, __s_first, __s_last, __pred, _IsVector{});
             },
@@ -1536,7 +1530,6 @@ __pattern_search(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first, _Ran
     // TODO is it correct that we check _RandomAccessIterator2 in __select_backend ?
     constexpr auto __dispatch_tag =
         oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
 
     if (__last - __first == __s_last - __s_first)
     {
@@ -1548,7 +1541,7 @@ __pattern_search(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first, _Ran
     {
         return __internal::__except_handler([&]() {
             return __internal::__parallel_find(
-                __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+                __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
                 [__last, __s_first, __s_last, __pred, __is_vector](_RandomAccessIterator1 __i,
                                                                    _RandomAccessIterator1 __j) {
                     return __internal::__find_subrange(__i, __j, __last, __s_first, __s_last, __pred, true,
@@ -1566,8 +1559,6 @@ __pattern_search(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _Ra
                  _RandomAccessIterator1 __last, _RandomAccessIterator2 __s_first, _RandomAccessIterator2 __s_last,
                  _BinaryPredicate __pred)
 {
-    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
-
     if (__last - __first == __s_last - __s_first)
     {
         const bool __res = __internal::__pattern_equal(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
@@ -1578,7 +1569,7 @@ __pattern_search(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _Ra
     {
         return __internal::__except_handler([&]() {
             return __internal::__parallel_find(
-                __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+                __tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
                 [__last, __s_first, __s_last, __pred](_RandomAccessIterator1 __i, _RandomAccessIterator1 __j) {
                     return __internal::__find_subrange(__i, __j, __last, __s_first, __s_last, __pred, true,
                                                        _IsVector{});
@@ -1636,7 +1627,6 @@ __pattern_search_n(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _Ra
 {
     constexpr auto __dispatch_tag =
         oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
 
     if (static_cast<_Size>(__last - __first) == __count)
     {
@@ -1650,7 +1640,7 @@ __pattern_search_n(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _Ra
     {
         return __internal::__except_handler([&__exec, __first, __last, __count, &__value, __pred, __is_vector]() {
             return __internal::__parallel_find(
-                __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+                __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
                 [__last, __count, &__value, __pred, __is_vector](_RandomAccessIterator __i, _RandomAccessIterator __j) {
                     return __internal::__find_subrange(__i, __j, __last, __count, __value, __pred, __is_vector);
                 },
@@ -1665,8 +1655,6 @@ _RandomAccessIterator
 __pattern_search_n(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
                    _RandomAccessIterator __last, _Size __count, const _Tp& __value, _BinaryPredicate __pred)
 {
-    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
-
     if (static_cast<_Size>(__last - __first) == __count)
     {
         const bool __result =
@@ -1676,9 +1664,9 @@ __pattern_search_n(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _
     }
     else
     {
-        return __internal::__except_handler([&__exec, __first, __last, __count, &__value, __pred]() {
+        return __internal::__except_handler([&__exec, __first, __last, __count, &__value, __pred, __tag]() {
             return __internal::__parallel_find(
-                __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+                __tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
                 [__last, __count, &__value, __pred](_RandomAccessIterator __i, _RandomAccessIterator __j) {
                     return __internal::__find_subrange(__i, __j, __last, __count, __value, __pred, _IsVector{});
                 },
@@ -6241,12 +6229,11 @@ __pattern_mismatch(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1, _
 {
     constexpr auto __dispatch_tag =
         oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
 
     return __internal::__except_handler([&]() {
         auto __n = ::std::min(__last1 - __first1, __last2 - __first2);
         auto __result = __internal::__parallel_find(
-            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first1, __first1 + __n,
+            __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __first1 + __n,
             [__first1, __first2, __pred, __is_vector](_RandomAccessIterator1 __i, _RandomAccessIterator1 __j) {
                 return __internal::__brick_mismatch(__i, __j, __first2 + (__i - __first1), __first2 + (__j - __first1),
                                                     __pred, __is_vector)
@@ -6264,12 +6251,10 @@ __pattern_mismatch(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _
                    _RandomAccessIterator1 __last1, _RandomAccessIterator2 __first2, _RandomAccessIterator2 __last2,
                    _Predicate __pred)
 {
-    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
-
     return __internal::__except_handler([&]() {
         auto __n = ::std::min(__last1 - __first1, __last2 - __first2);
         auto __result = __internal::__parallel_find(
-            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first1, __first1 + __n,
+            __tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __first1 + __n,
             [__first1, __first2, __pred](_RandomAccessIterator1 __i, _RandomAccessIterator1 __j) {
                 return __internal::__brick_mismatch(__i, __j, __first2 + (__i - __first1), __first2 + (__j - __first1),
                                                     __pred, _IsVector{})

From fff6053a93f5eb8c81557def2f35f1d9913dbcc7 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 14 Feb 2024 15:34:02 +0100
Subject: [PATCH 186/566] __parallel_reduce + tag impls

---
 include/oneapi/dpl/pstl/omp/parallel_reduce.h | 27 +++++++++++++++++++
 .../oneapi/dpl/pstl/parallel_backend_serial.h | 15 +++++++++++
 .../oneapi/dpl/pstl/parallel_backend_tbb.h    | 15 +++++++++++
 3 files changed, 57 insertions(+)

diff --git a/include/oneapi/dpl/pstl/omp/parallel_reduce.h b/include/oneapi/dpl/pstl/omp/parallel_reduce.h
index beefe09b738..a59a56d374f 100644
--- a/include/oneapi/dpl/pstl/omp/parallel_reduce.h
+++ b/include/oneapi/dpl/pstl/omp/parallel_reduce.h
@@ -77,6 +77,33 @@ __parallel_reduce(_ExecutionPolicy&&, _RandomAccessIterator __first, _RandomAcce
     return __res;
 }
 
+template <class _ExecutionPolicy, class _RandomAccessIterator, class _Value, typename _RealBody, typename _Reduction>
+_Value
+__parallel_reduce(oneapi::dpl::__internal::__omp_backend_tag, _ExecutionPolicy&&, _RandomAccessIterator __first,
+                  _RandomAccessIterator __last, _Value __identity, _RealBody __real_body, _Reduction __reduction)
+{
+    // We don't create a nested parallel region in an existing parallel region:
+    // just create tasks.
+    if (omp_in_parallel())
+    {
+        return oneapi::dpl::__omp_backend::__parallel_reduce_body(__first, __last, __identity, __real_body,
+                                                                  __reduction);
+    }
+
+    // In any case (nested or non-nested) one parallel region is created and only
+    // one thread creates a set of tasks.
+    _Value __res = __identity;
+
+    _PSTL_PRAGMA(omp parallel)
+    _PSTL_PRAGMA(omp single nowait)
+    {
+        __res =
+            oneapi::dpl::__omp_backend::__parallel_reduce_body(__first, __last, __identity, __real_body, __reduction);
+    }
+
+    return __res;
+}
+
 } // namespace __omp_backend
 } // namespace dpl
 } // namespace oneapi
diff --git a/include/oneapi/dpl/pstl/parallel_backend_serial.h b/include/oneapi/dpl/pstl/parallel_backend_serial.h
index 654859e7c6e..c6de6d97784 100644
--- a/include/oneapi/dpl/pstl/parallel_backend_serial.h
+++ b/include/oneapi/dpl/pstl/parallel_backend_serial.h
@@ -94,6 +94,21 @@ __parallel_reduce(_ExecutionPolicy&&, _Index __first, _Index __last, const _Valu
     }
 }
 
+template <class _ExecutionPolicy, class _Value, class _Index, typename _RealBody, typename _Reduction>
+_Value
+__parallel_reduce(oneapi::dpl::__internal::__serial_backend_tag, _ExecutionPolicy&&, _Index __first, _Index __last,
+                  const _Value& __identity, const _RealBody& __real_body, const _Reduction&)
+{
+    if (__first == __last)
+    {
+        return __identity;
+    }
+    else
+    {
+        return __real_body(__first, __last, __identity);
+    }
+}
+
 template <class _ExecutionPolicy, class _Index, class _UnaryOp, class _Tp, class _BinaryOp, class _Reduce>
 _Tp
 __parallel_transform_reduce(_ExecutionPolicy&&, _Index __first, _Index __last, _UnaryOp, _Tp __init, _BinaryOp,
diff --git a/include/oneapi/dpl/pstl/parallel_backend_tbb.h b/include/oneapi/dpl/pstl/parallel_backend_tbb.h
index fb2339808e3..77276c469ac 100644
--- a/include/oneapi/dpl/pstl/parallel_backend_tbb.h
+++ b/include/oneapi/dpl/pstl/parallel_backend_tbb.h
@@ -154,6 +154,21 @@ __parallel_reduce(_ExecutionPolicy&&, _Index __first, _Index __last, const _Valu
     });
 }
 
+template <class _ExecutionPolicy, class _Value, class _Index, typename _RealBody, typename _Reduction>
+_Value
+__parallel_reduce(oneapi::dpl::__internal::__tbb_backend_tag, _ExecutionPolicy&&, _Index __first, _Index __last,
+                  const _Value& __identity, const _RealBody& __real_body, const _Reduction& __reduction)
+{
+    return tbb::this_task_arena::isolate([__first, __last, &__identity, &__real_body, &__reduction]() -> _Value {
+        return tbb::parallel_reduce(
+            tbb::blocked_range<_Index>(__first, __last), __identity,
+            [__real_body](const tbb::blocked_range<_Index>& __r, const _Value& __value) -> _Value {
+                return __real_body(__r.begin(), __r.end(), __value);
+            },
+            __reduction);
+    });
+}
+
 //------------------------------------------------------------------------
 // parallel_transform_reduce
 //

From 373b48be9ec156e03ee85aecd08ebae2c6a06bfc Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 14 Feb 2024 15:38:08 +0100
Subject: [PATCH 187/566] __parallel_reduce + tag calls

---
 include/oneapi/dpl/pstl/algorithm_impl.h      | 74 +++++++++++++++++--
 .../experimental/internal/for_loop_impl.h     |  6 ++
 2 files changed, 73 insertions(+), 7 deletions(-)

diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index c759267de54..aa4c6b5b4d1 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -2082,9 +2082,13 @@ __pattern_count(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _Rando
     if (__first == __last)
         return _SizeType(0);
 
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     return __internal::__except_handler([&]() {
         return __par_backend::__parallel_reduce(
-            ::std::forward<_ExecutionPolicy>(__exec), __first, __last, _SizeType(0),
+            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, _SizeType(0),
             [__pred, __is_vector](_RandomAccessIterator __begin, _RandomAccessIterator __end, _SizeType __value)
                 -> _SizeType { return __value + __internal::__brick_count(__begin, __end, __pred, __is_vector); },
             ::std::plus<_SizeType>());
@@ -2096,6 +2100,8 @@ typename ::std::iterator_traits<_RandomAccessIterator>::difference_type
 __pattern_count(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
                 _RandomAccessIterator __last, _Predicate __pred)
 {
+    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
+
     typedef typename ::std::iterator_traits<_RandomAccessIterator>::difference_type _SizeType;
 
     //trivial pre-checks
@@ -2104,7 +2110,7 @@ __pattern_count(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAcc
 
     return __internal::__except_handler([&]() {
         return __par_backend::__parallel_reduce(
-            ::std::forward<_ExecutionPolicy>(__exec), __first, __last, _SizeType(0),
+            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, _SizeType(0),
             [__pred](_RandomAccessIterator __begin, _RandomAccessIterator __end, _SizeType __value) -> _SizeType {
                 return __value + __internal::__brick_count(__begin, __end, __pred, _IsVector{});
             },
@@ -2157,6 +2163,9 @@ _ForwardIterator
 __remove_elements(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last, _CalcMask __calc_mask,
                   _IsVector __is_vector)
 {
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     typedef typename ::std::iterator_traits<_ForwardIterator>::difference_type _DifferenceType;
     typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _Tp;
     _DifferenceType __n = __last - __first;
@@ -2165,7 +2174,7 @@ __remove_elements(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardI
     return __internal::__except_handler([&]() {
         bool* __mask = __mask_buf.get();
         _DifferenceType __min = __par_backend::__parallel_reduce(
-            ::std::forward<_ExecutionPolicy>(__exec), _DifferenceType(0), __n, __n,
+            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), _DifferenceType(0), __n, __n,
             [__first, __mask, &__calc_mask, __is_vector](_DifferenceType __i, _DifferenceType __j,
                                                          _DifferenceType __local_min) -> _DifferenceType {
                 // Create mask
@@ -3002,6 +3011,10 @@ oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, boo
 __pattern_is_partitioned(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIterator __last,
                          _UnaryPredicate __pred, _IsVector __is_vector, /*is_parallel=*/::std::true_type)
 {
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     //trivial pre-checks
     if (__first == __last)
         return true;
@@ -3039,7 +3052,7 @@ __pattern_is_partitioned(_ExecutionPolicy&& __exec, _RandomAccessIterator __firs
         const _ReduceType __identity{__not_init, __last};
 
         _ReduceType __result = __par_backend::__parallel_reduce(
-            ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __identity,
+            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __identity,
             [&__pred, __combine, __is_vector](_RandomAccessIterator __i, _RandomAccessIterator __j,
                                               _ReduceType __value) -> _ReduceType {
                 if (__value.__val == __broken)
@@ -3103,6 +3116,10 @@ __pattern_is_partitioned(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _
     if (__first == __last)
         return true;
 
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     return __internal::__except_handler([&]() {
         // State of current range:
         // broken     - current range is not partitioned by pred
@@ -3136,7 +3153,7 @@ __pattern_is_partitioned(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _
         const _ReduceType __identity{__not_init, __last};
 
         _ReduceType __result = __par_backend::__parallel_reduce(
-            ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __identity,
+            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __identity,
             [&__pred, __combine](_RandomAccessIterator __i, _RandomAccessIterator __j,
                                               _ReduceType __value) -> _ReduceType {
                 if (__value.__val == __broken)
@@ -3235,6 +3252,9 @@ oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _Ra
 __pattern_partition(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIterator __last,
                     _UnaryPredicate __pred, _IsVector __is_vector, /*is_parallel=*/::std::true_type)
 {
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
 
     // partitioned range: elements before pivot satisfy pred (true part),
     //                    elements after pivot don't satisfy pred (false part)
@@ -3286,6 +3306,7 @@ __pattern_partition(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _R
         };
 
         _PartitionRange __result = __par_backend::__parallel_reduce(
+            __backend_tag{},
             ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __init,
             [__pred, __is_vector, __reductor](_RandomAccessIterator __i, _RandomAccessIterator __j,
                                               _PartitionRange __value) -> _PartitionRange {
@@ -3360,8 +3381,12 @@ __pattern_partition(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Rando
             }
         };
 
+        constexpr auto __dispatch_tag =
+            oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
+        using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
         _PartitionRange __result = __par_backend::__parallel_reduce(
-            ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __init,
+            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __init,
             [__pred, __reductor](_RandomAccessIterator __i, _RandomAccessIterator __j,
                                  _PartitionRange __value) -> _PartitionRange {
                 //1. serial partition
@@ -3452,7 +3477,12 @@ __pattern_stable_partition(_ExecutionPolicy&& __exec, _RandomAccessIterator __fi
             }
         };
 
+        constexpr auto __dispatch_tag =
+            oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
+        using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
         _PartitionRange __result = __par_backend::__parallel_reduce(
+            __backend_tag{},
             ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __init,
             [&__pred, __is_vector, __reductor](_RandomAccessIterator __i, _RandomAccessIterator __j,
                                                _PartitionRange __value) -> _PartitionRange {
@@ -3503,8 +3533,12 @@ __pattern_stable_partition(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec,
             }
         };
 
+        constexpr auto __dispatch_tag =
+            oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
+        using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
         _PartitionRange __result = __par_backend::__parallel_reduce(
-            ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __init,
+            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __init,
             [&__pred, __reductor](_RandomAccessIterator __i, _RandomAccessIterator __j,
                                   _PartitionRange __value) -> _PartitionRange {
                 //1. serial stable_partition
@@ -4122,8 +4156,13 @@ __pattern_adjacent_find(_ExecutionPolicy&& __exec, _RandomAccessIterator __first
     if (__last - __first < 2)
         return __last;
 
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     return __internal::__except_handler([&]() {
         return __par_backend::__parallel_reduce(
+            __backend_tag{},
             ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __last,
             [__last, __pred, __is_vector, __or_semantic](_RandomAccessIterator __begin, _RandomAccessIterator __end,
                                                          _RandomAccessIterator __value) -> _RandomAccessIterator {
@@ -4166,8 +4205,13 @@ __pattern_adjacent_find(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _R
     if (__last - __first < 2)
         return __last;
 
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     return __internal::__except_handler([&]() {
         return __par_backend::__parallel_reduce(
+            __backend_tag{},
             ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __last,
             [__last, __pred, __or_semantic](_RandomAccessIterator __begin, _RandomAccessIterator __end,
                                             _RandomAccessIterator __value) -> _RandomAccessIterator {
@@ -5992,8 +6036,13 @@ __pattern_min_element(_ExecutionPolicy&& __exec, _RandomAccessIterator __first,
     if (__last - __first < 2)
         return __first;
 
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     return __internal::__except_handler([&]() {
         return __par_backend::__parallel_reduce(
+            __backend_tag{},
             ::std::forward<_ExecutionPolicy>(__exec), __first, __last, /*identity*/ __last,
             [=](_RandomAccessIterator __begin, _RandomAccessIterator __end,
                 _RandomAccessIterator __init) -> _RandomAccessIterator {
@@ -6023,8 +6072,11 @@ __pattern_min_element(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Ran
     if (__last - __first < 2)
         return __first;
 
+    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
+
     return __internal::__except_handler([&]() {
         return __par_backend::__parallel_reduce(
+            __backend_tag{},
             ::std::forward<_ExecutionPolicy>(__exec), __first, __last, /*identity*/ __last,
             [=](_RandomAccessIterator __begin, _RandomAccessIterator __end,
                 _RandomAccessIterator __init) -> _RandomAccessIterator {
@@ -6098,10 +6150,15 @@ __pattern_minmax_element(_ExecutionPolicy&& __exec, _RandomAccessIterator __firs
     if (__last - __first < 2)
         return ::std::make_pair(__first, __first);
 
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     return __internal::__except_handler([&]() {
         typedef ::std::pair<_RandomAccessIterator, _RandomAccessIterator> _Result;
 
         return __par_backend::__parallel_reduce(
+            __backend_tag{},
             ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
             /*identity*/ ::std::make_pair(__last, __last),
             [=, &__comp](_RandomAccessIterator __begin, _RandomAccessIterator __end, _Result __init) -> _Result {
@@ -6139,10 +6196,13 @@ __pattern_minmax_element(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _
     if (__last - __first < 2)
         return ::std::make_pair(__first, __first);
 
+    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
+
     return __internal::__except_handler([&]() {
         typedef ::std::pair<_RandomAccessIterator, _RandomAccessIterator> _Result;
 
         return __par_backend::__parallel_reduce(
+            __backend_tag{},
             ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
             /*identity*/ ::std::make_pair(__last, __last),
             [=, &__comp](_RandomAccessIterator __begin, _RandomAccessIterator __end, _Result __init) -> _Result {
diff --git a/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h b/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
index 170cc0f0d62..35e7433a09f 100644
--- a/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
+++ b/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
@@ -391,6 +391,8 @@ void
 __pattern_for_loop_n(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _Ip __first, _Size __n, _Function __f,
                      __single_stride_type, _Rest&&... __rest)
 {
+    using __backend_tag = typename decltype(__tag)::__backend_tag;
+
     using __pack_type = __reduction_pack<_Rest...>;
 
     // Create an identity pack object, operations are done on copies of it.
@@ -398,6 +400,7 @@ __pattern_for_loop_n(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec,
 
     oneapi::dpl::__internal::__except_handler([&]() {
         return __par_backend::__parallel_reduce(
+                   __backend_tag{},
                    ::std::forward<_ExecutionPolicy>(__exec), _Size(0), __n, __identity,
                    [__first, __f](_Size __i, _Size __j, __pack_type __value) {
                        const auto __subseq_start = __first + __i;
@@ -426,6 +429,8 @@ void
 __pattern_for_loop_n(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _Ip __first, _Size __n, _Function __f,
                      _Sp __stride, _Rest&&... __rest)
 {
+    using __backend_tag = typename decltype(__tag)::__backend_tag;
+
     using __pack_type = __reduction_pack<_Rest...>;
 
     // Create an identity pack object, operations are done on copies of it.
@@ -433,6 +438,7 @@ __pattern_for_loop_n(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec,
 
     oneapi::dpl::__internal::__except_handler([&]() {
         return __par_backend::__parallel_reduce(
+                   __backend_tag{},
                    ::std::forward<_ExecutionPolicy>(__exec), _Size(0), __n, __identity,
                    [__first, __f, __stride](_Size __i, _Size __j, __pack_type __value) {
                        const auto __subseq_start = __first + __i * __stride;

From adb4c5ecc63c50ab5f2ceea9091b4d4dec7682fa Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 14 Feb 2024 15:45:49 +0100
Subject: [PATCH 188/566] __parallel_strict_scan + tag impls

---
 include/oneapi/dpl/pstl/omp/parallel_scan.h   | 36 +++++++++++++++++
 .../oneapi/dpl/pstl/parallel_backend_serial.h | 13 ++++++
 .../oneapi/dpl/pstl/parallel_backend_tbb.h    | 40 +++++++++++++++++++
 3 files changed, 89 insertions(+)

diff --git a/include/oneapi/dpl/pstl/omp/parallel_scan.h b/include/oneapi/dpl/pstl/omp/parallel_scan.h
index 29c6c77be54..7001d4267b3 100644
--- a/include/oneapi/dpl/pstl/omp/parallel_scan.h
+++ b/include/oneapi/dpl/pstl/omp/parallel_scan.h
@@ -142,6 +142,42 @@ __parallel_strict_scan(_ExecutionPolicy&&, _Index __n, _Tp __initial, _Rp __redu
     }
 }
 
+template <class _ExecutionPolicy, typename _Index, typename _Tp, typename _Rp, typename _Cp, typename _Sp, typename _Ap>
+void
+__parallel_strict_scan(oneapi::dpl::__internal::__omp_backend_tag, _ExecutionPolicy&&, _Index __n, _Tp __initial,
+                       _Rp __reduce, _Cp __combine, _Sp __scan, _Ap __apex)
+{
+    if (__n <= __default_chunk_size)
+    {
+        _Tp __sum = __initial;
+        if (__n)
+        {
+            __sum = __combine(__sum, __reduce(_Index(0), __n));
+        }
+        __apex(__sum);
+        if (__n)
+        {
+            __scan(_Index(0), __n, __initial);
+        }
+        return;
+    }
+
+    if (omp_in_parallel())
+    {
+        oneapi::dpl::__omp_backend::__parallel_strict_scan_body<_ExecutionPolicy>(__n, __initial, __reduce, __combine,
+                                                                                  __scan, __apex);
+    }
+    else
+    {
+        _PSTL_PRAGMA(omp parallel)
+        _PSTL_PRAGMA(omp single nowait)
+        {
+            oneapi::dpl::__omp_backend::__parallel_strict_scan_body<_ExecutionPolicy>(__n, __initial, __reduce,
+                                                                                      __combine, __scan, __apex);
+        }
+    }
+}
+
 } // namespace __omp_backend
 } // namespace dpl
 } // namespace oneapi
diff --git a/include/oneapi/dpl/pstl/parallel_backend_serial.h b/include/oneapi/dpl/pstl/parallel_backend_serial.h
index c6de6d97784..6634e692033 100644
--- a/include/oneapi/dpl/pstl/parallel_backend_serial.h
+++ b/include/oneapi/dpl/pstl/parallel_backend_serial.h
@@ -138,6 +138,19 @@ __parallel_strict_scan(_ExecutionPolicy&&, _Index __n, _Tp __initial, _Rp __redu
         __scan(_Index(0), __n, __initial);
 }
 
+template <class _ExecutionPolicy, typename _Index, typename _Tp, typename _Rp, typename _Cp, typename _Sp, typename _Ap>
+void
+__parallel_strict_scan(oneapi::dpl::__internal::__serial_backend_tag, _ExecutionPolicy&&, _Index __n, _Tp __initial,
+                       _Rp __reduce, _Cp __combine, _Sp __scan, _Ap __apex)
+{
+    _Tp __sum = __initial;
+    if (__n)
+        __sum = __combine(__sum, __reduce(_Index(0), __n));
+    __apex(__sum);
+    if (__n)
+        __scan(_Index(0), __n, __initial);
+}
+
 template <class _ExecutionPolicy, class _Index, class _UnaryOp, class _Tp, class _BinaryOp, class _Reduce, class _Scan>
 _Tp
 __parallel_transform_scan(_ExecutionPolicy&&, _Index __n, _UnaryOp, _Tp __init, _BinaryOp, _Reduce, _Scan __scan)
diff --git a/include/oneapi/dpl/pstl/parallel_backend_tbb.h b/include/oneapi/dpl/pstl/parallel_backend_tbb.h
index 77276c469ac..98ac5da86cd 100644
--- a/include/oneapi/dpl/pstl/parallel_backend_tbb.h
+++ b/include/oneapi/dpl/pstl/parallel_backend_tbb.h
@@ -456,6 +456,46 @@ __parallel_strict_scan(_ExecutionPolicy&&, _Index __n, _Tp __initial, _Rp __redu
     });
 }
 
+template <class _ExecutionPolicy, typename _Index, typename _Tp, typename _Rp, typename _Cp, typename _Sp, typename _Ap>
+void
+__parallel_strict_scan(oneapi::dpl::__internal::__tbb_backend_tag, _ExecutionPolicy&&, _Index __n, _Tp __initial,
+                       _Rp __reduce, _Cp __combine, _Sp __scan, _Ap __apex)
+{
+    tbb::this_task_arena::isolate([=, &__combine]() {
+        if (__n > 1)
+        {
+            _Index __p = tbb::this_task_arena::max_concurrency();
+            const _Index __slack = 4;
+            _Index __tilesize = (__n - 1) / (__slack * __p) + 1;
+            _Index __m = (__n - 1) / __tilesize;
+            __tbb_backend::__buffer<_ExecutionPolicy, _Tp> __buf(__m + 1);
+            _Tp* __r = __buf.get();
+            __tbb_backend::__upsweep(_Index(0), _Index(__m + 1), __tilesize, __r, __n - __m * __tilesize, __reduce,
+                                     __combine);
+
+            // When __apex is a no-op and __combine has no side effects, a good optimizer
+            // should be able to eliminate all code between here and __apex.
+            // Alternatively, provide a default value for __apex that can be
+            // recognized by metaprogramming that conditionlly executes the following.
+            size_t __k = __m + 1;
+            _Tp __t = __r[__k - 1];
+            while ((__k &= __k - 1))
+                __t = __combine(__r[__k - 1], __t);
+            __apex(__combine(__initial, __t));
+            __tbb_backend::__downsweep(_Index(0), _Index(__m + 1), __tilesize, __r, __n - __m * __tilesize, __initial,
+                                       __combine, __scan);
+            return;
+        }
+        // Fewer than 2 elements in sequence, or out of memory.  Handle has single block.
+        _Tp __sum = __initial;
+        if (__n)
+            __sum = __combine(__sum, __reduce(_Index(0), __n));
+        __apex(__sum);
+        if (__n)
+            __scan(_Index(0), __n, __initial);
+    });
+}
+
 template <class _ExecutionPolicy, class _Index, class _Up, class _Tp, class _Cp, class _Rp, class _Sp>
 _Tp
 __parallel_transform_scan(_ExecutionPolicy&&, _Index __n, _Up __u, _Tp __init, _Cp __combine, _Rp __brick_reduce,

From a36888f1e160987930732490e1ad767bbfb4a664 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 14 Feb 2024 16:00:02 +0100
Subject: [PATCH 189/566] __parallel_strict_scan + tag calls

---
 include/oneapi/dpl/pstl/algorithm_impl.h | 31 ++++++++++++++++++++++++
 include/oneapi/dpl/pstl/numeric_impl.h   | 12 ++++++++-
 2 files changed, 42 insertions(+), 1 deletion(-)

diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index aa4c6b5b4d1..7c722594610 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -1968,6 +1968,10 @@ __pattern_copy_if(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first, _Ra
                   _RandomAccessIterator2 __result, _UnaryPredicate __pred, _IsVector __is_vector,
                   /*parallel=*/::std::true_type)
 {
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     typedef typename ::std::iterator_traits<_RandomAccessIterator1>::difference_type _DifferenceType;
     const _DifferenceType __n = __last - __first;
     if (_DifferenceType(1) < __n)
@@ -1977,6 +1981,7 @@ __pattern_copy_if(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first, _Ra
             bool* __mask = __mask_buf.get();
             _DifferenceType __m{};
             __par_backend::__parallel_strict_scan(
+                __backend_tag{},
                 ::std::forward<_ExecutionPolicy>(__exec), __n, _DifferenceType(0),
                 [=](_DifferenceType __i, _DifferenceType __len) { // Reduce
                     return __internal::__brick_calc_mask_1<_DifferenceType>(__first + __i, __first + (__i + __len),
@@ -2003,6 +2008,8 @@ _RandomAccessIterator2
 __pattern_copy_if(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator1 __first,
                   _RandomAccessIterator1 __last, _RandomAccessIterator2 __result, _UnaryPredicate __pred)
 {
+    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
+
     typedef typename ::std::iterator_traits<_RandomAccessIterator1>::difference_type _DifferenceType;
     const _DifferenceType __n = __last - __first;
     if (_DifferenceType(1) < __n)
@@ -2012,6 +2019,7 @@ __pattern_copy_if(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomA
             bool* __mask = __mask_buf.get();
             _DifferenceType __m{};
             __par_backend::__parallel_strict_scan(
+                __backend_tag{},
                 ::std::forward<_ExecutionPolicy>(__exec), __n, _DifferenceType(0),
                 [=](_DifferenceType __i, _DifferenceType __len) { // Reduce
                     return __internal::__brick_calc_mask_1<_DifferenceType>(__first + __i, __first + (__i + __len),
@@ -2212,6 +2220,7 @@ __remove_elements(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardI
         _DifferenceType __m{};
         // 2. Elements that doesn't satisfy pred are moved to result
         __par_backend::__parallel_strict_scan(
+            __backend_tag{},
             ::std::forward<_ExecutionPolicy>(__exec), __n, _DifferenceType(0),
             [__mask, __is_vector](_DifferenceType __i, _DifferenceType __len) {
                 return __internal::__brick_count(
@@ -2365,6 +2374,10 @@ __pattern_unique_copy(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first,
                       _RandomAccessIterator2 __result, _BinaryPredicate __pred, _IsVector __is_vector,
                       /*parallel=*/::std::true_type)
 {
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     typedef typename ::std::iterator_traits<_RandomAccessIterator1>::difference_type _DifferenceType;
     const _DifferenceType __n = __last - __first;
     if (_DifferenceType(2) < __n)
@@ -2376,6 +2389,7 @@ __pattern_unique_copy(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first,
                 bool* __mask = __mask_buf.get();
                 _DifferenceType __m{};
                 __par_backend::__parallel_strict_scan(
+                    __backend_tag{},
                     ::std::forward<_ExecutionPolicy>(__exec), __n, _DifferenceType(0),
                     [=](_DifferenceType __i, _DifferenceType __len) -> _DifferenceType { // Reduce
                         _DifferenceType __extra = 0;
@@ -2414,6 +2428,8 @@ _RandomAccessIterator2
 __pattern_unique_copy(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator1 __first,
                       _RandomAccessIterator1 __last, _RandomAccessIterator2 __result, _BinaryPredicate __pred)
 {
+    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
+
     typedef typename ::std::iterator_traits<_RandomAccessIterator1>::difference_type _DifferenceType;
     const _DifferenceType __n = __last - __first;
     if (_DifferenceType(2) < __n)
@@ -2425,6 +2441,7 @@ __pattern_unique_copy(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Ran
                 bool* __mask = __mask_buf.get();
                 _DifferenceType __m{};
                 __par_backend::__parallel_strict_scan(
+                    __backend_tag{},
                     ::std::forward<_ExecutionPolicy>(__exec), __n, _DifferenceType(0),
                     [=](_DifferenceType __i, _DifferenceType __len) -> _DifferenceType { // Reduce
                         _DifferenceType __extra = 0;
@@ -3609,6 +3626,11 @@ __pattern_partition_copy(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __fir
                          _RandomAccessIterator2 __out_true, _RandomAccessIterator3 __out_false, _UnaryPredicate __pred,
                          _IsVector __is_vector, /*is_parallelization=*/::std::true_type)
 {
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2,
+                                                  _RandomAccessIterator3>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     typedef typename ::std::iterator_traits<_RandomAccessIterator1>::difference_type _DifferenceType;
     typedef ::std::pair<_DifferenceType, _DifferenceType> _ReturnType;
     const _DifferenceType __n = __last - __first;
@@ -3620,6 +3642,7 @@ __pattern_partition_copy(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __fir
             bool* __mask = __mask_buf.get();
             _ReturnType __m{};
             __par_backend::__parallel_strict_scan(
+                __backend_tag{},
                 ::std::forward<_ExecutionPolicy>(__exec), __n, ::std::make_pair(_DifferenceType(0), _DifferenceType(0)),
                 [=](_DifferenceType __i, _DifferenceType __len) { // Reduce
                     return __internal::__brick_calc_mask_1<_DifferenceType>(__first + __i, __first + (__i + __len),
@@ -3648,6 +3671,8 @@ __pattern_partition_copy(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _
                          _RandomAccessIterator1 __last, _RandomAccessIterator2 __out_true,
                          _RandomAccessIterator3 __out_false, _UnaryPredicate __pred)
 {
+    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
+
     typedef typename ::std::iterator_traits<_RandomAccessIterator1>::difference_type _DifferenceType;
     typedef ::std::pair<_DifferenceType, _DifferenceType> _ReturnType;
     const _DifferenceType __n = __last - __first;
@@ -3658,6 +3683,7 @@ __pattern_partition_copy(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _
             bool* __mask = __mask_buf.get();
             _ReturnType __m{};
             __par_backend::__parallel_strict_scan(
+                __backend_tag{},
                 ::std::forward<_ExecutionPolicy>(__exec), __n, ::std::make_pair(_DifferenceType(0), _DifferenceType(0)),
                 [=](_DifferenceType __i, _DifferenceType __len) { // Reduce
                     return __internal::__brick_calc_mask_1<_DifferenceType>(__first + __i, __first + (__i + __len),
@@ -5068,6 +5094,10 @@ __parallel_set_op(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _Forwar
                   _ForwardIterator2 __first2, _ForwardIterator2 __last2, _OutputIterator __result, _Compare __comp,
                   _SizeFunction __size_func, _SetOP __set_op, _IsVector __is_vector)
 {
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1,
+                                                                              _ForwardIterator2, _OutputIterator>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     typedef typename ::std::iterator_traits<_ForwardIterator1>::difference_type _DifferenceType;
     typedef typename ::std::iterator_traits<_OutputIterator>::value_type _T;
 
@@ -5097,6 +5127,7 @@ __parallel_set_op(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _Forwar
                                                          __result + __s.__pos, __is_vector);
         };
         __par_backend::__parallel_strict_scan(
+            __backend_tag{},
             ::std::forward<_ExecutionPolicy>(__exec), __n1, _SetRange{0, 0, 0}, //-1, 0},
             [=](_DifferenceType __i, _DifferenceType __len) {                   // Reduce
                 //[__b; __e) - a subrange of the first sequence, to reduce
diff --git a/include/oneapi/dpl/pstl/numeric_impl.h b/include/oneapi/dpl/pstl/numeric_impl.h
index 1bf5ec91e9f..7a1c05f7af7 100644
--- a/include/oneapi/dpl/pstl/numeric_impl.h
+++ b/include/oneapi/dpl/pstl/numeric_impl.h
@@ -407,8 +407,14 @@ __pattern_transform_scan(_ExecutionPolicy&& __exec, _RandomAccessIterator __firs
     {
         return __result;
     }
+
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator, _OutputIterator>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     return __internal::__except_handler([&]() {
         __par_backend::__parallel_strict_scan(
+            __backend_tag{},
             ::std::forward<_ExecutionPolicy>(__exec), __n, __init,
             [__first, __unary_op, __binary_op, __result, __is_vector](_DifferenceType __i, _DifferenceType __len) {
                 return __internal::__brick_transform_scan(__first + __i, __first + (__i + __len), __result + __i,
@@ -432,7 +438,7 @@ __pattern_transform_scan(_ExecutionPolicy&& __exec, _RandomAccessIterator __firs
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _OutputIterator,
           class _UnaryOperation, class _Tp, class _BinaryOperation, class _Inclusive>
 ::std::enable_if_t<::std::is_floating_point_v<_Tp>, _OutputIterator>
-__pattern_transform_scan(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
+__pattern_transform_scan(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
                          _RandomAccessIterator __last, _OutputIterator __result, _UnaryOperation __unary_op, _Tp __init,
                          _BinaryOperation __binary_op, _Inclusive)
 {
@@ -443,8 +449,12 @@ __pattern_transform_scan(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __e
     {
         return __result;
     }
+
+    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
+
     return __internal::__except_handler([&]() {
         __par_backend::__parallel_strict_scan(
+            __backend_tag{},
             ::std::forward<_ExecutionPolicy>(__exec), __n, __init,
             [__first, __unary_op, __binary_op, __result](_DifferenceType __i, _DifferenceType __len) {
                 return __internal::__brick_transform_scan(__first + __i, __first + (__i + __len), __result + __i,

From 274def37334287edd0366203b299bd681ef2d424 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 14 Feb 2024 16:13:36 +0100
Subject: [PATCH 190/566] __parallel_transform_scan + tag impls

---
 .../pstl/hetero/dpcpp/parallel_backend_sycl.h | 59 +++++++++++++++++++
 .../hetero/dpcpp/parallel_backend_sycl_fpga.h | 13 ++++
 .../dpl/pstl/omp/parallel_transform_scan.h    |  9 +++
 .../oneapi/dpl/pstl/parallel_backend_serial.h |  8 +++
 .../oneapi/dpl/pstl/parallel_backend_tbb.h    | 12 ++++
 5 files changed, 101 insertions(+)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
index 0cc786fa94a..faf435a51ff 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
@@ -844,6 +844,65 @@ __parallel_transform_scan(_ExecutionPolicy&& __exec, _Range1&& __in_rng, _Range2
             .event());
 }
 
+template <typename _ExecutionPolicy, typename _Range1, typename _Range2, typename _UnaryOperation, typename _InitType,
+          typename _BinaryOperation, typename _Inclusive>
+auto
+__parallel_transform_scan(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&& __exec, _Range1&& __in_rng,
+                          _Range2&& __out_rng, ::std::size_t __n,
+                          _UnaryOperation __unary_op, _InitType __init, _BinaryOperation __binary_op, _Inclusive)
+{
+    using _Type = typename _InitType::__value_type;
+
+    // Next power of 2 greater than or equal to __n
+    auto __n_uniform = __n;
+    if ((__n_uniform & (__n_uniform - 1)) != 0)
+        __n_uniform = oneapi::dpl::__internal::__dpl_bit_floor(__n) << 1;
+
+    // Pessimistically only use half of the memory to take into account memory used by compiled kernel
+    const ::std::size_t __max_slm_size =
+        __exec.queue().get_device().template get_info<sycl::info::device::local_mem_size>() / 2;
+    const auto __req_slm_size = sizeof(_Type) * __n_uniform;
+
+    constexpr int __single_group_upper_limit = 16384;
+
+    constexpr bool __can_use_group_scan = unseq_backend::__has_known_identity<_BinaryOperation, _Type>::value;
+    if constexpr (__can_use_group_scan)
+    {
+        if (__n <= __single_group_upper_limit && __max_slm_size >= __req_slm_size)
+        {
+            return __parallel_transform_scan_single_group(
+                std::forward<_ExecutionPolicy>(__exec), ::std::forward<_Range1>(__in_rng),
+                ::std::forward<_Range2>(__out_rng), __n, __unary_op, __init, __binary_op, _Inclusive{});
+        }
+    }
+
+    // Either we can't use group scan or this input is too big for one workgroup
+    using _Assigner = unseq_backend::__scan_assigner;
+    using _NoAssign = unseq_backend::__scan_no_assign;
+    using _UnaryFunctor = unseq_backend::walk_n<_ExecutionPolicy, _UnaryOperation>;
+    using _NoOpFunctor = unseq_backend::walk_n<_ExecutionPolicy, oneapi::dpl::__internal::__no_op>;
+
+    _Assigner __assign_op;
+    _NoAssign __no_assign_op;
+    _NoOpFunctor __get_data_op;
+
+    return __future(
+        __parallel_transform_scan_base(
+            ::std::forward<_ExecutionPolicy>(__exec), ::std::forward<_Range1>(__in_rng),
+            ::std::forward<_Range2>(__out_rng), __binary_op, __init,
+            // local scan
+            unseq_backend::__scan<_Inclusive, _ExecutionPolicy, _BinaryOperation, _UnaryFunctor, _Assigner, _Assigner,
+                                  _NoOpFunctor, _InitType>{__binary_op, _UnaryFunctor{__unary_op}, __assign_op,
+                                                           __assign_op, __get_data_op},
+            // scan between groups
+            unseq_backend::__scan</*inclusive=*/::std::true_type, _ExecutionPolicy, _BinaryOperation, _NoOpFunctor,
+                                  _NoAssign, _Assigner, _NoOpFunctor, unseq_backend::__no_init_value<_Type>>{
+                __binary_op, _NoOpFunctor{}, __no_assign_op, __assign_op, __get_data_op},
+            // global scan
+            unseq_backend::__global_scan_functor<_Inclusive, _BinaryOperation, _InitType>{__binary_op, __init})
+            .event());
+}
+
 template <typename _SizeType>
 struct __invoke_single_group_copy_if
 {
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
index 64898115e0c..c554011a1a6 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
@@ -173,6 +173,19 @@ __parallel_transform_scan(_ExecutionPolicy&& __exec, _Range1&& __in_rng, _Range2
         __unary_op, __init, __binary_op, _Inclusive{});
 }
 
+template <typename _ExecutionPolicy, typename _Range1, typename _Range2, typename _UnaryOperation, typename _InitType,
+          typename _BinaryOperation, typename _Inclusive>
+auto
+__parallel_transform_scan(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionPolicy&& __exec, _Range1&& __in_rng,
+                          _Range2&& __out_rng, ::std::size_t __n, _UnaryOperation __unary_op, _InitType __init,
+                          _BinaryOperation __binary_op, _Inclusive)
+{
+    // workaround until we implement more performant version for patterns
+    return oneapi::dpl::__par_backend_hetero::__parallel_transform_scan(
+        oneapi::dpl::__internal::__device_backend_tag{}, __exec.__device_policy(), ::std::forward<_Range1>(__in_rng),
+        ::std::forward<_Range2>(__out_rng), __n, __unary_op, __init, __binary_op, _Inclusive{});
+}
+
 template <typename _ExecutionPolicy, typename _Range1, typename _Range2, typename _BinaryOperation, typename _InitType,
           typename _LocalScan, typename _GroupScan, typename _GlobalScan,
           oneapi::dpl::__internal::__enable_if_fpga_execution_policy<_ExecutionPolicy, int> = 0>
diff --git a/include/oneapi/dpl/pstl/omp/parallel_transform_scan.h b/include/oneapi/dpl/pstl/omp/parallel_transform_scan.h
index 98262635d1e..f0093ccab98 100644
--- a/include/oneapi/dpl/pstl/omp/parallel_transform_scan.h
+++ b/include/oneapi/dpl/pstl/omp/parallel_transform_scan.h
@@ -34,6 +34,15 @@ __parallel_transform_scan(_ExecutionPolicy&&, _Index __n, _Up /* __u */, _Tp __i
     return __scan(_Index(0), __n, __init);
 }
 
+template <class _ExecutionPolicy, class _Index, class _Up, class _Tp, class _Cp, class _Rp, class _Sp>
+_Tp
+__parallel_transform_scan(oneapi::dpl::__internal::__omp_backend_tag, _ExecutionPolicy&&, _Index __n, _Up /* __u */,
+                          _Tp __init, _Cp /* __combine */, _Rp /* __brick_reduce */, _Sp __scan)
+{
+    // TODO: parallelize this function.
+    return __scan(_Index(0), __n, __init);
+}
+
 } // namespace __omp_backend
 } // namespace dpl
 } // namespace oneapi
diff --git a/include/oneapi/dpl/pstl/parallel_backend_serial.h b/include/oneapi/dpl/pstl/parallel_backend_serial.h
index 6634e692033..590201fd16d 100644
--- a/include/oneapi/dpl/pstl/parallel_backend_serial.h
+++ b/include/oneapi/dpl/pstl/parallel_backend_serial.h
@@ -158,6 +158,14 @@ __parallel_transform_scan(_ExecutionPolicy&&, _Index __n, _UnaryOp, _Tp __init,
     return __scan(_Index(0), __n, __init);
 }
 
+template <class _ExecutionPolicy, class _Index, class _UnaryOp, class _Tp, class _BinaryOp, class _Reduce, class _Scan>
+_Tp
+__parallel_transform_scan(oneapi::dpl::__internal::__serial_backend_tag, _ExecutionPolicy&&, _Index __n, _UnaryOp,
+                          _Tp __init, _BinaryOp, _Reduce, _Scan __scan)
+{
+    return __scan(_Index(0), __n, __init);
+}
+
 template <class _ExecutionPolicy, typename _RandomAccessIterator, typename _Compare, typename _LeafSort>
 void
 __parallel_stable_sort(_ExecutionPolicy&&, _RandomAccessIterator __first, _RandomAccessIterator __last, _Compare __comp,
diff --git a/include/oneapi/dpl/pstl/parallel_backend_tbb.h b/include/oneapi/dpl/pstl/parallel_backend_tbb.h
index 98ac5da86cd..531858c9f67 100644
--- a/include/oneapi/dpl/pstl/parallel_backend_tbb.h
+++ b/include/oneapi/dpl/pstl/parallel_backend_tbb.h
@@ -507,6 +507,18 @@ __parallel_transform_scan(_ExecutionPolicy&&, _Index __n, _Up __u, _Tp __init, _
     return __body.sum();
 }
 
+template <class _ExecutionPolicy, class _Index, class _Up, class _Tp, class _Cp, class _Rp, class _Sp>
+_Tp
+__parallel_transform_scan(oneapi::dpl::__internal::__tbb_backend_tag, _ExecutionPolicy&&, _Index __n, _Up __u,
+                          _Tp __init, _Cp __combine, _Rp __brick_reduce, _Sp __scan)
+{
+    __trans_scan_body<_Index, _Up, _Tp, _Cp, _Rp, _Sp> __body(__u, __init, __combine, __brick_reduce, __scan);
+    auto __range = tbb::blocked_range<_Index>(0, __n);
+    tbb::this_task_arena::isolate([__range, &__body]() { tbb::parallel_scan(__range, __body); });
+    return __body.sum();
+}
+
+
 //------------------------------------------------------------------------
 // parallel_stable_sort
 //------------------------------------------------------------------------

From dbe64de6a7ca5d8fddd2d4490a2b720715c93f52 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 14 Feb 2024 16:15:14 +0100
Subject: [PATCH 191/566] __parallel_transform_scan + tag calls

---
 .../hetero/dpcpp/parallel_backend_sycl_fpga.h |  1 +
 .../dpl/pstl/hetero/numeric_impl_hetero.h     | 24 +++++++++++--------
 2 files changed, 15 insertions(+), 10 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
index c554011a1a6..bc1d53d4cea 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
@@ -169,6 +169,7 @@ __parallel_transform_scan(_ExecutionPolicy&& __exec, _Range1&& __in_rng, _Range2
 {
     // workaround until we implement more performant version for patterns
     return oneapi::dpl::__par_backend_hetero::__parallel_transform_scan(
+        oneapi::dpl::__internal::__device_backend_tag{},
         __exec.__device_policy(), ::std::forward<_Range1>(__in_rng), ::std::forward<_Range2>(__out_rng), __n,
         __unary_op, __init, __binary_op, _Inclusive{});
 }
diff --git a/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
index 3edccc031e6..16521a0835b 100644
--- a/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
@@ -185,6 +185,7 @@ __pattern_transform_scan_base(_ExecutionPolicy&& __exec, _Iterator1 __first, _It
         return __result;
 
     const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator1, _Iterator2>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
 
     const auto __n = __last - __first;
 
@@ -198,9 +199,9 @@ __pattern_transform_scan_base(_ExecutionPolicy&& __exec, _Iterator1 __first, _It
         auto __keep2 = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::write, _Iterator2>();
         auto __buf2 = __keep2(__result, __result + __n);
 
-        oneapi::dpl::__par_backend_hetero::__parallel_transform_scan(::std::forward<_ExecutionPolicy>(__exec),
-                                                                     __buf1.all_view(), __buf2.all_view(), __n,
-                                                                     __unary_op, __init, __binary_op, _Inclusive{})
+        oneapi::dpl::__par_backend_hetero::__parallel_transform_scan(
+            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __buf1.all_view(), __buf2.all_view(), __n,
+            __unary_op, __init, __binary_op, _Inclusive{})
             .wait();
     }
     else
@@ -223,8 +224,9 @@ __pattern_transform_scan_base(_ExecutionPolicy&& __exec, _Iterator1 __first, _It
         auto __buf2 = __keep2(__first_tmp, __last_tmp);
 
         // Run main algorithm and save data into temporary buffer
-        oneapi::dpl::__par_backend_hetero::__parallel_transform_scan(__policy, __buf1.all_view(), __buf2.all_view(),
-                                                                     __n, __unary_op, __init, __binary_op, _Inclusive{})
+        oneapi::dpl::__par_backend_hetero::__parallel_transform_scan(__backend_tag{}, __policy, __buf1.all_view(),
+                                                                     __buf2.all_view(), __n, __unary_op, __init,
+                                                                     __binary_op, _Inclusive{})
             .wait();
 
         // Move data from temporary buffer into results
@@ -263,6 +265,7 @@ __pattern_transform_scan_base(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __ex
         return __result;
 
     const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator1, _Iterator2>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
 
     const auto __n = __last - __first;
 
@@ -276,9 +279,9 @@ __pattern_transform_scan_base(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __ex
         auto __keep2 = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::write, _Iterator2>();
         auto __buf2 = __keep2(__result, __result + __n);
 
-        oneapi::dpl::__par_backend_hetero::__parallel_transform_scan(::std::forward<_ExecutionPolicy>(__exec),
-                                                                     __buf1.all_view(), __buf2.all_view(), __n,
-                                                                     __unary_op, __init, __binary_op, _Inclusive{})
+        oneapi::dpl::__par_backend_hetero::__parallel_transform_scan(
+            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __buf1.all_view(), __buf2.all_view(), __n,
+            __unary_op, __init, __binary_op, _Inclusive{})
             .wait();
     }
     else
@@ -301,8 +304,9 @@ __pattern_transform_scan_base(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __ex
         auto __buf2 = __keep2(__first_tmp, __last_tmp);
 
         // Run main algorithm and save data into temporary buffer
-        oneapi::dpl::__par_backend_hetero::__parallel_transform_scan(__policy, __buf1.all_view(), __buf2.all_view(),
-                                                                     __n, __unary_op, __init, __binary_op, _Inclusive{})
+        oneapi::dpl::__par_backend_hetero::__parallel_transform_scan(__backend_tag{}, __policy, __buf1.all_view(),
+                                                                     __buf2.all_view(), __n, __unary_op, __init,
+                                                                     __binary_op, _Inclusive{})
             .wait();
 
         // Move data from temporary buffer into results

From 3cc8b8da44d0845aefdfa385d20e9e1c172d2ec5 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 14 Feb 2024 16:20:47 +0100
Subject: [PATCH 192/566] __parallel_stable_sort + tag impls

---
 .../pstl/hetero/dpcpp/parallel_backend_sycl.h | 26 +++++++++++
 .../hetero/dpcpp/parallel_backend_sycl_fpga.h | 11 +++++
 .../dpl/pstl/omp/parallel_stable_sort.h       | 43 +++++++++++++++++++
 .../oneapi/dpl/pstl/parallel_backend_serial.h |  8 ++++
 .../oneapi/dpl/pstl/parallel_backend_tbb.h    | 25 +++++++++++
 5 files changed, 113 insertions(+)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
index faf435a51ff..c3e2fe8b483 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
@@ -2018,6 +2018,18 @@ __parallel_stable_sort(_ExecutionPolicy&& __exec, _Range&& __rng, _Compare, _Pro
     return __parallel_radix_sort<__internal::__is_comp_ascending<::std::decay_t<_Compare>>::value>(
         ::std::forward<_ExecutionPolicy>(__exec), ::std::forward<_Range>(__rng), __proj);
 }
+
+template <typename _ExecutionPolicy, typename _Range, typename _Compare, typename _Proj,
+          ::std::enable_if_t<
+              __is_radix_sort_usable_for_type<oneapi::dpl::__internal::__key_t<_Proj, _Range>, _Compare>::value,
+              int> = 0>
+auto
+__parallel_stable_sort(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&& __exec, _Range&& __rng,
+                       _Compare, _Proj __proj)
+{
+    return __parallel_radix_sort<__internal::__is_comp_ascending<::std::decay_t<_Compare>>::value>(
+        ::std::forward<_ExecutionPolicy>(__exec), ::std::forward<_Range>(__rng), __proj);
+}
 #endif
 
 template <typename _ExecutionPolicy, typename _Range, typename _Compare, typename _Proj,
@@ -2034,6 +2046,20 @@ __parallel_stable_sort(_ExecutionPolicy&& __exec, _Range&& __rng, _Compare __com
     return __parallel_sort_impl(::std::forward<_ExecutionPolicy>(__exec), ::std::forward<_Range>(__rng), __cmp_f);
 }
 
+template <typename _ExecutionPolicy, typename _Range, typename _Compare, typename _Proj,
+          ::std::enable_if_t<
+              !__is_radix_sort_usable_for_type<oneapi::dpl::__internal::__key_t<_Proj, _Range>, _Compare>::value,
+              int> = 0>
+auto
+__parallel_stable_sort(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&& __exec, _Range&& __rng,
+                       _Compare __comp, _Proj __proj)
+{
+    auto __cmp_f = [__comp, __proj](const auto& __a, const auto& __b) mutable {
+        return __comp(__proj(__a), __proj(__b));
+    };
+    return __parallel_sort_impl(::std::forward<_ExecutionPolicy>(__exec), ::std::forward<_Range>(__rng), __cmp_f);
+}
+
 //------------------------------------------------------------------------
 // parallel_partial_sort - async pattern
 //-----------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
index bc1d53d4cea..7f03369d7c4 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
@@ -339,6 +339,17 @@ __parallel_stable_sort(_ExecutionPolicy&& __exec, _Range&& __rng, _Compare __com
                                                                      ::std::forward<_Range>(__rng), __comp, __proj);
 }
 
+template <typename _ExecutionPolicy, typename _Range, typename _Compare, typename _Proj>
+auto
+__parallel_stable_sort(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionPolicy&& __exec, _Range&& __rng,
+                       _Compare __comp, _Proj __proj)
+{
+    // workaround until we implement more performant version for patterns
+    return oneapi::dpl::__par_backend_hetero::__parallel_stable_sort(oneapi::dpl::__internal::__device_backend_tag{},
+                                                                     __exec.__device_policy(),
+                                                                     ::std::forward<_Range>(__rng), __comp, __proj);
+}
+
 //------------------------------------------------------------------------
 // parallel_partial_sort
 //-----------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/omp/parallel_stable_sort.h b/include/oneapi/dpl/pstl/omp/parallel_stable_sort.h
index 14aa7b7bf04..5c05452a428 100644
--- a/include/oneapi/dpl/pstl/omp/parallel_stable_sort.h
+++ b/include/oneapi/dpl/pstl/omp/parallel_stable_sort.h
@@ -162,6 +162,49 @@ __parallel_stable_sort(_ExecutionPolicy&& /*__exec*/, _RandomAccessIterator __xs
     }
 }
 
+template <class _ExecutionPolicy, typename _RandomAccessIterator, typename _Compare, typename _LeafSort>
+void
+__parallel_stable_sort(oneapi::dpl::__internal::__omp_backend_tag, _ExecutionPolicy&& /*__exec*/,
+                       _RandomAccessIterator __xs, _RandomAccessIterator __xe, _Compare __comp, _LeafSort __leaf_sort,
+                       std::size_t __nsort = 0)
+{
+    auto __count = static_cast<std::size_t>(__xe - __xs);
+    if (__count <= __default_chunk_size || __nsort < __count)
+    {
+        __leaf_sort(__xs, __xe, __comp);
+        return;
+    }
+
+    // TODO: the partial sort implementation should
+    // be shared with the other backends.
+
+    if (omp_in_parallel())
+    {
+        if (__count <= __nsort)
+        {
+            oneapi::dpl::__omp_backend::__parallel_stable_sort_body(__xs, __xe, __comp, __leaf_sort);
+        }
+        else
+        {
+            oneapi::dpl::__omp_backend::__parallel_stable_partial_sort(__xs, __xe, __comp, __leaf_sort, __nsort);
+        }
+    }
+    else
+    {
+        _PSTL_PRAGMA(omp parallel)
+        _PSTL_PRAGMA(omp single nowait)
+        if (__count <= __nsort)
+        {
+            oneapi::dpl::__omp_backend::__parallel_stable_sort_body(__xs, __xe, __comp, __leaf_sort);
+        }
+        else
+        {
+            oneapi::dpl::__omp_backend::__parallel_stable_partial_sort(__xs, __xe, __comp, __leaf_sort, __nsort);
+        }
+    }
+}
+
+
 } // namespace __omp_backend
 } // namespace dpl
 } // namespace oneapi
diff --git a/include/oneapi/dpl/pstl/parallel_backend_serial.h b/include/oneapi/dpl/pstl/parallel_backend_serial.h
index 590201fd16d..ef9bdd7dfaf 100644
--- a/include/oneapi/dpl/pstl/parallel_backend_serial.h
+++ b/include/oneapi/dpl/pstl/parallel_backend_serial.h
@@ -174,6 +174,14 @@ __parallel_stable_sort(_ExecutionPolicy&&, _RandomAccessIterator __first, _Rando
     __leaf_sort(__first, __last, __comp);
 }
 
+template <class _ExecutionPolicy, typename _RandomAccessIterator, typename _Compare, typename _LeafSort>
+void
+__parallel_stable_sort(oneapi::dpl::__internal::__serial_backend_tag, _ExecutionPolicy&&, _RandomAccessIterator __first,
+                       _RandomAccessIterator __last, _Compare __comp, _LeafSort __leaf_sort, ::std::size_t = 0)
+{
+    __leaf_sort(__first, __last, __comp);
+}
+
 template <class _ExecutionPolicy, typename _RandomAccessIterator1, typename _RandomAccessIterator2,
           typename _RandomAccessIterator3, typename _Compare, typename _LeafMerge>
 void
diff --git a/include/oneapi/dpl/pstl/parallel_backend_tbb.h b/include/oneapi/dpl/pstl/parallel_backend_tbb.h
index 531858c9f67..a7c521ffe84 100644
--- a/include/oneapi/dpl/pstl/parallel_backend_tbb.h
+++ b/include/oneapi/dpl/pstl/parallel_backend_tbb.h
@@ -1296,6 +1296,31 @@ __parallel_stable_sort(_ExecutionPolicy&&, _RandomAccessIterator __xs, _RandomAc
     });
 }
 
+template <class _ExecutionPolicy, typename _RandomAccessIterator, typename _Compare, typename _LeafSort>
+void
+__parallel_stable_sort(oneapi::dpl::__internal::__tbb_backend_tag, _ExecutionPolicy&&, _RandomAccessIterator __xs,
+                       _RandomAccessIterator __xe, _Compare __comp, _LeafSort __leaf_sort, ::std::size_t __nsort)
+{
+    tbb::this_task_arena::isolate([=, &__nsort]() {
+        //sorting based on task tree and parallel merge
+        typedef typename ::std::iterator_traits<_RandomAccessIterator>::value_type _ValueType;
+        typedef typename ::std::iterator_traits<_RandomAccessIterator>::difference_type _DifferenceType;
+        const _DifferenceType __n = __xe - __xs;
+
+        const _DifferenceType __sort_cut_off = _ONEDPL_STABLE_SORT_CUT_OFF;
+        if (__n > __sort_cut_off)
+        {
+            __tbb_backend::__buffer<_ExecutionPolicy, _ValueType> __buf(__n);
+            __root_task<__stable_sort_func<_RandomAccessIterator, _ValueType*, _Compare, _LeafSort>> __root{
+                __xs, __xe, __buf.get(), true, __comp, __leaf_sort, __nsort, __xs, __buf.get()};
+            __task::spawn_root_and_wait(__root);
+            return;
+        }
+        //serial sort
+        __leaf_sort(__xs, __xe, __comp);
+    });
+}
+
 //------------------------------------------------------------------------
 // parallel_merge
 //------------------------------------------------------------------------

From fc62101e78787cd9bde3b58a67b745896a107816 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 14 Feb 2024 16:35:11 +0100
Subject: [PATCH 193/566] __parallel_stable_sort + tag calls

---
 .../dpl/internal/async_impl/glue_async_impl.h |  7 +++-
 include/oneapi/dpl/pstl/algorithm_impl.h      | 41 +++++++++++++++++++
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   |  8 +++-
 .../hetero/dpcpp/parallel_backend_sycl_fpga.h |  3 +-
 4 files changed, 54 insertions(+), 5 deletions(-)

diff --git a/include/oneapi/dpl/internal/async_impl/glue_async_impl.h b/include/oneapi/dpl/internal/async_impl/glue_async_impl.h
index 26eca467131..3cceec2f5cb 100644
--- a/include/oneapi/dpl/internal/async_impl/glue_async_impl.h
+++ b/include/oneapi/dpl/internal/async_impl/glue_async_impl.h
@@ -93,8 +93,11 @@ sort_async(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last, _Comp
     auto __keep = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read_write, _Iterator>();
     auto __buf = __keep(__first, __last);
 
-    return __par_backend_hetero::__parallel_stable_sort(::std::forward<_ExecutionPolicy>(__exec), __buf.all_view(),
-                                                        __comp, oneapi::dpl::identity{});
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
+    return __par_backend_hetero::__parallel_stable_sort(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
+                                                        __buf.all_view(), __comp, oneapi::dpl::identity{});
 }
 
 template <class _ExecutionPolicy, class _RandomAccessIterator, class... _Events,
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 7c722594610..260f1381d12 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -3733,8 +3733,13 @@ oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>
 __pattern_sort(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIterator __last, _Compare __comp,
                _IsVector /*is_vector*/, /*is_parallel=*/::std::true_type, /*is_move_constructible=*/::std::true_type)
 {
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     __internal::__except_handler([&]() {
         __par_backend::__parallel_stable_sort(
+            __backend_tag{},
             ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __comp,
             [](_RandomAccessIterator __first, _RandomAccessIterator __last, _Compare __comp) {
                 ::std::sort(__first, __last, __comp);
@@ -3749,8 +3754,11 @@ __pattern_sort(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAcce
                _RandomAccessIterator __last, _Compare __comp,
                /*is_move_constructible=*/::std::true_type)
 {
+    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
+
     __internal::__except_handler([&]() {
         __par_backend::__parallel_stable_sort(
+            __backend_tag{},
             ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __comp,
             [](_RandomAccessIterator __first, _RandomAccessIterator __last, _Compare __comp) {
                 ::std::sort(__first, __last, __comp);
@@ -3785,8 +3793,13 @@ oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>
 __pattern_stable_sort(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIterator __last,
                       _Compare __comp, _IsVector /*is_vector*/, /*is_parallel=*/::std::true_type)
 {
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     __internal::__except_handler([&]() {
         __par_backend::__parallel_stable_sort(
+            __backend_tag{},
             ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __comp,
             [](_RandomAccessIterator __first, _RandomAccessIterator __last, _Compare __comp) {
                 ::std::stable_sort(__first, __last, __comp);
@@ -3800,8 +3813,11 @@ void
 __pattern_stable_sort(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
                       _RandomAccessIterator __last, _Compare __comp)
 {
+    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
+
     __internal::__except_handler([&]() {
         __par_backend::__parallel_stable_sort(
+            __backend_tag{},
             ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __comp,
             [](_RandomAccessIterator __first, _RandomAccessIterator __last, _Compare __comp) {
                 ::std::stable_sort(__first, __last, __comp);
@@ -3866,8 +3882,13 @@ __pattern_sort_by_key(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __keys_f
         return __comp(::std::get<0>(__a), ::std::get<0>(__b));
     };
 
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     __internal::__except_handler([&]() {
         __par_backend::__parallel_stable_sort(
+            __backend_tag{},
             ::std::forward<_ExecutionPolicy>(__exec), __beg, __end, __cmp_f,
             [](auto __first, auto __last, auto __cmp) { ::std::sort(__first, __last, __cmp); }, __end - __beg);
     });
@@ -3890,8 +3911,11 @@ __pattern_sort_by_key(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Ran
         return __comp(::std::get<0>(__a), ::std::get<0>(__b));
     };
 
+    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
+
     __internal::__except_handler([&]() {
         __par_backend::__parallel_stable_sort(
+            __backend_tag{},
             ::std::forward<_ExecutionPolicy>(__exec), __beg, __end, __cmp_f,
             [](auto __first, auto __last, auto __cmp) { ::std::sort(__first, __last, __cmp); }, __end - __beg);
     });
@@ -3929,8 +3953,13 @@ __pattern_partial_sort(_ExecutionPolicy&& __exec, _RandomAccessIterator __first,
     if (__n == 0)
         return;
 
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     __except_handler([&]() {
         __par_backend::__parallel_stable_sort(
+            __backend_tag{},
             ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __comp,
             [__n](_RandomAccessIterator __begin, _RandomAccessIterator __end, _Compare __comp) {
                 if (__n < __end - __begin)
@@ -3951,8 +3980,11 @@ __pattern_partial_sort(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Ra
     if (__n == 0)
         return;
 
+    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
+
     __except_handler([&]() {
         __par_backend::__parallel_stable_sort(
+            __backend_tag{},
             ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __comp,
             [__n](_RandomAccessIterator __begin, _RandomAccessIterator __end, _Compare __comp) {
                 if (__n < __end - __begin)
@@ -3998,12 +4030,18 @@ __pattern_partial_sort_copy(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __
     {
         return __d_first;
     }
+
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     auto __n1 = __last - __first;
     auto __n2 = __d_last - __d_first;
     return __internal::__except_handler([&]() {
         if (__n2 >= __n1)
         {
             __par_backend::__parallel_stable_sort(
+                __backend_tag{},
                 ::std::forward<_ExecutionPolicy>(__exec), __d_first, __d_first + __n1, __comp,
                 [__first, __d_first, __is_vector](_RandomAccessIterator2 __i, _RandomAccessIterator2 __j,
                                                   _Compare __comp) {
@@ -4026,6 +4064,7 @@ __pattern_partial_sort_copy(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __
             _T1* __r = __buf.get();
 
             __par_backend::__parallel_stable_sort(
+                __backend_tag{},
                 ::std::forward<_ExecutionPolicy>(__exec), __r, __r + __n1, __comp,
                 [__n2, __first, __r](_T1* __i, _T1* __j, _Compare __comp) {
                     _RandomAccessIterator1 __it = __first + (__i - __r);
@@ -4080,6 +4119,7 @@ __pattern_partial_sort_copy(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec
         if (__n2 >= __n1)
         {
             __par_backend::__parallel_stable_sort(
+                __backend_tag{},
                 ::std::forward<_ExecutionPolicy>(__exec), __d_first, __d_first + __n1, __comp,
                 [__first, __d_first](_RandomAccessIterator2 __i, _RandomAccessIterator2 __j, _Compare __comp) {
                     _RandomAccessIterator1 __i1 = __first + (__i - __d_first);
@@ -4101,6 +4141,7 @@ __pattern_partial_sort_copy(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec
             _T1* __r = __buf.get();
 
             __par_backend::__parallel_stable_sort(
+                __backend_tag{},
                 ::std::forward<_ExecutionPolicy>(__exec), __r, __r + __n1, __comp,
                 [__n2, __first, __r](_T1* __i, _T1* __j, _Compare __comp) {
                     _RandomAccessIterator1 __it = __first + (__i - __r);
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 34707777978..e8e4856e3a1 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -2225,8 +2225,12 @@ __stable_sort_with_projection(_ExecutionPolicy&& __exec, _Iterator __first, _Ite
     auto __keep = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read_write, _Iterator>();
     auto __buf = __keep(__first, __last);
 
-    __par_backend_hetero::__parallel_stable_sort(
-        ::std::forward<_ExecutionPolicy>(__exec), __buf.all_view(), __comp, __proj).wait();
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
+    __par_backend_hetero::__parallel_stable_sort(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
+                                                 __buf.all_view(), __comp, __proj)
+        .wait();
 }
 
 template <typename _ExecutionPolicy, typename _Iterator, typename _Compare>
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
index 7f03369d7c4..0022ba737be 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
@@ -335,7 +335,8 @@ auto
 __parallel_stable_sort(_ExecutionPolicy&& __exec, _Range&& __rng, _Compare __comp, _Proj __proj)
 {
     // workaround until we implement more performant version for patterns
-    return oneapi::dpl::__par_backend_hetero::__parallel_stable_sort(__exec.__device_policy(),
+    return oneapi::dpl::__par_backend_hetero::__parallel_stable_sort(oneapi::dpl::__internal::__device_backend_tag{},
+                                                                     __exec.__device_policy(),
                                                                      ::std::forward<_Range>(__rng), __comp, __proj);
 }
 

From 579ea0266c22de170907e7148851f9e6ad26f470 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 14 Feb 2024 17:33:10 +0100
Subject: [PATCH 194/566] __pattern_histogram + tag impls

---
 include/oneapi/dpl/pstl/hetero/histogram_impl_hetero.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/include/oneapi/dpl/pstl/hetero/histogram_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/histogram_impl_hetero.h
index c594e5c25a0..491bc107951 100644
--- a/include/oneapi/dpl/pstl/hetero/histogram_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/histogram_impl_hetero.h
@@ -120,7 +120,7 @@ struct __hist_fill_zeros_wrapper
 template <typename _BackendTag, typename _ExecutionPolicy, typename _RandomAccessIterator1, typename _Size,
           typename _BinHash, typename _RandomAccessIterator2>
 void
-__pattern_histogram(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _RandomAccessIterator1 __first,
+__pattern_histogram(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _RandomAccessIterator1 __first,
                     _RandomAccessIterator1 __last, _Size __num_bins, _BinHash&& __func,
                     _RandomAccessIterator2 __histogram_first)
 {

From 1a897d26bb8ab650c41bece91fc9641863109260 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 14 Feb 2024 17:34:21 +0100
Subject: [PATCH 195/566] __pattern_rotate + tag impls

---
 include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index e8e4856e3a1..5df17e0e396 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -2982,7 +2982,7 @@ __pattern_rotate(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __new_f
 
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator>
 _Iterator
-__pattern_rotate(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __new_first,
+__pattern_rotate(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __new_first,
                  _Iterator __last)
 {
     auto __n = __last - __first;

From 4733da416a2750a68b376cd1be5314f3f07d56b6 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 14 Feb 2024 17:34:36 +0100
Subject: [PATCH 196/566] __pattern_rotate_copy + tag impls

---
 include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 5df17e0e396..1a88c92a564 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -3053,7 +3053,7 @@ __pattern_rotate_copy(_ExecutionPolicy&& __exec, _BidirectionalIterator __first,
 
 template <typename _BackendTag, typename _ExecutionPolicy, typename _BidirectionalIterator, typename _ForwardIterator>
 _ForwardIterator
-__pattern_rotate_copy(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _BidirectionalIterator __first,
+__pattern_rotate_copy(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _BidirectionalIterator __first,
                       _BidirectionalIterator __new_first, _BidirectionalIterator __last, _ForwardIterator __result)
 {
     auto __n = __last - __first;

From 2c66c75c8a5321acdf9bdab49e024ce8709d5ef2 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 14 Feb 2024 17:48:41 +0100
Subject: [PATCH 197/566] __parallel_merge + tag impls

---
 .../pstl/hetero/dpcpp/parallel_backend_sycl.h | 28 +++++++++++++++++
 .../hetero/dpcpp/parallel_backend_sycl_fpga.h | 14 +++++++++
 include/oneapi/dpl/pstl/omp/parallel_merge.h  | 30 +++++++++++++++++++
 .../oneapi/dpl/pstl/parallel_backend_serial.h | 10 +++++++
 .../oneapi/dpl/pstl/parallel_backend_tbb.h    | 30 +++++++++++++++++++
 5 files changed, 112 insertions(+)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
index c3e2fe8b483..5033f84a629 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
@@ -1741,6 +1741,34 @@ __parallel_merge(_ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&& __rng2,
     }
 }
 
+template <typename _ExecutionPolicy, typename _Range1, typename _Range2, typename _Range3, typename _Compare>
+auto
+__parallel_merge(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&& __exec, _Range1&& __rng1,
+                 _Range2&& __rng2, _Range3&& __rng3, _Compare __comp)
+{
+    using _CustomName = oneapi::dpl::__internal::__policy_kernel_name<_ExecutionPolicy>;
+
+    const auto __n = __rng1.size() + __rng2.size();
+    if (__n <= std::numeric_limits<::std::uint32_t>::max())
+    {
+        using _wi_index_type = ::std::uint32_t;
+        using _MergeKernel = oneapi::dpl::__par_backend_hetero::__internal::__kernel_name_provider<
+            __merge_kernel_name<_CustomName, _wi_index_type>>;
+        return __parallel_merge_submitter<_wi_index_type, _MergeKernel>()(
+            ::std::forward<_ExecutionPolicy>(__exec), ::std::forward<_Range1>(__rng1), ::std::forward<_Range2>(__rng2),
+            ::std::forward<_Range3>(__rng3), __comp);
+    }
+    else
+    {
+        using _wi_index_type = ::std::uint64_t;
+        using _MergeKernel = oneapi::dpl::__par_backend_hetero::__internal::__kernel_name_provider<
+            __merge_kernel_name<_CustomName, _wi_index_type>>;
+        return __parallel_merge_submitter<_wi_index_type, _MergeKernel>()(
+            ::std::forward<_ExecutionPolicy>(__exec), ::std::forward<_Range1>(__rng1), ::std::forward<_Range2>(__rng2),
+            ::std::forward<_Range3>(__rng3), __comp);
+    }
+}
+
 //-----------------------------------------------------------------------
 // parallel_sort: general implementation
 //-----------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
index 0022ba737be..32696ac7ea8 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
@@ -325,6 +325,20 @@ __parallel_merge(_ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&& __rng2,
         ::std::forward<_Range3>(__rng3), __comp);
 }
 
+template <typename _ExecutionPolicy, typename _Range1, typename _Range2, typename _Range3, typename _Compare>
+auto
+__parallel_merge(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionPolicy&& __exec, _Range1&& __rng1,
+                 _Range2&& __rng2, _Range3&& __rng3, _Compare __comp)
+    -> decltype(oneapi::dpl::__par_backend_hetero::__parallel_merge(
+        oneapi::dpl::__internal::__device_backend_tag{}, __exec.__device_policy(), ::std::forward<_Range1>(__rng1),
+        ::std::forward<_Range2>(__rng2), ::std::forward<_Range3>(__rng3), __comp))
+{
+    // workaround until we implement more performant version for patterns
+    return oneapi::dpl::__par_backend_hetero::__parallel_merge(
+        oneapi::dpl::__internal::__device_backend_tag{}, __exec.__device_policy(), ::std::forward<_Range1>(__rng1),
+        ::std::forward<_Range2>(__rng2), ::std::forward<_Range3>(__rng3), __comp);
+}
+
 //------------------------------------------------------------------------
 // parallel_stable_sort
 //-----------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/omp/parallel_merge.h b/include/oneapi/dpl/pstl/omp/parallel_merge.h
index 911d4b2643b..a708a502bd4 100644
--- a/include/oneapi/dpl/pstl/omp/parallel_merge.h
+++ b/include/oneapi/dpl/pstl/omp/parallel_merge.h
@@ -74,7 +74,37 @@ void
 __parallel_merge(_ExecutionPolicy&& /*__exec*/, _RandomAccessIterator1 __xs, _RandomAccessIterator1 __xe,
                  _RandomAccessIterator2 __ys, _RandomAccessIterator2 __ye, _RandomAccessIterator3 __zs, _Compare __comp,
                  _LeafMerge __leaf_merge)
+{
+    std::size_t __size_x = __xe - __xs;
+    std::size_t __size_y = __ye - __ys;
+
+    /*
+     * Run the merge in parallel by chunking it up. Use the smaller range (if any) as the iteration range, and the
+     * larger range as the search range.
+     */
 
+    if (omp_in_parallel())
+    {
+        oneapi::dpl::__omp_backend::__parallel_merge_body(__size_x, __size_y, __xs, __xe, __ys, __ye, __zs, __comp,
+                                                          __leaf_merge);
+    }
+    else
+    {
+        _PSTL_PRAGMA(omp parallel)
+        {
+            _PSTL_PRAGMA(omp single nowait)
+            oneapi::dpl::__omp_backend::__parallel_merge_body(__size_x, __size_y, __xs, __xe, __ys, __ye, __zs, __comp,
+                                                              __leaf_merge);
+        }
+    }
+}
+
+template <class _ExecutionPolicy, typename _RandomAccessIterator1, typename _RandomAccessIterator2,
+          typename _RandomAccessIterator3, typename _Compare, typename _LeafMerge>
+void
+__parallel_merge(oneapi::dpl::__internal::__omp_backend_tag, _ExecutionPolicy&& /*__exec*/,
+                 _RandomAccessIterator1 __xs, _RandomAccessIterator1 __xe, _RandomAccessIterator2 __ys,
+                 _RandomAccessIterator2 __ye, _RandomAccessIterator3 __zs, _Compare __comp, _LeafMerge __leaf_merge)
 {
     std::size_t __size_x = __xe - __xs;
     std::size_t __size_y = __ye - __ys;
diff --git a/include/oneapi/dpl/pstl/parallel_backend_serial.h b/include/oneapi/dpl/pstl/parallel_backend_serial.h
index ef9bdd7dfaf..6f69ef11416 100644
--- a/include/oneapi/dpl/pstl/parallel_backend_serial.h
+++ b/include/oneapi/dpl/pstl/parallel_backend_serial.h
@@ -192,6 +192,16 @@ __parallel_merge(_ExecutionPolicy&&, _RandomAccessIterator1 __first1, _RandomAcc
     __leaf_merge(__first1, __last1, __first2, __last2, __outit, __comp);
 }
 
+template <class _ExecutionPolicy, typename _RandomAccessIterator1, typename _RandomAccessIterator2,
+          typename _RandomAccessIterator3, typename _Compare, typename _LeafMerge>
+void
+__parallel_merge(oneapi::dpl::__internal::__serial_backend_tag, _ExecutionPolicy&&, _RandomAccessIterator1 __first1,
+                 _RandomAccessIterator1 __last1, _RandomAccessIterator2 __first2, _RandomAccessIterator2 __last2,
+                 _RandomAccessIterator3 __outit, _Compare __comp, _LeafMerge __leaf_merge)
+{
+    __leaf_merge(__first1, __last1, __first2, __last2, __outit, __comp);
+}
+
 template <class _ExecutionPolicy, typename _F1, typename _F2>
 void
 __parallel_invoke(_ExecutionPolicy&&, _F1&& __f1, _F2&& __f2)
diff --git a/include/oneapi/dpl/pstl/parallel_backend_tbb.h b/include/oneapi/dpl/pstl/parallel_backend_tbb.h
index a7c521ffe84..bc62641c9f7 100644
--- a/include/oneapi/dpl/pstl/parallel_backend_tbb.h
+++ b/include/oneapi/dpl/pstl/parallel_backend_tbb.h
@@ -1416,6 +1416,36 @@ __parallel_merge(_ExecutionPolicy&&, _RandomAccessIterator1 __xs, _RandomAccessI
     }
 }
 
+template <class _ExecutionPolicy, typename _RandomAccessIterator1, typename _RandomAccessIterator2,
+          typename _RandomAccessIterator3, typename _Compare, typename _LeafMerge>
+void
+__parallel_merge(oneapi::dpl::__internal::__tbb_backend_tag, _ExecutionPolicy&&, _RandomAccessIterator1 __xs,
+                 _RandomAccessIterator1 __xe,
+                 _RandomAccessIterator2 __ys, _RandomAccessIterator2 __ye, _RandomAccessIterator3 __zs, _Compare __comp,
+                 _LeafMerge __leaf_merge)
+{
+    typedef typename ::std::iterator_traits<_RandomAccessIterator1>::difference_type _DifferenceType1;
+    typedef typename ::std::iterator_traits<_RandomAccessIterator2>::difference_type _DifferenceType2;
+    typedef typename ::std::common_type_t<_DifferenceType1, _DifferenceType2> _SizeType;
+    const _SizeType __n = (__xe - __xs) + (__ye - __ys);
+    const _SizeType __merge_cut_off = _ONEDPL_MERGE_CUT_OFF;
+    if (__n <= __merge_cut_off)
+    {
+        // Fall back on serial merge
+        __leaf_merge(__xs, __xe, __ys, __ye, __zs, __comp);
+    }
+    else
+    {
+        tbb::this_task_arena::isolate([=]() {
+            typedef __merge_func_static<_RandomAccessIterator1, _RandomAccessIterator2, _RandomAccessIterator3,
+                                        _Compare, _LeafMerge>
+                _TaskType;
+            __root_task<_TaskType> __root{__xs, __xe, __ys, __ye, __zs, __comp, __leaf_merge};
+            __task::spawn_root_and_wait(__root);
+        });
+    }
+}
+
 //------------------------------------------------------------------------
 // parallel_invoke
 //------------------------------------------------------------------------

From cd4017313ff692e720900a1f6944d5d371191a5c Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 14 Feb 2024 17:49:47 +0100
Subject: [PATCH 198/566] __parallel_merge + tag calls

---
 include/oneapi/dpl/pstl/algorithm_impl.h              | 11 +++++++++++
 .../oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h    |  5 +++--
 2 files changed, 14 insertions(+), 2 deletions(-)

diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 260f1381d12..cd4d4eea729 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -4825,7 +4825,13 @@ __pattern_merge(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1, _Ran
                 _RandomAccessIterator2 __first2, _RandomAccessIterator2 __last2, _RandomAccessIterator3 __d_first,
                 _Compare __comp, _IsVector __is_vector, /* is_parallel = */ ::std::true_type)
 {
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2,
+                                                  _RandomAccessIterator3>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     __par_backend::__parallel_merge(
+        __backend_tag{},
         ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2, __d_first, __comp,
         [__is_vector](_RandomAccessIterator1 __f1, _RandomAccessIterator1 __l1, _RandomAccessIterator2 __f2,
                       _RandomAccessIterator2 __l2, _RandomAccessIterator3 __f3, _Compare __comp) {
@@ -4841,7 +4847,10 @@ __pattern_merge(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAcc
                 _RandomAccessIterator1 __last1, _RandomAccessIterator2 __first2, _RandomAccessIterator2 __last2,
                 _RandomAccessIterator3 __d_first, _Compare __comp)
 {
+    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
+
     __par_backend::__parallel_merge(
+        __backend_tag{},
         ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2, __d_first, __comp,
         [](_RandomAccessIterator1 __f1, _RandomAccessIterator1 __l1, _RandomAccessIterator2 __f2,
            _RandomAccessIterator2 __l2, _RandomAccessIterator3 __f3,
@@ -4920,6 +4929,7 @@ __pattern_inplace_merge(_ExecutionPolicy&& __exec, _RandomAccessIterator __first
         };
 
         __par_backend::__parallel_merge(
+            __backend_tag{},
             ::std::forward<_ExecutionPolicy>(__exec), __first, __middle, __middle, __last, __r, __comp,
             [__n, __move_values, __move_sequences](_RandomAccessIterator __f1, _RandomAccessIterator __l1,
                                                    _RandomAccessIterator __f2, _RandomAccessIterator __l2, _Tp* __f3,
@@ -4965,6 +4975,7 @@ __pattern_inplace_merge(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _R
         };
 
         __par_backend::__parallel_merge(
+            __backend_tag{},
             ::std::forward<_ExecutionPolicy>(__exec), __first, __middle, __middle, __last, __r, __comp,
             [__n, __move_values, __move_sequences](_RandomAccessIterator __f1, _RandomAccessIterator __l1,
                                                    _RandomAccessIterator __f2, _RandomAccessIterator __l2, _Tp* __f3,
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 1a88c92a564..1e1b026f34a 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -2075,6 +2075,7 @@ __pattern_merge(_ExecutionPolicy&& __exec, _Iterator1 __first1, _Iterator1 __las
         return __d_first;
 
     const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator1, _Iterator2, _Iterator3>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
 
     //To consider the direct copying pattern call in case just one of sequences is empty.
     if (__n1 == 0)
@@ -2099,7 +2100,7 @@ __pattern_merge(_ExecutionPolicy&& __exec, _Iterator1 __first1, _Iterator1 __las
         auto __keep3 = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::write, _Iterator3>();
         auto __buf3 = __keep3(__d_first, __d_first + __n);
 
-        __par_backend_hetero::__parallel_merge(::std::forward<_ExecutionPolicy>(__exec), __buf1.all_view(),
+        __par_backend_hetero::__parallel_merge(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __buf1.all_view(),
                                                __buf2.all_view(), __buf3.all_view(), __comp)
             .wait();
     }
@@ -2141,7 +2142,7 @@ __pattern_merge(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Ite
         auto __keep3 = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::write, _Iterator3>();
         auto __buf3 = __keep3(__d_first, __d_first + __n);
 
-        __par_backend_hetero::__parallel_merge(::std::forward<_ExecutionPolicy>(__exec), __buf1.all_view(),
+        __par_backend_hetero::__parallel_merge(_BackendTag{}, ::std::forward<_ExecutionPolicy>(__exec), __buf1.all_view(),
                                                __buf2.all_view(), __buf3.all_view(), __comp)
             .wait();
     }

From 3907f9960a789f56c2ef85901800a6bbe04ded9c Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 14 Feb 2024 18:01:39 +0100
Subject: [PATCH 199/566] __parallel_invoke + tag impls

---
 include/oneapi/dpl/pstl/omp/parallel_invoke.h    | 16 ++++++++++++++++
 .../oneapi/dpl/pstl/parallel_backend_serial.h    |  8 ++++++++
 include/oneapi/dpl/pstl/parallel_backend_tbb.h   |  9 +++++++++
 3 files changed, 33 insertions(+)

diff --git a/include/oneapi/dpl/pstl/omp/parallel_invoke.h b/include/oneapi/dpl/pstl/omp/parallel_invoke.h
index 32491ab9dfd..2520526a191 100644
--- a/include/oneapi/dpl/pstl/omp/parallel_invoke.h
+++ b/include/oneapi/dpl/pstl/omp/parallel_invoke.h
@@ -52,6 +52,22 @@ __parallel_invoke(_ExecutionPolicy&&, _F1&& __f1, _F2&& __f2)
     }
 }
 
+template <class _ExecutionPolicy, typename _F1, typename _F2>
+void
+__parallel_invoke(oneapi::dpl::__internal::__omp_backend_tag, _ExecutionPolicy&&, _F1&& __f1, _F2&& __f2)
+{
+    if (omp_in_parallel())
+    {
+        oneapi::dpl::__omp_backend::__parallel_invoke_body(std::forward<_F1>(__f1), std::forward<_F2>(__f2));
+    }
+    else
+    {
+        _PSTL_PRAGMA(omp parallel)
+        _PSTL_PRAGMA(omp single nowait)
+        oneapi::dpl::__omp_backend::__parallel_invoke_body(std::forward<_F1>(__f1), std::forward<_F2>(__f2));
+    }
+}
+
 } // namespace __omp_backend
 } // namespace dpl
 } // namespace oneapi
diff --git a/include/oneapi/dpl/pstl/parallel_backend_serial.h b/include/oneapi/dpl/pstl/parallel_backend_serial.h
index 6f69ef11416..264947fe550 100644
--- a/include/oneapi/dpl/pstl/parallel_backend_serial.h
+++ b/include/oneapi/dpl/pstl/parallel_backend_serial.h
@@ -210,6 +210,14 @@ __parallel_invoke(_ExecutionPolicy&&, _F1&& __f1, _F2&& __f2)
     ::std::forward<_F2>(__f2)();
 }
 
+template <class _ExecutionPolicy, typename _F1, typename _F2>
+void
+__parallel_invoke(oneapi::dpl::__internal::__serial_backend_tag, _ExecutionPolicy&&, _F1&& __f1, _F2&& __f2)
+{
+    ::std::forward<_F1>(__f1)();
+    ::std::forward<_F2>(__f2)();
+}
+
 template <class _ExecutionPolicy, class _ForwardIterator, class _Fp>
 void
 __parallel_for_each(_ExecutionPolicy&&, _ForwardIterator __begin, _ForwardIterator __end, _Fp __f)
diff --git a/include/oneapi/dpl/pstl/parallel_backend_tbb.h b/include/oneapi/dpl/pstl/parallel_backend_tbb.h
index bc62641c9f7..39caf6cbc21 100644
--- a/include/oneapi/dpl/pstl/parallel_backend_tbb.h
+++ b/include/oneapi/dpl/pstl/parallel_backend_tbb.h
@@ -1458,6 +1458,15 @@ __parallel_invoke(_ExecutionPolicy&&, _F1&& __f1, _F2&& __f2)
         [&]() { tbb::parallel_invoke(::std::forward<_F1>(__f1), ::std::forward<_F2>(__f2)); });
 }
 
+template <class _ExecutionPolicy, typename _F1, typename _F2>
+void
+__parallel_invoke(oneapi::dpl::__internal::__tbb_backend_tag, _ExecutionPolicy&&, _F1&& __f1, _F2&& __f2)
+{
+    //TODO: a version of tbb::this_task_arena::isolate with variadic arguments pack should be added in the future
+    tbb::this_task_arena::isolate(
+        [&]() { tbb::parallel_invoke(::std::forward<_F1>(__f1), ::std::forward<_F2>(__f2)); });
+}
+
 //------------------------------------------------------------------------
 // parallel_for_each
 //------------------------------------------------------------------------

From 9eeadbceb5878dd6a5ab5a912916523a7a2902c1 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 14 Feb 2024 18:01:56 +0100
Subject: [PATCH 200/566] __parallel_invoke + tag calls

---
 include/oneapi/dpl/pstl/algorithm_impl.h | 9 +++++----
 1 file changed, 5 insertions(+), 4 deletions(-)

diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index cd4d4eea729..26612aa8ae8 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -5245,6 +5245,7 @@ __parallel_set_union_op(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _
     typedef typename ::std::iterator_traits<_ForwardIterator1>::difference_type _DifferenceType;
 
     constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2, _OutputIterator>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
 
     const auto __n1 = __last1 - __first1;
     const auto __n2 = __last2 - __first2;
@@ -5268,7 +5269,7 @@ __parallel_set_union_op(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _
     {
         //{1} < {2}: seq2 is wholly greater than seq1, so, do parallel copying seq1 and seq2
         __par_backend::__parallel_invoke(
-            ::std::forward<_ExecutionPolicy>(__exec),
+            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
             [=] {
                 __internal::__pattern_walk2_brick(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first1,
                                                   __last1, __result, __copy_range);
@@ -5287,7 +5288,7 @@ __parallel_set_union_op(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _
     {
         //{2} < {1}: seq2 is wholly greater than seq1, so, do parallel copying seq1 and seq2
         __par_backend::__parallel_invoke(
-            ::std::forward<_ExecutionPolicy>(__exec),
+            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
             [=] {
                 __internal::__pattern_walk2_brick(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first2,
                                                   __last2, __result, __copy_range);
@@ -5305,7 +5306,7 @@ __parallel_set_union_op(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _
         auto __res_or = __result;
         __result += __m1; //we know proper offset due to [first1; left_bound_seq_1) < [first2; last2)
         __par_backend::__parallel_invoke(
-            ::std::forward<_ExecutionPolicy>(__exec),
+            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
             //do parallel copying of [first1; left_bound_seq_1)
             [=] {
                 __internal::__pattern_walk2_brick(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first1,
@@ -5327,7 +5328,7 @@ __parallel_set_union_op(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _
         auto __res_or = __result;
         __result += __m2; //we know proper offset due to [first2; left_bound_seq_2) < [first1; last1)
         __par_backend::__parallel_invoke(
-            ::std::forward<_ExecutionPolicy>(__exec),
+            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
             //do parallel copying of [first2; left_bound_seq_2)
             [=] {
                 __internal::__pattern_walk2_brick(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first2,

From 1222137d6b6219a00d44020034377850e7e7799c Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 15 Feb 2024 09:43:51 +0100
Subject: [PATCH 201/566] __parallel_partial_sort + tag impls

---
 .../pstl/hetero/dpcpp/parallel_backend_sycl.h    | 16 ++++++++++++++++
 .../hetero/dpcpp/parallel_backend_sycl_fpga.h    | 10 ++++++++++
 2 files changed, 26 insertions(+)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
index 5033f84a629..f0283088c2b 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
@@ -2110,6 +2110,22 @@ __parallel_partial_sort(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator
                                         __partial_merge_kernel<decltype(__mid_idx)>{__mid_idx}, __comp);
 }
 
+// TODO: check if it makes sense to move these wrappers out of backend to a common place
+// TODO: consider changing __partial_merge_kernel to make it compatible with
+//       __full_merge_kernel in order to use __parallel_sort_impl routine
+template <typename _ExecutionPolicy, typename _Iterator, typename _Compare>
+auto
+__parallel_partial_sort(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&& __exec, _Iterator __first,
+                        _Iterator __mid, _Iterator __last, _Compare __comp)
+{
+    const auto __mid_idx = __mid - __first;
+
+    auto __keep = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read_write, _Iterator>();
+    auto __buf = __keep(__first, __last);
+
+    return __parallel_partial_sort_impl(::std::forward<_ExecutionPolicy>(__exec), __buf.all_view(),
+                                        __partial_merge_kernel<decltype(__mid_idx)>{__mid_idx}, __comp);
+}
 } // namespace __par_backend_hetero
 } // namespace dpl
 } // namespace oneapi
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
index 32696ac7ea8..a82bb3df858 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
@@ -381,6 +381,16 @@ __parallel_partial_sort(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator
                                                                       __comp);
 }
 
+template <typename _ExecutionPolicy, typename _Iterator, typename _Compare>
+auto
+__parallel_partial_sort(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionPolicy&& __exec, _Iterator __first,
+                        _Iterator __mid, _Iterator __last, _Compare __comp)
+{
+    // workaround until we implement more performant version for patterns
+    return oneapi::dpl::__par_backend_hetero::__parallel_partial_sort(
+        oneapi::dpl::__internal::__device_backend_tag{}, __exec.__device_policy(), __first, __mid, __last, __comp);
+}
+
 //------------------------------------------------------------------------
 // parallel_histogram
 //-----------------------------------------------------------------------

From 3c699f7ce2a291c31fa819a311f1838b7795f394 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 15 Feb 2024 09:46:18 +0100
Subject: [PATCH 202/566] __parallel_partial_sort + tag calls

---
 .../dpl/pstl/hetero/algorithm_impl_hetero.h     | 17 +++++++++++++++++
 .../hetero/dpcpp/parallel_backend_sycl_fpga.h   |  4 ++--
 2 files changed, 19 insertions(+), 2 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 1e1b026f34a..b26d51e8958 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -2591,7 +2591,11 @@ __pattern_partial_sort(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator _
     if (__last - __first < 2)
         return;
 
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     __par_backend_hetero::__parallel_partial_sort(
+        __backend_tag{},
         ::std::forward<_ExecutionPolicy>(__exec),
         __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read_write>(__first),
         __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read_write>(__mid),
@@ -2608,6 +2612,7 @@ __pattern_partial_sort(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exe
         return;
 
     __par_backend_hetero::__parallel_partial_sort(
+        _BackendTag{},
         ::std::forward<_ExecutionPolicy>(__exec),
         __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read_write>(__first),
         __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read_write>(__mid),
@@ -2707,7 +2712,13 @@ __pattern_partial_sort_copy(_ExecutionPolicy&& __exec, _InIterator __first, _InI
 
         auto __buf_mid = __buf_first + __out_size;
 
+        constexpr auto __dispatch_tag11 =
+            oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, decltype(__buf_first), decltype(__buf_mid),
+                                                      decltype(__buf_last)>();
+        using __backend_tag11 = typename decltype(__dispatch_tag11)::__backend_tag;
+
         __par_backend_hetero::__parallel_partial_sort(
+            __backend_tag11{},
             __par_backend_hetero::make_wrapped_policy<__partial_sort_2>(__exec),
             __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read_write>(__buf_first),
             __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read_write>(__buf_mid),
@@ -2780,7 +2791,13 @@ __pattern_partial_sort_copy(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&&
 
         auto __buf_mid = __buf_first + __out_size;
 
+        constexpr auto __dispatch_tag11 =
+            oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, decltype(__buf_first), decltype(__buf_mid),
+                                                      decltype(__buf_last)>();
+        using __backend_tag11 = typename decltype(__dispatch_tag11)::__backend_tag;
+
         __par_backend_hetero::__parallel_partial_sort(
+            __backend_tag11{},
             __par_backend_hetero::make_wrapped_policy<__partial_sort_2>(__exec),
             __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read_write>(__buf_first),
             __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read_write>(__buf_mid),
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
index a82bb3df858..cb7de2993fc 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
@@ -377,8 +377,8 @@ __parallel_partial_sort(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator
                         _Compare __comp)
 {
     // workaround until we implement more performant version for patterns
-    return oneapi::dpl::__par_backend_hetero::__parallel_partial_sort(__exec.__device_policy(), __first, __mid, __last,
-                                                                      __comp);
+    return oneapi::dpl::__par_backend_hetero::__parallel_partial_sort(
+        oneapi::dpl::__internal::__device_backend_tag{}, __exec.__device_policy(), __first, __mid, __last, __comp);
 }
 
 template <typename _ExecutionPolicy, typename _Iterator, typename _Compare>

From e5c113ef4b1d86f6e69adfeae5ccf65c5283385a Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 15 Feb 2024 09:47:31 +0100
Subject: [PATCH 203/566] __pattern_partial_sort + tag impls

---
 include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index b26d51e8958..970f6e12f3c 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -2604,7 +2604,7 @@ __pattern_partial_sort(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator _
 }
 
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator, typename _Compare>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy>
+void
 __pattern_partial_sort(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __mid,
                        _Iterator __last, _Compare __comp)
 {

From 47a3f5cb3ec03743c51c04a171d4fd83c216989b Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 15 Feb 2024 10:10:44 +0100
Subject: [PATCH 204/566] __parallel_scan_copy + tag impls

---
 .../pstl/hetero/dpcpp/parallel_backend_sycl.h | 40 +++++++++++++++++++
 .../hetero/dpcpp/parallel_backend_sycl_fpga.h | 14 +++++++
 2 files changed, 54 insertions(+)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
index f0283088c2b..876fc496bb6 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
@@ -984,6 +984,46 @@ __parallel_scan_copy(_ExecutionPolicy&& __exec, _InRng&& __in_rng, _OutRng&& __o
         __copy_by_mask_op);
 }
 
+template <typename _ExecutionPolicy, typename _InRng, typename _OutRng, typename _Size, typename _CreateMaskOp,
+          typename _CopyByMaskOp>
+auto
+__parallel_scan_copy(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&& __exec, _InRng&& __in_rng,
+                     _OutRng&& __out_rng, _Size __n, _CreateMaskOp __create_mask_op, _CopyByMaskOp __copy_by_mask_op)
+{
+    using _ReduceOp = ::std::plus<_Size>;
+    using _Assigner = unseq_backend::__scan_assigner;
+    using _NoAssign = unseq_backend::__scan_no_assign;
+    using _MaskAssigner = unseq_backend::__mask_assigner<1>;
+    using _DataAcc = unseq_backend::walk_n<_ExecutionPolicy, oneapi::dpl::__internal::__no_op>;
+    using _InitType = unseq_backend::__no_init_value<_Size>;
+
+    _Assigner __assign_op;
+    _ReduceOp __reduce_op;
+    _DataAcc __get_data_op;
+    _MaskAssigner __add_mask_op;
+
+    // temporary buffer to store boolean mask
+    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, int32_t> __mask_buf(__exec, __n);
+
+    return __parallel_transform_scan_base(
+        ::std::forward<_ExecutionPolicy>(__exec),
+        oneapi::dpl::__ranges::make_zip_view(
+            ::std::forward<_InRng>(__in_rng),
+            oneapi::dpl::__ranges::all_view<int32_t, __par_backend_hetero::access_mode::read_write>(
+                __mask_buf.get_buffer())),
+        ::std::forward<_OutRng>(__out_rng), __reduce_op, _InitType{},
+        // local scan
+        unseq_backend::__scan</*inclusive*/ ::std::true_type, _ExecutionPolicy, _ReduceOp, _DataAcc, _Assigner,
+                              _MaskAssigner, _CreateMaskOp, _InitType>{__reduce_op, __get_data_op, __assign_op,
+                                                                       __add_mask_op, __create_mask_op},
+        // scan between groups
+        unseq_backend::__scan</*inclusive*/ ::std::true_type, _ExecutionPolicy, _ReduceOp, _DataAcc, _NoAssign,
+                              _Assigner, _DataAcc, _InitType>{__reduce_op, __get_data_op, _NoAssign{}, __assign_op,
+                                                              __get_data_op},
+        // global scan
+        __copy_by_mask_op);
+}
+
 template <typename _ExecutionPolicy, typename _InRng, typename _OutRng, typename _Size, typename _Pred,
           oneapi::dpl::__internal::__enable_if_device_execution_policy<_ExecutionPolicy, int> = 0>
 auto
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
index cb7de2993fc..78273a74876 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
@@ -224,6 +224,20 @@ __parallel_scan_copy(_ExecutionPolicy&& __exec, _InRng&& __in_rng, _OutRng&& __o
         __create_mask_op, __copy_by_mask_op);
 }
 
+template <typename _ExecutionPolicy, typename _InRng, typename _OutRng, typename _Size, typename _CreateMaskOp,
+          typename _CopyByMaskOp>
+auto
+__parallel_scan_copy(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionPolicy&& __exec, _InRng&& __in_rng,
+                     _OutRng&& __out_rng, _Size __n,
+                     _CreateMaskOp __create_mask_op, _CopyByMaskOp __copy_by_mask_op)
+{
+    // workaround until we implement more performant version for patterns
+    return oneapi::dpl::__par_backend_hetero::__parallel_scan_copy(
+        oneapi::dpl::__internal::__device_backend_tag{},
+        __exec.__device_policy(), ::std::forward<_InRng>(__in_rng), ::std::forward<_OutRng>(__out_rng), __n,
+        __create_mask_op, __copy_by_mask_op);
+}
+
 //------------------------------------------------------------------------
 // __parallel_find_or
 //-----------------------------------------------------------------------

From 480a31594b1d84884f9a18624df405afdfc136dc Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 15 Feb 2024 10:11:22 +0100
Subject: [PATCH 205/566] __parallel_scan_copy + tag calls

---
 .../oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h | 14 +++++++++-----
 .../dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h  |  3 ++-
 .../pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h |  1 +
 3 files changed, 12 insertions(+), 6 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 970f6e12f3c..ac5c5984511 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -1574,6 +1574,10 @@ oneapi::dpl::__internal::__enable_if_hetero_execution_policy<
 __pattern_scan_copy(_ExecutionPolicy&& __exec, _Iterator1 __first, _Iterator1 __last, _IteratorOrTuple __output_first,
                     _CreateMaskOp __create_mask_op, _CopyByMaskOp __copy_by_mask_op)
 {
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator1, _IteratorOrTuple>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     using _It1DifferenceType = typename ::std::iterator_traits<_Iterator1>::difference_type;
 
     if (__first == __last)
@@ -1587,8 +1591,8 @@ __pattern_scan_copy(_ExecutionPolicy&& __exec, _Iterator1 __first, _Iterator1 __
         oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::write, _IteratorOrTuple>();
     auto __buf2 = __keep2(__output_first, __output_first + __n);
 
-    auto __res =
-        __par_backend_hetero::__parallel_scan_copy(::std::forward<_ExecutionPolicy>(__exec), __buf1.all_view(),
+    auto __res = __par_backend_hetero::__parallel_scan_copy(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
+                                                            __buf1.all_view(),
                                                    __buf2.all_view(), __n, __create_mask_op, __copy_by_mask_op);
 
     ::std::size_t __num_copied = __res.get();
@@ -1614,9 +1618,9 @@ __pattern_scan_copy(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Itera
         oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::write, _IteratorOrTuple>();
     auto __buf2 = __keep2(__output_first, __output_first + __n);
 
-    auto __res =
-        __par_backend_hetero::__parallel_scan_copy(::std::forward<_ExecutionPolicy>(__exec), __buf1.all_view(),
-                                                   __buf2.all_view(), __n, __create_mask_op, __copy_by_mask_op);
+    auto __res = __par_backend_hetero::__parallel_scan_copy(_BackendTag{}, ::std::forward<_ExecutionPolicy>(__exec),
+                                                            __buf1.all_view(), __buf2.all_view(), __n, __create_mask_op,
+                                                            __copy_by_mask_op);
 
     ::std::size_t __num_copied = __res.get();
     return ::std::make_pair(__output_first + __n, __num_copied);
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
index 876fc496bb6..c641712a7f3 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
@@ -1062,7 +1062,8 @@ __parallel_copy_if(_ExecutionPolicy&& __exec, _InRng&& __in_rng, _OutRng&& __out
         using CopyOp = unseq_backend::__copy_by_mask<_ReduceOp, oneapi::dpl::__internal::__pstl_assign,
                                                      /*inclusive*/ ::std::true_type, 1>;
 
-        return __parallel_scan_copy(::std::forward<_ExecutionPolicy>(__exec), ::std::forward<_InRng>(__in_rng),
+        return __parallel_scan_copy(oneapi::dpl::__internal::__device_backend_tag{},
+                                    ::std::forward<_ExecutionPolicy>(__exec), ::std::forward<_InRng>(__in_rng),
                                     ::std::forward<_OutRng>(__out_rng), __n, CreateOp{__pred}, CopyOp{});
     }
 }
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
index 78273a74876..67099e7e041 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
@@ -220,6 +220,7 @@ __parallel_scan_copy(_ExecutionPolicy&& __exec, _InRng&& __in_rng, _OutRng&& __o
 {
     // workaround until we implement more performant version for patterns
     return oneapi::dpl::__par_backend_hetero::__parallel_scan_copy(
+        oneapi::dpl::__internal::__device_backend_tag{},
         __exec.__device_policy(), ::std::forward<_InRng>(__in_rng), ::std::forward<_OutRng>(__out_rng), __n,
         __create_mask_op, __copy_by_mask_op);
 }

From 057e6ec540b230b6dc6d92e1a18948fceec0a62e Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 15 Feb 2024 10:28:09 +0100
Subject: [PATCH 206/566] __parallel_copy_if + tag impls

---
 .../pstl/hetero/dpcpp/parallel_backend_sycl.h | 44 +++++++++++++++++++
 .../hetero/dpcpp/parallel_backend_sycl_fpga.h | 11 +++++
 2 files changed, 55 insertions(+)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
index c641712a7f3..dc104106485 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
@@ -1068,6 +1068,50 @@ __parallel_copy_if(_ExecutionPolicy&& __exec, _InRng&& __in_rng, _OutRng&& __out
     }
 }
 
+template <typename _ExecutionPolicy, typename _InRng, typename _OutRng, typename _Size, typename _Pred>
+auto
+__parallel_copy_if(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&& __exec, _InRng&& __in_rng,
+                   _OutRng&& __out_rng, _Size __n, _Pred __pred)
+{
+    using _SingleGroupInvoker = __invoke_single_group_copy_if<_Size>;
+
+    // Next power of 2 greater than or equal to __n
+    auto __n_uniform = ::oneapi::dpl::__internal::__dpl_bit_ceil(static_cast<::std::make_unsigned_t<_Size>>(__n));
+
+    // Pessimistically only use half of the memory to take into account memory used by compiled kernel
+    const ::std::size_t __max_slm_size =
+        __exec.queue().get_device().template get_info<sycl::info::device::local_mem_size>() / 2;
+
+    // The kernel stores n integers for the predicate and another n integers for the offsets
+    const auto __req_slm_size = sizeof(::std::uint16_t) * __n_uniform * 2;
+
+    constexpr ::std::uint16_t __single_group_upper_limit = 16384;
+
+    ::std::size_t __max_wg_size = oneapi::dpl::__internal::__max_work_group_size(__exec);
+
+    if (__n <= __single_group_upper_limit && __max_slm_size >= __req_slm_size &&
+        __max_wg_size >= _SingleGroupInvoker::__targeted_wg_size)
+    {
+        using _SizeBreakpoints =
+            ::std::integer_sequence<::std::uint16_t, 16, 32, 64, 128, 256, 512, 1024, 2048, 4096, 8192, 16384>;
+
+        return __par_backend_hetero::__static_monotonic_dispatcher<_SizeBreakpoints>::__dispatch(
+            _SingleGroupInvoker{}, __n, ::std::forward<_ExecutionPolicy>(__exec), __n, ::std::forward<_InRng>(__in_rng),
+            ::std::forward<_OutRng>(__out_rng), __pred);
+    }
+    else
+    {
+        using _ReduceOp = ::std::plus<_Size>;
+        using CreateOp = unseq_backend::__create_mask<_Pred, _Size>;
+        using CopyOp = unseq_backend::__copy_by_mask<_ReduceOp, oneapi::dpl::__internal::__pstl_assign,
+                                                     /*inclusive*/ ::std::true_type, 1>;
+
+        return __parallel_scan_copy(oneapi::dpl::__internal::__device_backend_tag{},
+                                    ::std::forward<_ExecutionPolicy>(__exec), ::std::forward<_InRng>(__in_rng),
+                                    ::std::forward<_OutRng>(__out_rng), __n, CreateOp{__pred}, CopyOp{});
+    }
+}
+
 //------------------------------------------------------------------------
 // find_or tags
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
index 67099e7e041..9376a5ff978 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
@@ -211,6 +211,17 @@ __parallel_copy_if(_ExecutionPolicy&& __exec, _InRng&& __in_rng, _OutRng&& __out
         __exec.__device_policy(), ::std::forward<_InRng>(__in_rng), ::std::forward<_OutRng>(__out_rng), __n, __pred);
 }
 
+template <typename _ExecutionPolicy, typename _InRng, typename _OutRng, typename _Size, typename _Pred>
+auto
+__parallel_copy_if(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionPolicy&& __exec, _InRng&& __in_rng,
+                   _OutRng&& __out_rng, _Size __n, _Pred __pred)
+{
+    // workaround until we implement more performant version for patterns
+    return oneapi::dpl::__par_backend_hetero::__parallel_copy_if(
+        oneapi::dpl::__internal::__device_backend_tag{},
+        __exec.__device_policy(), ::std::forward<_InRng>(__in_rng), ::std::forward<_OutRng>(__out_rng), __n, __pred);
+}
+
 template <typename _ExecutionPolicy, typename _InRng, typename _OutRng, typename _Size, typename _CreateMaskOp,
           typename _CopyByMaskOp,
           oneapi::dpl::__internal::__enable_if_fpga_execution_policy<_ExecutionPolicy, int> = 0>

From e25607832f01731e48dba3174cf73b9d3e6a0d36 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 15 Feb 2024 10:29:15 +0100
Subject: [PATCH 207/566] __parallel_copy_if + tag calls

---
 .../oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h   | 12 ++++++++----
 .../pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h   |  1 +
 2 files changed, 9 insertions(+), 4 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index ac5c5984511..fe3f635a07e 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -1631,6 +1631,10 @@ oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, _
 __pattern_copy_if(_ExecutionPolicy&& __exec, _Iterator1 __first, _Iterator1 __last, _Iterator2 __result_first,
                   _Predicate __pred, /*vector=*/::std::true_type, /*parallel=*/::std::true_type)
 {
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator1, _Iterator2>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     using _It1DifferenceType = typename ::std::iterator_traits<_Iterator1>::difference_type;
 
     if (__first == __last)
@@ -1643,8 +1647,8 @@ __pattern_copy_if(_ExecutionPolicy&& __exec, _Iterator1 __first, _Iterator1 __la
     auto __keep2 = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::write, _Iterator2>();
     auto __buf2 = __keep2(__result_first, __result_first + __n);
 
-    auto __res = __par_backend_hetero::__parallel_copy_if(::std::forward<_ExecutionPolicy>(__exec), __buf1.all_view(),
-                                                          __buf2.all_view(), __n, __pred);
+    auto __res = __par_backend_hetero::__parallel_copy_if(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
+                                                          __buf1.all_view(), __buf2.all_view(), __n, __pred);
 
     ::std::size_t __num_copied = __res.get();
     return __result_first + __num_copied;
@@ -1668,8 +1672,8 @@ __pattern_copy_if(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _I
     auto __keep2 = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::write, _Iterator2>();
     auto __buf2 = __keep2(__result_first, __result_first + __n);
 
-    auto __res = __par_backend_hetero::__parallel_copy_if(::std::forward<_ExecutionPolicy>(__exec), __buf1.all_view(),
-                                                          __buf2.all_view(), __n, __pred);
+    auto __res = __par_backend_hetero::__parallel_copy_if(_BackendTag{}, ::std::forward<_ExecutionPolicy>(__exec),
+                                                          __buf1.all_view(), __buf2.all_view(), __n, __pred);
 
     ::std::size_t __num_copied = __res.get();
     return __result_first + __num_copied;
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
index 9376a5ff978..88637fb9b7b 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
@@ -208,6 +208,7 @@ __parallel_copy_if(_ExecutionPolicy&& __exec, _InRng&& __in_rng, _OutRng&& __out
 {
     // workaround until we implement more performant version for patterns
     return oneapi::dpl::__par_backend_hetero::__parallel_copy_if(
+        oneapi::dpl::__internal::__device_backend_tag{},
         __exec.__device_policy(), ::std::forward<_InRng>(__in_rng), ::std::forward<_OutRng>(__out_rng), __n, __pred);
 }
 

From 6702cc6b4cdfeb451664973a2b5ae6e4a5856e64 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 15 Feb 2024 10:50:03 +0100
Subject: [PATCH 208/566] __parallel_find_or + tag impls

---
 .../pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h   | 12 ++++++++++++
 1 file changed, 12 insertions(+)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
index 88637fb9b7b..34f269a93d0 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
@@ -266,6 +266,18 @@ __parallel_find_or(_ExecutionPolicy&& __exec, _Brick __f, _BrickTag __brick_tag,
                                                                  ::std::forward<_Ranges>(__rngs)...);
 }
 
+template <typename _ExecutionPolicy, typename _Brick, typename _BrickTag, typename... _Ranges>
+::std::conditional_t<::std::is_same_v<_BrickTag, __parallel_or_tag>, bool,
+                        oneapi::dpl::__internal::__difference_t<
+                            typename oneapi::dpl::__ranges::__get_first_range_type<_Ranges...>::type>>
+__parallel_find_or(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionPolicy&& __exec, _Brick __f,
+                   _BrickTag __brick_tag, _Ranges&&... __rngs)
+{
+    return oneapi::dpl::__par_backend_hetero::__parallel_find_or(oneapi::dpl::__internal::__device_backend_tag{},
+                                                                 __exec.__device_policy(), __f, __brick_tag,
+                                                                 ::std::forward<_Ranges>(__rngs)...);
+}
+
 //------------------------------------------------------------------------
 // parallel_or
 //-----------------------------------------------------------------------

From f0353cac438c6f16dbed5b122211ed0df4140dc2 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 15 Feb 2024 10:50:49 +0100
Subject: [PATCH 209/566] __parallel_find_or + tag calls

---
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 19 +++++++++++++++++--
 .../pstl/hetero/dpcpp/parallel_backend_sycl.h | 19 +++++++++++++++++++
 .../hetero/dpcpp/parallel_backend_sycl_fpga.h |  3 ++-
 3 files changed, 38 insertions(+), 3 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index fe3f635a07e..ac0f22ebe10 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -931,6 +931,9 @@ __pattern_adjacent_find(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator
                         /*parallel*/ ::std::true_type, /*vector*/ ::std::true_type,
                         oneapi::dpl::__internal::__or_semantic)
 {
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     if (__last - __first < 2)
         return __last;
 
@@ -945,8 +948,8 @@ __pattern_adjacent_find(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator
     // TODO: in case of confilicting names
     // __par_backend_hetero::make_wrapped_policy<__par_backend_hetero::__or_policy_wrapper>()
     bool result = __par_backend_hetero::__parallel_find_or(
-        ::std::forward<_ExecutionPolicy>(__exec), _Predicate{adjacent_find_fn<_BinaryPredicate>{__predicate}},
-        __par_backend_hetero::__parallel_or_tag{},
+        __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
+        _Predicate{adjacent_find_fn<_BinaryPredicate>{__predicate}}, __par_backend_hetero::__parallel_or_tag{},
         oneapi::dpl::__ranges::make_zip_view(__buf1.all_view(), __buf2.all_view()));
 
     // inverted conditional because of
@@ -973,6 +976,7 @@ __pattern_adjacent_find(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __ex
     // TODO: in case of confilicting names
     // __par_backend_hetero::make_wrapped_policy<__par_backend_hetero::__or_policy_wrapper>()
     bool result = __par_backend_hetero::__parallel_find_or(
+        _BackendTag{},
         ::std::forward<_ExecutionPolicy>(__exec), _Predicate{adjacent_find_fn<_BinaryPredicate>{__predicate}},
         __par_backend_hetero::__parallel_or_tag{},
         oneapi::dpl::__ranges::make_zip_view(__buf1.all_view(), __buf2.all_view()));
@@ -1113,6 +1117,9 @@ oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, b
 __pattern_any_of(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last, _Pred __pred,
                  /*vector=*/::std::true_type, /*parallel=*/::std::true_type)
 {
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     if (__first == __last)
         return false;
 
@@ -1122,6 +1129,7 @@ __pattern_any_of(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last,
     auto __buf = __keep(__first, __last);
 
     return oneapi::dpl::__par_backend_hetero::__parallel_find_or(
+        __backend_tag{},
         __par_backend_hetero::make_wrapped_policy<__par_backend_hetero::__or_policy_wrapper>(
             ::std::forward<_ExecutionPolicy>(__exec)),
         _Predicate{__pred}, __par_backend_hetero::__parallel_or_tag{}, __buf.all_view());
@@ -1141,6 +1149,7 @@ __pattern_any_of(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _It
     auto __buf = __keep(__first, __last);
 
     return oneapi::dpl::__par_backend_hetero::__parallel_find_or(
+        _BackendTag{},
         __par_backend_hetero::make_wrapped_policy<__par_backend_hetero::__or_policy_wrapper>(
             ::std::forward<_ExecutionPolicy>(__exec)),
         _Predicate{__pred}, __par_backend_hetero::__parallel_or_tag{}, __buf.all_view());
@@ -1156,6 +1165,10 @@ __pattern_equal(_ExecutionPolicy&& __exec, _Iterator1 __first1, _Iterator1 __las
                 _Iterator2 __last2, _Pred __pred,
                 /*vector=*/::std::true_type, /*parallel=*/::std::true_type)
 {
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator1, _Iterator2>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     if (__last1 == __first1 || __last2 == __first2 || __last1 - __first1 != __last2 - __first2)
         return false;
 
@@ -1169,6 +1182,7 @@ __pattern_equal(_ExecutionPolicy&& __exec, _Iterator1 __first1, _Iterator1 __las
     // TODO: in case of confilicting names
     // __par_backend_hetero::make_wrapped_policy<__par_backend_hetero::__or_policy_wrapper>()
     return !__par_backend_hetero::__parallel_find_or(
+        __backend_tag{},
         ::std::forward<_ExecutionPolicy>(__exec), _Predicate{equal_predicate<_Pred>{__pred}},
         __par_backend_hetero::__parallel_or_tag{},
         oneapi::dpl::__ranges::make_zip_view(__buf1.all_view(), __buf2.all_view()));
@@ -1192,6 +1206,7 @@ __pattern_equal(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Ite
     // TODO: in case of confilicting names
     // __par_backend_hetero::make_wrapped_policy<__par_backend_hetero::__or_policy_wrapper>()
     return !__par_backend_hetero::__parallel_find_or(
+        _BackendTag{},
         ::std::forward<_ExecutionPolicy>(__exec), _Predicate{equal_predicate<_Pred>{__pred}},
         __par_backend_hetero::__parallel_or_tag{},
         oneapi::dpl::__ranges::make_zip_view(__buf1.all_view(), __buf2.all_view()));
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
index dc104106485..9a1a7ecd3f6 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
@@ -1466,12 +1466,17 @@ oneapi::dpl::__internal::__enable_if_device_execution_policy<_ExecutionPolicy, b
 __parallel_or(_ExecutionPolicy&& __exec, _Iterator1 __first, _Iterator1 __last, _Iterator2 __s_first,
               _Iterator2 __s_last, _Brick __f)
 {
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator1, _Iterator2>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     auto __keep = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _Iterator1>();
     auto __buf = __keep(__first, __last);
     auto __s_keep = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _Iterator2>();
     auto __s_buf = __s_keep(__s_first, __s_last);
 
     return oneapi::dpl::__par_backend_hetero::__parallel_find_or(
+        __backend_tag{},
         __par_backend_hetero::make_wrapped_policy<__or_policy_wrapper>(::std::forward<_ExecutionPolicy>(__exec)), __f,
         __parallel_or_tag{}, __buf.all_view(), __s_buf.all_view());
 }
@@ -1483,10 +1488,14 @@ template <typename _ExecutionPolicy, typename _Iterator, typename _Brick>
 oneapi::dpl::__internal::__enable_if_device_execution_policy<_ExecutionPolicy, bool>
 __parallel_or(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last, _Brick __f)
 {
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     auto __keep = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _Iterator>();
     auto __buf = __keep(__first, __last);
 
     return oneapi::dpl::__par_backend_hetero::__parallel_find_or(
+        __backend_tag{},
         __par_backend_hetero::make_wrapped_policy<__or_policy_wrapper>(::std::forward<_ExecutionPolicy>(__exec)), __f,
         __parallel_or_tag{}, __buf.all_view());
 }
@@ -1505,6 +1514,10 @@ oneapi::dpl::__internal::__enable_if_device_execution_policy<_ExecutionPolicy, _
 __parallel_find(_ExecutionPolicy&& __exec, _Iterator1 __first, _Iterator1 __last, _Iterator2 __s_first,
                 _Iterator2 __s_last, _Brick __f, _IsFirst)
 {
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator1, _Iterator2>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     auto __keep = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _Iterator1>();
     auto __buf = __keep(__first, __last);
     auto __s_keep = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _Iterator2>();
@@ -1513,6 +1526,7 @@ __parallel_find(_ExecutionPolicy&& __exec, _Iterator1 __first, _Iterator1 __last
     using _TagType = ::std::conditional_t<_IsFirst::value, __parallel_find_forward_tag<decltype(__buf.all_view())>,
                                           __parallel_find_backward_tag<decltype(__buf.all_view())>>;
     return __first + oneapi::dpl::__par_backend_hetero::__parallel_find_or(
+                         __backend_tag{},
                          __par_backend_hetero::make_wrapped_policy<__find_policy_wrapper>(
                              ::std::forward<_ExecutionPolicy>(__exec)),
                          __f, _TagType{}, __buf.all_view(), __s_buf.all_view());
@@ -1531,6 +1545,7 @@ __parallel_find(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&
     using _TagType = ::std::conditional_t<_IsFirst::value, __parallel_find_forward_tag<decltype(__buf.all_view())>,
                                           __parallel_find_backward_tag<decltype(__buf.all_view())>>;
     return __first + oneapi::dpl::__par_backend_hetero::__parallel_find_or(
+                         oneapi::dpl::__internal::__device_backend_tag{},
                          __par_backend_hetero::make_wrapped_policy<__find_policy_wrapper>(
                              ::std::forward<_ExecutionPolicy>(__exec)),
                          __f, _TagType{}, __buf.all_view(), __s_buf.all_view());
@@ -1544,12 +1559,16 @@ template <typename _ExecutionPolicy, typename _Iterator, typename _Brick, typena
 oneapi::dpl::__internal::__enable_if_device_execution_policy<_ExecutionPolicy, _Iterator>
 __parallel_find(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last, _Brick __f, _IsFirst)
 {
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     auto __keep = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _Iterator>();
     auto __buf = __keep(__first, __last);
 
     using _TagType = ::std::conditional_t<_IsFirst::value, __parallel_find_forward_tag<decltype(__buf.all_view())>,
                                           __parallel_find_backward_tag<decltype(__buf.all_view())>>;
     return __first + oneapi::dpl::__par_backend_hetero::__parallel_find_or(
+                         __backend_tag{},
                          __par_backend_hetero::make_wrapped_policy<__find_policy_wrapper>(
                              ::std::forward<_ExecutionPolicy>(__exec)),
                          __f, _TagType{}, __buf.all_view());
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
index 34f269a93d0..18b4d00b627 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
@@ -262,7 +262,8 @@ oneapi::dpl::__internal::__enable_if_fpga_execution_policy<
                              typename oneapi::dpl::__ranges::__get_first_range_type<_Ranges...>::type>>>
 __parallel_find_or(_ExecutionPolicy&& __exec, _Brick __f, _BrickTag __brick_tag, _Ranges&&... __rngs)
 {
-    return oneapi::dpl::__par_backend_hetero::__parallel_find_or(__exec.__device_policy(), __f, __brick_tag,
+    return oneapi::dpl::__par_backend_hetero::__parallel_find_or(oneapi::dpl::__internal::__device_backend_tag{},
+                                                                 __exec.__device_policy(), __f, __brick_tag,
                                                                  ::std::forward<_Ranges>(__rngs)...);
 }
 

From 500fe4d314d57a40fa49130f6d688faa5f02ffbf Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 15 Feb 2024 11:05:12 +0100
Subject: [PATCH 210/566] __parallel_or + tag impls

---
 .../pstl/hetero/dpcpp/parallel_backend_sycl.h | 33 +++++++++++++++++++
 .../hetero/dpcpp/parallel_backend_sycl_fpga.h | 21 ++++++++++++
 2 files changed, 54 insertions(+)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
index 9a1a7ecd3f6..22a07100d96 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
@@ -1481,6 +1481,22 @@ __parallel_or(_ExecutionPolicy&& __exec, _Iterator1 __first, _Iterator1 __last,
         __parallel_or_tag{}, __buf.all_view(), __s_buf.all_view());
 }
 
+template <typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2, typename _Brick>
+bool
+__parallel_or(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&& __exec, _Iterator1 __first,
+              _Iterator1 __last, _Iterator2 __s_first, _Iterator2 __s_last, _Brick __f)
+{
+    auto __keep = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _Iterator1>();
+    auto __buf = __keep(__first, __last);
+    auto __s_keep = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _Iterator2>();
+    auto __s_buf = __s_keep(__s_first, __s_last);
+
+    return oneapi::dpl::__par_backend_hetero::__parallel_find_or(
+        oneapi::dpl::__internal::__device_backend_tag{},
+        __par_backend_hetero::make_wrapped_policy<__or_policy_wrapper>(::std::forward<_ExecutionPolicy>(__exec)), __f,
+        __parallel_or_tag{}, __buf.all_view(), __s_buf.all_view());
+}
+
 // Special overload for single sequence cases.
 // TODO: check if similar pattern may apply to other algorithms. If so, these overloads should be moved out of
 // backend code.
@@ -1500,6 +1516,23 @@ __parallel_or(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last, _B
         __parallel_or_tag{}, __buf.all_view());
 }
 
+// Special overload for single sequence cases.
+// TODO: check if similar pattern may apply to other algorithms. If so, these overloads should be moved out of
+// backend code.
+template <typename _ExecutionPolicy, typename _Iterator, typename _Brick>
+bool
+__parallel_or(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&& __exec, _Iterator __first,
+              _Iterator __last, _Brick __f)
+{
+    auto __keep = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _Iterator>();
+    auto __buf = __keep(__first, __last);
+
+    return oneapi::dpl::__par_backend_hetero::__parallel_find_or(
+        oneapi::dpl::__internal::__device_backend_tag{},
+        __par_backend_hetero::make_wrapped_policy<__or_policy_wrapper>(::std::forward<_ExecutionPolicy>(__exec)), __f,
+        __parallel_or_tag{}, __buf.all_view());
+}
+
 //------------------------------------------------------------------------
 // parallel_find - sync pattern
 //-----------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
index 18b4d00b627..b6950c6f75c 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
@@ -292,6 +292,17 @@ __parallel_or(_ExecutionPolicy&& __exec, _Iterator1 __first, _Iterator1 __last,
                                                             __s_last, __f);
 }
 
+template <typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2, typename _Brick>
+bool
+__parallel_or(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionPolicy&& __exec, _Iterator1 __first,
+              _Iterator1 __last, _Iterator2 __s_first, _Iterator2 __s_last, _Brick __f)
+{
+    // workaround until we implement more performant version for patterns
+    return oneapi::dpl::__par_backend_hetero::__parallel_or(oneapi::dpl::__internal::__device_backend_tag{},
+                                                            __exec.__device_policy(), __first, __last, __s_first,
+                                                            __s_last, __f);
+}
+
 template <typename _ExecutionPolicy, typename _Iterator, typename _Brick>
 oneapi::dpl::__internal::__enable_if_fpga_execution_policy<_ExecutionPolicy, bool>
 __parallel_or(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last, _Brick __f)
@@ -300,6 +311,16 @@ __parallel_or(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last, _B
     return oneapi::dpl::__par_backend_hetero::__parallel_or(__exec.__device_policy(), __first, __last, __f);
 }
 
+template <typename _ExecutionPolicy, typename _Iterator, typename _Brick>
+bool
+__parallel_or(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionPolicy&& __exec, _Iterator __first,
+              _Iterator __last, _Brick __f)
+{
+    // workaround until we implement more performant version for patterns
+    return oneapi::dpl::__par_backend_hetero::__parallel_or(oneapi::dpl::__internal::__device_backend_tag{},
+                                                            __exec.__device_policy(), __first, __last, __f);
+}
+
 //------------------------------------------------------------------------
 // parallel_find
 //-----------------------------------------------------------------------

From 1284178d0eeb98e9135914d095a73910ae6230c0 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 15 Feb 2024 11:08:45 +0100
Subject: [PATCH 211/566] __parallel_or + tag calls

---
 include/oneapi/dpl/pstl/algorithm_impl.h      | 46 ++++++++++++++++---
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 12 +++++
 2 files changed, 52 insertions(+), 6 deletions(-)

diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 26612aa8ae8..f9ffbc6ac7d 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -85,8 +85,12 @@ oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, boo
 __pattern_any_of(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIterator __last, _Pred __pred,
                  _IsVector __is_vector, /*parallel=*/::std::true_type)
 {
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     return __internal::__except_handler([&]() {
-        return __internal::__parallel_or(::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+        return __internal::__parallel_or(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
                                          [__pred, __is_vector](_RandomAccessIterator __i, _RandomAccessIterator __j) {
                                              return __internal::__brick_any_of(__i, __j, __pred, __is_vector);
                                          });
@@ -98,8 +102,10 @@ bool
 __pattern_any_of(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
                  _RandomAccessIterator __last, _Pred __pred)
 {
+    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
+
     return __internal::__except_handler([&]() {
-        return __internal::__parallel_or(::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+        return __internal::__parallel_or(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
                                          [__pred](_RandomAccessIterator __i, _RandomAccessIterator __j) {
                                              return __internal::__brick_any_of(__i, __j, __pred, _IsVector{});
                                          });
@@ -1050,12 +1056,16 @@ __pattern_equal(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1, _Ran
                 _RandomAccessIterator2 __first2, _RandomAccessIterator2 __last2, _BinaryPredicate __p,
                 _IsVector __is_vector, /*is_parallel=*/::std::true_type)
 {
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     if (__last1 - __first1 != __last2 - __first2)
         return false;
 
     return __internal::__except_handler([&]() {
         return !__internal::__parallel_or(
-            ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1,
+            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1,
             [__first1, __first2, __p, __is_vector](_RandomAccessIterator1 __i, _RandomAccessIterator1 __j) {
                 return !__internal::__brick_equal(__i, __j, __first2 + (__i - __first1), __first2 + (__j - __first1),
                                                   __p, __is_vector);
@@ -1070,11 +1080,14 @@ __pattern_equal(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAcc
                 _RandomAccessIterator1 __last1, _RandomAccessIterator2 __first2, _RandomAccessIterator2 __last2,
                 _BinaryPredicate __p)
 {
+    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
+
     if (__last1 - __first1 != __last2 - __first2)
         return false;
 
     return __internal::__except_handler([&]() {
         return !__internal::__parallel_or(
+            __backend_tag{},
             ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1,
             [__first1, __first2, __p](_RandomAccessIterator1 __i, _RandomAccessIterator1 __j) {
                 return !__internal::__brick_equal(__i, __j, __first2 + (__i - __first1), __first2 + (__j - __first1),
@@ -1131,8 +1144,13 @@ __pattern_equal(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1, _Ran
                 _RandomAccessIterator2 __first2, _BinaryPredicate __p, _IsVector __is_vector,
                 /*is_parallel=*/::std::true_type)
 {
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     return __internal::__except_handler([&]() {
         return !__internal::__parallel_or(
+            __backend_tag{},
             ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1,
             [__first1, __first2, __p, __is_vector](_RandomAccessIterator1 __i, _RandomAccessIterator1 __j) {
                 return !__internal::__brick_equal(__i, __j, __first2 + (__i - __first1), __p, __is_vector);
@@ -1146,9 +1164,11 @@ bool
 __pattern_equal(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1,
                 _RandomAccessIterator1 __last1, _RandomAccessIterator2 __first2, _BinaryPredicate __p)
 {
+    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
+
     return __internal::__except_handler([&]() {
         return !__internal::__parallel_or(
-            ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1,
+            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1,
             [__first1, __first2, __p](_RandomAccessIterator1 __i, _RandomAccessIterator1 __j) {
                 return !__internal::__brick_equal(__i, __j, __first2 + (__i - __first1), __p, _IsVector{});
             });
@@ -5021,6 +5041,10 @@ __pattern_includes(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1, _
                    _RandomAccessIterator2 __first2, _RandomAccessIterator2 __last2, _Compare __comp, _IsVector,
                    /*is_parallel=*/::std::true_type)
 {
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     if (__first2 == __last2)
         return true;
 
@@ -5041,6 +5065,7 @@ __pattern_includes(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1, _
 
     return __internal::__except_handler([&]() {
         return !__internal::__parallel_or(
+            __backend_tag{},
             ::std::forward<_ExecutionPolicy>(__exec), __first2, __last2,
             [__first1, __last1, __first2, __last2, &__comp](_RandomAccessIterator2 __i, _RandomAccessIterator2 __j) {
                 assert(__j > __i);
@@ -5082,6 +5107,8 @@ __pattern_includes(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Random
                    _RandomAccessIterator1 __last1, _RandomAccessIterator2 __first2, _RandomAccessIterator2 __last2,
                    _Compare __comp)
 {
+    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
+
     if (__first2 == __last2)
         return true;
 
@@ -5102,6 +5129,7 @@ __pattern_includes(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Random
 
     return __internal::__except_handler([&]() {
         return !__internal::__parallel_or(
+            __backend_tag{},
             ::std::forward<_ExecutionPolicy>(__exec), __first2, __last2,
             [__first1, __last1, __first2, __last2, &__comp](_RandomAccessIterator2 __i, _RandomAccessIterator2 __j) {
                 assert(__j > __i);
@@ -6047,8 +6075,12 @@ oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, boo
 __pattern_is_heap(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIterator __last,
                   _Compare __comp, _IsVector __is_vector, /* is_parallel = */ ::std::true_type)
 {
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     return __internal::__except_handler([&]() {
-        return !__parallel_or(::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+        return !__parallel_or(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
                               [__first, __comp, __is_vector](_RandomAccessIterator __i, _RandomAccessIterator __j) {
                                   return !__internal::__is_heap_local(__first, __i - __first, __j - __first, __comp,
                                                                       __is_vector);
@@ -6061,8 +6093,10 @@ bool
 __pattern_is_heap(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
                   _RandomAccessIterator __last, _Compare __comp)
 {
+    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
+
     return __internal::__except_handler([&]() {
-        return !__parallel_or(::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+        return !__parallel_or(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
                               [__first, __comp](_RandomAccessIterator __i, _RandomAccessIterator __j) {
                                   return !__internal::__is_heap_local(__first, __i - __first, __j - __first, __comp,
                                                                       _IsVector{});
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index ac0f22ebe10..c6e23ae6808 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -2053,6 +2053,10 @@ oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, b
 __pattern_is_heap(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIterator __last,
                   _Compare __comp, /* vector */ ::std::true_type, /* parallel = */ ::std::true_type)
 {
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     if (__last - __first < 2)
         return true;
 
@@ -2060,6 +2064,7 @@ __pattern_is_heap(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _Ran
         oneapi::dpl::unseq_backend::single_match_pred_by_idx<_ExecutionPolicy, __is_heap_check<_Compare>>;
 
     return !__par_backend_hetero::__parallel_or(
+        __backend_tag{},
         ::std::forward<_ExecutionPolicy>(__exec),
         __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__first),
         __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__last), _Predicate{__comp});
@@ -2077,6 +2082,7 @@ __pattern_is_heap(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _R
         oneapi::dpl::unseq_backend::single_match_pred_by_idx<_ExecutionPolicy, __is_heap_check<_Compare>>;
 
     return !__par_backend_hetero::__parallel_or(
+        _BackendTag{},
         ::std::forward<_ExecutionPolicy>(__exec),
         __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__first),
         __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__last), _Predicate{__comp});
@@ -2553,6 +2559,10 @@ __pattern_includes(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _Forwa
                    _ForwardIterator2 __first2, _ForwardIterator2 __last2, _Compare __comp, /*vector=*/::std::true_type,
                    /*parallel=*/::std::true_type)
 {
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     //according to the spec
     if (__first2 == __last2)
         return true;
@@ -2567,6 +2577,7 @@ __pattern_includes(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _Forwa
 
     using __brick_include_type = unseq_backend::__brick_includes<_ExecutionPolicy, _Compare, _Size1, _Size2>;
     return !__par_backend_hetero::__parallel_or(
+        __backend_tag{},
         ::std::forward<_ExecutionPolicy>(__exec),
         __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__first2),
         __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__last2),
@@ -2595,6 +2606,7 @@ __pattern_includes(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _
 
     using __brick_include_type = unseq_backend::__brick_includes<_ExecutionPolicy, _Compare, _Size1, _Size2>;
     return !__par_backend_hetero::__parallel_or(
+        _BackendTag{},
         ::std::forward<_ExecutionPolicy>(__exec),
         __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__first2),
         __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__last2),

From 274acc73fc2d323441c927d85c4510513514cff1 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 15 Feb 2024 12:02:11 +0100
Subject: [PATCH 212/566] __internal::__parallel_or + tag calls

---
 include/oneapi/dpl/pstl/algorithm_impl.h | 45 ++++++++----------------
 1 file changed, 15 insertions(+), 30 deletions(-)

diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index f9ffbc6ac7d..7353936271e 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -87,10 +87,9 @@ __pattern_any_of(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _Rand
 {
     constexpr auto __dispatch_tag =
         oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
 
     return __internal::__except_handler([&]() {
-        return __internal::__parallel_or(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+        return __internal::__parallel_or(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
                                          [__pred, __is_vector](_RandomAccessIterator __i, _RandomAccessIterator __j) {
                                              return __internal::__brick_any_of(__i, __j, __pred, __is_vector);
                                          });
@@ -99,13 +98,11 @@ __pattern_any_of(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _Rand
 
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Pred>
 bool
-__pattern_any_of(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
+__pattern_any_of(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
                  _RandomAccessIterator __last, _Pred __pred)
 {
-    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
-
     return __internal::__except_handler([&]() {
-        return __internal::__parallel_or(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+        return __internal::__parallel_or(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
                                          [__pred](_RandomAccessIterator __i, _RandomAccessIterator __j) {
                                              return __internal::__brick_any_of(__i, __j, __pred, _IsVector{});
                                          });
@@ -1058,14 +1055,13 @@ __pattern_equal(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1, _Ran
 {
     constexpr auto __dispatch_tag =
         oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
 
     if (__last1 - __first1 != __last2 - __first2)
         return false;
 
     return __internal::__except_handler([&]() {
         return !__internal::__parallel_or(
-            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1,
+            __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1,
             [__first1, __first2, __p, __is_vector](_RandomAccessIterator1 __i, _RandomAccessIterator1 __j) {
                 return !__internal::__brick_equal(__i, __j, __first2 + (__i - __first1), __first2 + (__j - __first1),
                                                   __p, __is_vector);
@@ -1076,18 +1072,16 @@ __pattern_equal(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1, _Ran
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
           class _BinaryPredicate>
 bool
-__pattern_equal(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1,
+__pattern_equal(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1,
                 _RandomAccessIterator1 __last1, _RandomAccessIterator2 __first2, _RandomAccessIterator2 __last2,
                 _BinaryPredicate __p)
 {
-    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
-
     if (__last1 - __first1 != __last2 - __first2)
         return false;
 
     return __internal::__except_handler([&]() {
         return !__internal::__parallel_or(
-            __backend_tag{},
+            __tag,
             ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1,
             [__first1, __first2, __p](_RandomAccessIterator1 __i, _RandomAccessIterator1 __j) {
                 return !__internal::__brick_equal(__i, __j, __first2 + (__i - __first1), __first2 + (__j - __first1),
@@ -1146,11 +1140,10 @@ __pattern_equal(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1, _Ran
 {
     constexpr auto __dispatch_tag =
         oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
 
     return __internal::__except_handler([&]() {
         return !__internal::__parallel_or(
-            __backend_tag{},
+            __dispatch_tag,
             ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1,
             [__first1, __first2, __p, __is_vector](_RandomAccessIterator1 __i, _RandomAccessIterator1 __j) {
                 return !__internal::__brick_equal(__i, __j, __first2 + (__i - __first1), __p, __is_vector);
@@ -1161,14 +1154,12 @@ __pattern_equal(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1, _Ran
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
           class _BinaryPredicate>
 bool
-__pattern_equal(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1,
+__pattern_equal(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1,
                 _RandomAccessIterator1 __last1, _RandomAccessIterator2 __first2, _BinaryPredicate __p)
 {
-    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
-
     return __internal::__except_handler([&]() {
         return !__internal::__parallel_or(
-            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1,
+            __tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1,
             [__first1, __first2, __p](_RandomAccessIterator1 __i, _RandomAccessIterator1 __j) {
                 return !__internal::__brick_equal(__i, __j, __first2 + (__i - __first1), __p, _IsVector{});
             });
@@ -5043,7 +5034,6 @@ __pattern_includes(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1, _
 {
     constexpr auto __dispatch_tag =
         oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
 
     if (__first2 == __last2)
         return true;
@@ -5065,7 +5055,7 @@ __pattern_includes(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1, _
 
     return __internal::__except_handler([&]() {
         return !__internal::__parallel_or(
-            __backend_tag{},
+            __dispatch_tag,
             ::std::forward<_ExecutionPolicy>(__exec), __first2, __last2,
             [__first1, __last1, __first2, __last2, &__comp](_RandomAccessIterator2 __i, _RandomAccessIterator2 __j) {
                 assert(__j > __i);
@@ -5103,12 +5093,10 @@ __pattern_includes(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1, _
 
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _Compare>
 bool
-__pattern_includes(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1,
+__pattern_includes(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1,
                    _RandomAccessIterator1 __last1, _RandomAccessIterator2 __first2, _RandomAccessIterator2 __last2,
                    _Compare __comp)
 {
-    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
-
     if (__first2 == __last2)
         return true;
 
@@ -5129,7 +5117,7 @@ __pattern_includes(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Random
 
     return __internal::__except_handler([&]() {
         return !__internal::__parallel_or(
-            __backend_tag{},
+            __tag,
             ::std::forward<_ExecutionPolicy>(__exec), __first2, __last2,
             [__first1, __last1, __first2, __last2, &__comp](_RandomAccessIterator2 __i, _RandomAccessIterator2 __j) {
                 assert(__j > __i);
@@ -6077,10 +6065,9 @@ __pattern_is_heap(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _Ran
 {
     constexpr auto __dispatch_tag =
         oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
 
     return __internal::__except_handler([&]() {
-        return !__parallel_or(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+        return !__parallel_or(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
                               [__first, __comp, __is_vector](_RandomAccessIterator __i, _RandomAccessIterator __j) {
                                   return !__internal::__is_heap_local(__first, __i - __first, __j - __first, __comp,
                                                                       __is_vector);
@@ -6090,13 +6077,11 @@ __pattern_is_heap(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _Ran
 
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Compare>
 bool
-__pattern_is_heap(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
+__pattern_is_heap(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
                   _RandomAccessIterator __last, _Compare __comp)
 {
-    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
-
     return __internal::__except_handler([&]() {
-        return !__parallel_or(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+        return !__parallel_or(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
                               [__first, __comp](_RandomAccessIterator __i, _RandomAccessIterator __j) {
                                   return !__internal::__is_heap_local(__first, __i - __first, __j - __first, __comp,
                                                                       _IsVector{});

From 553f101046c18cdb6d4166576d4d32046d0fa709 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 15 Feb 2024 12:24:08 +0100
Subject: [PATCH 213/566] __pattern_set_intersection + tag impls

---
 include/oneapi/dpl/pstl/algorithm_impl.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 7353936271e..77365909771 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -5588,7 +5588,7 @@ __pattern_set_intersection(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __f
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
           class _RandomAccessIterator3, class _Compare>
 _RandomAccessIterator3
-__pattern_set_intersection(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1,
+__pattern_set_intersection(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1,
                            _RandomAccessIterator1 __last1, _RandomAccessIterator2 __first2,
                            _RandomAccessIterator2 __last2, _RandomAccessIterator3 __result, _Compare __comp)
 {

From f1179ba38f33cffdfa1f4a209c7292f1df60d51f Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 15 Feb 2024 12:19:40 +0100
Subject: [PATCH 214/566] __internal::__parallel_set_op + tag impls

---
 include/oneapi/dpl/pstl/algorithm_impl.h | 94 ++++++++++++++++++++++++
 1 file changed, 94 insertions(+)

diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 77365909771..158f65791d9 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -5250,6 +5250,100 @@ __parallel_set_op(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _Forwar
     });
 }
 
+template <class _IsVector, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2,
+          class _OutputIterator,
+          class _Compare, class _SizeFunction, class _SetOP>
+_OutputIterator
+__parallel_set_op(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _ForwardIterator1 __first1,
+                  _ForwardIterator1 __last1,
+                  _ForwardIterator2 __first2, _ForwardIterator2 __last2, _OutputIterator __result, _Compare __comp,
+                  _SizeFunction __size_func, _SetOP __set_op)
+{
+    using __backend_tag = typename decltype(__tag)::__backend_tag;
+
+    typedef typename ::std::iterator_traits<_ForwardIterator1>::difference_type _DifferenceType;
+    typedef typename ::std::iterator_traits<_OutputIterator>::value_type _T;
+
+    struct _SetRange
+    {
+        _DifferenceType __pos, __len, __buf_pos;
+        bool
+        empty() const
+        {
+            return __len == 0;
+        }
+    };
+
+    const _DifferenceType __n1 = __last1 - __first1;
+    const _DifferenceType __n2 = __last2 - __first2;
+
+    __par_backend::__buffer<_ExecutionPolicy, _T> __buf(__size_func(__n1, __n2));
+
+    return __internal::__except_handler([&__exec, __n1, __first1, __last1, __first2, __last2, __result,
+                                         __comp, __size_func, __set_op, &__buf]() {
+        auto __tmp_memory = __buf.get();
+        _DifferenceType __m{};
+        auto __scan = [=](_DifferenceType, _DifferenceType, const _SetRange& __s) { // Scan
+            if (!__s.empty())
+                __brick_move_destroy<_ExecutionPolicy>{}(__tmp_memory + __s.__buf_pos,
+                                                         __tmp_memory + (__s.__buf_pos + __s.__len),
+                                                         __result + __s.__pos, _IsVector{});
+        };
+        __par_backend::__parallel_strict_scan(
+            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __n1, _SetRange{0, 0, 0}, //-1, 0},
+            [=](_DifferenceType __i, _DifferenceType __len) {                                    // Reduce
+                //[__b; __e) - a subrange of the first sequence, to reduce
+                _ForwardIterator1 __b = __first1 + __i, __e = __first1 + (__i + __len);
+
+                //try searching for the first element which not equal to *__b
+                if (__b != __first1)
+                    __b = ::std::upper_bound(__b, __last1, *__b, __comp);
+
+                //try searching for the first element which not equal to *__e
+                if (__e != __last1)
+                    __e = ::std::upper_bound(__e, __last1, *__e, __comp);
+
+                //check is [__b; __e) empty
+                if (__e - __b < 1)
+                {
+                    _ForwardIterator2 __bb = __last2;
+                    if (__b != __last1)
+                        __bb = ::std::lower_bound(__first2, __last2, *__b, __comp);
+
+                    const _DifferenceType __buf_pos = __size_func((__b - __first1), (__bb - __first2));
+                    return _SetRange{0, 0, __buf_pos};
+                }
+
+                //try searching for "corresponding" subrange [__bb; __ee) in the second sequence
+                _ForwardIterator2 __bb = __first2;
+                if (__b != __first1)
+                    __bb = ::std::lower_bound(__first2, __last2, *__b, __comp);
+
+                _ForwardIterator2 __ee = __last2;
+                if (__e != __last1)
+                    __ee = ::std::lower_bound(__bb, __last2, *__e, __comp);
+
+                const _DifferenceType __buf_pos = __size_func((__b - __first1), (__bb - __first2));
+                auto __buffer_b = __tmp_memory + __buf_pos;
+                auto __res = __set_op(__b, __e, __bb, __ee, __buffer_b, __comp);
+
+                return _SetRange{0, __res - __buffer_b, __buf_pos};
+            },
+            [](const _SetRange& __a, const _SetRange& __b) { // Combine
+                if (__b.__buf_pos > __a.__buf_pos || ((__b.__buf_pos == __a.__buf_pos) && !__b.empty()))
+                    return _SetRange{__a.__pos + __a.__len + __b.__pos, __b.__len, __b.__buf_pos};
+                return _SetRange{__b.__pos + __b.__len + __a.__pos, __a.__len, __a.__buf_pos};
+            },
+            __scan,                                     // Scan
+            [&__m, &__scan](const _SetRange& __total) { // Apex
+                //final scan
+                __scan(0, 0, __total);
+                __m = __total.__pos + __total.__len;
+            });
+        return __result + __m;
+    });
+}
+
 //a shared parallel pattern for '__pattern_set_union' and '__pattern_set_symmetric_difference'
 template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _OutputIterator,
           class _Compare, class _SetUnionOp, class _IsVector>

From 96863214ec59393431d690e6d1df29c033029bdd Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 15 Feb 2024 12:23:01 +0100
Subject: [PATCH 215/566] __internal::__parallel_set_op + tag calls

---
 include/oneapi/dpl/pstl/algorithm_impl.h | 43 ++++++++++++++----------
 1 file changed, 25 insertions(+), 18 deletions(-)

diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 158f65791d9..db217b7b0a4 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -5424,9 +5424,9 @@ __parallel_set_union_op(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _
             },
             [=, &__result] {
                 __result = __internal::__parallel_set_op(
+                    __dispatch_tag,
                     ::std::forward<_ExecutionPolicy>(__exec), __left_bound_seq_1, __last1, __first2, __last2, __result,
-                    __comp, [](_DifferenceType __n, _DifferenceType __m) { return __n + __m; }, __set_union_op,
-                    __is_vector);
+                    __comp, [](_DifferenceType __n, _DifferenceType __m) { return __n + __m; }, __set_union_op);
             });
         return __result;
     }
@@ -5446,16 +5446,17 @@ __parallel_set_union_op(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _
             },
             [=, &__result] {
                 __result = __internal::__parallel_set_op(
+                    __dispatch_tag,
                     ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __left_bound_seq_2, __last2, __result,
-                    __comp, [](_DifferenceType __n, _DifferenceType __m) { return __n + __m; }, __set_union_op,
-                    __is_vector);
+                    __comp, [](_DifferenceType __n, _DifferenceType __m) { return __n + __m; }, __set_union_op);
             });
         return __result;
     }
 
     return __internal::__parallel_set_op(
+        __dispatch_tag,
         ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2, __result, __comp,
-        [](_DifferenceType __n, _DifferenceType __m) { return __n + __m; }, __set_union_op, __is_vector);
+        [](_DifferenceType __n, _DifferenceType __m) { return __n + __m; }, __set_union_op);
 }
 
 //------------------------------------------------------------------------
@@ -5622,6 +5623,10 @@ __pattern_set_intersection(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __f
                            _RandomAccessIterator3 __result, _Compare __comp, _IsVector __is_vector,
                            /*is_parallel=*/::std::true_type)
 {
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2,
+                                                  _RandomAccessIterator3>();
+
     typedef typename ::std::iterator_traits<_RandomAccessIterator3>::value_type _T;
     typedef typename ::std::iterator_traits<_RandomAccessIterator1>::difference_type _DifferenceType;
 
@@ -5649,14 +5654,14 @@ __pattern_set_intersection(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __f
     {
         //we know proper offset due to [first1; left_bound_seq_1) < [first2; last2)
         return __internal::__parallel_set_op(
+            __dispatch_tag,
             ::std::forward<_ExecutionPolicy>(__exec), __left_bound_seq_1, __last1, __first2, __last2, __result, __comp,
             [](_DifferenceType __n, _DifferenceType __m) { return ::std::min(__n, __m); },
             [](_RandomAccessIterator1 __first1, _RandomAccessIterator1 __last1, _RandomAccessIterator2 __first2,
                _RandomAccessIterator2 __last2, _T* __result, _Compare __comp) {
                 return oneapi::dpl::__utils::__set_intersection_construct(__first1, __last1, __first2, __last2,
                                                                           __result, __comp);
-            },
-            __is_vector);
+            });
     }
 
     const auto __m2 = __last2 - __left_bound_seq_2 + __n1;
@@ -5664,14 +5669,14 @@ __pattern_set_intersection(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __f
     {
         //we know proper offset due to [first2; left_bound_seq_2) < [first1; last1)
         __result = __internal::__parallel_set_op(
+            __dispatch_tag,
             ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __left_bound_seq_2, __last2, __result, __comp,
             [](_DifferenceType __n, _DifferenceType __m) { return ::std::min(__n, __m); },
             [](_RandomAccessIterator1 __first1, _RandomAccessIterator1 __last1, _RandomAccessIterator2 __first2,
                _RandomAccessIterator2 __last2, _T* __result, _Compare __comp) {
                 return oneapi::dpl::__utils::__set_intersection_construct(__first2, __last2, __first1, __last1,
                                                                           __result, __comp);
-            },
-            __is_vector);
+            });
         return __result;
     }
 
@@ -5713,14 +5718,14 @@ __pattern_set_intersection(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& _
     {
         //we know proper offset due to [first1; left_bound_seq_1) < [first2; last2)
         return __internal::__parallel_set_op(
+            __tag,
             ::std::forward<_ExecutionPolicy>(__exec), __left_bound_seq_1, __last1, __first2, __last2, __result, __comp,
             [](_DifferenceType __n, _DifferenceType __m) { return ::std::min(__n, __m); },
             [](_RandomAccessIterator1 __first1, _RandomAccessIterator1 __last1, _RandomAccessIterator2 __first2,
                _RandomAccessIterator2 __last2, _T* __result, _Compare __comp) {
                 return oneapi::dpl::__utils::__set_intersection_construct(__first1, __last1, __first2, __last2,
                                                                           __result, __comp);
-            },
-            _IsVector{});
+            });
     }
 
     const auto __m2 = __last2 - __left_bound_seq_2 + __n1;
@@ -5728,14 +5733,14 @@ __pattern_set_intersection(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& _
     {
         //we know proper offset due to [first2; left_bound_seq_2) < [first1; last1)
         __result = __internal::__parallel_set_op(
+            __tag,
             ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __left_bound_seq_2, __last2, __result, __comp,
             [](_DifferenceType __n, _DifferenceType __m) { return ::std::min(__n, __m); },
             [](_RandomAccessIterator1 __first1, _RandomAccessIterator1 __last1, _RandomAccessIterator2 __first2,
                _RandomAccessIterator2 __last2, _T* __result, _Compare __comp) {
                 return oneapi::dpl::__utils::__set_intersection_construct(__first2, __last2, __first1, __last1,
                                                                           __result, __comp);
-            },
-            _IsVector{});
+            });
         return __result;
     }
 
@@ -5800,7 +5805,9 @@ __pattern_set_difference(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __fir
     typedef typename ::std::iterator_traits<_RandomAccessIterator3>::value_type _T;
     typedef typename ::std::iterator_traits<_RandomAccessIterator1>::difference_type _DifferenceType;
 
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2, _RandomAccessIterator3>();
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2,
+                                                  _RandomAccessIterator3>();
 
     const auto __n1 = __last1 - __first1;
     const auto __n2 = __last2 - __first2;
@@ -5830,14 +5837,14 @@ __pattern_set_difference(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __fir
 
     if (__n1 + __n2 > __set_algo_cut_off)
         return __parallel_set_op(
+            __dispatch_tag,
             ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2, __result, __comp,
             [](_DifferenceType __n, _DifferenceType) { return __n; },
             [](_RandomAccessIterator1 __first1, _RandomAccessIterator1 __last1, _RandomAccessIterator2 __first2,
                _RandomAccessIterator2 __last2, _T* __result, _Compare __comp) {
                 return oneapi::dpl::__utils::__set_difference_construct(__first1, __last1, __first2, __last2, __result,
                                                                         __comp, __BrickCopyConstruct<_IsVector>());
-            },
-            __is_vector);
+            });
 
     // use serial algorithm
     return ::std::set_difference(__first1, __last1, __first2, __last2, __result, __comp);
@@ -5882,14 +5889,14 @@ __pattern_set_difference(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __e
 
     if (__n1 + __n2 > __set_algo_cut_off)
         return __parallel_set_op(
+            __tag,
             ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2, __result, __comp,
             [](_DifferenceType __n, _DifferenceType) { return __n; },
             [](_RandomAccessIterator1 __first1, _RandomAccessIterator1 __last1, _RandomAccessIterator2 __first2,
                _RandomAccessIterator2 __last2, _T* __result, _Compare __comp) {
                 return oneapi::dpl::__utils::__set_difference_construct(__first1, __last1, __first2, __last2, __result,
                                                                         __comp, __BrickCopyConstruct<_IsVector>());
-            },
-            _IsVector{});
+            });
 
     // use serial algorithm
     return ::std::set_difference(__first1, __last1, __first2, __last2, __result, __comp);

From c6998fa5b12b0a5101a560087765e098f896dcd4 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 15 Feb 2024 12:31:37 +0100
Subject: [PATCH 216/566] __pattern_set_union + tag impls

---
 include/oneapi/dpl/pstl/algorithm_impl.h | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index db217b7b0a4..8ff4c73d8e6 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -5546,7 +5546,7 @@ __pattern_set_union(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1,
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
           class _OutputIterator, class _Compare>
 _OutputIterator
-__pattern_set_union(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1,
+__pattern_set_union(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1,
                     _RandomAccessIterator1 __last1, _RandomAccessIterator2 __first2, _RandomAccessIterator2 __last2,
                     _OutputIterator __result, _Compare __comp)
 {
@@ -5980,7 +5980,7 @@ __pattern_set_symmetric_difference(_ExecutionPolicy&& __exec, _RandomAccessItera
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
           class _RandomAccessIterator3, class _Compare>
 _RandomAccessIterator3
-__pattern_set_symmetric_difference(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec,
+__pattern_set_symmetric_difference(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec,
                                    _RandomAccessIterator1 __first1, _RandomAccessIterator1 __last1,
                                    _RandomAccessIterator2 __first2, _RandomAccessIterator2 __last2,
                                    _RandomAccessIterator3 __result, _Compare __comp)

From a2b414e70b632472d3d988e5f51d1b8c05063aab Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 15 Feb 2024 12:29:56 +0100
Subject: [PATCH 217/566] __internal::__parallel_set_union_op + tag impls

---
 include/oneapi/dpl/pstl/algorithm_impl.h | 113 +++++++++++++++++++++++
 1 file changed, 113 insertions(+)

diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 8ff4c73d8e6..796c743056c 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -5459,6 +5459,119 @@ __parallel_set_union_op(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _
         [](_DifferenceType __n, _DifferenceType __m) { return __n + __m; }, __set_union_op);
 }
 
+//a shared parallel pattern for '__pattern_set_union' and '__pattern_set_symmetric_difference'
+template <class _IsVector, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2,
+          class _OutputIterator, class _Compare, class _SetUnionOp>
+_OutputIterator
+__parallel_set_union_op(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _ForwardIterator1 __first1,
+                        _ForwardIterator1 __last1, _ForwardIterator2 __first2, _ForwardIterator2 __last2,
+                        _OutputIterator __result, _Compare __comp, _SetUnionOp __set_union_op)
+{
+    typedef typename ::std::iterator_traits<_ForwardIterator1>::difference_type _DifferenceType;
+
+    using __backend_tag = typename decltype(__tag)::__backend_tag;
+
+    const auto __n1 = __last1 - __first1;
+    const auto __n2 = __last2 - __first2;
+
+    __brick_copy<_ExecutionPolicy> __copy_range{};
+
+    // {1} {}: parallel copying just first sequence
+    if (__n2 == 0)
+        return __internal::__pattern_walk2_brick(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first1,
+                                                 __last1, __result, __copy_range);
+
+    // {} {2}: parallel copying justmake  second sequence
+    if (__n1 == 0)
+        return __internal::__pattern_walk2_brick(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first2,
+                                                 __last2, __result, __copy_range);
+
+    // testing  whether the sequences are intersected
+    _ForwardIterator1 __left_bound_seq_1 = ::std::lower_bound(__first1, __last1, *__first2, __comp);
+
+    if (__left_bound_seq_1 == __last1)
+    {
+        //{1} < {2}: seq2 is wholly greater than seq1, so, do parallel copying seq1 and seq2
+        __par_backend::__parallel_invoke(
+            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
+            [=] {
+                __internal::__pattern_walk2_brick(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first1,
+                                                  __last1, __result, __copy_range);
+            },
+            [=] {
+                __internal::__pattern_walk2_brick(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first2,
+                                                  __last2, __result + __n1, __copy_range);
+            });
+        return __result + __n1 + __n2;
+    }
+
+    // testing  whether the sequences are intersected
+    _ForwardIterator2 __left_bound_seq_2 = ::std::lower_bound(__first2, __last2, *__first1, __comp);
+
+    if (__left_bound_seq_2 == __last2)
+    {
+        //{2} < {1}: seq2 is wholly greater than seq1, so, do parallel copying seq1 and seq2
+        __par_backend::__parallel_invoke(
+            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
+            [=] {
+                __internal::__pattern_walk2_brick(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first2,
+                                                  __last2, __result, __copy_range);
+            },
+            [=] {
+                __internal::__pattern_walk2_brick(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first1,
+                                                  __last1, __result + __n2, __copy_range);
+            });
+        return __result + __n1 + __n2;
+    }
+
+    const auto __m1 = __left_bound_seq_1 - __first1;
+    if (__m1 > __set_algo_cut_off)
+    {
+        auto __res_or = __result;
+        __result += __m1; //we know proper offset due to [first1; left_bound_seq_1) < [first2; last2)
+        __par_backend::__parallel_invoke(
+            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
+            //do parallel copying of [first1; left_bound_seq_1)
+            [=] {
+                __internal::__pattern_walk2_brick(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first1,
+                                                  __left_bound_seq_1, __res_or, __copy_range);
+            },
+            [=, &__result] {
+                __result = __internal::__parallel_set_op(
+                    __tag, ::std::forward<_ExecutionPolicy>(__exec), __left_bound_seq_1, __last1, __first2,
+                    __last2, __result, __comp, [](_DifferenceType __n, _DifferenceType __m) { return __n + __m; },
+                    __set_union_op);
+            });
+        return __result;
+    }
+
+    const auto __m2 = __left_bound_seq_2 - __first2;
+    assert(__m1 == 0 || __m2 == 0);
+    if (__m2 > __set_algo_cut_off)
+    {
+        auto __res_or = __result;
+        __result += __m2; //we know proper offset due to [first2; left_bound_seq_2) < [first1; last1)
+        __par_backend::__parallel_invoke(
+            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
+            //do parallel copying of [first2; left_bound_seq_2)
+            [=] {
+                __internal::__pattern_walk2_brick(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first2,
+                                                  __left_bound_seq_2, __res_or, __copy_range);
+            },
+            [=, &__result] {
+                __result = __internal::__parallel_set_op(
+                    __tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __left_bound_seq_2,
+                    __last2, __result, __comp, [](_DifferenceType __n, _DifferenceType __m) { return __n + __m; },
+                    __set_union_op);
+            });
+        return __result;
+    }
+
+    return __internal::__parallel_set_op(
+        __tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2, __result,
+        __comp, [](_DifferenceType __n, _DifferenceType __m) { return __n + __m; }, __set_union_op);
+}
+
 //------------------------------------------------------------------------
 // set_union
 //------------------------------------------------------------------------

From 16f7eb5b814996c8e01ee90e57ecaac875f3696f Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 15 Feb 2024 12:31:11 +0100
Subject: [PATCH 218/566] __internal::__parallel_set_union_op + tag calls

---
 include/oneapi/dpl/pstl/algorithm_impl.h | 22 ++++++++++++++--------
 1 file changed, 14 insertions(+), 8 deletions(-)

diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 796c743056c..92ec5572b52 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -5637,6 +5637,9 @@ __pattern_set_union(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1,
                     _RandomAccessIterator2 __first2, _RandomAccessIterator2 __last2, _OutputIterator __result,
                     _Compare __comp, _IsVector __is_vector, /*__is_parallel=*/::std::true_type)
 {
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2,
+                                                  _OutputIterator>();
 
     const auto __n1 = __last1 - __first1;
     const auto __n2 = __last2 - __first2;
@@ -5647,13 +5650,13 @@ __pattern_set_union(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1,
 
     typedef typename ::std::iterator_traits<_OutputIterator>::value_type _Tp;
     return __parallel_set_union_op(
+        __dispatch_tag,
         ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2, __result, __comp,
         [](_RandomAccessIterator1 __first1, _RandomAccessIterator1 __last1, _RandomAccessIterator2 __first2,
            _RandomAccessIterator2 __last2, _Tp* __result, _Compare __comp) {
             return oneapi::dpl::__utils::__set_union_construct(__first1, __last1, __first2, __last2, __result, __comp,
                                                                __BrickCopyConstruct<_IsVector>());
-        },
-        __is_vector);
+        });
 }
 
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
@@ -5672,13 +5675,13 @@ __pattern_set_union(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec,
 
     typedef typename ::std::iterator_traits<_OutputIterator>::value_type _Tp;
     return __parallel_set_union_op(
+        __tag,
         ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2, __result, __comp,
         [](_RandomAccessIterator1 __first1, _RandomAccessIterator1 __last1, _RandomAccessIterator2 __first2,
            _RandomAccessIterator2 __last2, _Tp* __result, _Compare __comp) {
             return oneapi::dpl::__utils::__set_union_construct(__first1, __last1, __first2, __last2, __result, __comp,
                                                                __BrickCopyConstruct<_IsVector>());
-        },
-        _IsVector{});
+        });
 }
 
 //------------------------------------------------------------------------
@@ -6071,6 +6074,9 @@ __pattern_set_symmetric_difference(_ExecutionPolicy&& __exec, _RandomAccessItera
                                    _RandomAccessIterator2 __last2, _RandomAccessIterator3 __result, _Compare __comp,
                                    _IsVector __is_vector, /*is_parallel=*/::std::true_type)
 {
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2,
+                                                  _RandomAccessIterator3>();
 
     const auto __n1 = __last1 - __first1;
     const auto __n2 = __last2 - __first2;
@@ -6081,13 +6087,13 @@ __pattern_set_symmetric_difference(_ExecutionPolicy&& __exec, _RandomAccessItera
 
     typedef typename ::std::iterator_traits<_RandomAccessIterator3>::value_type _T;
     return __internal::__parallel_set_union_op(
+        __dispatch_tag,
         ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2, __result, __comp,
         [](_RandomAccessIterator1 __first1, _RandomAccessIterator1 __last1, _RandomAccessIterator2 __first2,
            _RandomAccessIterator2 __last2, _T* __result, _Compare __comp) {
             return oneapi::dpl::__utils::__set_symmetric_difference_construct(
                 __first1, __last1, __first2, __last2, __result, __comp, __BrickCopyConstruct<_IsVector>());
-        },
-        __is_vector);
+        });
 }
 
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
@@ -6107,13 +6113,13 @@ __pattern_set_symmetric_difference(__parallel_tag<_IsVector> __tag, _ExecutionPo
 
     typedef typename ::std::iterator_traits<_RandomAccessIterator3>::value_type _T;
     return __internal::__parallel_set_union_op(
+        __tag,
         ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2, __result, __comp,
         [](_RandomAccessIterator1 __first1, _RandomAccessIterator1 __last1, _RandomAccessIterator2 __first2,
            _RandomAccessIterator2 __last2, _T* __result, _Compare __comp) {
             return oneapi::dpl::__utils::__set_symmetric_difference_construct(
                 __first1, __last1, __first2, __last2, __result, __comp, __BrickCopyConstruct<_IsVector>());
-        },
-        _IsVector{});
+        });
 }
 
 //------------------------------------------------------------------------

From 9d151e62e2cf812bac84150df9a79d0254737cab Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 15 Feb 2024 14:05:24 +0100
Subject: [PATCH 219/566] __pattern_search + tag impls

---
 include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index c6e23ae6808..43a2d40c716 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -1422,7 +1422,7 @@ __pattern_search(_ExecutionPolicy&& __exec, _Iterator1 __first, _Iterator1 __las
 
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2, typename _Pred>
 _Iterator1
-__pattern_search(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Iterator1 __first, _Iterator1 __last,
+__pattern_search(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator1 __first, _Iterator1 __last,
                  _Iterator2 __s_first, _Iterator2 __s_last, _Pred __pred)
 {
     if (__s_last == __s_first)

From 94bcc7c0e5fa93e444e3f39ae28bde428ce39181 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 15 Feb 2024 13:55:57 +0100
Subject: [PATCH 220/566] Move declarations of __serial_backend_tag,
 __tbb_backend_tag and __omp_backend_tag

---
 include/oneapi/dpl/pstl/execution_defs.h | 16 ++++++++++++++++
 include/oneapi/dpl/pstl/execution_impl.h | 12 ------------
 2 files changed, 16 insertions(+), 12 deletions(-)

diff --git a/include/oneapi/dpl/pstl/execution_defs.h b/include/oneapi/dpl/pstl/execution_defs.h
index d16a030b216..5001ec9fd0f 100644
--- a/include/oneapi/dpl/pstl/execution_defs.h
+++ b/include/oneapi/dpl/pstl/execution_defs.h
@@ -213,6 +213,22 @@ __check_size(...) -> typename ::std::iterator_traits<_It>::difference_type;
 template <typename _R>
 using __difference_t = ::std::make_signed_t<decltype(__check_size<_R>(0))>;
 
+//------------------------------------------------------------------------
+// backend tags
+//------------------------------------------------------------------------
+
+struct __serial_backend_tag
+{
+};
+
+struct __tbb_backend_tag
+{
+};
+
+struct __omp_backend_tag
+{
+};
+
 } // namespace __internal
 
 } // namespace dpl
diff --git a/include/oneapi/dpl/pstl/execution_impl.h b/include/oneapi/dpl/pstl/execution_impl.h
index 33ada982ff7..dae807121cf 100644
--- a/include/oneapi/dpl/pstl/execution_impl.h
+++ b/include/oneapi/dpl/pstl/execution_impl.h
@@ -97,18 +97,6 @@ __is_parallelization_preferred()
 // backend selector with tags
 //------------------------------------------------------------------------
 
-struct __serial_backend_tag
-{
-};
-
-struct __tbb_backend_tag
-{
-};
-
-struct __omp_backend_tag
-{
-};
-
 #if _ONEDPL_PAR_BACKEND_TBB
 using __par_backend_tag = __tbb_backend_tag;
 #elif _ONEDPL_PAR_BACKEND_OPENMP

From cdc3cb91dcc09eac755b4dabb999c12ca9591e29 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 15 Feb 2024 14:42:52 +0100
Subject: [PATCH 221/566] include/oneapi/dpl/pstl/parallel_backend.h - fix
 compile error

---
 include/oneapi/dpl/pstl/parallel_backend.h | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/include/oneapi/dpl/pstl/parallel_backend.h b/include/oneapi/dpl/pstl/parallel_backend.h
index 1e78d1f635b..b243e8fb492 100644
--- a/include/oneapi/dpl/pstl/parallel_backend.h
+++ b/include/oneapi/dpl/pstl/parallel_backend.h
@@ -18,14 +18,14 @@
 
 // Select a parallel backend
 #if ONEDPL_USE_TBB_BACKEND || (!defined(ONEDPL_USE_TBB_BACKEND) && !ONEDPL_USE_OPENMP_BACKEND && _ONEDPL_TBB_AVAILABLE)
-#    include "parallel_backend_tbb.h"
 #    define _ONEDPL_PAR_BACKEND_TBB 1
+#    include "parallel_backend_tbb.h"
 #elif ONEDPL_USE_OPENMP_BACKEND || (!defined(ONEDPL_USE_OPENMP_BACKEND) && _ONEDPL_OPENMP_AVAILABLE)
-#    include "parallel_backend_omp.h"
 #    define _ONEDPL_PAR_BACKEND_OPENMP 1
+#    include "parallel_backend_omp.h"
 #else
-#    include "parallel_backend_serial.h"
 #    define _ONEDPL_PAR_BACKEND_SERIAL 1
+#    include "parallel_backend_serial.h"
 #endif
 
 #if _ONEDPL_BACKEND_SYCL

From f564fd904f62cc16e1ec5459588bf796a7d0fc02 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 15 Feb 2024 15:07:48 +0100
Subject: [PATCH 222/566] __pattern_is_heap_until + tag impls

---
 include/oneapi/dpl/pstl/algorithm_impl.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 92ec5572b52..9b85f235ddb 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -6206,7 +6206,7 @@ __pattern_is_heap_until(_ExecutionPolicy&& __exec, _RandomAccessIterator __first
 
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Compare>
 _RandomAccessIterator
-__pattern_is_heap_until(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
+__pattern_is_heap_until(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
                         _RandomAccessIterator __last, _Compare __comp)
 {
     return __internal::__except_handler([&]() {

From 347751509e2aae1bfc23ef51341f99097c4fe4b8 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 15 Feb 2024 16:16:58 +0100
Subject: [PATCH 223/566] Apply GitHUB clang format

---
 include/oneapi/dpl/pstl/algorithm_impl.h      | 359 ++++++++----------
 .../experimental/internal/for_loop_impl.h     |  24 +-
 include/oneapi/dpl/pstl/glue_algorithm_impl.h |  42 +-
 include/oneapi/dpl/pstl/glue_memory_impl.h    |  20 +-
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   |  93 ++---
 .../pstl/hetero/dpcpp/parallel_backend_sycl.h |  22 +-
 .../hetero/dpcpp/parallel_backend_sycl_fpga.h |  35 +-
 .../dpl/pstl/hetero/numeric_impl_hetero.h     |  10 +-
 include/oneapi/dpl/pstl/histogram_impl.h      |   3 +-
 include/oneapi/dpl/pstl/numeric_impl.h        |  12 +-
 include/oneapi/dpl/pstl/omp/parallel_merge.h  |   6 +-
 .../dpl/pstl/omp/parallel_stable_sort.h       |   1 -
 .../oneapi/dpl/pstl/parallel_backend_tbb.h    |   6 +-
 13 files changed, 283 insertions(+), 350 deletions(-)

diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 9b85f235ddb..8790cc64f25 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -194,8 +194,8 @@ __pattern_walk1(__parallel_forward_tag, _ExecutionPolicy&& __exec, _ForwardItera
 
 template <class _IsVector, class _ExecutionPolicy, class _ForwardIterator, class _Function>
 void
-__pattern_walk1(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _ForwardIterator __first,
-                _ForwardIterator __last, _Function __f)
+__pattern_walk1(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last,
+                _Function __f)
 {
     using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
 
@@ -348,7 +348,8 @@ __pattern_walk1_n(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _Siz
                   _IsVector __is_vector,
                   /*is_parallel=*/::std::true_type)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
 
     oneapi::dpl::__internal::__pattern_walk1(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first,
                                              __first + __n, __f);
@@ -360,8 +361,8 @@ _RandomAccessIterator
 __pattern_walk1_n(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _RandomAccessIterator __first, _Size __n,
                   _Function __f)
 {
-    oneapi::dpl::__internal::__pattern_walk1(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first,
-                                             __first + __n, __f);
+    oneapi::dpl::__internal::__pattern_walk1(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __first + __n,
+                                             __f);
     return __first + __n;
 }
 
@@ -499,7 +500,8 @@ __pattern_walk2(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1, _Ran
     });
 }
 
-template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _Function>
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
+          class _Function>
 _RandomAccessIterator2
 __pattern_walk2(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1,
                 _RandomAccessIterator1 __last1, _RandomAccessIterator2 __first2, _Function __f)
@@ -597,7 +599,8 @@ oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _Ra
 __pattern_walk2_n(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1, _Size __n,
                   _RandomAccessIterator2 __first2, _Function __f, _IsVector __is_vector, /*parallel=*/::std::true_type)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2>();
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2>();
 
     return __internal::__pattern_walk2(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first1,
                                        __first1 + __n, __first2, __f);
@@ -649,8 +652,7 @@ __pattern_walk2_brick(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1
 
     return __except_handler([&]() {
         __par_backend::__parallel_for(
-            __backend_tag{},
-            ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1,
+            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1,
             [&__is_vector, __first1, __first2, __brick](_RandomAccessIterator1 __i, _RandomAccessIterator1 __j) {
                 __brick(__i, __j, __first2 + (__i - __first1), __is_vector);
             });
@@ -749,8 +751,7 @@ __pattern_walk2_brick_n(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __firs
 
     return __except_handler([&]() {
         __par_backend::__parallel_for(
-            __backend_tag{},
-            ::std::forward<_ExecutionPolicy>(__exec), __first1, __first1 + __n,
+            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first1, __first1 + __n,
             [&__is_vector, __first1, __first2, __brick](_RandomAccessIterator1 __i, _RandomAccessIterator1 __j) {
                 __brick(__i, __j - __i, __first2 + (__i - __first1), __is_vector);
             });
@@ -758,7 +759,8 @@ __pattern_walk2_brick_n(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __firs
     });
 }
 
-template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _Size, class _RandomAccessIterator2, class _Brick>
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _Size,
+          class _RandomAccessIterator2, class _Brick>
 _RandomAccessIterator2
 __pattern_walk2_brick_n(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1,
                         _Size __n, _RandomAccessIterator2 __first2, _Brick __brick)
@@ -767,8 +769,7 @@ __pattern_walk2_brick_n(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _R
 
     return __except_handler([&]() {
         __par_backend::__parallel_for(
-            __backend_tag{},
-            ::std::forward<_ExecutionPolicy>(__exec), __first1, __first1 + __n,
+            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first1, __first1 + __n,
             [__first1, __first2, __brick](_RandomAccessIterator1 __i, _RandomAccessIterator1 __j) {
                 __brick(__i, __j - __i, __first2 + (__i - __first1), _IsVector{});
             });
@@ -858,8 +859,7 @@ __pattern_walk3(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1, _Ran
 
     return __internal::__except_handler([&]() {
         __par_backend::__parallel_for(
-            __backend_tag{},
-            ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1,
+            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1,
             [__f, __first1, __first2, __first3, __is_vector](_RandomAccessIterator1 __i, _RandomAccessIterator1 __j) {
                 __internal::__brick_walk3(__i, __j, __first2 + (__i - __first1), __first3 + (__i - __first1), __f,
                                           __is_vector);
@@ -871,15 +871,15 @@ __pattern_walk3(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1, _Ran
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
           class _RandomAccessIterator3, class _Function>
 _RandomAccessIterator3
-__pattern_walk3(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1, _RandomAccessIterator1 __last1,
-                _RandomAccessIterator2 __first2, _RandomAccessIterator3 __first3, _Function __f)
+__pattern_walk3(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1,
+                _RandomAccessIterator1 __last1, _RandomAccessIterator2 __first2, _RandomAccessIterator3 __first3,
+                _Function __f)
 {
     using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
 
     return __internal::__except_handler([&]() {
         __par_backend::__parallel_for(
-            __backend_tag{},
-            ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1,
+            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1,
             [__f, __first1, __first2, __first3](_RandomAccessIterator1 __i, _RandomAccessIterator1 __j) {
                 __internal::__brick_walk3(__i, __j, __first2 + (__i - __first1), __first3 + (__i - __first1), __f,
                                           _IsVector{});
@@ -969,8 +969,8 @@ __pattern_walk2_transform_if(_ExecutionPolicy&& __exec, _ForwardIterator1 __firs
 template <class _Tag, typename _ExecutionPolicy, typename _ForwardIterator1, typename _ForwardIterator2,
           typename _Function>
 _ForwardIterator2
-__pattern_walk2_transform_if(_Tag __tag, _ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
-                             _ForwardIterator2 __first2, _Function __func) noexcept
+__pattern_walk2_transform_if(_Tag __tag, _ExecutionPolicy&& __exec, _ForwardIterator1 __first1,
+                             _ForwardIterator1 __last1, _ForwardIterator2 __first2, _Function __func) noexcept
 {
     static_assert(__is_backend_tag_v<_Tag>);
 
@@ -1038,8 +1038,7 @@ __pattern_equal(_ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _BinaryPredicate>
 bool
 __pattern_equal(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
-                _ForwardIterator2 __first2,
-                _ForwardIterator2 __last2, _BinaryPredicate __p) noexcept
+                _ForwardIterator2 __first2, _ForwardIterator2 __last2, _BinaryPredicate __p) noexcept
 {
     static_assert(__is_backend_tag_v<_Tag>);
 
@@ -1081,8 +1080,7 @@ __pattern_equal(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _Ran
 
     return __internal::__except_handler([&]() {
         return !__internal::__parallel_or(
-            __tag,
-            ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1,
+            __tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1,
             [__first1, __first2, __p](_RandomAccessIterator1 __i, _RandomAccessIterator1 __j) {
                 return !__internal::__brick_equal(__i, __j, __first2 + (__i - __first1), __first2 + (__j - __first1),
                                                   __p, _IsVector{});
@@ -1090,7 +1088,6 @@ __pattern_equal(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _Ran
     });
 }
 
-
 //------------------------------------------------------------------------
 // equal version for sequences with equal length
 //------------------------------------------------------------------------
@@ -1143,8 +1140,7 @@ __pattern_equal(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1, _Ran
 
     return __internal::__except_handler([&]() {
         return !__internal::__parallel_or(
-            __dispatch_tag,
-            ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1,
+            __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1,
             [__first1, __first2, __p, __is_vector](_RandomAccessIterator1 __i, _RandomAccessIterator1 __j) {
                 return !__internal::__brick_equal(__i, __j, __first2 + (__i - __first1), __p, __is_vector);
             });
@@ -1343,15 +1339,14 @@ __pattern_find_end(_ExecutionPolicy&&, _ForwardIterator1 __first, _ForwardIterat
 
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _BinaryPredicate>
 _ForwardIterator1
-__pattern_find_end(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first, _ForwardIterator1 __last, _ForwardIterator2 __s_first,
-                   _ForwardIterator2 __s_last, _BinaryPredicate __pred) noexcept
+__pattern_find_end(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first, _ForwardIterator1 __last,
+                   _ForwardIterator2 __s_first, _ForwardIterator2 __s_last, _BinaryPredicate __pred) noexcept
 {
     static_assert(__is_backend_tag_v<_Tag>);
 
     return __internal::__brick_find_end(__first, __last, __s_first, __s_last, __pred, typename _Tag::__is_vector{});
 }
 
-
 template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _BinaryPredicate,
           class _IsVector>
 oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator1>
@@ -1388,13 +1383,13 @@ template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1,
           class _BinaryPredicate>
 _RandomAccessIterator1
 __pattern_find_end(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _RandomAccessIterator1 __first,
-                   _RandomAccessIterator1 __last,
-                   _RandomAccessIterator2 __s_first, _RandomAccessIterator2 __s_last, _BinaryPredicate __pred)
+                   _RandomAccessIterator1 __last, _RandomAccessIterator2 __s_first, _RandomAccessIterator2 __s_last,
+                   _BinaryPredicate __pred)
 {
     if (__last - __first == __s_last - __s_first)
     {
-        const bool __res = __internal::__pattern_equal(__tag, ::std::forward<_ExecutionPolicy>(__exec),
-                                                       __first, __last, __s_first, __pred);
+        const bool __res = __internal::__pattern_equal(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+                                                       __s_first, __pred);
         return __res ? __first : __last;
     }
     else
@@ -1411,7 +1406,6 @@ __pattern_find_end(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _
     }
 }
 
-
 //------------------------------------------------------------------------
 // find_first_of
 //------------------------------------------------------------------------
@@ -1686,7 +1680,6 @@ __pattern_search_n(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _
     }
 }
 
-
 //------------------------------------------------------------------------
 // copy_n
 //------------------------------------------------------------------------
@@ -1992,8 +1985,7 @@ __pattern_copy_if(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first, _Ra
             bool* __mask = __mask_buf.get();
             _DifferenceType __m{};
             __par_backend::__parallel_strict_scan(
-                __backend_tag{},
-                ::std::forward<_ExecutionPolicy>(__exec), __n, _DifferenceType(0),
+                __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __n, _DifferenceType(0),
                 [=](_DifferenceType __i, _DifferenceType __len) { // Reduce
                     return __internal::__brick_calc_mask_1<_DifferenceType>(__first + __i, __first + (__i + __len),
                                                                             __mask + __i, __pred, __is_vector)
@@ -2030,8 +2022,7 @@ __pattern_copy_if(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomA
             bool* __mask = __mask_buf.get();
             _DifferenceType __m{};
             __par_backend::__parallel_strict_scan(
-                __backend_tag{},
-                ::std::forward<_ExecutionPolicy>(__exec), __n, _DifferenceType(0),
+                __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __n, _DifferenceType(0),
                 [=](_DifferenceType __i, _DifferenceType __len) { // Reduce
                     return __internal::__brick_calc_mask_1<_DifferenceType>(__first + __i, __first + (__i + __len),
                                                                             __mask + __i, __pred, _IsVector{})
@@ -2168,7 +2159,8 @@ __pattern_unique(_ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator
 
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _BinaryPredicate>
 _ForwardIterator
-__pattern_unique(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _BinaryPredicate __pred) noexcept
+__pattern_unique(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
+                 _BinaryPredicate __pred) noexcept
 {
     static_assert(__is_backend_tag_v<_Tag>);
 
@@ -2231,8 +2223,7 @@ __remove_elements(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardI
         _DifferenceType __m{};
         // 2. Elements that doesn't satisfy pred are moved to result
         __par_backend::__parallel_strict_scan(
-            __backend_tag{},
-            ::std::forward<_ExecutionPolicy>(__exec), __n, _DifferenceType(0),
+            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __n, _DifferenceType(0),
             [__mask, __is_vector](_DifferenceType __i, _DifferenceType __len) {
                 return __internal::__brick_count(
                     __mask + __i, __mask + __i + __len, [](bool __val) { return __val; }, __is_vector);
@@ -2400,8 +2391,7 @@ __pattern_unique_copy(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first,
                 bool* __mask = __mask_buf.get();
                 _DifferenceType __m{};
                 __par_backend::__parallel_strict_scan(
-                    __backend_tag{},
-                    ::std::forward<_ExecutionPolicy>(__exec), __n, _DifferenceType(0),
+                    __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __n, _DifferenceType(0),
                     [=](_DifferenceType __i, _DifferenceType __len) -> _DifferenceType { // Reduce
                         _DifferenceType __extra = 0;
                         if (__i == 0)
@@ -2452,8 +2442,7 @@ __pattern_unique_copy(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Ran
                 bool* __mask = __mask_buf.get();
                 _DifferenceType __m{};
                 __par_backend::__parallel_strict_scan(
-                    __backend_tag{},
-                    ::std::forward<_ExecutionPolicy>(__exec), __n, _DifferenceType(0),
+                    __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __n, _DifferenceType(0),
                     [=](_DifferenceType __i, _DifferenceType __len) -> _DifferenceType { // Reduce
                         _DifferenceType __extra = 0;
                         if (__i == 0)
@@ -2728,7 +2717,8 @@ __pattern_rotate(_ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator
 
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator>
 _ForwardIterator
-__pattern_rotate(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __middle, _ForwardIterator __last) noexcept
+__pattern_rotate(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __middle,
+                 _ForwardIterator __last) noexcept
 {
     static_assert(__is_backend_tag_v<_Tag>);
 
@@ -2920,8 +2910,7 @@ __pattern_rotate_copy(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first,
     using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
 
     __par_backend::__parallel_for(
-        __backend_tag{},
-        ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+        __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
         [__first, __last, __middle, __result, __is_vector](_RandomAccessIterator1 __b, _RandomAccessIterator1 __e) {
             __internal::__brick_copy<_ExecutionPolicy> __copy{};
             if (__b > __middle)
@@ -2953,8 +2942,7 @@ __pattern_rotate_copy(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Ran
     using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
 
     __par_backend::__parallel_for(
-        __backend_tag{},
-        ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+        __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
         [__first, __last, __middle, __result](_RandomAccessIterator1 __b, _RandomAccessIterator1 __e) {
             __internal::__brick_copy<_ExecutionPolicy> __copy{};
             if (__b > __middle)
@@ -3027,7 +3015,8 @@ __pattern_is_partitioned(_ExecutionPolicy&&, _ForwardIterator __first, _ForwardI
 
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _UnaryPredicate>
 bool
-__pattern_is_partitioned(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _UnaryPredicate __pred) noexcept
+__pattern_is_partitioned(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
+                         _UnaryPredicate __pred) noexcept
 {
     static_assert(__is_backend_tag_v<_Tag>);
 
@@ -3183,7 +3172,7 @@ __pattern_is_partitioned(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _
         _ReduceType __result = __par_backend::__parallel_reduce(
             __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __identity,
             [&__pred, __combine](_RandomAccessIterator __i, _RandomAccessIterator __j,
-                                              _ReduceType __value) -> _ReduceType {
+                                 _ReduceType __value) -> _ReduceType {
                 if (__value.__val == __broken)
                     return _ReduceType{__broken, __i};
 
@@ -3312,8 +3301,7 @@ __pattern_partition(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _R
             else if (__size2 > __size1)
             {
                 __par_backend::__parallel_for(
-                    __backend_tag{},
-                    ::std::forward<_ExecutionPolicy>(__exec), __val1.__pivot, __val1.__pivot + __size1,
+                    __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __val1.__pivot, __val1.__pivot + __size1,
                     [__val1, __val2, __size1, __is_vector](_RandomAccessIterator __i, _RandomAccessIterator __j) {
                         __internal::__brick_swap_ranges(__i, __j, (__val2.__pivot - __size1) + (__i - __val1.__pivot),
                                                         __is_vector);
@@ -3324,8 +3312,7 @@ __pattern_partition(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _R
             else
             {
                 __par_backend::__parallel_for(
-                    __backend_tag{},
-                    ::std::forward<_ExecutionPolicy>(__exec), __val1.__pivot, __val1.__pivot + __size2,
+                    __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __val1.__pivot, __val1.__pivot + __size2,
                     [__val1, __val2, __is_vector](_RandomAccessIterator __i, _RandomAccessIterator __j) {
                         __internal::__brick_swap_ranges(__i, __j, __val2.__begin + (__i - __val1.__pivot), __is_vector);
                     });
@@ -3334,8 +3321,7 @@ __pattern_partition(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _R
         };
 
         _PartitionRange __result = __par_backend::__parallel_reduce(
-            __backend_tag{},
-            ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __init,
+            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __init,
             [__pred, __is_vector, __reductor](_RandomAccessIterator __i, _RandomAccessIterator __j,
                                               _PartitionRange __value) -> _PartitionRange {
                 //1. serial partition
@@ -3381,12 +3367,13 @@ __pattern_partition(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Rando
             // then we should swap the false part of left range and last part of true part of right range
             else if (__size2 > __size1)
             {
-                constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, decltype(__val1.__pivot), decltype(__val1.__pivot + __size1)>();
+                constexpr auto __dispatch_tag =
+                    oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, decltype(__val1.__pivot),
+                                                              decltype(__val1.__pivot + __size1)>();
                 using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
 
                 __par_backend::__parallel_for(
-                    __backend_tag{},
-                    ::std::forward<_ExecutionPolicy>(__exec), __val1.__pivot, __val1.__pivot + __size1,
+                    __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __val1.__pivot, __val1.__pivot + __size1,
                     [__val1, __val2, __size1](_RandomAccessIterator __i, _RandomAccessIterator __j) {
                         __internal::__brick_swap_ranges(__i, __j, (__val2.__pivot - __size1) + (__i - __val1.__pivot),
                                                         _IsVector{});
@@ -3396,12 +3383,13 @@ __pattern_partition(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Rando
             // else we should swap the first part of false part of left range and true part of right range
             else
             {
-                constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, decltype(__val1.__pivot), decltype(__val1.__pivot + __size2)>();
+                constexpr auto __dispatch_tag =
+                    oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, decltype(__val1.__pivot),
+                                                              decltype(__val1.__pivot + __size2)>();
                 using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
 
                 __par_backend::__parallel_for(
-                    __backend_tag{},
-                    ::std::forward<_ExecutionPolicy>(__exec), __val1.__pivot, __val1.__pivot + __size2,
+                    __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __val1.__pivot, __val1.__pivot + __size2,
                     [__val1, __val2](_RandomAccessIterator __i, _RandomAccessIterator __j) {
                         __internal::__brick_swap_ranges(__i, __j, __val2.__begin + (__i - __val1.__pivot), _IsVector{});
                     });
@@ -3510,8 +3498,7 @@ __pattern_stable_partition(_ExecutionPolicy&& __exec, _RandomAccessIterator __fi
         using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
 
         _PartitionRange __result = __par_backend::__parallel_reduce(
-            __backend_tag{},
-            ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __init,
+            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __init,
             [&__pred, __is_vector, __reductor](_RandomAccessIterator __i, _RandomAccessIterator __j,
                                                _PartitionRange __value) -> _PartitionRange {
                 //1. serial stable_partition
@@ -3653,8 +3640,8 @@ __pattern_partition_copy(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __fir
             bool* __mask = __mask_buf.get();
             _ReturnType __m{};
             __par_backend::__parallel_strict_scan(
-                __backend_tag{},
-                ::std::forward<_ExecutionPolicy>(__exec), __n, ::std::make_pair(_DifferenceType(0), _DifferenceType(0)),
+                __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __n,
+                ::std::make_pair(_DifferenceType(0), _DifferenceType(0)),
                 [=](_DifferenceType __i, _DifferenceType __len) { // Reduce
                     return __internal::__brick_calc_mask_1<_DifferenceType>(__first + __i, __first + (__i + __len),
                                                                             __mask + __i, __pred, __is_vector);
@@ -3694,8 +3681,8 @@ __pattern_partition_copy(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _
             bool* __mask = __mask_buf.get();
             _ReturnType __m{};
             __par_backend::__parallel_strict_scan(
-                __backend_tag{},
-                ::std::forward<_ExecutionPolicy>(__exec), __n, ::std::make_pair(_DifferenceType(0), _DifferenceType(0)),
+                __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __n,
+                ::std::make_pair(_DifferenceType(0), _DifferenceType(0)),
                 [=](_DifferenceType __i, _DifferenceType __len) { // Reduce
                     return __internal::__brick_calc_mask_1<_DifferenceType>(__first + __i, __first + (__i + __len),
                                                                             __mask + __i, __pred, _IsVector{});
@@ -3750,8 +3737,7 @@ __pattern_sort(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _Random
 
     __internal::__except_handler([&]() {
         __par_backend::__parallel_stable_sort(
-            __backend_tag{},
-            ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __comp,
+            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __comp,
             [](_RandomAccessIterator __first, _RandomAccessIterator __last, _Compare __comp) {
                 ::std::sort(__first, __last, __comp);
             },
@@ -3769,8 +3755,7 @@ __pattern_sort(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAcce
 
     __internal::__except_handler([&]() {
         __par_backend::__parallel_stable_sort(
-            __backend_tag{},
-            ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __comp,
+            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __comp,
             [](_RandomAccessIterator __first, _RandomAccessIterator __last, _Compare __comp) {
                 ::std::sort(__first, __last, __comp);
             },
@@ -3792,7 +3777,8 @@ __pattern_stable_sort(_ExecutionPolicy&&, _RandomAccessIterator __first, _Random
 
 template <class _Tag, class _ExecutionPolicy, class _RandomAccessIterator, class _Compare>
 void
-__pattern_stable_sort(_Tag, _ExecutionPolicy&&, _RandomAccessIterator __first, _RandomAccessIterator __last, _Compare __comp) noexcept
+__pattern_stable_sort(_Tag, _ExecutionPolicy&&, _RandomAccessIterator __first, _RandomAccessIterator __last,
+                      _Compare __comp) noexcept
 {
     static_assert(__is_backend_tag_v<_Tag>);
 
@@ -3810,8 +3796,7 @@ __pattern_stable_sort(_ExecutionPolicy&& __exec, _RandomAccessIterator __first,
 
     __internal::__except_handler([&]() {
         __par_backend::__parallel_stable_sort(
-            __backend_tag{},
-            ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __comp,
+            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __comp,
             [](_RandomAccessIterator __first, _RandomAccessIterator __last, _Compare __comp) {
                 ::std::stable_sort(__first, __last, __comp);
             },
@@ -3828,8 +3813,7 @@ __pattern_stable_sort(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Ran
 
     __internal::__except_handler([&]() {
         __par_backend::__parallel_stable_sort(
-            __backend_tag{},
-            ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __comp,
+            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __comp,
             [](_RandomAccessIterator __first, _RandomAccessIterator __last, _Compare __comp) {
                 ::std::stable_sort(__first, __last, __comp);
             },
@@ -3899,8 +3883,7 @@ __pattern_sort_by_key(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __keys_f
 
     __internal::__except_handler([&]() {
         __par_backend::__parallel_stable_sort(
-            __backend_tag{},
-            ::std::forward<_ExecutionPolicy>(__exec), __beg, __end, __cmp_f,
+            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __beg, __end, __cmp_f,
             [](auto __first, auto __last, auto __cmp) { ::std::sort(__first, __last, __cmp); }, __end - __beg);
     });
 }
@@ -3926,8 +3909,7 @@ __pattern_sort_by_key(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Ran
 
     __internal::__except_handler([&]() {
         __par_backend::__parallel_stable_sort(
-            __backend_tag{},
-            ::std::forward<_ExecutionPolicy>(__exec), __beg, __end, __cmp_f,
+            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __beg, __end, __cmp_f,
             [](auto __first, auto __last, auto __cmp) { ::std::sort(__first, __last, __cmp); }, __end - __beg);
     });
 }
@@ -3970,8 +3952,7 @@ __pattern_partial_sort(_ExecutionPolicy&& __exec, _RandomAccessIterator __first,
 
     __except_handler([&]() {
         __par_backend::__parallel_stable_sort(
-            __backend_tag{},
-            ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __comp,
+            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __comp,
             [__n](_RandomAccessIterator __begin, _RandomAccessIterator __end, _Compare __comp) {
                 if (__n < __end - __begin)
                     ::std::partial_sort(__begin, __begin + __n, __end, __comp);
@@ -3995,8 +3976,7 @@ __pattern_partial_sort(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Ra
 
     __except_handler([&]() {
         __par_backend::__parallel_stable_sort(
-            __backend_tag{},
-            ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __comp,
+            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __comp,
             [__n](_RandomAccessIterator __begin, _RandomAccessIterator __end, _Compare __comp) {
                 if (__n < __end - __begin)
                     ::std::partial_sort(__begin, __begin + __n, __end, __comp);
@@ -4052,8 +4032,7 @@ __pattern_partial_sort_copy(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __
         if (__n2 >= __n1)
         {
             __par_backend::__parallel_stable_sort(
-                __backend_tag{},
-                ::std::forward<_ExecutionPolicy>(__exec), __d_first, __d_first + __n1, __comp,
+                __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __d_first, __d_first + __n1, __comp,
                 [__first, __d_first, __is_vector](_RandomAccessIterator2 __i, _RandomAccessIterator2 __j,
                                                   _Compare __comp) {
                     _RandomAccessIterator1 __i1 = __first + (__i - __d_first);
@@ -4075,8 +4054,7 @@ __pattern_partial_sort_copy(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __
             _T1* __r = __buf.get();
 
             __par_backend::__parallel_stable_sort(
-                __backend_tag{},
-                ::std::forward<_ExecutionPolicy>(__exec), __r, __r + __n1, __comp,
+                __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __r, __r + __n1, __comp,
                 [__n2, __first, __r](_T1* __i, _T1* __j, _Compare __comp) {
                     _RandomAccessIterator1 __it = __first + (__i - __r);
 
@@ -4130,8 +4108,7 @@ __pattern_partial_sort_copy(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec
         if (__n2 >= __n1)
         {
             __par_backend::__parallel_stable_sort(
-                __backend_tag{},
-                ::std::forward<_ExecutionPolicy>(__exec), __d_first, __d_first + __n1, __comp,
+                __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __d_first, __d_first + __n1, __comp,
                 [__first, __d_first](_RandomAccessIterator2 __i, _RandomAccessIterator2 __j, _Compare __comp) {
                     _RandomAccessIterator1 __i1 = __first + (__i - __d_first);
                     _RandomAccessIterator1 __j1 = __first + (__j - __d_first);
@@ -4152,8 +4129,7 @@ __pattern_partial_sort_copy(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec
             _T1* __r = __buf.get();
 
             __par_backend::__parallel_stable_sort(
-                __backend_tag{},
-                ::std::forward<_ExecutionPolicy>(__exec), __r, __r + __n1, __comp,
+                __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __r, __r + __n1, __comp,
                 [__n2, __first, __r](_T1* __i, _T1* __j, _Compare __comp) {
                     _RandomAccessIterator1 __it = __first + (__i - __r);
 
@@ -4240,8 +4216,7 @@ __pattern_adjacent_find(_ExecutionPolicy&& __exec, _RandomAccessIterator __first
 
     return __internal::__except_handler([&]() {
         return __par_backend::__parallel_reduce(
-            __backend_tag{},
-            ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __last,
+            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __last,
             [__last, __pred, __is_vector, __or_semantic](_RandomAccessIterator __begin, _RandomAccessIterator __end,
                                                          _RandomAccessIterator __value) -> _RandomAccessIterator {
                 // TODO: investigate performance benefits from the use of shared variable for the result,
@@ -4289,8 +4264,7 @@ __pattern_adjacent_find(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _R
 
     return __internal::__except_handler([&]() {
         return __par_backend::__parallel_reduce(
-            __backend_tag{},
-            ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __last,
+            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __last,
             [__last, __pred, __or_semantic](_RandomAccessIterator __begin, _RandomAccessIterator __end,
                                             _RandomAccessIterator __value) -> _RandomAccessIterator {
                 // TODO: investigate performance benefits from the use of shared variable for the result,
@@ -4358,7 +4332,8 @@ __pattern_nth_element(_ExecutionPolicy&& __exec, _RandomAccessIterator __first,
         return;
     }
 
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
 
     using ::std::iter_swap;
     typedef typename ::std::iterator_traits<_RandomAccessIterator>::value_type _Tp;
@@ -4487,8 +4462,7 @@ __pattern_fill(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _Random
 
     return __internal::__except_handler([&__exec, __first, __last, &__value, __is_vector]() {
         __par_backend::__parallel_for(
-            __backend_tag{},
-            ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
             [&__value, __is_vector](_RandomAccessIterator __begin, _RandomAccessIterator __end) {
                 __internal::__brick_fill<_Tp, _ExecutionPolicy>{__value}(__begin, __end, __is_vector);
             });
@@ -4558,7 +4532,8 @@ _RandomAccessIterator
 __pattern_fill_n(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _Size __count, const _Tp& __value,
                  /*is_parallel=*/::std::true_type, _IsVector /*__is_vector*/)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
 
     return __internal::__pattern_fill(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first,
                                       __first + __count, __value);
@@ -4820,8 +4795,9 @@ __pattern_merge(_ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _OutputIterator,
           class _Compare>
 _OutputIterator
-__pattern_merge(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1 __last1, _ForwardIterator2 __first2,
-                _ForwardIterator2 __last2, _OutputIterator __d_first, _Compare __comp) noexcept
+__pattern_merge(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
+                _ForwardIterator2 __first2, _ForwardIterator2 __last2, _OutputIterator __d_first,
+                _Compare __comp) noexcept
 {
     static_assert(__is_backend_tag_v<_Tag>);
 
@@ -4842,8 +4818,8 @@ __pattern_merge(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1, _Ran
     using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
 
     __par_backend::__parallel_merge(
-        __backend_tag{},
-        ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2, __d_first, __comp,
+        __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2, __d_first,
+        __comp,
         [__is_vector](_RandomAccessIterator1 __f1, _RandomAccessIterator1 __l1, _RandomAccessIterator2 __f2,
                       _RandomAccessIterator2 __l2, _RandomAccessIterator3 __f3, _Compare __comp) {
             return __internal::__brick_merge(__f1, __l1, __f2, __l2, __f3, __comp, __is_vector);
@@ -4861,8 +4837,8 @@ __pattern_merge(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAcc
     using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
 
     __par_backend::__parallel_merge(
-        __backend_tag{},
-        ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2, __d_first, __comp,
+        __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2, __d_first,
+        __comp,
         [](_RandomAccessIterator1 __f1, _RandomAccessIterator1 __l1, _RandomAccessIterator2 __f2,
            _RandomAccessIterator2 __l2, _RandomAccessIterator3 __f3,
            _Compare __comp) { return __internal::__brick_merge(__f1, __l1, __f2, __l2, __f3, __comp, _IsVector{}); });
@@ -4940,8 +4916,7 @@ __pattern_inplace_merge(_ExecutionPolicy&& __exec, _RandomAccessIterator __first
         };
 
         __par_backend::__parallel_merge(
-            __backend_tag{},
-            ::std::forward<_ExecutionPolicy>(__exec), __first, __middle, __middle, __last, __r, __comp,
+            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __middle, __middle, __last, __r, __comp,
             [__n, __move_values, __move_sequences](_RandomAccessIterator __f1, _RandomAccessIterator __l1,
                                                    _RandomAccessIterator __f2, _RandomAccessIterator __l2, _Tp* __f3,
                                                    _Compare __comp) {
@@ -4986,8 +4961,7 @@ __pattern_inplace_merge(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _R
         };
 
         __par_backend::__parallel_merge(
-            __backend_tag{},
-            ::std::forward<_ExecutionPolicy>(__exec), __first, __middle, __middle, __last, __r, __comp,
+            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __middle, __middle, __last, __r, __comp,
             [__n, __move_values, __move_sequences](_RandomAccessIterator __f1, _RandomAccessIterator __l1,
                                                    _RandomAccessIterator __f2, _RandomAccessIterator __l2, _Tp* __f3,
                                                    _Compare __comp) {
@@ -4997,7 +4971,8 @@ __pattern_inplace_merge(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _R
             });
         __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __r, __r + __n,
                                       [__r, __first](_Tp* __i, _Tp* __j) {
-                                          __brick_move_destroy<_ExecutionPolicy>{}(__i, __j, __first + (__i - __r), _IsVector{});
+                                          __brick_move_destroy<_ExecutionPolicy>{}(__i, __j, __first + (__i - __r),
+                                                                                   _IsVector{});
                                       });
     });
 }
@@ -5055,8 +5030,7 @@ __pattern_includes(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1, _
 
     return __internal::__except_handler([&]() {
         return !__internal::__parallel_or(
-            __dispatch_tag,
-            ::std::forward<_ExecutionPolicy>(__exec), __first2, __last2,
+            __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first2, __last2,
             [__first1, __last1, __first2, __last2, &__comp](_RandomAccessIterator2 __i, _RandomAccessIterator2 __j) {
                 assert(__j > __i);
                 //assert(__j - __i > 1);
@@ -5091,7 +5065,8 @@ __pattern_includes(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1, _
     });
 }
 
-template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _Compare>
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
+          class _Compare>
 bool
 __pattern_includes(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1,
                    _RandomAccessIterator1 __last1, _RandomAccessIterator2 __first2, _RandomAccessIterator2 __last2,
@@ -5117,8 +5092,7 @@ __pattern_includes(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _
 
     return __internal::__except_handler([&]() {
         return !__internal::__parallel_or(
-            __tag,
-            ::std::forward<_ExecutionPolicy>(__exec), __first2, __last2,
+            __tag, ::std::forward<_ExecutionPolicy>(__exec), __first2, __last2,
             [__first1, __last1, __first2, __last2, &__comp](_RandomAccessIterator2 __i, _RandomAccessIterator2 __j) {
                 assert(__j > __i);
                 //assert(__j - __i > 1);
@@ -5195,9 +5169,8 @@ __parallel_set_op(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _Forwar
                                                          __result + __s.__pos, __is_vector);
         };
         __par_backend::__parallel_strict_scan(
-            __backend_tag{},
-            ::std::forward<_ExecutionPolicy>(__exec), __n1, _SetRange{0, 0, 0}, //-1, 0},
-            [=](_DifferenceType __i, _DifferenceType __len) {                   // Reduce
+            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __n1, _SetRange{0, 0, 0}, //-1, 0},
+            [=](_DifferenceType __i, _DifferenceType __len) {                                    // Reduce
                 //[__b; __e) - a subrange of the first sequence, to reduce
                 _ForwardIterator1 __b = __first1 + __i, __e = __first1 + (__i + __len);
 
@@ -5251,13 +5224,11 @@ __parallel_set_op(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _Forwar
 }
 
 template <class _IsVector, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2,
-          class _OutputIterator,
-          class _Compare, class _SizeFunction, class _SetOP>
+          class _OutputIterator, class _Compare, class _SizeFunction, class _SetOP>
 _OutputIterator
 __parallel_set_op(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _ForwardIterator1 __first1,
-                  _ForwardIterator1 __last1,
-                  _ForwardIterator2 __first2, _ForwardIterator2 __last2, _OutputIterator __result, _Compare __comp,
-                  _SizeFunction __size_func, _SetOP __set_op)
+                  _ForwardIterator1 __last1, _ForwardIterator2 __first2, _ForwardIterator2 __last2,
+                  _OutputIterator __result, _Compare __comp, _SizeFunction __size_func, _SetOP __set_op)
 {
     using __backend_tag = typename decltype(__tag)::__backend_tag;
 
@@ -5279,8 +5250,8 @@ __parallel_set_op(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _F
 
     __par_backend::__buffer<_ExecutionPolicy, _T> __buf(__size_func(__n1, __n2));
 
-    return __internal::__except_handler([&__exec, __n1, __first1, __last1, __first2, __last2, __result,
-                                         __comp, __size_func, __set_op, &__buf]() {
+    return __internal::__except_handler([&__exec, __n1, __first1, __last1, __first2, __last2, __result, __comp,
+                                         __size_func, __set_op, &__buf]() {
         auto __tmp_memory = __buf.get();
         _DifferenceType __m{};
         auto __scan = [=](_DifferenceType, _DifferenceType, const _SetRange& __s) { // Scan
@@ -5354,7 +5325,8 @@ __parallel_set_union_op(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _
 {
     typedef typename ::std::iterator_traits<_ForwardIterator1>::difference_type _DifferenceType;
 
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2, _OutputIterator>();
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1,
+                                                                              _ForwardIterator2, _OutputIterator>();
     using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
 
     const auto __n1 = __last1 - __first1;
@@ -5424,9 +5396,9 @@ __parallel_set_union_op(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _
             },
             [=, &__result] {
                 __result = __internal::__parallel_set_op(
-                    __dispatch_tag,
-                    ::std::forward<_ExecutionPolicy>(__exec), __left_bound_seq_1, __last1, __first2, __last2, __result,
-                    __comp, [](_DifferenceType __n, _DifferenceType __m) { return __n + __m; }, __set_union_op);
+                    __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __left_bound_seq_1, __last1, __first2,
+                    __last2, __result, __comp, [](_DifferenceType __n, _DifferenceType __m) { return __n + __m; },
+                    __set_union_op);
             });
         return __result;
     }
@@ -5446,17 +5418,16 @@ __parallel_set_union_op(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _
             },
             [=, &__result] {
                 __result = __internal::__parallel_set_op(
-                    __dispatch_tag,
-                    ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __left_bound_seq_2, __last2, __result,
-                    __comp, [](_DifferenceType __n, _DifferenceType __m) { return __n + __m; }, __set_union_op);
+                    __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __left_bound_seq_2,
+                    __last2, __result, __comp, [](_DifferenceType __n, _DifferenceType __m) { return __n + __m; },
+                    __set_union_op);
             });
         return __result;
     }
 
     return __internal::__parallel_set_op(
-        __dispatch_tag,
-        ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2, __result, __comp,
-        [](_DifferenceType __n, _DifferenceType __m) { return __n + __m; }, __set_union_op);
+        __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2, __result,
+        __comp, [](_DifferenceType __n, _DifferenceType __m) { return __n + __m; }, __set_union_op);
 }
 
 //a shared parallel pattern for '__pattern_set_union' and '__pattern_set_symmetric_difference'
@@ -5478,13 +5449,13 @@ __parallel_set_union_op(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __ex
 
     // {1} {}: parallel copying just first sequence
     if (__n2 == 0)
-        return __internal::__pattern_walk2_brick(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first1,
-                                                 __last1, __result, __copy_range);
+        return __internal::__pattern_walk2_brick(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1,
+                                                 __result, __copy_range);
 
     // {} {2}: parallel copying justmake  second sequence
     if (__n1 == 0)
-        return __internal::__pattern_walk2_brick(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first2,
-                                                 __last2, __result, __copy_range);
+        return __internal::__pattern_walk2_brick(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first2, __last2,
+                                                 __result, __copy_range);
 
     // testing  whether the sequences are intersected
     _ForwardIterator1 __left_bound_seq_1 = ::std::lower_bound(__first1, __last1, *__first2, __comp);
@@ -5495,12 +5466,12 @@ __parallel_set_union_op(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __ex
         __par_backend::__parallel_invoke(
             __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
             [=] {
-                __internal::__pattern_walk2_brick(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first1,
-                                                  __last1, __result, __copy_range);
+                __internal::__pattern_walk2_brick(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1,
+                                                  __result, __copy_range);
             },
             [=] {
-                __internal::__pattern_walk2_brick(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first2,
-                                                  __last2, __result + __n1, __copy_range);
+                __internal::__pattern_walk2_brick(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first2, __last2,
+                                                  __result + __n1, __copy_range);
             });
         return __result + __n1 + __n2;
     }
@@ -5514,12 +5485,12 @@ __parallel_set_union_op(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __ex
         __par_backend::__parallel_invoke(
             __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
             [=] {
-                __internal::__pattern_walk2_brick(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first2,
-                                                  __last2, __result, __copy_range);
+                __internal::__pattern_walk2_brick(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first2, __last2,
+                                                  __result, __copy_range);
             },
             [=] {
-                __internal::__pattern_walk2_brick(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first1,
-                                                  __last1, __result + __n2, __copy_range);
+                __internal::__pattern_walk2_brick(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1,
+                                                  __result + __n2, __copy_range);
             });
         return __result + __n1 + __n2;
     }
@@ -5538,8 +5509,8 @@ __parallel_set_union_op(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __ex
             },
             [=, &__result] {
                 __result = __internal::__parallel_set_op(
-                    __tag, ::std::forward<_ExecutionPolicy>(__exec), __left_bound_seq_1, __last1, __first2,
-                    __last2, __result, __comp, [](_DifferenceType __n, _DifferenceType __m) { return __n + __m; },
+                    __tag, ::std::forward<_ExecutionPolicy>(__exec), __left_bound_seq_1, __last1, __first2, __last2,
+                    __result, __comp, [](_DifferenceType __n, _DifferenceType __m) { return __n + __m; },
                     __set_union_op);
             });
         return __result;
@@ -5560,16 +5531,16 @@ __parallel_set_union_op(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __ex
             },
             [=, &__result] {
                 __result = __internal::__parallel_set_op(
-                    __tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __left_bound_seq_2,
-                    __last2, __result, __comp, [](_DifferenceType __n, _DifferenceType __m) { return __n + __m; },
+                    __tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __left_bound_seq_2, __last2,
+                    __result, __comp, [](_DifferenceType __n, _DifferenceType __m) { return __n + __m; },
                     __set_union_op);
             });
         return __result;
     }
 
     return __internal::__parallel_set_op(
-        __tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2, __result,
-        __comp, [](_DifferenceType __n, _DifferenceType __m) { return __n + __m; }, __set_union_op);
+        __tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2, __result, __comp,
+        [](_DifferenceType __n, _DifferenceType __m) { return __n + __m; }, __set_union_op);
 }
 
 //------------------------------------------------------------------------
@@ -5650,8 +5621,8 @@ __pattern_set_union(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1,
 
     typedef typename ::std::iterator_traits<_OutputIterator>::value_type _Tp;
     return __parallel_set_union_op(
-        __dispatch_tag,
-        ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2, __result, __comp,
+        __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2, __result,
+        __comp,
         [](_RandomAccessIterator1 __first1, _RandomAccessIterator1 __last1, _RandomAccessIterator2 __first2,
            _RandomAccessIterator2 __last2, _Tp* __result, _Compare __comp) {
             return oneapi::dpl::__utils::__set_union_construct(__first1, __last1, __first2, __last2, __result, __comp,
@@ -5675,8 +5646,7 @@ __pattern_set_union(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec,
 
     typedef typename ::std::iterator_traits<_OutputIterator>::value_type _Tp;
     return __parallel_set_union_op(
-        __tag,
-        ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2, __result, __comp,
+        __tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2, __result, __comp,
         [](_RandomAccessIterator1 __first1, _RandomAccessIterator1 __last1, _RandomAccessIterator2 __first2,
            _RandomAccessIterator2 __last2, _Tp* __result, _Compare __comp) {
             return oneapi::dpl::__utils::__set_union_construct(__first1, __last1, __first2, __last2, __result, __comp,
@@ -5770,9 +5740,8 @@ __pattern_set_intersection(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __f
     {
         //we know proper offset due to [first1; left_bound_seq_1) < [first2; last2)
         return __internal::__parallel_set_op(
-            __dispatch_tag,
-            ::std::forward<_ExecutionPolicy>(__exec), __left_bound_seq_1, __last1, __first2, __last2, __result, __comp,
-            [](_DifferenceType __n, _DifferenceType __m) { return ::std::min(__n, __m); },
+            __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __left_bound_seq_1, __last1, __first2, __last2,
+            __result, __comp, [](_DifferenceType __n, _DifferenceType __m) { return ::std::min(__n, __m); },
             [](_RandomAccessIterator1 __first1, _RandomAccessIterator1 __last1, _RandomAccessIterator2 __first2,
                _RandomAccessIterator2 __last2, _T* __result, _Compare __comp) {
                 return oneapi::dpl::__utils::__set_intersection_construct(__first1, __last1, __first2, __last2,
@@ -5785,9 +5754,8 @@ __pattern_set_intersection(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __f
     {
         //we know proper offset due to [first2; left_bound_seq_2) < [first1; last1)
         __result = __internal::__parallel_set_op(
-            __dispatch_tag,
-            ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __left_bound_seq_2, __last2, __result, __comp,
-            [](_DifferenceType __n, _DifferenceType __m) { return ::std::min(__n, __m); },
+            __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __left_bound_seq_2, __last2,
+            __result, __comp, [](_DifferenceType __n, _DifferenceType __m) { return ::std::min(__n, __m); },
             [](_RandomAccessIterator1 __first1, _RandomAccessIterator1 __last1, _RandomAccessIterator2 __first2,
                _RandomAccessIterator2 __last2, _T* __result, _Compare __comp) {
                 return oneapi::dpl::__utils::__set_intersection_construct(__first2, __last2, __first1, __last1,
@@ -5834,9 +5802,8 @@ __pattern_set_intersection(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& _
     {
         //we know proper offset due to [first1; left_bound_seq_1) < [first2; last2)
         return __internal::__parallel_set_op(
-            __tag,
-            ::std::forward<_ExecutionPolicy>(__exec), __left_bound_seq_1, __last1, __first2, __last2, __result, __comp,
-            [](_DifferenceType __n, _DifferenceType __m) { return ::std::min(__n, __m); },
+            __tag, ::std::forward<_ExecutionPolicy>(__exec), __left_bound_seq_1, __last1, __first2, __last2, __result,
+            __comp, [](_DifferenceType __n, _DifferenceType __m) { return ::std::min(__n, __m); },
             [](_RandomAccessIterator1 __first1, _RandomAccessIterator1 __last1, _RandomAccessIterator2 __first2,
                _RandomAccessIterator2 __last2, _T* __result, _Compare __comp) {
                 return oneapi::dpl::__utils::__set_intersection_construct(__first1, __last1, __first2, __last2,
@@ -5849,9 +5816,8 @@ __pattern_set_intersection(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& _
     {
         //we know proper offset due to [first2; left_bound_seq_2) < [first1; last1)
         __result = __internal::__parallel_set_op(
-            __tag,
-            ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __left_bound_seq_2, __last2, __result, __comp,
-            [](_DifferenceType __n, _DifferenceType __m) { return ::std::min(__n, __m); },
+            __tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __left_bound_seq_2, __last2, __result,
+            __comp, [](_DifferenceType __n, _DifferenceType __m) { return ::std::min(__n, __m); },
             [](_RandomAccessIterator1 __first1, _RandomAccessIterator1 __last1, _RandomAccessIterator2 __first2,
                _RandomAccessIterator2 __last2, _T* __result, _Compare __comp) {
                 return oneapi::dpl::__utils::__set_intersection_construct(__first2, __last2, __first1, __last1,
@@ -5953,9 +5919,8 @@ __pattern_set_difference(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __fir
 
     if (__n1 + __n2 > __set_algo_cut_off)
         return __parallel_set_op(
-            __dispatch_tag,
-            ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2, __result, __comp,
-            [](_DifferenceType __n, _DifferenceType) { return __n; },
+            __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2, __result,
+            __comp, [](_DifferenceType __n, _DifferenceType) { return __n; },
             [](_RandomAccessIterator1 __first1, _RandomAccessIterator1 __last1, _RandomAccessIterator2 __first2,
                _RandomAccessIterator2 __last2, _T* __result, _Compare __comp) {
                 return oneapi::dpl::__utils::__set_difference_construct(__first1, __last1, __first2, __last2, __result,
@@ -5970,9 +5935,8 @@ template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1,
           class _RandomAccessIterator3, class _Compare>
 _RandomAccessIterator3
 __pattern_set_difference(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1,
-                         _RandomAccessIterator1 __last1,
-                         _RandomAccessIterator2 __first2, _RandomAccessIterator2 __last2,
-                         _RandomAccessIterator3 __result, _Compare __comp)
+                         _RandomAccessIterator1 __last1, _RandomAccessIterator2 __first2,
+                         _RandomAccessIterator2 __last2, _RandomAccessIterator3 __result, _Compare __comp)
 {
     typedef typename ::std::iterator_traits<_RandomAccessIterator3>::value_type _T;
     typedef typename ::std::iterator_traits<_RandomAccessIterator1>::difference_type _DifferenceType;
@@ -6005,8 +5969,7 @@ __pattern_set_difference(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __e
 
     if (__n1 + __n2 > __set_algo_cut_off)
         return __parallel_set_op(
-            __tag,
-            ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2, __result, __comp,
+            __tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2, __result, __comp,
             [](_DifferenceType __n, _DifferenceType) { return __n; },
             [](_RandomAccessIterator1 __first1, _RandomAccessIterator1 __last1, _RandomAccessIterator2 __first2,
                _RandomAccessIterator2 __last2, _T* __result, _Compare __comp) {
@@ -6087,8 +6050,8 @@ __pattern_set_symmetric_difference(_ExecutionPolicy&& __exec, _RandomAccessItera
 
     typedef typename ::std::iterator_traits<_RandomAccessIterator3>::value_type _T;
     return __internal::__parallel_set_union_op(
-        __dispatch_tag,
-        ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2, __result, __comp,
+        __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2, __result,
+        __comp,
         [](_RandomAccessIterator1 __first1, _RandomAccessIterator1 __last1, _RandomAccessIterator2 __first2,
            _RandomAccessIterator2 __last2, _T* __result, _Compare __comp) {
             return oneapi::dpl::__utils::__set_symmetric_difference_construct(
@@ -6113,8 +6076,7 @@ __pattern_set_symmetric_difference(__parallel_tag<_IsVector> __tag, _ExecutionPo
 
     typedef typename ::std::iterator_traits<_RandomAccessIterator3>::value_type _T;
     return __internal::__parallel_set_union_op(
-        __tag,
-        ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2, __result, __comp,
+        __tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2, __result, __comp,
         [](_RandomAccessIterator1 __first1, _RandomAccessIterator1 __last1, _RandomAccessIterator2 __first2,
            _RandomAccessIterator2 __last2, _T* __result, _Compare __comp) {
             return oneapi::dpl::__utils::__set_symmetric_difference_construct(
@@ -6219,7 +6181,6 @@ __pattern_is_heap_until(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __ex
     });
 }
 
-
 //------------------------------------------------------------------------
 // is_heap
 //------------------------------------------------------------------------
@@ -6271,7 +6232,8 @@ __pattern_is_heap(_ExecutionPolicy&&, _RandomAccessIterator __first, _RandomAcce
 
 template <class _Tag, class _ExecutionPolicy, class _RandomAccessIterator, class _Compare>
 bool
-__pattern_is_heap(_Tag, _ExecutionPolicy&&, _RandomAccessIterator __first, _RandomAccessIterator __last, _Compare __comp) noexcept
+__pattern_is_heap(_Tag, _ExecutionPolicy&&, _RandomAccessIterator __first, _RandomAccessIterator __last,
+                  _Compare __comp) noexcept
 {
     static_assert(__is_backend_tag_v<_Tag>);
 
@@ -6343,7 +6305,8 @@ __pattern_min_element(_ExecutionPolicy&&, _ForwardIterator __first, _ForwardIter
 
 template <class _Tag, typename _ExecutionPolicy, typename _ForwardIterator, typename _Compare>
 _ForwardIterator
-__pattern_min_element(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _Compare __comp) noexcept
+__pattern_min_element(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
+                      _Compare __comp) noexcept
 {
     static_assert(__is_backend_tag_v<_Tag>);
 
@@ -6365,8 +6328,7 @@ __pattern_min_element(_ExecutionPolicy&& __exec, _RandomAccessIterator __first,
 
     return __internal::__except_handler([&]() {
         return __par_backend::__parallel_reduce(
-            __backend_tag{},
-            ::std::forward<_ExecutionPolicy>(__exec), __first, __last, /*identity*/ __last,
+            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, /*identity*/ __last,
             [=](_RandomAccessIterator __begin, _RandomAccessIterator __end,
                 _RandomAccessIterator __init) -> _RandomAccessIterator {
                 const _RandomAccessIterator __subresult =
@@ -6399,8 +6361,7 @@ __pattern_min_element(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Ran
 
     return __internal::__except_handler([&]() {
         return __par_backend::__parallel_reduce(
-            __backend_tag{},
-            ::std::forward<_ExecutionPolicy>(__exec), __first, __last, /*identity*/ __last,
+            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, /*identity*/ __last,
             [=](_RandomAccessIterator __begin, _RandomAccessIterator __end,
                 _RandomAccessIterator __init) -> _RandomAccessIterator {
                 const _RandomAccessIterator __subresult =
@@ -6481,8 +6442,7 @@ __pattern_minmax_element(_ExecutionPolicy&& __exec, _RandomAccessIterator __firs
         typedef ::std::pair<_RandomAccessIterator, _RandomAccessIterator> _Result;
 
         return __par_backend::__parallel_reduce(
-            __backend_tag{},
-            ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
             /*identity*/ ::std::make_pair(__last, __last),
             [=, &__comp](_RandomAccessIterator __begin, _RandomAccessIterator __end, _Result __init) -> _Result {
                 const _Result __subresult = __internal::__brick_minmax_element(__begin, __end, __comp, __is_vector);
@@ -6525,8 +6485,7 @@ __pattern_minmax_element(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _
         typedef ::std::pair<_RandomAccessIterator, _RandomAccessIterator> _Result;
 
         return __par_backend::__parallel_reduce(
-            __backend_tag{},
-            ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
             /*identity*/ ::std::make_pair(__last, __last),
             [=, &__comp](_RandomAccessIterator __begin, _RandomAccessIterator __end, _Result __init) -> _Result {
                 const _Result __subresult = __internal::__brick_minmax_element(__begin, __end, __comp, _IsVector{});
diff --git a/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h b/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
index 35e7433a09f..e86beeb276d 100644
--- a/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
+++ b/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
@@ -262,7 +262,8 @@ inline constexpr bool __is_random_access_or_integral_v = __is_random_access_or_i
 // Vectorized version of for_loop
 template <class _Tag, typename _ExecutionPolicy, typename _Ip, typename _Function, typename _Sp, typename... _Rest>
 void
-__pattern_for_loop(_Tag __tag, _ExecutionPolicy&& __exec, _Ip __first, _Ip __last, _Function __f, _Sp __stride, _Rest&&... __rest) noexcept
+__pattern_for_loop(_Tag __tag, _ExecutionPolicy&& __exec, _Ip __first, _Ip __last, _Function __f, _Sp __stride,
+                   _Rest&&... __rest) noexcept
 {
     static_assert(__is_backend_tag_serial_v<_Tag>);
 
@@ -362,7 +363,8 @@ __execute_loop_strided(_Ip __first, _Ip __last, _Function __f, _Sp __stride, _Pa
 // Sequenced version of for_loop for non-RAI and non-integral types
 template <class _Tag, typename _ExecutionPolicy, typename _Ip, typename _Function, typename... _Rest>
 ::std::enable_if_t<!__is_random_access_or_integral_v<_Ip>>
-__pattern_for_loop(_Tag, _ExecutionPolicy&&, _Ip __first, _Ip __last, _Function __f, __single_stride_type, _Rest&&... __rest) noexcept
+__pattern_for_loop(_Tag, _ExecutionPolicy&&, _Ip __first, _Ip __last, _Function __f, __single_stride_type,
+                   _Rest&&... __rest) noexcept
 {
     static_assert(__is_backend_tag_serial_v<_Tag>);
 
@@ -400,8 +402,7 @@ __pattern_for_loop_n(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec,
 
     oneapi::dpl::__internal::__except_handler([&]() {
         return __par_backend::__parallel_reduce(
-                   __backend_tag{},
-                   ::std::forward<_ExecutionPolicy>(__exec), _Size(0), __n, __identity,
+                   __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), _Size(0), __n, __identity,
                    [__first, __f](_Size __i, _Size __j, __pack_type __value) {
                        const auto __subseq_start = __first + __i;
                        const auto __length = __j - __i;
@@ -438,8 +439,7 @@ __pattern_for_loop_n(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec,
 
     oneapi::dpl::__internal::__except_handler([&]() {
         return __par_backend::__parallel_reduce(
-                   __backend_tag{},
-                   ::std::forward<_ExecutionPolicy>(__exec), _Size(0), __n, __identity,
+                   __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), _Size(0), __n, __identity,
                    [__first, __f, __stride](_Size __i, _Size __j, __pack_type __value) {
                        const auto __subseq_start = __first + __i * __stride;
                        const auto __length = __j - __i;
@@ -517,16 +517,14 @@ struct __use_par_vec_helper<_Ip, ::std::enable_if_t<!::std::is_integral_v<_Ip>>>
 // Special versions for for_loop: handles both iterators and integral types(treated as random access iterators)
 template <typename _ExecutionPolicy, typename _Ip>
 constexpr auto
-__use_vectorization()
-    -> decltype(__use_par_vec_helper<_Ip>::template __use_vector<_ExecutionPolicy>())
+__use_vectorization() -> decltype(__use_par_vec_helper<_Ip>::template __use_vector<_ExecutionPolicy>())
 {
     return __use_par_vec_helper<_Ip>::template __use_vector<_ExecutionPolicy>();
 }
 
 template <typename _ExecutionPolicy, typename _Ip>
 constexpr auto
-__use_parallelization()
-    -> decltype(__use_par_vec_helper<_Ip>::template __use_parallel<_ExecutionPolicy>())
+__use_parallelization() -> decltype(__use_par_vec_helper<_Ip>::template __use_parallel<_ExecutionPolicy>())
 {
     return __use_par_vec_helper<_Ip>::template __use_parallel<_ExecutionPolicy>();
 }
@@ -547,9 +545,9 @@ __for_loop_impl(_ExecutionPolicy&& __exec, _Ip __start, _Ip __finish, _Fp&& __f,
     }
     else
     {
-        oneapi::dpl::__internal::__pattern_for_loop(__serial_tag<_IsVector>{},
-                                                    ::std::forward<_ExecutionPolicy>(__exec), __start, __finish, __f,
-                                                    __stride, ::std::get<_Is>(::std::move(__t))...);
+        oneapi::dpl::__internal::__pattern_for_loop(__serial_tag<_IsVector>{}, ::std::forward<_ExecutionPolicy>(__exec),
+                                                    __start, __finish, __f, __stride,
+                                                    ::std::get<_Is>(::std::move(__t))...);
     }
 }
 
diff --git a/include/oneapi/dpl/pstl/glue_algorithm_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
index 19b0208acca..495ff06be8c 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
@@ -187,10 +187,9 @@ adjacent_find(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardItera
 {
     constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
 
-    return oneapi::dpl::__internal::__pattern_adjacent_find(
-        __dispatch_tag,
-        ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __pred,
-        oneapi::dpl::__internal::__first_semantic());
+    return oneapi::dpl::__internal::__pattern_adjacent_find(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
+                                                            __first, __last, __pred,
+                                                            oneapi::dpl::__internal::__first_semantic());
 }
 
 // [alg.count]
@@ -271,7 +270,8 @@ template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterato
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _ForwardIterator2>
 copy(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _ForwardIterator1 __last, _ForwardIterator2 __result)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
 
     return oneapi::dpl::__internal::__pattern_walk2_brick(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
                                                           __first, __last, __result,
@@ -284,11 +284,11 @@ copy_n(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _Size __n, _Forward
 {
     using _DecayedExecutionPolicy = ::std::decay_t<_ExecutionPolicy>;
 
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
 
     return oneapi::dpl::__internal::__pattern_walk2_brick_n(
-        __dispatch_tag,
-        ::std::forward<_ExecutionPolicy>(__exec), __first, __n, __result,
+        __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __n, __result,
         oneapi::dpl::__internal::__brick_copy_n<_DecayedExecutionPolicy>{});
 }
 
@@ -331,11 +331,11 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _Forward
 transform(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _ForwardIterator1 __last, _ForwardIterator2 __result,
           _UnaryOperation __op)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
 
     return oneapi::dpl::__internal::__pattern_walk2(
-        __dispatch_tag, 
-        ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result,
+        __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result,
         oneapi::dpl::__internal::__transform_functor<_UnaryOperation>{::std::move(__op)});
 }
 
@@ -346,11 +346,11 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _Forward
 transform(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 __last1, _ForwardIterator2 __first2,
           _ForwardIterator __result, _BinaryOperation __op)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2, _ForwardIterator>();
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1,
+                                                                              _ForwardIterator2, _ForwardIterator>();
 
     return oneapi::dpl::__internal::__pattern_walk3(
-        __dispatch_tag,
-        ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __result,
+        __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __result,
         oneapi::dpl::__internal::__transform_functor<_BinaryOperation>(::std::move(__op)));
 }
 
@@ -416,7 +416,8 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _Forward
 replace_copy_if(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _ForwardIterator1 __last,
                 _ForwardIterator2 __result, _UnaryPredicate __pred, const _Tp& __new_value)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
 
     return oneapi::dpl::__internal::__pattern_walk2(
         __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result,
@@ -681,10 +682,8 @@ sort(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIter
 
     typedef typename ::std::iterator_traits<_RandomAccessIterator>::value_type _InputType;
 
-    oneapi::dpl::__internal::__pattern_sort(
-        __dispatch_tag,
-        ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __comp,
-        typename ::std::is_move_constructible<_InputType>::type());
+    oneapi::dpl::__internal::__pattern_sort(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+                                            __comp, typename ::std::is_move_constructible<_InputType>::type());
 }
 
 template <class _ExecutionPolicy, class _RandomAccessIterator>
@@ -793,7 +792,7 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, bool>
 equal(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 __last1, _ForwardIterator2 __first2,
       _BinaryPredicate __p)
 {
-   // TODO is it correct that we check _ForwardIterator2 in __select_backend ?
+    // TODO is it correct that we check _ForwardIterator2 in __select_backend ?
     constexpr auto __dispatch_tag =
         oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
 
@@ -838,7 +837,8 @@ move(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _ForwardIterator1 __l
 {
     using _DecayedExecutionPolicy = ::std::decay_t<_ExecutionPolicy>;
 
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
 
     return oneapi::dpl::__internal::__pattern_walk2_brick(
         __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __d_first,
diff --git a/include/oneapi/dpl/pstl/glue_memory_impl.h b/include/oneapi/dpl/pstl/glue_memory_impl.h
index cc46b189b0f..9058b0dc07c 100644
--- a/include/oneapi/dpl/pstl/glue_memory_impl.h
+++ b/include/oneapi/dpl/pstl/glue_memory_impl.h
@@ -45,7 +45,8 @@ uninitialized_copy(_ExecutionPolicy&& __exec, _InputIterator __first, _InputIter
     typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _ValueType2;
     typedef ::std::decay_t<_ExecutionPolicy> _DecayedExecutionPolicy;
 
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _InputIterator, _ForwardIterator>();
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _InputIterator, _ForwardIterator>();
 
     if constexpr (::std::is_trivial_v<_ValueType1> && ::std::is_trivial_v<_ValueType2>)
     {
@@ -69,7 +70,8 @@ uninitialized_copy_n(_ExecutionPolicy&& __exec, _InputIterator __first, _Size __
     typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _ValueType2;
     typedef ::std::decay_t<_ExecutionPolicy> _DecayedExecutionPolicy;
 
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _InputIterator, _ForwardIterator>();
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _InputIterator, _ForwardIterator>();
 
     if constexpr (::std::is_trivial_v<_ValueType1> && ::std::is_trivial_v<_ValueType2>)
     {
@@ -95,7 +97,8 @@ uninitialized_move(_ExecutionPolicy&& __exec, _InputIterator __first, _InputIter
     typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _ValueType2;
     typedef ::std::decay_t<_ExecutionPolicy> _DecayedExecutionPolicy;
 
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _InputIterator, _ForwardIterator>();
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _InputIterator, _ForwardIterator>();
 
     if constexpr (::std::is_trivial_v<_ValueType1> && ::std::is_trivial_v<_ValueType2>)
     {
@@ -119,7 +122,8 @@ uninitialized_move_n(_ExecutionPolicy&& __exec, _InputIterator __first, _Size __
     typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _ValueType2;
     typedef ::std::decay_t<_ExecutionPolicy> _DecayedExecutionPolicy;
 
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _InputIterator, _ForwardIterator>();
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _InputIterator, _ForwardIterator>();
 
     if constexpr (::std::is_trivial_v<_ValueType1> && ::std::is_trivial_v<_ValueType2>)
     {
@@ -149,15 +153,13 @@ uninitialized_fill(_ExecutionPolicy&& __exec, _ForwardIterator __first, _Forward
     if constexpr (::std::is_arithmetic_v<_ValueType>)
     {
         oneapi::dpl::__internal::__pattern_walk_brick(
-            __dispatch_tag,
-            ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+            __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
             oneapi::dpl::__internal::__brick_fill<_ValueType, _DecayedExecutionPolicy>{_ValueType(__value)});
     }
     else
     {
         oneapi::dpl::__internal::__pattern_walk1(
-            __dispatch_tag,
-            ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+            __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
             oneapi::dpl::__internal::__op_uninitialized_fill<_Tp, _DecayedExecutionPolicy>{__value});
     }
 }
@@ -224,7 +226,7 @@ destroy(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __
 #endif // _PSTL_ICPX_OMP_SIMD_DESTROY_WINDOWS_BROKEN || _ONEDPL_ICPX_OMP_SIMD_DESTROY_WINDOWS_BROKEN
 
         constexpr auto __dispatch_tag =
-                oneapi::dpl::__internal::__select_backend<_ExecutionPolicyDest, _ForwardIterator>();
+            oneapi::dpl::__internal::__select_backend<_ExecutionPolicyDest, _ForwardIterator>();
 
         oneapi::dpl::__internal::__pattern_walk1(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first,
                                                  __last, [](_ReferenceType __val) { __val.~_ValueType(); });
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 43a2d40c716..cd2bebea667 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -205,8 +205,7 @@ __pattern_swap(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIt
 
     using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
 
-    return __pattern_walk2<__backend_tag, /*_IsSync=*/::std::true_type,
-                           __par_backend_hetero::access_mode::read_write,
+    return __pattern_walk2<__backend_tag, /*_IsSync=*/::std::true_type, __par_backend_hetero::access_mode::read_write,
                            __par_backend_hetero::access_mode::read_write>(
         __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __f);
 }
@@ -259,8 +258,7 @@ __pattern_walk3(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardI
     return __first3 + __n;
 }
 
-template <typename _BackendTag,
-          __par_backend_hetero::access_mode __acc_mode1 = __par_backend_hetero::access_mode::read,
+template <typename _BackendTag, __par_backend_hetero::access_mode __acc_mode1 = __par_backend_hetero::access_mode::read,
           __par_backend_hetero::access_mode __acc_mode2 = __par_backend_hetero::access_mode::read,
           __par_backend_hetero::access_mode __acc_mode3 = __par_backend_hetero::access_mode::write,
           typename _ExecutionPolicy, typename _ForwardIterator1, typename _ForwardIterator2, typename _ForwardIterator3,
@@ -459,8 +457,7 @@ __pattern_walk2_transform_if(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&&
 {
     // Require `read_write` access mode for output sequence to force a copy in for host iterators to capture incoming
     // values of the output sequence for elements where the predicate is false.
-    return __pattern_walk2<_BackendTag, /*_IsSync=*/::std::true_type,
-                           __par_backend_hetero::access_mode::read,
+    return __pattern_walk2<_BackendTag, /*_IsSync=*/::std::true_type, __par_backend_hetero::access_mode::read,
                            __par_backend_hetero::access_mode::read_write>(
         __tag,
         __par_backend_hetero::make_wrapped_policy<__walk2_transform_if_wrapper>(
@@ -563,7 +560,6 @@ __pattern_fill_n(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Fo
                              __count, fill_functor<_T>{__value});
 }
 
-
 //------------------------------------------------------------------------
 // generate
 //------------------------------------------------------------------------
@@ -976,9 +972,8 @@ __pattern_adjacent_find(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __ex
     // TODO: in case of confilicting names
     // __par_backend_hetero::make_wrapped_policy<__par_backend_hetero::__or_policy_wrapper>()
     bool result = __par_backend_hetero::__parallel_find_or(
-        _BackendTag{},
-        ::std::forward<_ExecutionPolicy>(__exec), _Predicate{adjacent_find_fn<_BinaryPredicate>{__predicate}},
-        __par_backend_hetero::__parallel_or_tag{},
+        _BackendTag{}, ::std::forward<_ExecutionPolicy>(__exec),
+        _Predicate{adjacent_find_fn<_BinaryPredicate>{__predicate}}, __par_backend_hetero::__parallel_or_tag{},
         oneapi::dpl::__ranges::make_zip_view(__buf1.all_view(), __buf2.all_view()));
 
     // inverted conditional because of
@@ -1021,8 +1016,7 @@ __pattern_adjacent_find(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator, typename _BinaryPredicate>
 _Iterator
 __pattern_adjacent_find(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last,
-                        _BinaryPredicate __predicate,
-                        oneapi::dpl::__internal::__first_semantic)
+                        _BinaryPredicate __predicate, oneapi::dpl::__internal::__first_semantic)
 {
     if (__last - __first < 2)
         return __last;
@@ -1182,8 +1176,7 @@ __pattern_equal(_ExecutionPolicy&& __exec, _Iterator1 __first1, _Iterator1 __las
     // TODO: in case of confilicting names
     // __par_backend_hetero::make_wrapped_policy<__par_backend_hetero::__or_policy_wrapper>()
     return !__par_backend_hetero::__parallel_find_or(
-        __backend_tag{},
-        ::std::forward<_ExecutionPolicy>(__exec), _Predicate{equal_predicate<_Pred>{__pred}},
+        __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), _Predicate{equal_predicate<_Pred>{__pred}},
         __par_backend_hetero::__parallel_or_tag{},
         oneapi::dpl::__ranges::make_zip_view(__buf1.all_view(), __buf2.all_view()));
 }
@@ -1206,13 +1199,11 @@ __pattern_equal(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Ite
     // TODO: in case of confilicting names
     // __par_backend_hetero::make_wrapped_policy<__par_backend_hetero::__or_policy_wrapper>()
     return !__par_backend_hetero::__parallel_find_or(
-        _BackendTag{},
-        ::std::forward<_ExecutionPolicy>(__exec), _Predicate{equal_predicate<_Pred>{__pred}},
+        _BackendTag{}, ::std::forward<_ExecutionPolicy>(__exec), _Predicate{equal_predicate<_Pred>{__pred}},
         __par_backend_hetero::__parallel_or_tag{},
         oneapi::dpl::__ranges::make_zip_view(__buf1.all_view(), __buf2.all_view()));
 }
 
-
 //------------------------------------------------------------------------
 // equal version for sequences with equal length
 //------------------------------------------------------------------------
@@ -1375,7 +1366,6 @@ __pattern_find_first_of(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __ex
         ::std::true_type{});
 }
 
-
 //------------------------------------------------------------------------
 // search
 //------------------------------------------------------------------------
@@ -1607,8 +1597,8 @@ __pattern_scan_copy(_ExecutionPolicy&& __exec, _Iterator1 __first, _Iterator1 __
     auto __buf2 = __keep2(__output_first, __output_first + __n);
 
     auto __res = __par_backend_hetero::__parallel_scan_copy(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
-                                                            __buf1.all_view(),
-                                                   __buf2.all_view(), __n, __create_mask_op, __copy_by_mask_op);
+                                                            __buf1.all_view(), __buf2.all_view(), __n, __create_mask_op,
+                                                            __copy_by_mask_op);
 
     ::std::size_t __num_copied = __res.get();
     return ::std::make_pair(__output_first + __n, __num_copied);
@@ -2064,8 +2054,7 @@ __pattern_is_heap(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _Ran
         oneapi::dpl::unseq_backend::single_match_pred_by_idx<_ExecutionPolicy, __is_heap_check<_Compare>>;
 
     return !__par_backend_hetero::__parallel_or(
-        __backend_tag{},
-        ::std::forward<_ExecutionPolicy>(__exec),
+        __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
         __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__first),
         __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__last), _Predicate{__comp});
 }
@@ -2082,8 +2071,7 @@ __pattern_is_heap(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _R
         oneapi::dpl::unseq_backend::single_match_pred_by_idx<_ExecutionPolicy, __is_heap_check<_Compare>>;
 
     return !__par_backend_hetero::__parallel_or(
-        _BackendTag{},
-        ::std::forward<_ExecutionPolicy>(__exec),
+        _BackendTag{}, ::std::forward<_ExecutionPolicy>(__exec),
         __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__first),
         __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__last), _Predicate{__comp});
 }
@@ -2103,7 +2091,8 @@ __pattern_merge(_ExecutionPolicy&& __exec, _Iterator1 __first1, _Iterator1 __las
     if (__n == 0)
         return __d_first;
 
-    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator1, _Iterator2, _Iterator3>();
+    const auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator1, _Iterator2, _Iterator3>();
     using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
 
     //To consider the direct copying pattern call in case just one of sequences is empty.
@@ -2129,8 +2118,8 @@ __pattern_merge(_ExecutionPolicy&& __exec, _Iterator1 __first1, _Iterator1 __las
         auto __keep3 = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::write, _Iterator3>();
         auto __buf3 = __keep3(__d_first, __d_first + __n);
 
-        __par_backend_hetero::__parallel_merge(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __buf1.all_view(),
-                                               __buf2.all_view(), __buf3.all_view(), __comp)
+        __par_backend_hetero::__parallel_merge(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
+                                               __buf1.all_view(), __buf2.all_view(), __buf3.all_view(), __comp)
             .wait();
     }
     return __d_first + __n;
@@ -2171,8 +2160,8 @@ __pattern_merge(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Ite
         auto __keep3 = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::write, _Iterator3>();
         auto __buf3 = __keep3(__d_first, __d_first + __n);
 
-        __par_backend_hetero::__parallel_merge(_BackendTag{}, ::std::forward<_ExecutionPolicy>(__exec), __buf1.all_view(),
-                                               __buf2.all_view(), __buf3.all_view(), __comp)
+        __par_backend_hetero::__parallel_merge(_BackendTag{}, ::std::forward<_ExecutionPolicy>(__exec),
+                                               __buf1.all_view(), __buf2.all_view(), __buf3.all_view(), __comp)
             .wait();
     }
     return __d_first + __n;
@@ -2438,7 +2427,8 @@ __pattern_partition(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __la
     constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator>();
 
     //TODO: consider nonstable approaches
-    return __pattern_stable_partition(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __pred);
+    return __pattern_stable_partition(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+                                      __pred);
 }
 
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator, typename _UnaryPredicate>
@@ -2449,7 +2439,8 @@ __pattern_partition(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec,
     constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator>();
 
     //TODO: consider nonstable approaches
-    return __pattern_stable_partition(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __pred);
+    return __pattern_stable_partition(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+                                      __pred);
 }
 
 //------------------------------------------------------------------------
@@ -2577,8 +2568,7 @@ __pattern_includes(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _Forwa
 
     using __brick_include_type = unseq_backend::__brick_includes<_ExecutionPolicy, _Compare, _Size1, _Size2>;
     return !__par_backend_hetero::__parallel_or(
-        __backend_tag{},
-        ::std::forward<_ExecutionPolicy>(__exec),
+        __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
         __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__first2),
         __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__last2),
         __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__first1),
@@ -2606,8 +2596,7 @@ __pattern_includes(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _
 
     using __brick_include_type = unseq_backend::__brick_includes<_ExecutionPolicy, _Compare, _Size1, _Size2>;
     return !__par_backend_hetero::__parallel_or(
-        _BackendTag{},
-        ::std::forward<_ExecutionPolicy>(__exec),
+        _BackendTag{}, ::std::forward<_ExecutionPolicy>(__exec),
         __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__first2),
         __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__last2),
         __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__first1),
@@ -2630,8 +2619,7 @@ __pattern_partial_sort(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator _
     using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
 
     __par_backend_hetero::__parallel_partial_sort(
-        __backend_tag{},
-        ::std::forward<_ExecutionPolicy>(__exec),
+        __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
         __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read_write>(__first),
         __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read_write>(__mid),
         __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read_write>(__last), __comp)
@@ -2647,8 +2635,7 @@ __pattern_partial_sort(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exe
         return;
 
     __par_backend_hetero::__parallel_partial_sort(
-        _BackendTag{},
-        ::std::forward<_ExecutionPolicy>(__exec),
+        _BackendTag{}, ::std::forward<_ExecutionPolicy>(__exec),
         __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read_write>(__first),
         __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read_write>(__mid),
         __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read_write>(__last), __comp)
@@ -2753,8 +2740,7 @@ __pattern_partial_sort_copy(_ExecutionPolicy&& __exec, _InIterator __first, _InI
         using __backend_tag11 = typename decltype(__dispatch_tag11)::__backend_tag;
 
         __par_backend_hetero::__parallel_partial_sort(
-            __backend_tag11{},
-            __par_backend_hetero::make_wrapped_policy<__partial_sort_2>(__exec),
+            __backend_tag11{}, __par_backend_hetero::make_wrapped_policy<__partial_sort_2>(__exec),
             __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read_write>(__buf_first),
             __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read_write>(__buf_mid),
             __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read_write>(__buf_last), __comp);
@@ -2799,7 +2785,8 @@ __pattern_partial_sort_copy(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&&
         constexpr auto __dispatch_tag =
             oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, decltype(__out_first), decltype(__out_end)>();
 
-        __pattern_sort(__dispatch_tag,
+        __pattern_sort(
+            __dispatch_tag,
             __par_backend_hetero::make_wrapped_policy<__partial_sort_1>(::std::forward<_ExecutionPolicy>(__exec)),
             __out_first, __out_end, __comp, ::std::true_type{});
 
@@ -2832,8 +2819,7 @@ __pattern_partial_sort_copy(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&&
         using __backend_tag11 = typename decltype(__dispatch_tag11)::__backend_tag;
 
         __par_backend_hetero::__parallel_partial_sort(
-            __backend_tag11{},
-            __par_backend_hetero::make_wrapped_policy<__partial_sort_2>(__exec),
+            __backend_tag11{}, __par_backend_hetero::make_wrapped_policy<__partial_sort_2>(__exec),
             __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read_write>(__buf_first),
             __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read_write>(__buf_mid),
             __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read_write>(__buf_last), __comp);
@@ -2866,7 +2852,8 @@ __pattern_nth_element(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __
     // TODO: check partition-based implementation
     // - try to avoid host dereference issue
     // - measure performance of the issue-free implementation
-    __pattern_partial_sort(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __nth + 1, __last, __comp);
+    __pattern_partial_sort(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __nth + 1, __last,
+                           __comp);
 }
 
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator, typename _Compare>
@@ -3060,8 +3047,8 @@ __pattern_rotate(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Iterator
     using _Function = __brick_move<_ExecutionPolicy>;
     auto __brick = unseq_backend::walk_n<_ExecutionPolicy, _Function>{_Function{}};
 
-    oneapi::dpl::__par_backend_hetero::__parallel_for(_BackendTag{}, ::std::forward<_ExecutionPolicy>(__exec),
-                                                      __brick, __n, __temp_rng, __buf.all_view())
+    oneapi::dpl::__par_backend_hetero::__parallel_for(_BackendTag{}, ::std::forward<_ExecutionPolicy>(__exec), __brick,
+                                                      __n, __temp_rng, __buf.all_view())
         .wait();
 
     return __first + (__last - __new_first);
@@ -3094,8 +3081,7 @@ __pattern_rotate_copy(_ExecutionPolicy&& __exec, _BidirectionalIterator __first,
     const auto __shift = __new_first - __first;
 
     oneapi::dpl::__par_backend_hetero::__parallel_for(
-        __backend_tag{},
-        ::std::forward<_ExecutionPolicy>(__exec),
+        __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
         unseq_backend::__rotate_copy<typename ::std::iterator_traits<_BidirectionalIterator>::difference_type>{__n,
                                                                                                                __shift},
         __n, __buf1.all_view(), __buf2.all_view())
@@ -3313,7 +3299,8 @@ __pattern_set_difference(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1,
     // {1} \ {}: the difference is {1}
     if (__first2 == __last2)
     {
-        const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2, _OutputIterator>();
+        const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1,
+                                                                              _ForwardIterator2, _OutputIterator>();
 
         return oneapi::dpl::__internal::__pattern_walk2_brick(
             __dispatch_tag,
@@ -3372,7 +3359,8 @@ __pattern_set_union(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _Forw
     if (__first1 == __last1 && __first2 == __last2)
         return __result;
 
-    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2, _OutputIterator>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1,
+                                                                          _ForwardIterator2, _OutputIterator>();
 
     //{1} is empty
     if (__first1 == __last1)
@@ -3509,7 +3497,8 @@ __pattern_set_symmetric_difference(_ExecutionPolicy&& __exec, _ForwardIterator1
     if (__first1 == __last1 && __first2 == __last2)
         return __result;
 
-    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2, _OutputIterator>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1,
+                                                                          _ForwardIterator2, _OutputIterator>();
 
     //{1} is empty
     if (__first1 == __last1)
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
index 22a07100d96..0187a91a2b8 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
@@ -848,8 +848,8 @@ template <typename _ExecutionPolicy, typename _Range1, typename _Range2, typenam
           typename _BinaryOperation, typename _Inclusive>
 auto
 __parallel_transform_scan(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&& __exec, _Range1&& __in_rng,
-                          _Range2&& __out_rng, ::std::size_t __n,
-                          _UnaryOperation __unary_op, _InitType __init, _BinaryOperation __binary_op, _Inclusive)
+                          _Range2&& __out_rng, ::std::size_t __n, _UnaryOperation __unary_op, _InitType __init,
+                          _BinaryOperation __binary_op, _Inclusive)
 {
     using _Type = typename _InitType::__value_type;
 
@@ -1451,7 +1451,6 @@ __parallel_find_or(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPoli
         return __result != __init_value ? __result : __rng_n;
 }
 
-
 //------------------------------------------------------------------------
 // parallel_or - sync pattern
 //------------------------------------------------------------------------
@@ -1584,7 +1583,6 @@ __parallel_find(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&
                          __f, _TagType{}, __buf.all_view(), __s_buf.all_view());
 }
 
-
 // Special overload for single sequence cases.
 // TODO: check if similar pattern may apply to other algorithms. If so, these overloads should be moved out of
 // backend code.
@@ -2184,10 +2182,10 @@ __parallel_stable_sort(_ExecutionPolicy&& __exec, _Range&& __rng, _Compare, _Pro
         ::std::forward<_ExecutionPolicy>(__exec), ::std::forward<_Range>(__rng), __proj);
 }
 
-template <typename _ExecutionPolicy, typename _Range, typename _Compare, typename _Proj,
-          ::std::enable_if_t<
-              __is_radix_sort_usable_for_type<oneapi::dpl::__internal::__key_t<_Proj, _Range>, _Compare>::value,
-              int> = 0>
+template <
+    typename _ExecutionPolicy, typename _Range, typename _Compare, typename _Proj,
+    ::std::enable_if_t<
+        __is_radix_sort_usable_for_type<oneapi::dpl::__internal::__key_t<_Proj, _Range>, _Compare>::value, int> = 0>
 auto
 __parallel_stable_sort(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&& __exec, _Range&& __rng,
                        _Compare, _Proj __proj)
@@ -2211,10 +2209,10 @@ __parallel_stable_sort(_ExecutionPolicy&& __exec, _Range&& __rng, _Compare __com
     return __parallel_sort_impl(::std::forward<_ExecutionPolicy>(__exec), ::std::forward<_Range>(__rng), __cmp_f);
 }
 
-template <typename _ExecutionPolicy, typename _Range, typename _Compare, typename _Proj,
-          ::std::enable_if_t<
-              !__is_radix_sort_usable_for_type<oneapi::dpl::__internal::__key_t<_Proj, _Range>, _Compare>::value,
-              int> = 0>
+template <
+    typename _ExecutionPolicy, typename _Range, typename _Compare, typename _Proj,
+    ::std::enable_if_t<
+        !__is_radix_sort_usable_for_type<oneapi::dpl::__internal::__key_t<_Proj, _Range>, _Compare>::value, int> = 0>
 auto
 __parallel_stable_sort(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&& __exec, _Range&& __rng,
                        _Compare __comp, _Proj __proj)
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
index b6950c6f75c..4d2568f72c1 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
@@ -169,9 +169,8 @@ __parallel_transform_scan(_ExecutionPolicy&& __exec, _Range1&& __in_rng, _Range2
 {
     // workaround until we implement more performant version for patterns
     return oneapi::dpl::__par_backend_hetero::__parallel_transform_scan(
-        oneapi::dpl::__internal::__device_backend_tag{},
-        __exec.__device_policy(), ::std::forward<_Range1>(__in_rng), ::std::forward<_Range2>(__out_rng), __n,
-        __unary_op, __init, __binary_op, _Inclusive{});
+        oneapi::dpl::__internal::__device_backend_tag{}, __exec.__device_policy(), ::std::forward<_Range1>(__in_rng),
+        ::std::forward<_Range2>(__out_rng), __n, __unary_op, __init, __binary_op, _Inclusive{});
 }
 
 template <typename _ExecutionPolicy, typename _Range1, typename _Range2, typename _UnaryOperation, typename _InitType,
@@ -208,8 +207,8 @@ __parallel_copy_if(_ExecutionPolicy&& __exec, _InRng&& __in_rng, _OutRng&& __out
 {
     // workaround until we implement more performant version for patterns
     return oneapi::dpl::__par_backend_hetero::__parallel_copy_if(
-        oneapi::dpl::__internal::__device_backend_tag{},
-        __exec.__device_policy(), ::std::forward<_InRng>(__in_rng), ::std::forward<_OutRng>(__out_rng), __n, __pred);
+        oneapi::dpl::__internal::__device_backend_tag{}, __exec.__device_policy(), ::std::forward<_InRng>(__in_rng),
+        ::std::forward<_OutRng>(__out_rng), __n, __pred);
 }
 
 template <typename _ExecutionPolicy, typename _InRng, typename _OutRng, typename _Size, typename _Pred>
@@ -219,8 +218,8 @@ __parallel_copy_if(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionPolicy
 {
     // workaround until we implement more performant version for patterns
     return oneapi::dpl::__par_backend_hetero::__parallel_copy_if(
-        oneapi::dpl::__internal::__device_backend_tag{},
-        __exec.__device_policy(), ::std::forward<_InRng>(__in_rng), ::std::forward<_OutRng>(__out_rng), __n, __pred);
+        oneapi::dpl::__internal::__device_backend_tag{}, __exec.__device_policy(), ::std::forward<_InRng>(__in_rng),
+        ::std::forward<_OutRng>(__out_rng), __n, __pred);
 }
 
 template <typename _ExecutionPolicy, typename _InRng, typename _OutRng, typename _Size, typename _CreateMaskOp,
@@ -232,23 +231,20 @@ __parallel_scan_copy(_ExecutionPolicy&& __exec, _InRng&& __in_rng, _OutRng&& __o
 {
     // workaround until we implement more performant version for patterns
     return oneapi::dpl::__par_backend_hetero::__parallel_scan_copy(
-        oneapi::dpl::__internal::__device_backend_tag{},
-        __exec.__device_policy(), ::std::forward<_InRng>(__in_rng), ::std::forward<_OutRng>(__out_rng), __n,
-        __create_mask_op, __copy_by_mask_op);
+        oneapi::dpl::__internal::__device_backend_tag{}, __exec.__device_policy(), ::std::forward<_InRng>(__in_rng),
+        ::std::forward<_OutRng>(__out_rng), __n, __create_mask_op, __copy_by_mask_op);
 }
 
 template <typename _ExecutionPolicy, typename _InRng, typename _OutRng, typename _Size, typename _CreateMaskOp,
           typename _CopyByMaskOp>
 auto
 __parallel_scan_copy(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionPolicy&& __exec, _InRng&& __in_rng,
-                     _OutRng&& __out_rng, _Size __n,
-                     _CreateMaskOp __create_mask_op, _CopyByMaskOp __copy_by_mask_op)
+                     _OutRng&& __out_rng, _Size __n, _CreateMaskOp __create_mask_op, _CopyByMaskOp __copy_by_mask_op)
 {
     // workaround until we implement more performant version for patterns
     return oneapi::dpl::__par_backend_hetero::__parallel_scan_copy(
-        oneapi::dpl::__internal::__device_backend_tag{},
-        __exec.__device_policy(), ::std::forward<_InRng>(__in_rng), ::std::forward<_OutRng>(__out_rng), __n,
-        __create_mask_op, __copy_by_mask_op);
+        oneapi::dpl::__internal::__device_backend_tag{}, __exec.__device_policy(), ::std::forward<_InRng>(__in_rng),
+        ::std::forward<_OutRng>(__out_rng), __n, __create_mask_op, __copy_by_mask_op);
 }
 
 //------------------------------------------------------------------------
@@ -268,9 +264,9 @@ __parallel_find_or(_ExecutionPolicy&& __exec, _Brick __f, _BrickTag __brick_tag,
 }
 
 template <typename _ExecutionPolicy, typename _Brick, typename _BrickTag, typename... _Ranges>
-::std::conditional_t<::std::is_same_v<_BrickTag, __parallel_or_tag>, bool,
-                        oneapi::dpl::__internal::__difference_t<
-                            typename oneapi::dpl::__ranges::__get_first_range_type<_Ranges...>::type>>
+::std::conditional_t<
+    ::std::is_same_v<_BrickTag, __parallel_or_tag>, bool,
+    oneapi::dpl::__internal::__difference_t<typename oneapi::dpl::__ranges::__get_first_range_type<_Ranges...>::type>>
 __parallel_find_or(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionPolicy&& __exec, _Brick __f,
                    _BrickTag __brick_tag, _Ranges&&... __rngs)
 {
@@ -340,8 +336,7 @@ __parallel_find(_ExecutionPolicy&& __exec, _Iterator1 __first, _Iterator1 __last
 template <typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2, typename _Brick, typename _IsFirst>
 _Iterator1
 __parallel_find(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionPolicy&& __exec, _Iterator1 __first,
-                _Iterator1 __last, _Iterator2 __s_first,
-                _Iterator2 __s_last, _Brick __f, _IsFirst __is_first)
+                _Iterator1 __last, _Iterator2 __s_first, _Iterator2 __s_last, _Brick __f, _IsFirst __is_first)
 {
     // workaround until we implement more performant version for patterns
     return oneapi::dpl::__par_backend_hetero::__parallel_find(oneapi::dpl::__internal::__device_backend_tag{},
diff --git a/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
index 16521a0835b..dde4eb10f58 100644
--- a/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
@@ -320,12 +320,11 @@ __pattern_transform_scan_base(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __ex
     return __result + __n;
 }
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2,
-          typename _UnaryOperation, typename _Type,
-          typename _BinaryOperation, typename _Inclusive>
+          typename _UnaryOperation, typename _Type, typename _BinaryOperation, typename _Inclusive>
 _Iterator2
 __pattern_transform_scan(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator1 __first,
-                         _Iterator1 __last, _Iterator2 __result,
-                         _UnaryOperation __unary_op, _Type __init, _BinaryOperation __binary_op, _Inclusive)
+                         _Iterator1 __last, _Iterator2 __result, _UnaryOperation __unary_op, _Type __init,
+                         _BinaryOperation __binary_op, _Inclusive)
 {
     using _RepackedType = __par_backend_hetero::__repacked_tuple_t<_Type>;
     using _InitType = unseq_backend::__init_value<_RepackedType>;
@@ -385,7 +384,8 @@ __pattern_adjacent_difference(_ExecutionPolicy&& __exec, _ForwardIterator1 __fir
     if (__n <= 0)
         return __d_first;
 
-    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
+    const auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
     using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
 
     using _It1ValueT = typename ::std::iterator_traits<_ForwardIterator1>::value_type;
diff --git a/include/oneapi/dpl/pstl/histogram_impl.h b/include/oneapi/dpl/pstl/histogram_impl.h
index a67455df13e..b9fe40422f9 100644
--- a/include/oneapi/dpl/pstl/histogram_impl.h
+++ b/include/oneapi/dpl/pstl/histogram_impl.h
@@ -56,8 +56,7 @@ histogram(_ExecutionPolicy&& exec, _RandomAccessIterator1 first, _RandomAccessIt
         oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2>();
 
     oneapi::dpl::__internal::__pattern_histogram(
-        __dispatch_tag,
-        ::std::forward<_ExecutionPolicy>(exec), first, last, num_bins,
+        __dispatch_tag, ::std::forward<_ExecutionPolicy>(exec), first, last, num_bins,
         oneapi::dpl::__internal::__evenly_divided_binhash<_ValueType>(first_bin_min_val, last_bin_max_val, num_bins),
         histogram_first);
     return histogram_first + num_bins;
diff --git a/include/oneapi/dpl/pstl/numeric_impl.h b/include/oneapi/dpl/pstl/numeric_impl.h
index 7a1c05f7af7..6c099bb4689 100644
--- a/include/oneapi/dpl/pstl/numeric_impl.h
+++ b/include/oneapi/dpl/pstl/numeric_impl.h
@@ -382,8 +382,7 @@ __pattern_transform_scan(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __e
                                                             __unary_op,
                                                             /*__is_vector*/ ::std::false_type());
             },
-            [__first, __unary_op, __binary_op, __result](_DifferenceType __i, _DifferenceType __j,
-                                                                      _Tp __init) {
+            [__first, __unary_op, __binary_op, __result](_DifferenceType __i, _DifferenceType __j, _Tp __init) {
                 return __internal::__brick_transform_scan(__first + __i, __first + __j, __result + __i, __unary_op,
                                                           __init, __binary_op, _Inclusive(), _IsVector{})
                     .second;
@@ -414,8 +413,7 @@ __pattern_transform_scan(_ExecutionPolicy&& __exec, _RandomAccessIterator __firs
 
     return __internal::__except_handler([&]() {
         __par_backend::__parallel_strict_scan(
-            __backend_tag{},
-            ::std::forward<_ExecutionPolicy>(__exec), __n, __init,
+            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __n, __init,
             [__first, __unary_op, __binary_op, __result, __is_vector](_DifferenceType __i, _DifferenceType __len) {
                 return __internal::__brick_transform_scan(__first + __i, __first + (__i + __len), __result + __i,
                                                           __unary_op, _Tp{}, __binary_op, _Inclusive(), __is_vector)
@@ -454,8 +452,7 @@ __pattern_transform_scan(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _
 
     return __internal::__except_handler([&]() {
         __par_backend::__parallel_strict_scan(
-            __backend_tag{},
-            ::std::forward<_ExecutionPolicy>(__exec), __n, __init,
+            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __n, __init,
             [__first, __unary_op, __binary_op, __result](_DifferenceType __i, _DifferenceType __len) {
                 return __internal::__brick_transform_scan(__first + __i, __first + (__i + __len), __result + __i,
                                                           __unary_op, _Tp{}, __binary_op, _Inclusive(), _IsVector{})
@@ -493,8 +490,7 @@ __pattern_transform_scan(_ExecutionPolicy&& __exec, _ForwardIterator __first, _F
             oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator, _OutputIterator>();
 
         return __pattern_transform_scan(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), ++__first, __last,
-                                        ++__result,
-                                        __unary_op, __tmp, __binary_op, _Inclusive());
+                                        ++__result, __unary_op, __tmp, __binary_op, _Inclusive());
     }
     else
     {
diff --git a/include/oneapi/dpl/pstl/omp/parallel_merge.h b/include/oneapi/dpl/pstl/omp/parallel_merge.h
index a708a502bd4..39e8606d4ea 100644
--- a/include/oneapi/dpl/pstl/omp/parallel_merge.h
+++ b/include/oneapi/dpl/pstl/omp/parallel_merge.h
@@ -102,9 +102,9 @@ __parallel_merge(_ExecutionPolicy&& /*__exec*/, _RandomAccessIterator1 __xs, _Ra
 template <class _ExecutionPolicy, typename _RandomAccessIterator1, typename _RandomAccessIterator2,
           typename _RandomAccessIterator3, typename _Compare, typename _LeafMerge>
 void
-__parallel_merge(oneapi::dpl::__internal::__omp_backend_tag, _ExecutionPolicy&& /*__exec*/,
-                 _RandomAccessIterator1 __xs, _RandomAccessIterator1 __xe, _RandomAccessIterator2 __ys,
-                 _RandomAccessIterator2 __ye, _RandomAccessIterator3 __zs, _Compare __comp, _LeafMerge __leaf_merge)
+__parallel_merge(oneapi::dpl::__internal::__omp_backend_tag, _ExecutionPolicy&& /*__exec*/, _RandomAccessIterator1 __xs,
+                 _RandomAccessIterator1 __xe, _RandomAccessIterator2 __ys, _RandomAccessIterator2 __ye,
+                 _RandomAccessIterator3 __zs, _Compare __comp, _LeafMerge __leaf_merge)
 {
     std::size_t __size_x = __xe - __xs;
     std::size_t __size_y = __ye - __ys;
diff --git a/include/oneapi/dpl/pstl/omp/parallel_stable_sort.h b/include/oneapi/dpl/pstl/omp/parallel_stable_sort.h
index 5c05452a428..2c5c760f584 100644
--- a/include/oneapi/dpl/pstl/omp/parallel_stable_sort.h
+++ b/include/oneapi/dpl/pstl/omp/parallel_stable_sort.h
@@ -204,7 +204,6 @@ __parallel_stable_sort(oneapi::dpl::__internal::__omp_backend_tag, _ExecutionPol
     }
 }
 
-
 } // namespace __omp_backend
 } // namespace dpl
 } // namespace oneapi
diff --git a/include/oneapi/dpl/pstl/parallel_backend_tbb.h b/include/oneapi/dpl/pstl/parallel_backend_tbb.h
index 39caf6cbc21..7639551a1c6 100644
--- a/include/oneapi/dpl/pstl/parallel_backend_tbb.h
+++ b/include/oneapi/dpl/pstl/parallel_backend_tbb.h
@@ -518,7 +518,6 @@ __parallel_transform_scan(oneapi::dpl::__internal::__tbb_backend_tag, _Execution
     return __body.sum();
 }
 
-
 //------------------------------------------------------------------------
 // parallel_stable_sort
 //------------------------------------------------------------------------
@@ -1420,9 +1419,8 @@ template <class _ExecutionPolicy, typename _RandomAccessIterator1, typename _Ran
           typename _RandomAccessIterator3, typename _Compare, typename _LeafMerge>
 void
 __parallel_merge(oneapi::dpl::__internal::__tbb_backend_tag, _ExecutionPolicy&&, _RandomAccessIterator1 __xs,
-                 _RandomAccessIterator1 __xe,
-                 _RandomAccessIterator2 __ys, _RandomAccessIterator2 __ye, _RandomAccessIterator3 __zs, _Compare __comp,
-                 _LeafMerge __leaf_merge)
+                 _RandomAccessIterator1 __xe, _RandomAccessIterator2 __ys, _RandomAccessIterator2 __ye,
+                 _RandomAccessIterator3 __zs, _Compare __comp, _LeafMerge __leaf_merge)
 {
     typedef typename ::std::iterator_traits<_RandomAccessIterator1>::difference_type _DifferenceType1;
     typedef typename ::std::iterator_traits<_RandomAccessIterator2>::difference_type _DifferenceType2;

From 56406dfe64df2f20ce038f78d9032bb952728a07 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 15 Feb 2024 17:05:00 +0100
Subject: [PATCH 224/566] Fix review comment: I would propose to add a
 parameter _BackendType to this template definition...

---
 include/oneapi/dpl/pstl/algorithm_fwd.h  | 3 ++-
 include/oneapi/dpl/pstl/execution_impl.h | 7 ++++---
 include/oneapi/dpl/pstl/numeric_fwd.h    | 2 +-
 3 files changed, 7 insertions(+), 5 deletions(-)

diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index ec2cd79a135..e456f7a5093 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -26,9 +26,10 @@ namespace dpl
 namespace __internal
 {
 
-template <class _IsVector>
+template <class _IsVector, typename _BackendType = __par_backend_tag>
 struct __parallel_tag;
 
+template <typename _BackendType = __par_backend_tag>
 struct __parallel_forward_tag;
 
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/execution_impl.h b/include/oneapi/dpl/pstl/execution_impl.h
index dae807121cf..cb2b71225b1 100644
--- a/include/oneapi/dpl/pstl/execution_impl.h
+++ b/include/oneapi/dpl/pstl/execution_impl.h
@@ -113,21 +113,22 @@ struct __serial_tag
     using __is_vector = _IsVector;
 };
 
-template <class _IsVector>
+template <class _IsVector, typename _BackendType = __par_backend_tag>
 struct __parallel_tag
 {
     using __is_vector = _IsVector;
     // backend tag can be change depending on
     // TBB availability in the environment
-    using __backend_tag = __par_backend_tag;
+    using __backend_tag = _BackendType;
 };
 
+template <typename _BackendType = __par_backend_tag>
 struct __parallel_forward_tag
 {
     using __is_vector = ::std::false_type;
     // backend tag can be change depending on
     // TBB availability in the environment
-    using __backend_tag = __par_backend_tag;
+    using __backend_tag = _BackendType;
 };
 
 template <class _IsVector, class... _IteratorTypes>
diff --git a/include/oneapi/dpl/pstl/numeric_fwd.h b/include/oneapi/dpl/pstl/numeric_fwd.h
index fdd4928c230..0e8ba2a96e7 100644
--- a/include/oneapi/dpl/pstl/numeric_fwd.h
+++ b/include/oneapi/dpl/pstl/numeric_fwd.h
@@ -25,7 +25,7 @@ namespace dpl
 {
 namespace __internal
 {
-template <class _IsVector>
+template <class _IsVector, typename _BackendType = __par_backend_tag>
 struct __parallel_tag;
 
 //------------------------------------------------------------------------

From ae2208d6364bd3bba343bc467be0781298ef8fac Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 09:53:27 +0100
Subject: [PATCH 225/566] __parallel_transform_scan + tag calls

---
 include/oneapi/dpl/internal/async_impl/async_impl_hetero.h | 1 +
 include/oneapi/dpl/pstl/numeric_impl.h                     | 2 ++
 2 files changed, 3 insertions(+)

diff --git a/include/oneapi/dpl/internal/async_impl/async_impl_hetero.h b/include/oneapi/dpl/internal/async_impl/async_impl_hetero.h
index 232957a9b4a..d14f2aed022 100644
--- a/include/oneapi/dpl/internal/async_impl/async_impl_hetero.h
+++ b/include/oneapi/dpl/internal/async_impl/async_impl_hetero.h
@@ -225,6 +225,7 @@ __pattern_transform_scan_base_async(_ExecutionPolicy&& __exec, _Iterator1 __firs
     auto __buf2 = __keep2(__result, __result + __n);
 
     auto __res = oneapi::dpl::__par_backend_hetero::__parallel_transform_scan(
+        oneapi::dpl::__internal::__fpga_backend_tag{},
         ::std::forward<_ExecutionPolicy>(__exec), __buf1.all_view(), __buf2.all_view(), __n, __unary_op, __init,
         __binary_op, _Inclusive{});
     return __res.__make_future(__result + __n);
diff --git a/include/oneapi/dpl/pstl/numeric_impl.h b/include/oneapi/dpl/pstl/numeric_impl.h
index 6c099bb4689..216f313ffbc 100644
--- a/include/oneapi/dpl/pstl/numeric_impl.h
+++ b/include/oneapi/dpl/pstl/numeric_impl.h
@@ -343,6 +343,7 @@ __pattern_transform_scan(_ExecutionPolicy&& __exec, _RandomAccessIterator __firs
 
     return __internal::__except_handler([&]() {
         __par_backend::__parallel_transform_scan(
+            oneapi::dpl::__internal::__serial_backend_tag{},
             ::std::forward<_ExecutionPolicy>(__exec), __last - __first,
             [__first, __unary_op](_DifferenceType __i) mutable { return __unary_op(__first[__i]); }, __init,
             __binary_op,
@@ -373,6 +374,7 @@ __pattern_transform_scan(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __e
 
     return __internal::__except_handler([&]() {
         __par_backend::__parallel_transform_scan(
+            oneapi::dpl::__internal::__serial_backend_tag{},
             ::std::forward<_ExecutionPolicy>(__exec), __last - __first,
             [__first, __unary_op](_DifferenceType __i) mutable { return __unary_op(__first[__i]); }, __init,
             __binary_op,

From 4937245cc45553021f66f6461b8a94ac2ed2bdc6 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 10:35:29 +0100
Subject: [PATCH 226/566] __internal::__stable_sort_with_projection + tag impls

---
 include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index cd2bebea667..65471caa67d 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -2233,10 +2233,10 @@ __pattern_inplace_merge(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __ex
 //------------------------------------------------------------------------
 // sort
 //------------------------------------------------------------------------
-template <typename _ExecutionPolicy, typename _Iterator, typename _Compare, typename _Proj>
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator, typename _Compare, typename _Proj>
 void
-__stable_sort_with_projection(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last, _Compare __comp,
-                              _Proj __proj)
+__stable_sort_with_projection(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator __first,
+                              _Iterator __last, _Compare __comp, _Proj __proj)
 {
     if (__last - __first < 2)
         return;

From 33c9cf63f1418bd3360112ce3532462a4c422b7e Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 10:36:20 +0100
Subject: [PATCH 227/566] __internal::__stable_sort_with_projection + tag calls

---
 include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 65471caa67d..623087f919e 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -2266,7 +2266,7 @@ void
 __pattern_sort(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last,
                _Compare __comp, /*is_move_constructible=*/::std::true_type)
 {
-    __stable_sort_with_projection(::std::forward<_ExecutionPolicy>(__exec), __first, __last, __comp,
+    __stable_sort_with_projection(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __comp,
                                   oneapi::dpl::identity{});
 }
 
@@ -2287,7 +2287,7 @@ void
 __pattern_stable_sort(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last,
                       _Compare __comp)
 {
-    __stable_sort_with_projection(::std::forward<_ExecutionPolicy>(__exec), __first, __last, __comp,
+    __stable_sort_with_projection(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __comp,
                                   oneapi::dpl::identity{});
 }
 
@@ -2318,7 +2318,7 @@ __pattern_sort_by_key(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec
 
     auto __beg = oneapi::dpl::make_zip_iterator(__keys_first, __values_first);
     auto __end = __beg + (__keys_last - __keys_first);
-    __stable_sort_with_projection(::std::forward<_ExecutionPolicy>(__exec), __beg, __end, __comp,
+    __stable_sort_with_projection(__tag, ::std::forward<_ExecutionPolicy>(__exec), __beg, __end, __comp,
                                   [](const auto& __a) { return ::std::get<0>(__a); });
 }
 

From 5b3988096f14eec13517f46d4ad304eb45e3f6e7 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 10:45:38 +0100
Subject: [PATCH 228/566] __parallel_stable_sort + tag calls

---
 include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h | 5 +----
 1 file changed, 1 insertion(+), 4 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 623087f919e..7e2cd9e51a2 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -2244,10 +2244,7 @@ __stable_sort_with_projection(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&
     auto __keep = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read_write, _Iterator>();
     auto __buf = __keep(__first, __last);
 
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
-    __par_backend_hetero::__parallel_stable_sort(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
+    __par_backend_hetero::__parallel_stable_sort(_BackendTag{}, ::std::forward<_ExecutionPolicy>(__exec),
                                                  __buf.all_view(), __comp, __proj)
         .wait();
 }

From 73e913a7978cfdb49df3aab4413d6b4ef4d143b6 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 10:50:29 +0100
Subject: [PATCH 229/566] __pattern_sort + tag calls

---
 include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h | 5 +----
 1 file changed, 1 insertion(+), 4 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 7e2cd9e51a2..d4726b3cf29 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -2779,11 +2779,8 @@ __pattern_partial_sort_copy(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&&
             __brick_copy<_ExecutionPolicy>{});
 
         // Use regular sort as partial_sort isn't required to be stable
-        constexpr auto __dispatch_tag =
-            oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, decltype(__out_first), decltype(__out_end)>();
-
         __pattern_sort(
-            __dispatch_tag,
+            __tag,
             __par_backend_hetero::make_wrapped_policy<__partial_sort_1>(::std::forward<_ExecutionPolicy>(__exec)),
             __out_first, __out_end, __comp, ::std::true_type{});
 

From 2a2736ede8d279cedaf46e0435a0f0656cfaa8b3 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 10:53:17 +0100
Subject: [PATCH 230/566] __pattern_shift_left + tag calls

---
 include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index d4726b3cf29..4bbce19b233 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -3678,8 +3678,8 @@ __pattern_shift_left(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __l
     auto __keep = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read_write, _Iterator>();
     auto __buf = __keep(__first, __last);
 
-    auto __res =
-        oneapi::dpl::__internal::__pattern_shift_left(::std::forward<_ExecutionPolicy>(__exec), __buf.all_view(), __n);
+    auto __res = oneapi::dpl::__internal::__pattern_shift_left(__tag, ::std::forward<_ExecutionPolicy>(__exec),
+                                                               __buf.all_view(), __n);
     return __first + __res;
 }
 
@@ -3723,7 +3723,7 @@ __pattern_shift_right(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec
 
     //A shift right is the shift left with a reverse logic.
     auto __rng = oneapi::dpl::__ranges::reverse_view_simple<decltype(__buf.all_view())>{__buf.all_view()};
-    auto __res = oneapi::dpl::__internal::__pattern_shift_left(::std::forward<_ExecutionPolicy>(__exec), __rng, __n);
+    auto __res = oneapi::dpl::__internal::__pattern_shift_left(__tag, ::std::forward<_ExecutionPolicy>(__exec), __rng, __n);
 
     return __last - __res;
 }

From 5e548881ec5ba8d9434455936c71b420104f0f04 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 10:55:44 +0100
Subject: [PATCH 231/566] __parallel_transform_scan_single_group + tag impls

---
 .../oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h   | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
index 0187a91a2b8..43b6079a658 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
@@ -692,9 +692,10 @@ struct __parallel_copy_if_static_single_group_submitter<_Size, _ElemsPerItem, _W
 template <typename _ExecutionPolicy, typename _InRng, typename _OutRng, typename _UnaryOperation, typename _InitType,
           typename _BinaryOperation, typename _Inclusive>
 auto
-__parallel_transform_scan_single_group(_ExecutionPolicy&& __exec, _InRng&& __in_rng, _OutRng&& __out_rng,
-                                       ::std::size_t __n, _UnaryOperation __unary_op, _InitType __init,
-                                       _BinaryOperation __binary_op, _Inclusive)
+__parallel_transform_scan_single_group(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&& __exec,
+                                       _InRng&& __in_rng, _OutRng&& __out_rng, ::std::size_t __n,
+                                       _UnaryOperation __unary_op, _InitType __init, _BinaryOperation __binary_op,
+                                       _Inclusive)
 {
     using _CustomName = oneapi::dpl::__internal::__policy_kernel_name<_ExecutionPolicy>;
 

From 6df1714c7d9a8d63942ff153807ae572c27ce572 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 10:55:56 +0100
Subject: [PATCH 232/566] __parallel_transform_scan_single_group + tag calls

---
 include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
index 43b6079a658..e7ba291b27d 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
@@ -872,8 +872,9 @@ __parallel_transform_scan(oneapi::dpl::__internal::__device_backend_tag, _Execut
         if (__n <= __single_group_upper_limit && __max_slm_size >= __req_slm_size)
         {
             return __parallel_transform_scan_single_group(
-                std::forward<_ExecutionPolicy>(__exec), ::std::forward<_Range1>(__in_rng),
-                ::std::forward<_Range2>(__out_rng), __n, __unary_op, __init, __binary_op, _Inclusive{});
+                oneapi::dpl::__internal::__device_backend_tag{}, std::forward<_ExecutionPolicy>(__exec),
+                ::std::forward<_Range1>(__in_rng), ::std::forward<_Range2>(__out_rng), __n, __unary_op, __init,
+                __binary_op, _Inclusive{});
         }
     }
 

From e2578c53a4cc60853b4067f75d5fc05f7040d18f Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 10:59:21 +0100
Subject: [PATCH 233/566] __parallel_transform_scan_base + tag impls

---
 .../oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h | 9 ++++-----
 .../dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h   | 9 ++++-----
 2 files changed, 8 insertions(+), 10 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
index e7ba291b27d..838973de76a 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
@@ -769,12 +769,11 @@ __parallel_transform_scan_single_group(oneapi::dpl::__internal::__device_backend
 }
 
 template <typename _ExecutionPolicy, typename _Range1, typename _Range2, typename _BinaryOperation, typename _InitType,
-          typename _LocalScan, typename _GroupScan, typename _GlobalScan,
-          oneapi::dpl::__internal::__enable_if_device_execution_policy<_ExecutionPolicy, int> = 0>
+          typename _LocalScan, typename _GroupScan, typename _GlobalScan>
 auto
-__parallel_transform_scan_base(_ExecutionPolicy&& __exec, _Range1&& __in_rng, _Range2&& __out_rng,
-                               _BinaryOperation __binary_op, _InitType __init, _LocalScan __local_scan,
-                               _GroupScan __group_scan, _GlobalScan __global_scan)
+__parallel_transform_scan_base(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&& __exec,
+                               _Range1&& __in_rng, _Range2&& __out_rng, _BinaryOperation __binary_op, _InitType __init,
+                               _LocalScan __local_scan, _GroupScan __group_scan, _GlobalScan __global_scan)
 {
     using _CustomName = oneapi::dpl::__internal::__policy_kernel_name<_ExecutionPolicy>;
 
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
index e5a24731ede..8f3f7debd7c 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
@@ -187,12 +187,11 @@ __parallel_transform_scan(oneapi::dpl::__internal::__fpga_backend_tag, _Executio
 }
 
 template <typename _ExecutionPolicy, typename _Range1, typename _Range2, typename _BinaryOperation, typename _InitType,
-          typename _LocalScan, typename _GroupScan, typename _GlobalScan,
-          oneapi::dpl::__internal::__enable_if_fpga_execution_policy<_ExecutionPolicy, int> = 0>
+          typename _LocalScan, typename _GroupScan, typename _GlobalScan>
 auto
-__parallel_transform_scan_base(_ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&& __rng2,
-                               _BinaryOperation __binary_op, _InitType __init, _LocalScan __local_scan,
-                               _GroupScan __group_scan, _GlobalScan __global_scan)
+__parallel_transform_scan_base(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionPolicy&& __exec, _Range1&& __rng1,
+                               _Range2&& __rng2, _BinaryOperation __binary_op, _InitType __init,
+                               _LocalScan __local_scan, _GroupScan __group_scan, _GlobalScan __global_scan)
 {
     // workaround until we implement more performant version for patterns
     return oneapi::dpl::__par_backend_hetero::__parallel_transform_scan_base(

From 027a881b3c24981f0192a8905f5691e49c7fbd8d Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 11:01:09 +0100
Subject: [PATCH 234/566] __parallel_transform_scan_base + tag calls

---
 include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h        | 1 +
 include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h  | 3 +++
 .../oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h | 4 ++--
 3 files changed, 6 insertions(+), 2 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 4bbce19b233..2b7108c7ba3 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -3236,6 +3236,7 @@ __pattern_hetero_set_op(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _F
 
     auto __result_size =
         __par_backend_hetero::__parallel_transform_scan_base(
+            _BackendTag{},
             ::std::forward<_ExecutionPolicy>(__exec),
             oneapi::dpl::__ranges::make_zip_view(
                 __buf1.all_view(), __buf2.all_view(),
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
index 838973de76a..bde1afc65a8 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
@@ -889,6 +889,7 @@ __parallel_transform_scan(oneapi::dpl::__internal::__device_backend_tag, _Execut
 
     return __future(
         __parallel_transform_scan_base(
+            oneapi::dpl::__internal::__device_backend_tag{},
             ::std::forward<_ExecutionPolicy>(__exec), ::std::forward<_Range1>(__in_rng),
             ::std::forward<_Range2>(__out_rng), __binary_op, __init,
             // local scan
@@ -967,6 +968,7 @@ __parallel_scan_copy(_ExecutionPolicy&& __exec, _InRng&& __in_rng, _OutRng&& __o
     oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, int32_t> __mask_buf(__exec, __n);
 
     return __parallel_transform_scan_base(
+        oneapi::dpl::__internal::__device_backend_tag{},
         ::std::forward<_ExecutionPolicy>(__exec),
         oneapi::dpl::__ranges::make_zip_view(
             ::std::forward<_InRng>(__in_rng),
@@ -1007,6 +1009,7 @@ __parallel_scan_copy(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPo
     oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, int32_t> __mask_buf(__exec, __n);
 
     return __parallel_transform_scan_base(
+        oneapi::dpl::__internal::__device_backend_tag{},
         ::std::forward<_ExecutionPolicy>(__exec),
         oneapi::dpl::__ranges::make_zip_view(
             ::std::forward<_InRng>(__in_rng),
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
index 8f3f7debd7c..de69c3db67d 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
@@ -195,8 +195,8 @@ __parallel_transform_scan_base(oneapi::dpl::__internal::__fpga_backend_tag, _Exe
 {
     // workaround until we implement more performant version for patterns
     return oneapi::dpl::__par_backend_hetero::__parallel_transform_scan_base(
-        __exec.__device_policy(), ::std::forward<_Range1>(__rng1), ::std::forward<_Range2>(__rng2), __binary_op, __init,
-        __local_scan, __group_scan, __global_scan);
+        oneapi::dpl::__internal::__device_backend_tag{}, __exec.__device_policy(), ::std::forward<_Range1>(__rng1),
+        ::std::forward<_Range2>(__rng2), __binary_op, __init, __local_scan, __group_scan, __global_scan);
 }
 
 template <typename _ExecutionPolicy, typename _InRng, typename _OutRng, typename _Size, typename _Pred,

From 0ed4560ad4e9259e6114481ed89e0b4152a90546 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 11:13:38 +0100
Subject: [PATCH 235/566] __parallel_or + tag impls

---
 .../dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h      | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
index de69c3db67d..5e4b84c0888 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
@@ -278,9 +278,9 @@ __parallel_find_or(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionPolicy
 // parallel_or
 //-----------------------------------------------------------------------
 template <typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2, typename _Brick>
-oneapi::dpl::__internal::__enable_if_fpga_execution_policy<_ExecutionPolicy, bool>
-__parallel_or(_ExecutionPolicy&& __exec, _Iterator1 __first, _Iterator1 __last, _Iterator2 __s_first,
-              _Iterator2 __s_last, _Brick __f)
+bool
+__parallel_or(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionPolicy&& __exec, _Iterator1 __first,
+              _Iterator1 __last, _Iterator2 __s_first, _Iterator2 __s_last, _Brick __f)
 {
     // workaround until we implement more performant version for patterns
     return oneapi::dpl::__par_backend_hetero::__parallel_or(__exec.__device_policy(), __first, __last, __s_first,

From 964d983ba44abe65de13424c39cdc2ff647182ab Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 11:14:16 +0100
Subject: [PATCH 236/566] __parallel_or + tag calls

---
 .../oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h  | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
index 5e4b84c0888..fd4a9f7697a 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
@@ -283,7 +283,8 @@ __parallel_or(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionPolicy&& __
               _Iterator1 __last, _Iterator2 __s_first, _Iterator2 __s_last, _Brick __f)
 {
     // workaround until we implement more performant version for patterns
-    return oneapi::dpl::__par_backend_hetero::__parallel_or(__exec.__device_policy(), __first, __last, __s_first,
+    return oneapi::dpl::__par_backend_hetero::__parallel_or(oneapi::dpl::__internal::__device_backend_tag{},
+                                                            __exec.__device_policy(), __first, __last, __s_first,
                                                             __s_last, __f);
 }
 

From 27f4a95ec1307279bef67209e44796cf6836bc48 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 11:30:06 +0100
Subject: [PATCH 237/566] __parallel_sort_impl + tag impls

---
 .../oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h    | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
index bde1afc65a8..79e30906ec8 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
@@ -2035,10 +2035,10 @@ struct __parallel_sort_submitter<_IdType, __internal::__optional_kernel_name<_Le
     }
 };
 
-template <typename _ExecutionPolicy, typename _Range, typename _Compare,
-          oneapi::dpl::__internal::__enable_if_device_execution_policy<_ExecutionPolicy, int> = 0>
+template <typename _ExecutionPolicy, typename _Range, typename _Compare>
 auto
-__parallel_sort_impl(_ExecutionPolicy&& __exec, _Range&& __rng, _Compare __comp)
+__parallel_sort_impl(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&& __exec, _Range&& __rng,
+                     _Compare __comp)
 {
     using _CustomName = oneapi::dpl::__internal::__policy_kernel_name<_ExecutionPolicy>;
 

From bb8aa37ea9c4e169a436bfbaec196b9fce25527d Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 11:30:26 +0100
Subject: [PATCH 238/566] __parallel_sort_impl + tag calls

---
 include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
index 79e30906ec8..7698ce0c85a 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
@@ -2224,7 +2224,8 @@ __parallel_stable_sort(oneapi::dpl::__internal::__device_backend_tag, _Execution
     auto __cmp_f = [__comp, __proj](const auto& __a, const auto& __b) mutable {
         return __comp(__proj(__a), __proj(__b));
     };
-    return __parallel_sort_impl(::std::forward<_ExecutionPolicy>(__exec), ::std::forward<_Range>(__rng), __cmp_f);
+    return __parallel_sort_impl(oneapi::dpl::__internal::__device_backend_tag{},
+                                ::std::forward<_ExecutionPolicy>(__exec), ::std::forward<_Range>(__rng), __cmp_f);
 }
 
 //------------------------------------------------------------------------

From 84d5ada9ecfccefc8b925ae652cfc9c2fcf157ca Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 11:32:32 +0100
Subject: [PATCH 239/566] __parallel_partial_sort_impl + tag impls

---
 .../oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h    | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
index 7698ce0c85a..5027ebace7a 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
@@ -2142,10 +2142,10 @@ struct __parallel_partial_sort_submitter<__internal::__optional_kernel_name<_Glo
     }
 };
 
-template <typename _ExecutionPolicy, typename _Range, typename _Merge, typename _Compare,
-          oneapi::dpl::__internal::__enable_if_device_execution_policy<_ExecutionPolicy, int> = 0>
+template <typename _ExecutionPolicy, typename _Range, typename _Merge, typename _Compare>
 auto
-__parallel_partial_sort_impl(_ExecutionPolicy&& __exec, _Range&& __rng, _Merge __merge, _Compare __comp)
+__parallel_partial_sort_impl(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&& __exec, _Range&& __rng,
+                             _Merge __merge, _Compare __comp)
 {
     using _CustomName = oneapi::dpl::__internal::__policy_kernel_name<_ExecutionPolicy>;
     using _GlobalSortKernel =

From 0bb8b080666a45f0c6bafe9ee9732fb166dde5db Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 11:32:40 +0100
Subject: [PATCH 240/566] __parallel_partial_sort_impl + tag calls

---
 .../oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h    | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
index 5027ebace7a..cbd63bc9259 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
@@ -2246,7 +2246,8 @@ __parallel_partial_sort(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator
     auto __keep = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read_write, _Iterator>();
     auto __buf = __keep(__first, __last);
 
-    return __parallel_partial_sort_impl(::std::forward<_ExecutionPolicy>(__exec), __buf.all_view(),
+    return __parallel_partial_sort_impl(oneapi::dpl::__internal::__device_backend_tag{},
+                                        ::std::forward<_ExecutionPolicy>(__exec), __buf.all_view(),
                                         __partial_merge_kernel<decltype(__mid_idx)>{__mid_idx}, __comp);
 }
 
@@ -2263,7 +2264,8 @@ __parallel_partial_sort(oneapi::dpl::__internal::__device_backend_tag, _Executio
     auto __keep = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read_write, _Iterator>();
     auto __buf = __keep(__first, __last);
 
-    return __parallel_partial_sort_impl(::std::forward<_ExecutionPolicy>(__exec), __buf.all_view(),
+    return __parallel_partial_sort_impl(oneapi::dpl::__internal::__device_backend_tag{},
+                                        ::std::forward<_ExecutionPolicy>(__exec), __buf.all_view(),
                                         __partial_merge_kernel<decltype(__mid_idx)>{__mid_idx}, __comp);
 }
 } // namespace __par_backend_hetero

From 4c6ffbff0a243ea2948cb81df6265c341bac0caa Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 11:37:57 +0100
Subject: [PATCH 241/566] __parallel_histogram + tag impls

---
 .../dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h      | 6 +++---
 .../dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h | 3 ++-
 2 files changed, 5 insertions(+), 4 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
index fd4a9f7697a..1f85484f809 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
@@ -452,10 +452,10 @@ __parallel_partial_sort(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionP
 //-----------------------------------------------------------------------
 
 // TODO: check if it makes sense to move these wrappers out of backend to a common place
-template <typename _ExecutionPolicy, typename _Event, typename _Range1, typename _Range2, typename _BinHashMgr,
-          oneapi::dpl::__internal::__enable_if_fpga_execution_policy<_ExecutionPolicy, int> = 0>
+template <typename _ExecutionPolicy, typename _Event, typename _Range1, typename _Range2, typename _BinHashMgr>
 auto
-__parallel_histogram(_ExecutionPolicy&& __exec, const _Event& __init_event, _Range1&& __input, _Range2&& __bins,
+__parallel_histogram(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionPolicy&& __exec, const _Event& __init_event,
+                     _Range1&& __input, _Range2&& __bins,
                      const _BinHashMgr& __binhash_manager)
 {
     if constexpr (sizeof(oneapi::dpl::__internal::__value_t<_Range2>) <= sizeof(::std::uint32_t))
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h
index fdcf06ad984..ae569b96544 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h
@@ -540,7 +540,8 @@ __parallel_histogram_select_kernel(_ExecutionPolicy&& __exec, const sycl::event&
 
 template <typename _ExecutionPolicy, typename _Range1, typename _Range2, typename _BinHashMgr>
 auto
-__parallel_histogram(_ExecutionPolicy&& __exec, const sycl::event& __init_event, _Range1&& __input, _Range2&& __bins,
+__parallel_histogram(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&& __exec,
+                     const sycl::event& __init_event, _Range1&& __input, _Range2&& __bins,
                      const _BinHashMgr& __binhash_manager)
 {
     if (__input.size() < 1048576) // 2^20

From 6138b635ea4f11d48b0d39f728370c56ee8ec4e1 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 11:38:11 +0100
Subject: [PATCH 242/566] __parallel_histogram + tag calls

---
 .../oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h | 1 +
 include/oneapi/dpl/pstl/hetero/histogram_impl_hetero.h        | 4 ++--
 2 files changed, 3 insertions(+), 2 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
index 1f85484f809..224e6b6f9fe 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
@@ -462,6 +462,7 @@ __parallel_histogram(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionPoli
     {
         // workaround until we implement more performant version for patterns
         return oneapi::dpl::__par_backend_hetero::__parallel_histogram(
+            oneapi::dpl::__internal::__device_backend_tag{},
             __exec.__device_policy(), __init_event, ::std::forward<_Range1>(__input), ::std::forward<_Range2>(__bins),
             __binhash_manager);
     }
diff --git a/include/oneapi/dpl/pstl/hetero/histogram_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/histogram_impl_hetero.h
index 491bc107951..87d22e9a0a7 100644
--- a/include/oneapi/dpl/pstl/hetero/histogram_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/histogram_impl_hetero.h
@@ -157,8 +157,8 @@ __pattern_histogram(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Rando
                                                         _RandomAccessIterator1>();
             auto __input_buf = __keep_input(__first, __last);
 
-            __parallel_histogram(::std::forward<_ExecutionPolicy>(__exec), __init_event, __input_buf.all_view(),
-                                 ::std::move(__bins), __binhash_manager)
+            __parallel_histogram(_BackendTag{}, ::std::forward<_ExecutionPolicy>(__exec), __init_event,
+                                 __input_buf.all_view(), ::std::move(__bins), __binhash_manager)
                 .wait();
         }
         else

From 9dee971c9f51313e311cd0db07008d917591dcab Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 11:41:56 +0100
Subject: [PATCH 243/566] __histogram_general_registers_local_reduction + tag
 impls

---
 .../dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h
index ae569b96544..e113ea38e8a 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h
@@ -285,9 +285,9 @@ struct __histogram_general_registers_local_reduction_submitter<__iters_per_work_
 template <::std::uint16_t __iters_per_work_item, ::std::uint8_t __bins_per_work_item, typename _ExecutionPolicy,
           typename _Range1, typename _Range2, typename _BinHashMgr>
 auto
-__histogram_general_registers_local_reduction(_ExecutionPolicy&& __exec, const sycl::event& __init_event,
-                                              ::std::uint16_t __work_group_size, _Range1&& __input, _Range2&& __bins,
-                                              const _BinHashMgr& __binhash_manager)
+__histogram_general_registers_local_reduction(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&& __exec,
+                                              const sycl::event& __init_event, ::std::uint16_t __work_group_size,
+                                              _Range1&& __input, _Range2&& __bins, const _BinHashMgr& __binhash_manager)
 {
     using _kernel_base_name = typename ::std::decay_t<_ExecutionPolicy>::kernel_name;
 

From 1f54f6d463f08de9b6b321aa6bf5684de0088df1 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 11:42:05 +0100
Subject: [PATCH 244/566] __histogram_general_registers_local_reduction + tag
 calls

---
 .../dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h      | 1 +
 1 file changed, 1 insertion(+)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h
index e113ea38e8a..b90ab4a4015 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h
@@ -513,6 +513,7 @@ __parallel_histogram_select_kernel(_ExecutionPolicy&& __exec, const sycl::event&
     {
         return __future(
             __histogram_general_registers_local_reduction<__iters_per_work_item, __max_work_item_private_bins>(
+                oneapi::dpl::__internal::__device_backend_tag{},
                 ::std::forward<_ExecutionPolicy>(__exec), __init_event, __work_group_size,
                 ::std::forward<_Range1>(__input), ::std::forward<_Range2>(__bins), __binhash_manager));
     }

From 6a31d18c6f9ab946da8a0899de2e35d2ac15e211 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 11:43:35 +0100
Subject: [PATCH 245/566] __histogram_general_local_atomics + tag impls

---
 .../dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h
index b90ab4a4015..a85ef4f90ce 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h
@@ -380,9 +380,9 @@ struct __histogram_general_local_atomics_submitter<__iters_per_work_item,
 template <::std::uint16_t __iters_per_work_item, typename _ExecutionPolicy, typename _Range1, typename _Range2,
           typename _BinHashMgr>
 auto
-__histogram_general_local_atomics(_ExecutionPolicy&& __exec, const sycl::event& __init_event,
-                                  ::std::uint16_t __work_group_size, _Range1&& __input, _Range2&& __bins,
-                                  const _BinHashMgr& __binhash_manager)
+__histogram_general_local_atomics(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&& __exec,
+                                  const sycl::event& __init_event, ::std::uint16_t __work_group_size, _Range1&& __input,
+                                  _Range2&& __bins, const _BinHashMgr& __binhash_manager)
 {
     using _kernel_base_name = typename ::std::decay_t<_ExecutionPolicy>::kernel_name;
 

From 59429cb063621dac7fdb78749dbdd551ee76b9c0 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 11:43:42 +0100
Subject: [PATCH 246/566] __histogram_general_local_atomics + tag calls

---
 .../dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h   | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h
index a85ef4f90ce..65174273bb6 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h
@@ -523,8 +523,8 @@ __parallel_histogram_select_kernel(_ExecutionPolicy&& __exec, const sycl::event&
              __local_mem_size)
     {
         return __future(__histogram_general_local_atomics<__iters_per_work_item>(
-            ::std::forward<_ExecutionPolicy>(__exec), __init_event, __work_group_size, ::std::forward<_Range1>(__input),
-            ::std::forward<_Range2>(__bins), __binhash_manager));
+            oneapi::dpl::__internal::__device_backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __init_event,
+            __work_group_size, ::std::forward<_Range1>(__input), ::std::forward<_Range2>(__bins), __binhash_manager));
     }
     else // otherwise, use global atomics (private copies per workgroup)
     {

From 901d67bd4e39950bb2f2aa6e6eb94fdc16f02c45 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 11:44:40 +0100
Subject: [PATCH 247/566] __histogram_general_private_global_atomics + tag
 impls

---
 .../pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h    | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h
index 65174273bb6..7c56570bc29 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h
@@ -477,9 +477,10 @@ struct __histogram_general_private_global_atomics_submitter<__internal::__option
 };
 template <typename _ExecutionPolicy, typename _Range1, typename _Range2, typename _BinHashMgr>
 auto
-__histogram_general_private_global_atomics(_ExecutionPolicy&& __exec, const sycl::event& __init_event,
-                                           ::std::uint16_t __min_iters_per_work_item, ::std::uint16_t __work_group_size,
-                                           _Range1&& __input, _Range2&& __bins, const _BinHashMgr& __binhash_manager)
+__histogram_general_private_global_atomics(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&& __exec,
+                                           const sycl::event& __init_event, ::std::uint16_t __min_iters_per_work_item,
+                                           ::std::uint16_t __work_group_size, _Range1&& __input, _Range2&& __bins,
+                                           const _BinHashMgr& __binhash_manager)
 {
     using _kernel_base_name = typename ::std::decay_t<_ExecutionPolicy>::kernel_name;
 

From 9bc97a282017165b6fbad5079a1da5a27b2596e3 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 11:44:47 +0100
Subject: [PATCH 248/566] __histogram_general_private_global_atomics + tag
 calls

---
 .../dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h  | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h
index 7c56570bc29..119a5b49bed 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h
@@ -535,8 +535,9 @@ __parallel_histogram_select_kernel(_ExecutionPolicy&& __exec, const sycl::event&
         // private copies of the histogram bins in global memory.  No unrolling is taken advantage of here because it
         // is a runtime argument.
         return __future(__histogram_general_private_global_atomics(
-            ::std::forward<_ExecutionPolicy>(__exec), __init_event, __iters_per_work_item, __work_group_size,
-            ::std::forward<_Range1>(__input), ::std::forward<_Range2>(__bins), __binhash_manager));
+            oneapi::dpl::__internal::__device_backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __init_event,
+            __iters_per_work_item, __work_group_size, ::std::forward<_Range1>(__input), ::std::forward<_Range2>(__bins),
+            __binhash_manager));
     }
 }
 

From 141a922594bed90857ba54a85c7d40f0f79398b2 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 11:45:57 +0100
Subject: [PATCH 249/566] __parallel_histogram_select_kernel + tag impls

---
 .../dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h  | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h
index 119a5b49bed..508fd1f9a18 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h
@@ -495,8 +495,9 @@ __histogram_general_private_global_atomics(oneapi::dpl::__internal::__device_bac
 template <::std::uint16_t __iters_per_work_item, typename _ExecutionPolicy, typename _Range1, typename _Range2,
           typename _BinHashMgr>
 auto
-__parallel_histogram_select_kernel(_ExecutionPolicy&& __exec, const sycl::event& __init_event, _Range1&& __input,
-                                   _Range2&& __bins, const _BinHashMgr& __binhash_manager)
+__parallel_histogram_select_kernel(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&& __exec,
+                                   const sycl::event& __init_event, _Range1&& __input, _Range2&& __bins,
+                                   const _BinHashMgr& __binhash_manager)
 {
     using _private_histogram_type = ::std::uint16_t;
     using _local_histogram_type = ::std::uint32_t;

From fb5d2b88ca8506714748b4dd71b43ed20442cb0f Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 11:46:04 +0100
Subject: [PATCH 250/566] __parallel_histogram_select_kernel + tag calls

---
 .../dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h     | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h
index 508fd1f9a18..8fbe0ec5af6 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h
@@ -551,12 +551,14 @@ __parallel_histogram(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPo
     if (__input.size() < 1048576) // 2^20
     {
         return __parallel_histogram_select_kernel</*iters_per_workitem = */ 4>(
+            oneapi::dpl::__internal::__device_backend_tag{},
             ::std::forward<_ExecutionPolicy>(__exec), __init_event, ::std::forward<_Range1>(__input),
             ::std::forward<_Range2>(__bins), __binhash_manager);
     }
     else
     {
         return __parallel_histogram_select_kernel</*iters_per_workitem = */ 32>(
+            oneapi::dpl::__internal::__device_backend_tag{},
             ::std::forward<_ExecutionPolicy>(__exec), __init_event, ::std::forward<_Range1>(__input),
             ::std::forward<_Range2>(__bins), __binhash_manager);
     }

From f1aad06b0ad94fe4807fb17551f24a54ff8bc432 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 11:53:21 +0100
Subject: [PATCH 251/566] __parallel_radix_sort + tag impls

---
 .../dpl/pstl/hetero/dpcpp/parallel_backend_sycl_radix_sort.h   | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_radix_sort.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_radix_sort.h
index 9b9f1bf9d5c..df262d54dfc 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_radix_sort.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_radix_sort.h
@@ -743,7 +743,8 @@ struct __parallel_radix_sort_iteration
 //-----------------------------------------------------------------------
 template <bool __is_ascending, typename _Range, typename _ExecutionPolicy, typename _Proj>
 auto
-__parallel_radix_sort(_ExecutionPolicy&& __exec, _Range&& __in_rng, _Proj __proj)
+__parallel_radix_sort(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&& __exec, _Range&& __in_rng,
+                      _Proj __proj)
 {
     const ::std::size_t __n = __in_rng.size();
     assert(__n > 1);

From 9fd8b676876cd0f73b75c8009e1ba4607617ca71 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 11:53:28 +0100
Subject: [PATCH 252/566] __parallel_radix_sort + tag calls

---
 include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h | 1 +
 1 file changed, 1 insertion(+)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
index cbd63bc9259..7ee5ba6fb9c 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
@@ -2195,6 +2195,7 @@ __parallel_stable_sort(oneapi::dpl::__internal::__device_backend_tag, _Execution
                        _Compare, _Proj __proj)
 {
     return __parallel_radix_sort<__internal::__is_comp_ascending<::std::decay_t<_Compare>>::value>(
+        oneapi::dpl::__internal::__device_backend_tag{},
         ::std::forward<_ExecutionPolicy>(__exec), ::std::forward<_Range>(__rng), __proj);
 }
 #endif

From b126b3388786fdc69040206b1085afaed5c57922 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 11:55:24 +0100
Subject: [PATCH 253/566] __parallel_transform_reduce_small_impl + tag impls

---
 .../dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h   | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h
index 6eab6ac03ec..178a802e5ea 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h
@@ -138,10 +138,11 @@ struct __parallel_transform_reduce_small_submitter<_Tp, __work_group_size, __ite
 
 template <typename _Tp, ::std::uint16_t __work_group_size, ::std::uint8_t __iters_per_work_item, typename _Commutative,
           typename _ExecutionPolicy, typename _Size, typename _ReduceOp, typename _TransformOp, typename _InitType,
-          oneapi::dpl::__internal::__enable_if_device_execution_policy<_ExecutionPolicy, int> = 0, typename... _Ranges>
+          typename... _Ranges>
 auto
-__parallel_transform_reduce_small_impl(_ExecutionPolicy&& __exec, const _Size __n, _ReduceOp __reduce_op,
-                                       _TransformOp __transform_op, _InitType __init, _Ranges&&... __rngs)
+__parallel_transform_reduce_small_impl(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&& __exec,
+                                       const _Size __n, _ReduceOp __reduce_op, _TransformOp __transform_op,
+                                       _InitType __init, _Ranges&&... __rngs)
 {
     using _CustomName = oneapi::dpl::__internal::__policy_kernel_name<_ExecutionPolicy>;
     using _ReduceKernel = oneapi::dpl::__par_backend_hetero::__internal::__kernel_name_provider<

From 10da4e77e77025f48ccf4116c9ecea8c0e0013df Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 11:55:34 +0100
Subject: [PATCH 254/566] __parallel_transform_reduce_small_impl + tag calls

---
 .../dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h    | 6 ++++++
 1 file changed, 6 insertions(+)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h
index 178a802e5ea..d2ebe069483 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h
@@ -532,36 +532,42 @@ __parallel_transform_reduce(oneapi::dpl::__internal::__device_backend_tag, _Exec
         if (__n <= 256)
         {
             return __parallel_transform_reduce_small_impl<_Tp, 256, 1, _Commutative>(
+                oneapi::dpl::__internal::__device_backend_tag{},
                 ::std::forward<_ExecutionPolicy>(__exec), __n, __reduce_op, __transform_op, __init,
                 ::std::forward<_Ranges>(__rngs)...);
         }
         else if (__n <= 512)
         {
             return __parallel_transform_reduce_small_impl<_Tp, 256, 2, _Commutative>(
+                oneapi::dpl::__internal::__device_backend_tag{},
                 ::std::forward<_ExecutionPolicy>(__exec), __n, __reduce_op, __transform_op, __init,
                 ::std::forward<_Ranges>(__rngs)...);
         }
         else if (__n <= 1024)
         {
             return __parallel_transform_reduce_small_impl<_Tp, 256, 4, _Commutative>(
+                oneapi::dpl::__internal::__device_backend_tag{},
                 ::std::forward<_ExecutionPolicy>(__exec), __n, __reduce_op, __transform_op, __init,
                 ::std::forward<_Ranges>(__rngs)...);
         }
         else if (__n <= 2048)
         {
             return __parallel_transform_reduce_small_impl<_Tp, 256, 8, _Commutative>(
+                oneapi::dpl::__internal::__device_backend_tag{},
                 ::std::forward<_ExecutionPolicy>(__exec), __n, __reduce_op, __transform_op, __init,
                 ::std::forward<_Ranges>(__rngs)...);
         }
         else if (__n <= 4096)
         {
             return __parallel_transform_reduce_small_impl<_Tp, 256, 16, _Commutative>(
+                oneapi::dpl::__internal::__device_backend_tag{},
                 ::std::forward<_ExecutionPolicy>(__exec), __n, __reduce_op, __transform_op, __init,
                 ::std::forward<_Ranges>(__rngs)...);
         }
         else if (__n <= 8192)
         {
             return __parallel_transform_reduce_small_impl<_Tp, 256, 32, _Commutative>(
+                oneapi::dpl::__internal::__device_backend_tag{},
                 ::std::forward<_ExecutionPolicy>(__exec), __n, __reduce_op, __transform_op, __init,
                 ::std::forward<_Ranges>(__rngs)...);
         }

From 53f0693780e72c73ce25486441dc59445f5fad06 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 11:57:10 +0100
Subject: [PATCH 255/566] __parallel_transform_reduce_mid_impl + tag impls

---
 .../dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h  | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h
index d2ebe069483..7dcb52bd767 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h
@@ -256,11 +256,11 @@ struct __parallel_transform_reduce_work_group_kernel_submitter<
 
 template <typename _Tp, ::std::uint16_t __work_group_size, ::std::uint8_t __iters_per_work_item_device_kernel,
           ::std::uint8_t __iters_per_work_item_work_group_kernel, typename _Commutative, typename _ExecutionPolicy,
-          typename _Size, typename _ReduceOp, typename _TransformOp, typename _InitType,
-          oneapi::dpl::__internal::__enable_if_device_execution_policy<_ExecutionPolicy, int> = 0, typename... _Ranges>
+          typename _Size, typename _ReduceOp, typename _TransformOp, typename _InitType, typename... _Ranges>
 auto
-__parallel_transform_reduce_mid_impl(_ExecutionPolicy&& __exec, _Size __n, _ReduceOp __reduce_op,
-                                     _TransformOp __transform_op, _InitType __init, _Ranges&&... __rngs)
+__parallel_transform_reduce_mid_impl(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&& __exec,
+                                     _Size __n, _ReduceOp __reduce_op, _TransformOp __transform_op, _InitType __init,
+                                     _Ranges&&... __rngs)
 {
     using _CustomName = oneapi::dpl::__internal::__policy_kernel_name<_ExecutionPolicy>;
 

From 39b069615a97b002b7d2b29927000d01ea181f89 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 11:57:17 +0100
Subject: [PATCH 256/566] __parallel_transform_reduce_mid_impl + tag calls

---
 .../dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h    | 6 ++++++
 1 file changed, 6 insertions(+)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h
index 7dcb52bd767..f1f8c04cac2 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h
@@ -578,36 +578,42 @@ __parallel_transform_reduce(oneapi::dpl::__internal::__device_backend_tag, _Exec
         else if (__n <= 2097152)
         {
             return __parallel_transform_reduce_mid_impl<_Tp, 256, 32, 1, _Commutative>(
+                oneapi::dpl::__internal::__device_backend_tag{},
                 ::std::forward<_ExecutionPolicy>(__exec), __n, __reduce_op, __transform_op, __init,
                 ::std::forward<_Ranges>(__rngs)...);
         }
         else if (__n <= 4194304)
         {
             return __parallel_transform_reduce_mid_impl<_Tp, 256, 32, 2, _Commutative>(
+                oneapi::dpl::__internal::__device_backend_tag{},
                 ::std::forward<_ExecutionPolicy>(__exec), __n, __reduce_op, __transform_op, __init,
                 ::std::forward<_Ranges>(__rngs)...);
         }
         else if (__n <= 8388608)
         {
             return __parallel_transform_reduce_mid_impl<_Tp, 256, 32, 4, _Commutative>(
+                oneapi::dpl::__internal::__device_backend_tag{},
                 ::std::forward<_ExecutionPolicy>(__exec), __n, __reduce_op, __transform_op, __init,
                 ::std::forward<_Ranges>(__rngs)...);
         }
         else if (__n <= 16777216)
         {
             return __parallel_transform_reduce_mid_impl<_Tp, 256, 32, 8, _Commutative>(
+                oneapi::dpl::__internal::__device_backend_tag{},
                 ::std::forward<_ExecutionPolicy>(__exec), __n, __reduce_op, __transform_op, __init,
                 ::std::forward<_Ranges>(__rngs)...);
         }
         else if (__n <= 33554432)
         {
             return __parallel_transform_reduce_mid_impl<_Tp, 256, 32, 16, _Commutative>(
+                oneapi::dpl::__internal::__device_backend_tag{},
                 ::std::forward<_ExecutionPolicy>(__exec), __n, __reduce_op, __transform_op, __init,
                 ::std::forward<_Ranges>(__rngs)...);
         }
         else if (__n <= 67108864)
         {
             return __parallel_transform_reduce_mid_impl<_Tp, 256, 32, 32, _Commutative>(
+                oneapi::dpl::__internal::__device_backend_tag{},
                 ::std::forward<_ExecutionPolicy>(__exec), __n, __reduce_op, __transform_op, __init,
                 ::std::forward<_Ranges>(__rngs)...);
         }

From 901d99085889be93d0005743cfa54251dbfe03ff Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 12:32:16 +0100
Subject: [PATCH 257/566] __internal::__remove_elements + tag impls

---
 include/oneapi/dpl/pstl/algorithm_impl.h | 17 ++++++++---------
 1 file changed, 8 insertions(+), 9 deletions(-)

diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 8790cc64f25..e9d589d2efe 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -2169,13 +2169,12 @@ __pattern_unique(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIte
 
 // That function is shared between two algorithms - remove_if (__pattern_remove_if) and unique (pattern unique). But a mask calculation is different.
 // So, a caller passes _CalcMask brick into remove_elements.
-template <class _ExecutionPolicy, class _ForwardIterator, class _CalcMask, class _IsVector>
+template <class _IsVector, class _ExecutionPolicy, class _ForwardIterator, class _CalcMask>
 _ForwardIterator
-__remove_elements(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last, _CalcMask __calc_mask,
-                  _IsVector __is_vector)
+__remove_elements(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _ForwardIterator __first,
+                  _ForwardIterator __last, _CalcMask __calc_mask)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+    using __backend_tag = typename decltype(__tag)::__backend_tag;
 
     typedef typename ::std::iterator_traits<_ForwardIterator>::difference_type _DifferenceType;
     typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _Tp;
@@ -2186,7 +2185,7 @@ __remove_elements(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardI
         bool* __mask = __mask_buf.get();
         _DifferenceType __min = __par_backend::__parallel_reduce(
             __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), _DifferenceType(0), __n, __n,
-            [__first, __mask, &__calc_mask, __is_vector](_DifferenceType __i, _DifferenceType __j,
+            [__first, __mask, &__calc_mask](_DifferenceType __i, _DifferenceType __j,
                                                          _DifferenceType __local_min) -> _DifferenceType {
                 // Create mask
                 __calc_mask(__mask + __i, __mask + __j, __first + __i);
@@ -2198,7 +2197,7 @@ __remove_elements(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardI
                 }
                 // find first iterator that should be removed
                 bool* __result = __internal::__brick_find_if(
-                    __mask + __i, __mask + __j, [](bool __val) { return !__val; }, __is_vector);
+                    __mask + __i, __mask + __j, [](bool __val) { return !__val; }, _IsVector{});
                 if (__result - __mask == __j)
                 {
                     return __local_min;
@@ -2224,9 +2223,9 @@ __remove_elements(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardI
         // 2. Elements that doesn't satisfy pred are moved to result
         __par_backend::__parallel_strict_scan(
             __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __n, _DifferenceType(0),
-            [__mask, __is_vector](_DifferenceType __i, _DifferenceType __len) {
+            [__mask](_DifferenceType __i, _DifferenceType __len) {
                 return __internal::__brick_count(
-                    __mask + __i, __mask + __i + __len, [](bool __val) { return __val; }, __is_vector);
+                    __mask + __i, __mask + __i + __len, [](bool __val) { return __val; }, _IsVector{});
             },
             ::std::plus<_DifferenceType>(),
             [=](_DifferenceType __i, _DifferenceType __len, _DifferenceType __initial) {

From 8c60029e6355753989418ca64896211bdab71378 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 12:32:43 +0100
Subject: [PATCH 258/566] __internal::__remove_elements + tag calls

---
 include/oneapi/dpl/pstl/algorithm_impl.h | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)

diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index e9d589d2efe..8c4569daf09 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -2279,7 +2279,7 @@ __pattern_unique(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _Rand
 
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _BinaryPredicate>
 _RandomAccessIterator
-__pattern_unique(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
+__pattern_unique(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
                  _RandomAccessIterator __last, _BinaryPredicate __pred)
 {
     typedef typename ::std::iterator_traits<_RandomAccessIterator>::reference _ReferenceType;
@@ -2294,6 +2294,7 @@ __pattern_unique(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAc
         return __internal::__brick_unique(__first, __last, __pred, _IsVector{});
     }
     return __internal::__remove_elements(
+        __tag,
         ::std::forward<_ExecutionPolicy>(__exec), ++__first, __last,
         [&__pred](bool* __b, bool* __e, _RandomAccessIterator __it) {
             __internal::__brick_walk3(
@@ -4738,7 +4739,7 @@ __pattern_remove_if(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _R
 
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _UnaryPredicate>
 _RandomAccessIterator
-__pattern_remove_if(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
+__pattern_remove_if(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
                     _RandomAccessIterator __last, _UnaryPredicate __pred)
 {
     typedef typename ::std::iterator_traits<_RandomAccessIterator>::reference _ReferenceType;
@@ -4750,6 +4751,7 @@ __pattern_remove_if(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Rando
     }
 
     return __internal::__remove_elements(
+        __tag,
         ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
         [&__pred](bool* __b, bool* __e, _RandomAccessIterator __it) {
             __internal::__brick_walk2(

From ed3f0e93e90b819710ba821c11f974f4a9cdabee Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 13:22:37 +0100
Subject: [PATCH 259/566] internal::lower_bound_impl + tag impls

---
 .../oneapi/dpl/internal/binary_search_impl.h   | 18 ++++++++++--------
 1 file changed, 10 insertions(+), 8 deletions(-)

diff --git a/include/oneapi/dpl/internal/binary_search_impl.h b/include/oneapi/dpl/internal/binary_search_impl.h
index 3fd155f03a0..cd5b384a46b 100644
--- a/include/oneapi/dpl/internal/binary_search_impl.h
+++ b/include/oneapi/dpl/internal/binary_search_impl.h
@@ -68,12 +68,14 @@ struct custom_brick
     }
 };
 
-template <typename Policy, typename InputIterator1, typename InputIterator2, typename OutputIterator,
+template <class _Tag, typename Policy, typename InputIterator1, typename InputIterator2, typename OutputIterator,
           typename StrictWeakOrdering>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<Policy, OutputIterator>
-lower_bound_impl(Policy&& policy, InputIterator1 start, InputIterator1 end, InputIterator2 value_start,
+OutputIterator
+lower_bound_impl(_Tag, Policy&& policy, InputIterator1 start, InputIterator1 end, InputIterator2 value_start,
                  InputIterator2 value_end, OutputIterator result, StrictWeakOrdering comp)
 {
+    static_assert(__is_backend_tag_v<_Tag>);
+
     return oneapi::dpl::transform(policy, value_start, value_end, result,
                                   [=](typename ::std::iterator_traits<InputIterator2>::reference val) {
                                       return ::std::lower_bound(start, end, val, comp) - start;
@@ -105,11 +107,11 @@ binary_search_impl(Policy&& policy, InputIterator1 start, InputIterator1 end, In
 }
 
 #if _ONEDPL_BACKEND_SYCL
-template <typename Policy, typename InputIterator1, typename InputIterator2, typename OutputIterator,
-          typename StrictWeakOrdering>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<Policy, OutputIterator>
-lower_bound_impl(Policy&& policy, InputIterator1 start, InputIterator1 end, InputIterator2 value_start,
-                 InputIterator2 value_end, OutputIterator result, StrictWeakOrdering comp)
+template <typename _BackendTag, typename Policy, typename InputIterator1, typename InputIterator2,
+          typename OutputIterator, typename StrictWeakOrdering>
+OutputIterator
+lower_bound_impl(__hetero_tag<_BackendTag> __tag, Policy&& policy, InputIterator1 start, InputIterator1 end,
+                 InputIterator2 value_start, InputIterator2 value_end, OutputIterator result, StrictWeakOrdering comp)
 {
     constexpr auto __dispatch_tag =
         oneapi::dpl::__internal::__select_backend<Policy, InputIterator1, InputIterator2, OutputIterator>();

From 925553c2fd27082ac4a5253cc1d9bfc27d0eb82b Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 13:22:45 +0100
Subject: [PATCH 260/566] internal::lower_bound_impl + tag calls

---
 include/oneapi/dpl/internal/binary_search_impl.h | 13 ++++++++++---
 1 file changed, 10 insertions(+), 3 deletions(-)

diff --git a/include/oneapi/dpl/internal/binary_search_impl.h b/include/oneapi/dpl/internal/binary_search_impl.h
index cd5b384a46b..ffa1de8896d 100644
--- a/include/oneapi/dpl/internal/binary_search_impl.h
+++ b/include/oneapi/dpl/internal/binary_search_impl.h
@@ -218,8 +218,11 @@ oneapi::dpl::__internal::__enable_if_execution_policy<Policy, OutputIterator>
 lower_bound(Policy&& policy, InputIterator1 start, InputIterator1 end, InputIterator2 value_start,
             InputIterator2 value_end, OutputIterator result)
 {
-    return internal::lower_bound_impl(::std::forward<Policy>(policy), start, end, value_start, value_end, result,
-                                      oneapi::dpl::__internal::__pstl_less());
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<Policy, InputIterator1, InputIterator2, OutputIterator>();
+
+    return internal::lower_bound_impl(__dispatch_tag, ::std::forward<Policy>(policy), start, end, value_start,
+                                      value_end, result, oneapi::dpl::__internal::__pstl_less());
 }
 
 template <typename Policy, typename InputIterator1, typename InputIterator2, typename OutputIterator,
@@ -228,7 +231,11 @@ oneapi::dpl::__internal::__enable_if_execution_policy<Policy, OutputIterator>
 lower_bound(Policy&& policy, InputIterator1 start, InputIterator1 end, InputIterator2 value_start,
             InputIterator2 value_end, OutputIterator result, StrictWeakOrdering comp)
 {
-    return internal::lower_bound_impl(::std::forward<Policy>(policy), start, end, value_start, value_end, result, comp);
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<Policy, InputIterator1, InputIterator2, OutputIterator>();
+
+    return internal::lower_bound_impl(__dispatch_tag, ::std::forward<Policy>(policy), start, end, value_start,
+                                      value_end, result, comp);
 }
 //Lower Bound end
 

From ce75efc204ff8252e737b85b4f6b448313a10701 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 13:26:46 +0100
Subject: [PATCH 261/566] internal::binary_search_impl + tag impls

---
 .../oneapi/dpl/internal/binary_search_impl.h  | 22 +++++++++----------
 1 file changed, 11 insertions(+), 11 deletions(-)

diff --git a/include/oneapi/dpl/internal/binary_search_impl.h b/include/oneapi/dpl/internal/binary_search_impl.h
index ffa1de8896d..2c9873c7962 100644
--- a/include/oneapi/dpl/internal/binary_search_impl.h
+++ b/include/oneapi/dpl/internal/binary_search_impl.h
@@ -94,12 +94,14 @@ upper_bound_impl(Policy&& policy, InputIterator1 start, InputIterator1 end, Inpu
                                   });
 }
 
-template <typename Policy, typename InputIterator1, typename InputIterator2, typename OutputIterator,
+template <class _Tag, typename Policy, typename InputIterator1, typename InputIterator2, typename OutputIterator,
           typename StrictWeakOrdering>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<Policy, OutputIterator>
-binary_search_impl(Policy&& policy, InputIterator1 start, InputIterator1 end, InputIterator2 value_start,
+OutputIterator
+binary_search_impl(_Tag, Policy&& policy, InputIterator1 start, InputIterator1 end, InputIterator2 value_start,
                    InputIterator2 value_end, OutputIterator result, StrictWeakOrdering comp)
 {
+    static_assert(__is_backend_tag_v<_Tag>);
+
     return oneapi::dpl::transform(policy, value_start, value_end, result,
                                   [=](typename ::std::iterator_traits<InputIterator2>::reference val) {
                                       return ::std::binary_search(start, end, val, comp);
@@ -175,15 +177,13 @@ upper_bound_impl(Policy&& policy, InputIterator1 start, InputIterator1 end, Inpu
     return result + value_size;
 }
 
-template <typename Policy, typename InputIterator1, typename InputIterator2, typename OutputIterator,
-          typename StrictWeakOrdering>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<Policy, OutputIterator>
-binary_search_impl(Policy&& policy, InputIterator1 start, InputIterator1 end, InputIterator2 value_start,
-                   InputIterator2 value_end, OutputIterator result, StrictWeakOrdering comp)
+template <typename _BackendTag, typename Policy, typename InputIterator1, typename InputIterator2,
+          typename OutputIterator, typename StrictWeakOrdering>
+OutputIterator
+binary_search_impl(__hetero_tag<_BackendTag> __tag, Policy&& policy, InputIterator1 start, InputIterator1 end,
+                   InputIterator2 value_start, InputIterator2 value_end, OutputIterator result, StrictWeakOrdering comp)
 {
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<Policy, InputIterator1, InputIterator2, OutputIterator>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+    using __backend_tag = typename decltype(__tag)::__backend_tag;
 
     namespace __bknd = __par_backend_hetero;
     const auto size = ::std::distance(start, end);

From 08e00e8d35733b4863a3887150bc5dd650ada9f4 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 13:26:52 +0100
Subject: [PATCH 262/566] internal::binary_search_impl + tag calls

---
 include/oneapi/dpl/internal/binary_search_impl.h | 14 ++++++++++----
 1 file changed, 10 insertions(+), 4 deletions(-)

diff --git a/include/oneapi/dpl/internal/binary_search_impl.h b/include/oneapi/dpl/internal/binary_search_impl.h
index 2c9873c7962..198b9c27170 100644
--- a/include/oneapi/dpl/internal/binary_search_impl.h
+++ b/include/oneapi/dpl/internal/binary_search_impl.h
@@ -268,8 +268,11 @@ oneapi::dpl::__internal::__enable_if_execution_policy<Policy, OutputIterator>
 binary_search(Policy&& policy, InputIterator1 start, InputIterator1 end, InputIterator2 value_start,
               InputIterator2 value_end, OutputIterator result)
 {
-    return internal::binary_search_impl(::std::forward<Policy>(policy), start, end, value_start, value_end, result,
-                                        oneapi::dpl::__internal::__pstl_less());
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<Policy, InputIterator1, InputIterator2, OutputIterator>();
+
+    return internal::binary_search_impl(__dispatch_tag, ::std::forward<Policy>(policy), start, end, value_start,
+                                        value_end, result, oneapi::dpl::__internal::__pstl_less());
 }
 
 template <typename Policy, typename InputIterator1, typename InputIterator2, typename OutputIterator,
@@ -278,8 +281,11 @@ oneapi::dpl::__internal::__enable_if_execution_policy<Policy, OutputIterator>
 binary_search(Policy&& policy, InputIterator1 start, InputIterator1 end, InputIterator2 value_start,
               InputIterator2 value_end, OutputIterator result, StrictWeakOrdering comp)
 {
-    return internal::binary_search_impl(::std::forward<Policy>(policy), start, end, value_start, value_end, result,
-                                        comp);
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<Policy, InputIterator1, InputIterator2, OutputIterator>();
+
+    return internal::binary_search_impl(__dispatch_tag, ::std::forward<Policy>(policy), start, end, value_start,
+                                        value_end, result, comp);
 }
 
 //Binary search end

From 6b831c6734a026382286f349d39562703de8a2c5 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 13:30:27 +0100
Subject: [PATCH 263/566] internal::pattern_exclusive_scan_by_segment + tag
 impls

---
 .../internal/exclusive_scan_by_segment_impl.h | 24 +++++++++++--------
 1 file changed, 14 insertions(+), 10 deletions(-)

diff --git a/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h b/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h
index 79d8239ea12..ee000e175aa 100644
--- a/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h
@@ -37,12 +37,15 @@ class ExclusiveScan1;
 template <typename Name>
 class ExclusiveScan2;
 
-template <typename Policy, typename InputIterator1, typename InputIterator2, typename OutputIterator, typename T,
-          typename BinaryPredicate, typename Operator>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<Policy, OutputIterator>
-pattern_exclusive_scan_by_segment(Policy&& policy, InputIterator1 first1, InputIterator1 last1, InputIterator2 first2,
-                                  OutputIterator result, T init, BinaryPredicate binary_pred, Operator binary_op)
+template <class _Tag, typename Policy, typename InputIterator1, typename InputIterator2, typename OutputIterator,
+          typename T, typename BinaryPredicate, typename Operator>
+OutputIterator
+pattern_exclusive_scan_by_segment(_Tag, Policy&& policy, InputIterator1 first1, InputIterator1 last1,
+                                  InputIterator2 first2, OutputIterator result, T init, BinaryPredicate binary_pred,
+                                  Operator binary_op)
 {
+    static_assert(__is_backend_tag_v<_Tag>);
+
     const auto n = ::std::distance(first1, last1);
 
     // Check for empty and single element ranges
@@ -160,11 +163,12 @@ exclusive_scan_by_segment_impl(Policy&& policy, InputIterator1 first1, InputIter
     return result + n;
 }
 
-template <typename Policy, typename InputIterator1, typename InputIterator2, typename OutputIterator, typename T,
-          typename BinaryPredicate, typename Operator>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<Policy, OutputIterator>
-pattern_exclusive_scan_by_segment(Policy&& policy, InputIterator1 first1, InputIterator1 last1, InputIterator2 first2,
-                                  OutputIterator result, T init, BinaryPredicate binary_pred, Operator binary_op)
+template <typename _BackendTag, typename Policy, typename InputIterator1, typename InputIterator2,
+          typename OutputIterator, typename T, typename BinaryPredicate, typename Operator>
+OutputIterator
+pattern_exclusive_scan_by_segment(__hetero_tag<_BackendTag> __tag, Policy&& policy, InputIterator1 first1,
+                                  InputIterator1 last1, InputIterator2 first2, OutputIterator result, T init,
+                                  BinaryPredicate binary_pred, Operator binary_op)
 {
     return internal::exclusive_scan_by_segment_impl(
         ::std::forward<Policy>(policy), first1, last1, first2, result, init, binary_pred, binary_op,

From c88c653a6fe9ae4d06e06e0c01b26e8283e5c02f Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 13:30:36 +0100
Subject: [PATCH 264/566] internal::pattern_exclusive_scan_by_segment + tag
 calls

---
 .../oneapi/dpl/internal/exclusive_scan_by_segment_impl.h   | 7 +++++--
 1 file changed, 5 insertions(+), 2 deletions(-)

diff --git a/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h b/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h
index ee000e175aa..7e25849a09f 100644
--- a/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h
@@ -185,8 +185,11 @@ oneapi::dpl::__internal::__enable_if_execution_policy<Policy, OutputIterator>
 exclusive_scan_by_segment(Policy&& policy, InputIterator1 first1, InputIterator1 last1, InputIterator2 first2,
                           OutputIterator result, T init, BinaryPredicate binary_pred, Operator binary_op)
 {
-    return internal::pattern_exclusive_scan_by_segment(::std::forward<Policy>(policy), first1, last1, first2, result,
-                                                       init, binary_pred, binary_op);
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<Policy, InputIterator1, InputIterator2, OutputIterator>();
+
+    return internal::pattern_exclusive_scan_by_segment(__dispatch_tag, ::std::forward<Policy>(policy), first1, last1,
+                                                       first2, result, init, binary_pred, binary_op);
 }
 
 template <typename Policy, typename InputIterator1, typename InputIterator2, typename OutputIterator, typename T,

From 3d750426f0f1ce60e3d3fca8bd5d24f379cbbe7e Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 13:32:54 +0100
Subject: [PATCH 265/566] internal::pattern_inclusive_scan_by_segment + tag
 impls

---
 .../internal/inclusive_scan_by_segment_impl.h | 22 +++++++++++--------
 1 file changed, 13 insertions(+), 9 deletions(-)

diff --git a/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h b/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h
index 0f0433b2bb8..c92b6127b70 100644
--- a/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h
@@ -36,12 +36,15 @@ namespace internal
 template <typename Name>
 class InclusiveScan1;
 
-template <typename Policy, typename InputIterator1, typename InputIterator2, typename OutputIterator,
+template <class _Tag, typename Policy, typename InputIterator1, typename InputIterator2, typename OutputIterator,
           typename BinaryPredicate, typename BinaryOperator>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<Policy, OutputIterator>
-pattern_inclusive_scan_by_segment(Policy&& policy, InputIterator1 first1, InputIterator1 last1, InputIterator2 first2,
-                                  OutputIterator result, BinaryPredicate binary_pred, BinaryOperator binary_op)
+OutputIterator
+pattern_inclusive_scan_by_segment(_Tag, Policy&& policy, InputIterator1 first1, InputIterator1 last1,
+                                  InputIterator2 first2, OutputIterator result, BinaryPredicate binary_pred,
+                                  BinaryOperator binary_op)
 {
+    static_assert(__is_backend_tag_v<_Tag>);
+
     const auto n = ::std::distance(first1, last1);
 
     // Check for empty and single element ranges
@@ -123,11 +126,12 @@ inclusive_scan_by_segment_impl(Policy&& policy, InputIterator1 first1, InputIter
     return result + n;
 }
 
-template <typename Policy, typename InputIterator1, typename InputIterator2, typename OutputIterator,
-          typename BinaryPredicate, typename BinaryOperator>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<Policy, OutputIterator>
-pattern_inclusive_scan_by_segment(Policy&& policy, InputIterator1 first1, InputIterator1 last1, InputIterator2 first2,
-                                  OutputIterator result, BinaryPredicate binary_pred, BinaryOperator binary_op)
+template <typename _BackendTag, typename Policy, typename InputIterator1, typename InputIterator2,
+          typename OutputIterator, typename BinaryPredicate, typename BinaryOperator>
+OutputIterator
+pattern_inclusive_scan_by_segment(__hetero_tag<_BackendTag> __tag, Policy&& policy, InputIterator1 first1,
+                                  InputIterator1 last1, InputIterator2 first2, OutputIterator result,
+                                  BinaryPredicate binary_pred, BinaryOperator binary_op)
 {
     return internal::inclusive_scan_by_segment_impl(
         ::std::forward<Policy>(policy), first1, last1, first2, result, binary_pred, binary_op,

From 9c1447355ac2de2a409a19ece8d0e346c3925956 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 13:33:00 +0100
Subject: [PATCH 266/566] internal::pattern_inclusive_scan_by_segment + tag
 calls

---
 .../oneapi/dpl/internal/inclusive_scan_by_segment_impl.h   | 7 +++++--
 1 file changed, 5 insertions(+), 2 deletions(-)

diff --git a/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h b/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h
index c92b6127b70..a44ebd6b617 100644
--- a/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h
@@ -148,8 +148,11 @@ oneapi::dpl::__internal::__enable_if_execution_policy<Policy, OutputIterator>
 inclusive_scan_by_segment(Policy&& policy, InputIterator1 first1, InputIterator1 last1, InputIterator2 first2,
                           OutputIterator result, BinaryPredicate binary_pred, BinaryOperator binary_op)
 {
-    return internal::pattern_inclusive_scan_by_segment(::std::forward<Policy>(policy), first1, last1, first2, result,
-                                                       binary_pred, binary_op);
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<Policy, InputIterator1, InputIterator2, OutputIterator>();
+
+    return internal::pattern_inclusive_scan_by_segment(__dispatch_tag, ::std::forward<Policy>(policy), first1, last1,
+                                                       first2, result, binary_pred, binary_op);
 }
 
 template <typename Policy, typename InputIter1, typename InputIter2, typename OutputIter, typename BinaryPredicate>

From b6b67966ade6a0a7dd8212c686fbca1d1307ad57 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 13:39:53 +0100
Subject: [PATCH 267/566] internal::reduce_by_segment_impl + tag impls

---
 .../dpl/internal/reduce_by_segment_impl.h      | 18 ++++++++++--------
 1 file changed, 10 insertions(+), 8 deletions(-)

diff --git a/include/oneapi/dpl/internal/reduce_by_segment_impl.h b/include/oneapi/dpl/internal/reduce_by_segment_impl.h
index 0102d6c2925..e664e563b4d 100644
--- a/include/oneapi/dpl/internal/reduce_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/reduce_by_segment_impl.h
@@ -78,13 +78,15 @@ class Reduce3;
 template <typename Name>
 class Reduce4;
 
-template <typename Policy, typename InputIterator1, typename InputIterator2, typename OutputIterator1,
+template <class _Tag, typename Policy, typename InputIterator1, typename InputIterator2, typename OutputIterator1,
           typename OutputIterator2, typename BinaryPred, typename BinaryOperator>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<Policy, ::std::pair<OutputIterator1, OutputIterator2>>
-reduce_by_segment_impl(Policy&& policy, InputIterator1 first1, InputIterator1 last1, InputIterator2 first2,
+::std::pair<OutputIterator1, OutputIterator2>
+reduce_by_segment_impl(_Tag, Policy&& policy, InputIterator1 first1, InputIterator1 last1, InputIterator2 first2,
                        OutputIterator1 result1, OutputIterator2 result2, BinaryPred binary_pred,
                        BinaryOperator binary_op)
 {
+    static_assert(__is_backend_tag_v<_Tag>);
+
     // The algorithm reduces values in [first2, first2 + (last1-first1)) where the associated
     // keys for the values are equal to the adjacent key. This function's implementation is a derivative work
     // and responsible for the second copyright notice in this header.
@@ -570,11 +572,11 @@ __sycl_reduce_by_segment(_ExecutionPolicy&& __exec, _Range1&& __keys, _Range2&&
     return __end_idx.get_host_access()[0] + 1;
 }
 
-template <typename Policy, typename InputIterator1, typename InputIterator2, typename OutputIterator1,
-          typename OutputIterator2, typename BinaryPred, typename BinaryOperator>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<Policy, ::std::pair<OutputIterator1, OutputIterator2>>
-reduce_by_segment_impl(Policy&& policy, InputIterator1 first1, InputIterator1 last1, InputIterator2 first2,
-                       OutputIterator1 result1, OutputIterator2 result2, BinaryPred binary_pred,
+template <typename _BackendTag, typename Policy, typename InputIterator1, typename InputIterator2,
+          typename OutputIterator1, typename OutputIterator2, typename BinaryPred, typename BinaryOperator>
+::std::pair<OutputIterator1, OutputIterator2>
+reduce_by_segment_impl(__hetero_tag<_BackendTag>, Policy&& policy, InputIterator1 first1, InputIterator1 last1,
+                       InputIterator2 first2, OutputIterator1 result1, OutputIterator2 result2, BinaryPred binary_pred,
                        BinaryOperator binary_op)
 {
     // The algorithm reduces values in [first2, first2 + (last1-first1)) where the associated

From 4209762b15686a9f0af2f95d82873e3420436a06 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 13:40:01 +0100
Subject: [PATCH 268/566] internal::reduce_by_segment_impl + tag calls

---
 include/oneapi/dpl/internal/reduce_by_segment_impl.h | 8 ++++++--
 1 file changed, 6 insertions(+), 2 deletions(-)

diff --git a/include/oneapi/dpl/internal/reduce_by_segment_impl.h b/include/oneapi/dpl/internal/reduce_by_segment_impl.h
index e664e563b4d..8a44c0836bb 100644
--- a/include/oneapi/dpl/internal/reduce_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/reduce_by_segment_impl.h
@@ -626,8 +626,12 @@ oneapi::dpl::__internal::__enable_if_execution_policy<Policy, ::std::pair<Output
 reduce_by_segment(Policy&& policy, InputIterator1 first1, InputIterator1 last1, InputIterator2 first2,
                   OutputIterator1 result1, OutputIterator2 result2, BinaryPred binary_pred, BinaryOperator binary_op)
 {
-    return internal::reduce_by_segment_impl(::std::forward<Policy>(policy), first1, last1, first2, result1, result2,
-                                            binary_pred, binary_op);
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, InputIterator1, InputIterator2, OutputIterator1,
+                                                  OutputIterator2>();
+
+    return internal::reduce_by_segment_impl(__dispatch_tag, ::std::forward<Policy>(policy), first1, last1, first2,
+                                            result1, result2, binary_pred, binary_op);
 }
 
 template <typename Policy, typename InputIt1, typename InputIt2, typename OutputIt1, typename OutputIt2,

From 878f070fb3249d132fe4746773be5d32cd5f714d Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 13:44:14 +0100
Subject: [PATCH 269/566] internal::upper_bound_impl + tag impls

---
 .../oneapi/dpl/internal/binary_search_impl.h  | 22 +++++++++----------
 1 file changed, 11 insertions(+), 11 deletions(-)

diff --git a/include/oneapi/dpl/internal/binary_search_impl.h b/include/oneapi/dpl/internal/binary_search_impl.h
index 198b9c27170..252e7f37009 100644
--- a/include/oneapi/dpl/internal/binary_search_impl.h
+++ b/include/oneapi/dpl/internal/binary_search_impl.h
@@ -82,12 +82,14 @@ lower_bound_impl(_Tag, Policy&& policy, InputIterator1 start, InputIterator1 end
                                   });
 }
 
-template <typename Policy, typename InputIterator1, typename InputIterator2, typename OutputIterator,
+template <class _Tag, typename Policy, typename InputIterator1, typename InputIterator2, typename OutputIterator,
           typename StrictWeakOrdering>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<Policy, OutputIterator>
-upper_bound_impl(Policy&& policy, InputIterator1 start, InputIterator1 end, InputIterator2 value_start,
+OutputIterator
+upper_bound_impl(_Tag, Policy&& policy, InputIterator1 start, InputIterator1 end, InputIterator2 value_start,
                  InputIterator2 value_end, OutputIterator result, StrictWeakOrdering comp)
 {
+    static_assert(__is_backend_tag_v<_Tag>);
+
     return oneapi::dpl::transform(policy, value_start, value_end, result,
                                   [=](typename ::std::iterator_traits<InputIterator2>::reference val) {
                                       return ::std::upper_bound(start, end, val, comp) - start;
@@ -143,15 +145,13 @@ lower_bound_impl(__hetero_tag<_BackendTag> __tag, Policy&& policy, InputIterator
     return result + value_size;
 }
 
-template <typename Policy, typename InputIterator1, typename InputIterator2, typename OutputIterator,
-          typename StrictWeakOrdering>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<Policy, OutputIterator>
-upper_bound_impl(Policy&& policy, InputIterator1 start, InputIterator1 end, InputIterator2 value_start,
-                 InputIterator2 value_end, OutputIterator result, StrictWeakOrdering comp)
+template <typename _BackendTag, typename Policy, typename InputIterator1, typename InputIterator2,
+          typename OutputIterator, typename StrictWeakOrdering>
+OutputIterator
+upper_bound_impl(__hetero_tag<_BackendTag> __tag, Policy&& policy, InputIterator1 start, InputIterator1 end,
+                 InputIterator2 value_start, InputIterator2 value_end, OutputIterator result, StrictWeakOrdering comp)
 {
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<Policy, InputIterator1, InputIterator2, OutputIterator>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+    using __backend_tag = typename decltype(__tag)::__backend_tag;
 
     namespace __bknd = __par_backend_hetero;
     const auto size = ::std::distance(start, end);

From 1ffe860eb495b986580a72f7b5c469b071dedc03 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 13:44:22 +0100
Subject: [PATCH 270/566] internal::upper_bound_impl + tag calls

---
 include/oneapi/dpl/internal/binary_search_impl.h | 13 ++++++++++---
 1 file changed, 10 insertions(+), 3 deletions(-)

diff --git a/include/oneapi/dpl/internal/binary_search_impl.h b/include/oneapi/dpl/internal/binary_search_impl.h
index 252e7f37009..8a6f6b2a910 100644
--- a/include/oneapi/dpl/internal/binary_search_impl.h
+++ b/include/oneapi/dpl/internal/binary_search_impl.h
@@ -246,8 +246,11 @@ oneapi::dpl::__internal::__enable_if_execution_policy<Policy, OutputIterator>
 upper_bound(Policy&& policy, InputIterator1 start, InputIterator1 end, InputIterator2 value_start,
             InputIterator2 value_end, OutputIterator result)
 {
-    return internal::upper_bound_impl(::std::forward<Policy>(policy), start, end, value_start, value_end, result,
-                                      oneapi::dpl::__internal::__pstl_less());
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<Policy, InputIterator1, InputIterator2, OutputIterator>();
+
+    return internal::upper_bound_impl(__dispatch_tag, ::std::forward<Policy>(policy), start, end, value_start,
+                                      value_end, result, oneapi::dpl::__internal::__pstl_less());
 }
 
 template <typename Policy, typename InputIterator1, typename InputIterator2, typename OutputIterator,
@@ -256,7 +259,11 @@ oneapi::dpl::__internal::__enable_if_execution_policy<Policy, OutputIterator>
 upper_bound(Policy&& policy, InputIterator1 start, InputIterator1 end, InputIterator2 value_start,
             InputIterator2 value_end, OutputIterator result, StrictWeakOrdering comp)
 {
-    return internal::upper_bound_impl(::std::forward<Policy>(policy), start, end, value_start, value_end, result, comp);
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<Policy, InputIterator1, InputIterator2, OutputIterator>();
+
+    return internal::upper_bound_impl(__dispatch_tag, ::std::forward<Policy>(policy), start, end, value_start,
+                                      value_end, result, comp);
 }
 
 //Upper Bound end

From 1dcc5f23c5907f04a855193fa20853235b4581b0 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 13:54:17 +0100
Subject: [PATCH 271/566] internal::exclusive_scan_by_segment_impl + tag impls

---
 .../internal/exclusive_scan_by_segment_impl.h | 24 +++++++++----------
 1 file changed, 12 insertions(+), 12 deletions(-)

diff --git a/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h b/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h
index 7e25849a09f..ddc9b774abb 100644
--- a/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h
@@ -94,23 +94,23 @@ pattern_exclusive_scan_by_segment(_Tag, Policy&& policy, InputIterator1 first1,
 }
 
 #if _ONEDPL_BACKEND_SYCL
-template <typename Policy, typename InputIterator1, typename InputIterator2, typename OutputIterator, typename T,
-          typename BinaryPredicate, typename Operator>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<Policy, OutputIterator>
-exclusive_scan_by_segment_impl(Policy&& policy, InputIterator1 first1, InputIterator1 last1, InputIterator2 first2,
-                               OutputIterator result, T init, BinaryPredicate binary_pred, Operator binary_op,
-                               ::std::true_type /* has_known_identity*/)
+template <typename _BackendTag, typename Policy, typename InputIterator1, typename InputIterator2,
+          typename OutputIterator, typename T, typename BinaryPredicate, typename Operator>
+OutputIterator
+exclusive_scan_by_segment_impl(__hetero_tag<_BackendTag>, Policy&& policy, InputIterator1 first1, InputIterator1 last1,
+                               InputIterator2 first2, OutputIterator result, T init, BinaryPredicate binary_pred,
+                               Operator binary_op, ::std::true_type /* has_known_identity*/)
 {
     return internal::__scan_by_segment_impl_common(::std::forward<Policy>(policy), first1, last1, first2, result, init,
                                                    binary_pred, binary_op, ::std::false_type{});
 }
 
-template <typename Policy, typename InputIterator1, typename InputIterator2, typename OutputIterator, typename T,
-          typename BinaryPredicate, typename Operator>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<Policy, OutputIterator>
-exclusive_scan_by_segment_impl(Policy&& policy, InputIterator1 first1, InputIterator1 last1, InputIterator2 first2,
-                               OutputIterator result, T init, BinaryPredicate binary_pred, Operator binary_op,
-                               ::std::false_type /* has_known_identity*/)
+template <typename _BackendTag, typename Policy, typename InputIterator1, typename InputIterator2,
+          typename OutputIterator, typename T, typename BinaryPredicate, typename Operator>
+OutputIterator
+exclusive_scan_by_segment_impl(__hetero_tag<_BackendTag>, Policy&& policy, InputIterator1 first1, InputIterator1 last1,
+                               InputIterator2 first2, OutputIterator result, T init, BinaryPredicate binary_pred,
+                               Operator binary_op, ::std::false_type /* has_known_identity*/)
 {
 
     const auto n = ::std::distance(first1, last1);

From 5543b62163c14c9300e02124957c0132bf6e08ec Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 13:54:47 +0100
Subject: [PATCH 272/566] internal::exclusive_scan_by_segment_impl + tag calls

---
 include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h b/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h
index ddc9b774abb..fb1beb9b3eb 100644
--- a/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h
@@ -171,7 +171,7 @@ pattern_exclusive_scan_by_segment(__hetero_tag<_BackendTag> __tag, Policy&& poli
                                   BinaryPredicate binary_pred, Operator binary_op)
 {
     return internal::exclusive_scan_by_segment_impl(
-        ::std::forward<Policy>(policy), first1, last1, first2, result, init, binary_pred, binary_op,
+        __tag, ::std::forward<Policy>(policy), first1, last1, first2, result, init, binary_pred, binary_op,
         typename unseq_backend::__has_known_identity<
             Operator, typename ::std::iterator_traits<InputIterator2>::value_type>::type{});
 }

From 4a23e3e39722ec5551c25f436d7c09ba69d52dc4 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 13:59:03 +0100
Subject: [PATCH 273/566] internal::inclusive_scan_by_segment_impl + tag impls

---
 .../internal/inclusive_scan_by_segment_impl.h | 24 +++++++++----------
 1 file changed, 12 insertions(+), 12 deletions(-)

diff --git a/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h b/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h
index a44ebd6b617..64a77e8bd5d 100644
--- a/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h
@@ -75,12 +75,12 @@ pattern_inclusive_scan_by_segment(_Tag, Policy&& policy, InputIterator1 first1,
 }
 
 #if _ONEDPL_BACKEND_SYCL
-template <typename Policy, typename InputIterator1, typename InputIterator2, typename OutputIterator,
-          typename BinaryPredicate, typename BinaryOperator>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<Policy, OutputIterator>
-inclusive_scan_by_segment_impl(Policy&& policy, InputIterator1 first1, InputIterator1 last1, InputIterator2 first2,
-                               OutputIterator result, BinaryPredicate binary_pred, BinaryOperator binary_op,
-                               ::std::true_type /* has_known_identity */)
+template <typename _BackendTag, typename Policy, typename InputIterator1, typename InputIterator2,
+          typename OutputIterator, typename BinaryPredicate, typename BinaryOperator>
+OutputIterator
+inclusive_scan_by_segment_impl(__hetero_tag<_BackendTag>, Policy&& policy, InputIterator1 first1, InputIterator1 last1,
+                               InputIterator2 first2, OutputIterator result, BinaryPredicate binary_pred,
+                               BinaryOperator binary_op, ::std::true_type /* has_known_identity */)
 {
     using iter_value_t = typename ::std::iterator_traits<InputIterator2>::value_type;
     iter_value_t identity = unseq_backend::__known_identity<BinaryOperator, iter_value_t>;
@@ -88,12 +88,12 @@ inclusive_scan_by_segment_impl(Policy&& policy, InputIterator1 first1, InputIter
                                                    identity, binary_pred, binary_op, ::std::true_type{});
 }
 
-template <typename Policy, typename InputIterator1, typename InputIterator2, typename OutputIterator,
-          typename BinaryPredicate, typename BinaryOperator>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<Policy, OutputIterator>
-inclusive_scan_by_segment_impl(Policy&& policy, InputIterator1 first1, InputIterator1 last1, InputIterator2 first2,
-                               OutputIterator result, BinaryPredicate binary_pred, BinaryOperator binary_op,
-                               ::std::false_type /* has_known_identity */)
+template <typename _BackendTag, typename Policy, typename InputIterator1, typename InputIterator2,
+          typename OutputIterator, typename BinaryPredicate, typename BinaryOperator>
+OutputIterator
+inclusive_scan_by_segment_impl(__hetero_tag<_BackendTag>, Policy&& policy, InputIterator1 first1, InputIterator1 last1,
+                               InputIterator2 first2, OutputIterator result, BinaryPredicate binary_pred,
+                               BinaryOperator binary_op, ::std::false_type /* has_known_identity */)
 {
 
     typedef unsigned int FlagType;

From 40cc8d4632bf9dc9fc936a736b1d63a7cd899f29 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 13:59:14 +0100
Subject: [PATCH 274/566] internal::inclusive_scan_by_segment_impl + tag calls

---
 include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h b/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h
index 64a77e8bd5d..08d92bed504 100644
--- a/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h
@@ -134,7 +134,7 @@ pattern_inclusive_scan_by_segment(__hetero_tag<_BackendTag> __tag, Policy&& poli
                                   BinaryPredicate binary_pred, BinaryOperator binary_op)
 {
     return internal::inclusive_scan_by_segment_impl(
-        ::std::forward<Policy>(policy), first1, last1, first2, result, binary_pred, binary_op,
+        __tag, ::std::forward<Policy>(policy), first1, last1, first2, result, binary_pred, binary_op,
         typename unseq_backend::__has_known_identity<
             BinaryOperator, typename ::std::iterator_traits<InputIterator2>::value_type>::type{});
 }

From fe6bc53e40f8214b0a1e51c1ae86c03f3f7eb845 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 14:07:21 +0100
Subject: [PATCH 275/566] internal::__sycl_reduce_by_segment + tag impls

---
 .../dpl/internal/reduce_by_segment_impl.h     | 26 +++++++++----------
 1 file changed, 12 insertions(+), 14 deletions(-)

diff --git a/include/oneapi/dpl/internal/reduce_by_segment_impl.h b/include/oneapi/dpl/internal/reduce_by_segment_impl.h
index 8a44c0836bb..35465ffc5d1 100644
--- a/include/oneapi/dpl/internal/reduce_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/reduce_by_segment_impl.h
@@ -190,26 +190,24 @@ template <typename... _Name>
 using _SegReducePrefixPhase = __seg_reduce_prefix_kernel<_Name...>;
 } // namespace
 
-template <typename _ExecutionPolicy, typename _Range1, typename _Range2, typename _Range3, typename _Range4,
-          typename _BinaryPredicate, typename _BinaryOperator>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy,
-                                                             oneapi::dpl::__internal::__difference_t<_Range3>>
-__sycl_reduce_by_segment(_ExecutionPolicy&& __exec, _Range1&& __keys, _Range2&& __values, _Range3&& __out_keys,
-                         _Range4&& __out_values, _BinaryPredicate __binary_pred, _BinaryOperator __binary_op,
-                         ::std::false_type /* has_known_identity */)
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Range1, typename _Range2, typename _Range3,
+          typename _Range4, typename _BinaryPredicate, typename _BinaryOperator>
+oneapi::dpl::__internal::__difference_t<_Range3>
+__sycl_reduce_by_segment(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Range1&& __keys, _Range2&& __values,
+                         _Range3&& __out_keys, _Range4&& __out_values, _BinaryPredicate __binary_pred,
+                         _BinaryOperator __binary_op, ::std::false_type /* has_known_identity */)
 {
     return oneapi::dpl::experimental::ranges::reduce_by_segment(
         ::std::forward<_ExecutionPolicy>(__exec), ::std::forward<_Range1>(__keys), ::std::forward<_Range2>(__values),
         ::std::forward<_Range3>(__out_keys), ::std::forward<_Range4>(__out_values), __binary_pred, __binary_op);
 }
 
-template <typename _ExecutionPolicy, typename _Range1, typename _Range2, typename _Range3, typename _Range4,
-          typename _BinaryPredicate, typename _BinaryOperator>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy,
-                                                             oneapi::dpl::__internal::__difference_t<_Range3>>
-__sycl_reduce_by_segment(_ExecutionPolicy&& __exec, _Range1&& __keys, _Range2&& __values, _Range3&& __out_keys,
-                         _Range4&& __out_values, _BinaryPredicate __binary_pred, _BinaryOperator __binary_op,
-                         ::std::true_type /* has_known_identity */)
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Range1, typename _Range2, typename _Range3,
+          typename _Range4, typename _BinaryPredicate, typename _BinaryOperator>
+oneapi::dpl::__internal::__difference_t<_Range3>
+__sycl_reduce_by_segment(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Range1&& __keys, _Range2&& __values,
+                         _Range3&& __out_keys, _Range4&& __out_values, _BinaryPredicate __binary_pred,
+                         _BinaryOperator __binary_op, ::std::true_type /* has_known_identity */)
 {
     using _CustomName = oneapi::dpl::__internal::__policy_kernel_name<_ExecutionPolicy>;
 

From 6697736479e95532b4696d7e8079a7f484fc3fef Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 14:07:47 +0100
Subject: [PATCH 276/566] internal::__sycl_reduce_by_segment + tag calls

---
 include/oneapi/dpl/internal/reduce_by_segment_impl.h | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/include/oneapi/dpl/internal/reduce_by_segment_impl.h b/include/oneapi/dpl/internal/reduce_by_segment_impl.h
index 35465ffc5d1..860664bde67 100644
--- a/include/oneapi/dpl/internal/reduce_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/reduce_by_segment_impl.h
@@ -573,7 +573,7 @@ __sycl_reduce_by_segment(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _
 template <typename _BackendTag, typename Policy, typename InputIterator1, typename InputIterator2,
           typename OutputIterator1, typename OutputIterator2, typename BinaryPred, typename BinaryOperator>
 ::std::pair<OutputIterator1, OutputIterator2>
-reduce_by_segment_impl(__hetero_tag<_BackendTag>, Policy&& policy, InputIterator1 first1, InputIterator1 last1,
+reduce_by_segment_impl(__hetero_tag<_BackendTag> __tag, Policy&& policy, InputIterator1 first1, InputIterator1 last1,
                        InputIterator2 first2, OutputIterator1 result1, OutputIterator2 result2, BinaryPred binary_pred,
                        BinaryOperator binary_op)
 {
@@ -609,9 +609,9 @@ reduce_by_segment_impl(__hetero_tag<_BackendTag>, Policy&& policy, InputIterator
                                                      typename ::std::iterator_traits<InputIterator2>::value_type>::type;
 
     // number of unique keys
-    _CountType __n = __sycl_reduce_by_segment(::std::forward<Policy>(policy), key_buf.all_view(), value_buf.all_view(),
-                                              key_output_buf.all_view(), value_output_buf.all_view(), binary_pred,
-                                              binary_op, has_known_identity{});
+    _CountType __n = __sycl_reduce_by_segment(
+        __tag, ::std::forward<Policy>(policy), key_buf.all_view(), value_buf.all_view(), key_output_buf.all_view(),
+        value_output_buf.all_view(), binary_pred, binary_op, has_known_identity{});
 
     return ::std::make_pair(result1 + __n, result2 + __n);
 }

From c195783439bb99b94088548f49483173b382b95e Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 14:10:00 +0100
Subject: [PATCH 277/566] internal::__scan_by_segment_impl_common + tag impls

---
 include/oneapi/dpl/internal/scan_by_segment_impl.h | 11 ++++++-----
 1 file changed, 6 insertions(+), 5 deletions(-)

diff --git a/include/oneapi/dpl/internal/scan_by_segment_impl.h b/include/oneapi/dpl/internal/scan_by_segment_impl.h
index 2d2cafaa038..98cc9e1d93a 100644
--- a/include/oneapi/dpl/internal/scan_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/scan_by_segment_impl.h
@@ -364,11 +364,12 @@ struct __sycl_scan_by_segment_impl
     }
 };
 
-template <typename Policy, typename InputIterator1, typename InputIterator2, typename OutputIterator, typename T,
-          typename BinaryPredicate, typename Operator, typename Inclusive>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<Policy, OutputIterator>
-__scan_by_segment_impl_common(Policy&& policy, InputIterator1 first1, InputIterator1 last1, InputIterator2 first2,
-                              OutputIterator result, T init, BinaryPredicate binary_pred, Operator binary_op, Inclusive)
+template <typename _BackendTag, typename Policy, typename InputIterator1, typename InputIterator2,
+          typename OutputIterator, typename T, typename BinaryPredicate, typename Operator, typename Inclusive>
+OutputIterator
+__scan_by_segment_impl_common(__hetero_tag<_BackendTag>, Policy&& policy, InputIterator1 first1, InputIterator1 last1,
+                              InputIterator2 first2, OutputIterator result, T init, BinaryPredicate binary_pred,
+                              Operator binary_op, Inclusive)
 {
     const auto n = ::std::distance(first1, last1);
 

From 41980d370a24c7423e94d6a3754e05a159811c83 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 14:10:07 +0100
Subject: [PATCH 278/566] internal::__scan_by_segment_impl_common + tag calls

---
 .../oneapi/dpl/internal/exclusive_scan_by_segment_impl.h    | 6 +++---
 .../oneapi/dpl/internal/inclusive_scan_by_segment_impl.h    | 4 ++--
 2 files changed, 5 insertions(+), 5 deletions(-)

diff --git a/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h b/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h
index fb1beb9b3eb..07d5b930c2f 100644
--- a/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h
@@ -97,12 +97,12 @@ pattern_exclusive_scan_by_segment(_Tag, Policy&& policy, InputIterator1 first1,
 template <typename _BackendTag, typename Policy, typename InputIterator1, typename InputIterator2,
           typename OutputIterator, typename T, typename BinaryPredicate, typename Operator>
 OutputIterator
-exclusive_scan_by_segment_impl(__hetero_tag<_BackendTag>, Policy&& policy, InputIterator1 first1, InputIterator1 last1,
+exclusive_scan_by_segment_impl(__hetero_tag<_BackendTag> __tag, Policy&& policy, InputIterator1 first1, InputIterator1 last1,
                                InputIterator2 first2, OutputIterator result, T init, BinaryPredicate binary_pred,
                                Operator binary_op, ::std::true_type /* has_known_identity*/)
 {
-    return internal::__scan_by_segment_impl_common(::std::forward<Policy>(policy), first1, last1, first2, result, init,
-                                                   binary_pred, binary_op, ::std::false_type{});
+    return internal::__scan_by_segment_impl_common(__tag, ::std::forward<Policy>(policy), first1, last1, first2, result,
+                                                   init, binary_pred, binary_op, ::std::false_type{});
 }
 
 template <typename _BackendTag, typename Policy, typename InputIterator1, typename InputIterator2,
diff --git a/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h b/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h
index 08d92bed504..43310cd063b 100644
--- a/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h
@@ -78,13 +78,13 @@ pattern_inclusive_scan_by_segment(_Tag, Policy&& policy, InputIterator1 first1,
 template <typename _BackendTag, typename Policy, typename InputIterator1, typename InputIterator2,
           typename OutputIterator, typename BinaryPredicate, typename BinaryOperator>
 OutputIterator
-inclusive_scan_by_segment_impl(__hetero_tag<_BackendTag>, Policy&& policy, InputIterator1 first1, InputIterator1 last1,
+inclusive_scan_by_segment_impl(__hetero_tag<_BackendTag> __tag, Policy&& policy, InputIterator1 first1, InputIterator1 last1,
                                InputIterator2 first2, OutputIterator result, BinaryPredicate binary_pred,
                                BinaryOperator binary_op, ::std::true_type /* has_known_identity */)
 {
     using iter_value_t = typename ::std::iterator_traits<InputIterator2>::value_type;
     iter_value_t identity = unseq_backend::__known_identity<BinaryOperator, iter_value_t>;
-    return internal::__scan_by_segment_impl_common(::std::forward<Policy>(policy), first1, last1, first2, result,
+    return internal::__scan_by_segment_impl_common(__tag, ::std::forward<Policy>(policy), first1, last1, first2, result,
                                                    identity, binary_pred, binary_op, ::std::true_type{});
 }
 

From 1e2b03485c1330fd3abd6c852820cab14175d86b Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 14:18:37 +0100
Subject: [PATCH 279/566] internal::__pattern_transform_scan_base_async + tag
 impls

---
 .../dpl/internal/async_impl/async_impl_hetero.h       | 11 +++++------
 1 file changed, 5 insertions(+), 6 deletions(-)

diff --git a/include/oneapi/dpl/internal/async_impl/async_impl_hetero.h b/include/oneapi/dpl/internal/async_impl/async_impl_hetero.h
index d14f2aed022..6c151386e03 100644
--- a/include/oneapi/dpl/internal/async_impl/async_impl_hetero.h
+++ b/include/oneapi/dpl/internal/async_impl/async_impl_hetero.h
@@ -208,13 +208,12 @@ __pattern_fill_async(_ExecutionPolicy&& __exec, _ForwardIterator __first, _Forwa
 // transform_scan
 //------------------------------------------------------------------------
 
-template <typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2, typename _UnaryOperation,
-          typename _InitType, typename _BinaryOperation, typename _Inclusive,
-          oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, int> = 0>
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2,
+          typename _UnaryOperation, typename _InitType, typename _BinaryOperation, typename _Inclusive>
 auto
-__pattern_transform_scan_base_async(_ExecutionPolicy&& __exec, _Iterator1 __first, _Iterator1 __last,
-                                    _Iterator2 __result, _UnaryOperation __unary_op, _InitType __init,
-                                    _BinaryOperation __binary_op, _Inclusive)
+__pattern_transform_scan_base_async(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Iterator1 __first,
+                                    _Iterator1 __last, _Iterator2 __result, _UnaryOperation __unary_op,
+                                    _InitType __init, _BinaryOperation __binary_op, _Inclusive)
 {
     assert(__first < __last);
 

From 868e686d09a200117fdd652bd6e256e53880d2a6 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 14:18:48 +0100
Subject: [PATCH 280/566] internal::__pattern_transform_scan_base_async + tag
 calls

---
 .../dpl/internal/async_impl/async_impl_hetero.h | 17 ++++++++++++-----
 1 file changed, 12 insertions(+), 5 deletions(-)

diff --git a/include/oneapi/dpl/internal/async_impl/async_impl_hetero.h b/include/oneapi/dpl/internal/async_impl/async_impl_hetero.h
index 6c151386e03..c8b40ce399e 100644
--- a/include/oneapi/dpl/internal/async_impl/async_impl_hetero.h
+++ b/include/oneapi/dpl/internal/async_impl/async_impl_hetero.h
@@ -224,7 +224,7 @@ __pattern_transform_scan_base_async(__hetero_tag<_BackendTag>, _ExecutionPolicy&
     auto __buf2 = __keep2(__result, __result + __n);
 
     auto __res = oneapi::dpl::__par_backend_hetero::__parallel_transform_scan(
-        oneapi::dpl::__internal::__fpga_backend_tag{},
+        _BackendTag{},
         ::std::forward<_ExecutionPolicy>(__exec), __buf1.all_view(), __buf2.all_view(), __n, __unary_op, __init,
         __binary_op, _Inclusive{});
     return __res.__make_future(__result + __n);
@@ -240,8 +240,12 @@ __pattern_transform_scan_async(_ExecutionPolicy&& __exec, _Iterator1 __first, _I
     using _RepackedType = __par_backend_hetero::__repacked_tuple_t<_Type>;
     using _InitType = unseq_backend::__init_value<_RepackedType>;
 
-    return __pattern_transform_scan_base_async(::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result,
-                                               __unary_op, _InitType{__init}, __binary_op, _Inclusive{});
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator1, _Iterator2>();
+
+    return __pattern_transform_scan_base_async(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first,
+                                               __last, __result, __unary_op, _InitType{__init}, __binary_op,
+                                               _Inclusive{});
 }
 
 // scan without initial element
@@ -256,8 +260,11 @@ __pattern_transform_scan_async(_ExecutionPolicy&& __exec, _Iterator1 __first, _I
     using _RepackedType = __par_backend_hetero::__repacked_tuple_t<_ValueType>;
     using _InitType = unseq_backend::__no_init_value<_RepackedType>;
 
-    return __pattern_transform_scan_base_async(::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result,
-                                               __unary_op, _InitType{}, __binary_op, _Inclusive{});
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator1, _Iterator2>();
+
+    return __pattern_transform_scan_base_async(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first,
+                                               __last, __result, __unary_op, _InitType{}, __binary_op, _Inclusive{});
 }
 
 } // namespace __internal

From a728717e678c9dbebcc2a5bf1fe4689c820c52ca Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 14:24:02 +0100
Subject: [PATCH 281/566] internal::__pattern_transform_scan_async + tag impls

---
 .../internal/async_impl/async_impl_hetero.h   | 35 ++++++++-----------
 1 file changed, 14 insertions(+), 21 deletions(-)

diff --git a/include/oneapi/dpl/internal/async_impl/async_impl_hetero.h b/include/oneapi/dpl/internal/async_impl/async_impl_hetero.h
index c8b40ce399e..dc08e7ba2d1 100644
--- a/include/oneapi/dpl/internal/async_impl/async_impl_hetero.h
+++ b/include/oneapi/dpl/internal/async_impl/async_impl_hetero.h
@@ -230,41 +230,34 @@ __pattern_transform_scan_base_async(__hetero_tag<_BackendTag>, _ExecutionPolicy&
     return __res.__make_future(__result + __n);
 }
 
-template <typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2, typename _UnaryOperation, typename _Type,
-          typename _BinaryOperation, typename _Inclusive,
-          oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, int> = 0>
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2,
+          typename _UnaryOperation, typename _Type, typename _BinaryOperation, typename _Inclusive>
 auto
-__pattern_transform_scan_async(_ExecutionPolicy&& __exec, _Iterator1 __first, _Iterator1 __last, _Iterator2 __result,
-                               _UnaryOperation __unary_op, _Type __init, _BinaryOperation __binary_op, _Inclusive)
+__pattern_transform_scan_async(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator1 __first,
+                               _Iterator1 __last, _Iterator2 __result, _UnaryOperation __unary_op, _Type __init,
+                               _BinaryOperation __binary_op, _Inclusive)
 {
     using _RepackedType = __par_backend_hetero::__repacked_tuple_t<_Type>;
     using _InitType = unseq_backend::__init_value<_RepackedType>;
 
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator1, _Iterator2>();
-
-    return __pattern_transform_scan_base_async(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first,
-                                               __last, __result, __unary_op, _InitType{__init}, __binary_op,
-                                               _Inclusive{});
+    return __pattern_transform_scan_base_async(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+                                               __result, __unary_op, _InitType{__init}, __binary_op, _Inclusive{});
 }
 
 // scan without initial element
-template <typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2, typename _UnaryOperation,
-          typename _BinaryOperation, typename _Inclusive,
-          oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, int> = 0>
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2,
+          typename _UnaryOperation, typename _BinaryOperation, typename _Inclusive>
 auto
-__pattern_transform_scan_async(_ExecutionPolicy&& __exec, _Iterator1 __first, _Iterator1 __last, _Iterator2 __result,
-                               _UnaryOperation __unary_op, _BinaryOperation __binary_op, _Inclusive)
+__pattern_transform_scan_async(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator1 __first,
+                               _Iterator1 __last, _Iterator2 __result, _UnaryOperation __unary_op,
+                               _BinaryOperation __binary_op, _Inclusive)
 {
     using _ValueType = typename ::std::iterator_traits<_Iterator1>::value_type;
     using _RepackedType = __par_backend_hetero::__repacked_tuple_t<_ValueType>;
     using _InitType = unseq_backend::__no_init_value<_RepackedType>;
 
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator1, _Iterator2>();
-
-    return __pattern_transform_scan_base_async(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first,
-                                               __last, __result, __unary_op, _InitType{}, __binary_op, _Inclusive{});
+    return __pattern_transform_scan_base_async(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+                                               __result, __unary_op, _InitType{}, __binary_op, _Inclusive{});
 }
 
 } // namespace __internal

From 38e5f66b20d1689bef95e45372c6d88cd5fecc5b Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 14:24:20 +0100
Subject: [PATCH 282/566] internal::__pattern_transform_scan_async + tag calls

---
 .../dpl/internal/async_impl/glue_async_impl.h | 49 ++++++++++++++-----
 1 file changed, 36 insertions(+), 13 deletions(-)

diff --git a/include/oneapi/dpl/internal/async_impl/glue_async_impl.h b/include/oneapi/dpl/internal/async_impl/glue_async_impl.h
index 3cceec2f5cb..d89da19493e 100644
--- a/include/oneapi/dpl/internal/async_impl/glue_async_impl.h
+++ b/include/oneapi/dpl/internal/async_impl/glue_async_impl.h
@@ -234,10 +234,13 @@ auto
 inclusive_scan_async(_ExecutionPolicy&& __exec, _ForwardIt1 __first1, _ForwardIt1 __last1, _ForwardIt2 __first2,
                      _BinaryOperation __binary_op, _Events&&... __dependencies)
 {
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIt1, _ForwardIt2>();
+
     wait_for_all(::std::forward<_Events>(__dependencies)...);
     return oneapi::dpl::__internal::__pattern_transform_scan_async(
-        ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, oneapi::dpl::__internal::__no_op(),
-        __binary_op, /*inclusive=*/::std::true_type());
+        __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2,
+        oneapi::dpl::__internal::__no_op(), __binary_op, /*inclusive=*/::std::true_type());
 }
 
 template <class _ExecutionPolicy, class _ForwardIt1, class _ForwardIt2, class _BinaryOperation, class _T,
@@ -248,10 +251,13 @@ auto
 inclusive_scan_async(_ExecutionPolicy&& __exec, _ForwardIt1 __first1, _ForwardIt1 __last1, _ForwardIt2 __first2,
                      _BinaryOperation __binary_op, _T __init, _Events&&... __dependencies)
 {
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIt1, _ForwardIt2>();
+
     wait_for_all(::std::forward<_Events>(__dependencies)...);
     return oneapi::dpl::__internal::__pattern_transform_scan_async(
-        ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, oneapi::dpl::__internal::__no_op(),
-        __init, __binary_op, /*inclusive=*/::std::true_type());
+        __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2,
+        oneapi::dpl::__internal::__no_op(), __init, __binary_op, /*inclusive=*/::std::true_type());
 }
 
 template <class _ExecutionPolicy, class _ForwardIt1, class _ForwardIt2, class _T, class... _Events,
@@ -260,10 +266,13 @@ auto
 exclusive_scan_async(_ExecutionPolicy&& __exec, _ForwardIt1 __first1, _ForwardIt1 __last1, _ForwardIt2 __first2,
                      _T __init, _Events&&... __dependencies)
 {
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIt1, _ForwardIt2>();
+
     wait_for_all(::std::forward<_Events>(__dependencies)...);
     return oneapi::dpl::__internal::__pattern_transform_scan_async(
-        ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, oneapi::dpl::__internal::__no_op(),
-        __init, ::std::plus<_T>(), /*exclusive=*/::std::false_type());
+        __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2,
+        oneapi::dpl::__internal::__no_op(), __init, ::std::plus<_T>(), /*exclusive=*/::std::false_type());
 }
 
 template <class _ExecutionPolicy, class _ForwardIt1, class _ForwardIt2, class _T, class _BinaryOperation,
@@ -274,10 +283,13 @@ auto
 exclusive_scan_async(_ExecutionPolicy&& __exec, _ForwardIt1 __first1, _ForwardIt1 __last1, _ForwardIt2 __first2,
                      _T __init, _BinaryOperation __binary_op, _Events&&... __dependencies)
 {
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIt1, _ForwardIt2>();
+
     wait_for_all(::std::forward<_Events>(__dependencies)...);
     return oneapi::dpl::__internal::__pattern_transform_scan_async(
-        ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, oneapi::dpl::__internal::__no_op(),
-        __init, __binary_op, /*exclusive=*/::std::false_type());
+        __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2,
+        oneapi::dpl::__internal::__no_op(), __init, __binary_op, /*exclusive=*/::std::false_type());
 }
 
 template <class _ExecutionPolicy, class _ForwardIt1, class _ForwardIt2, class _T, class _BinaryOperation,
@@ -288,8 +300,12 @@ transform_exclusive_scan_async(_ExecutionPolicy&& __exec, _ForwardIt1 __first1,
                                _ForwardIt2 __first2, _T __init, _BinaryOperation __binary_op,
                                _UnaryOperation __unary_op, _Events&&... __dependencies)
 {
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIt1, _ForwardIt2>();
+
     wait_for_all(::std::forward<_Events>(__dependencies)...);
-    return oneapi::dpl::__internal::__pattern_transform_scan_async(::std::forward<_ExecutionPolicy>(__exec), __first1,
+    return oneapi::dpl::__internal::__pattern_transform_scan_async(__dispatch_tag,
+                                                                   ::std::forward<_ExecutionPolicy>(__exec), __first1,
                                                                    __last1, __first2, __unary_op, __init, __binary_op,
                                                                    /*exclusive=*/::std::false_type());
 }
@@ -302,10 +318,13 @@ transform_inclusive_scan_async(_ExecutionPolicy&& __exec, _ForwardIt1 __first1,
                                _ForwardIt2 __first2, _BinaryOperation __binary_op, _UnaryOperation __unary_op,
                                _Events&&... __dependencies)
 {
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIt1, _ForwardIt2>();
+
     wait_for_all(::std::forward<_Events>(__dependencies)...);
-    return oneapi::dpl::__internal::__pattern_transform_scan_async(::std::forward<_ExecutionPolicy>(__exec), __first1,
-                                                                   __last1, __first2, __unary_op, __binary_op,
-                                                                   /*inclusive=*/::std::true_type());
+    return oneapi::dpl::__internal::__pattern_transform_scan_async(
+        __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __unary_op, __binary_op,
+        /*inclusive=*/::std::true_type());
 }
 
 template <class _ExecutionPolicy, class _ForwardIt1, class _ForwardIt2, class _BinaryOperation, class _UnaryOperation,
@@ -317,8 +336,12 @@ transform_inclusive_scan_async(_ExecutionPolicy&& __exec, _ForwardIt1 __first1,
                                _ForwardIt2 __first2, _BinaryOperation __binary_op, _UnaryOperation __unary_op,
                                _T __init, _Events&&... __dependencies)
 {
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIt1, _ForwardIt2>();
+
     wait_for_all(::std::forward<_Events>(__dependencies)...);
-    return oneapi::dpl::__internal::__pattern_transform_scan_async(::std::forward<_ExecutionPolicy>(__exec), __first1,
+    return oneapi::dpl::__internal::__pattern_transform_scan_async(__dispatch_tag,
+                                                                   ::std::forward<_ExecutionPolicy>(__exec), __first1,
                                                                    __last1, __first2, __unary_op, __init, __binary_op,
                                                                    /*inclusive=*/::std::true_type());
 }

From 665980179d1742d06be5095b0bfef9a7e755fc18 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 15 Feb 2024 19:05:50 +0100
Subject: [PATCH 283/566] Remove __parallel_for implementations without tags

---
 .../pstl/hetero/dpcpp/parallel_backend_sycl.h | 12 ------------
 .../hetero/dpcpp/parallel_backend_sycl_fpga.h | 14 +-------------
 include/oneapi/dpl/pstl/omp/parallel_for.h    | 19 -------------------
 .../oneapi/dpl/pstl/parallel_backend_serial.h |  7 -------
 .../oneapi/dpl/pstl/parallel_backend_tbb.h    | 11 -----------
 5 files changed, 1 insertion(+), 62 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
index 7ee5ba6fb9c..af3724db1e5 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
@@ -263,18 +263,6 @@ struct __parallel_for_submitter<__internal::__optional_kernel_name<_Name...>>
 
 //General version of parallel_for, one additional parameter - __count of iterations of loop __cgh.parallel_for,
 //for some algorithms happens that size of processing range is n, but amount of iterations is n/2.
-template <typename _ExecutionPolicy, typename _Fp, typename _Index,
-          oneapi::dpl::__internal::__enable_if_device_execution_policy<_ExecutionPolicy, int> = 0, typename... _Ranges>
-auto
-__parallel_for(_ExecutionPolicy&& __exec, _Fp __brick, _Index __count, _Ranges&&... __rngs)
-{
-    using _CustomName = oneapi::dpl::__internal::__policy_kernel_name<_ExecutionPolicy>;
-    using _ForKernel = oneapi::dpl::__par_backend_hetero::__internal::__kernel_name_provider<_CustomName>;
-
-    return __parallel_for_submitter<_ForKernel>()(::std::forward<_ExecutionPolicy>(__exec), __brick, __count,
-                                                  ::std::forward<_Ranges>(__rngs)...);
-}
-
 template <typename _ExecutionPolicy, typename _Fp, typename _Index, typename... _Ranges>
 auto
 __parallel_for(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&& __exec, _Fp __brick, _Index __count,
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
index 224e6b6f9fe..769e5abe4af 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
@@ -76,7 +76,7 @@ struct __parallel_for_fpga_submitter<__internal::__optional_kernel_name<_Name...
         return __future(__event);
     }
 
-// KSATODO is this define check really required here?
+// TODO is this define check really required here?
 #if _ONEDPL_FPGA_DEVICE
     template <typename _ExecutionPolicy, typename _Fp, typename _Index, typename... _Ranges>
     auto
@@ -104,18 +104,6 @@ struct __parallel_for_fpga_submitter<__internal::__optional_kernel_name<_Name...
 #endif // _ONEDPL_FPGA_DEVICE
 };
 
-template <typename _ExecutionPolicy, typename _Fp, typename _Index, typename... _Ranges,
-          oneapi::dpl::__internal::__enable_if_fpga_execution_policy<_ExecutionPolicy, int> = 0>
-auto
-__parallel_for(_ExecutionPolicy&& __exec, _Fp __brick, _Index __count, _Ranges&&... __rngs)
-{
-    using _CustomName = oneapi::dpl::__internal::__policy_kernel_name<_ExecutionPolicy>;
-    using __parallel_for_name = __internal::__kernel_name_provider<_CustomName>;
-
-    return __parallel_for_fpga_submitter<__parallel_for_name>()(std::forward<_ExecutionPolicy>(__exec), __brick,
-                                                                __count, std::forward<_Ranges>(__rngs)...);
-}
-
 template <typename _ExecutionPolicy, typename _Fp, typename _Index, typename... _Ranges>
 auto
 __parallel_for(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionPolicy&& __exec, _Fp __brick, _Index __count,
diff --git a/include/oneapi/dpl/pstl/omp/parallel_for.h b/include/oneapi/dpl/pstl/omp/parallel_for.h
index 6f870dc4b4a..1a0ea24d798 100644
--- a/include/oneapi/dpl/pstl/omp/parallel_for.h
+++ b/include/oneapi/dpl/pstl/omp/parallel_for.h
@@ -47,25 +47,6 @@ __parallel_for_body(_Index __first, _Index __last, _Fp __f)
 // Evaluation of brick f[i,j) for each subrange [i,j) of [first, last)
 //------------------------------------------------------------------------
 
-template <class _ExecutionPolicy, class _Index, class _Fp>
-void
-__parallel_for(_ExecutionPolicy&&, _Index __first, _Index __last, _Fp __f)
-{
-    if (omp_in_parallel())
-    {
-        // we don't create a nested parallel region in an existing parallel
-        // region: just create tasks
-        oneapi::dpl::__omp_backend::__parallel_for_body(__first, __last, __f);
-    }
-    else
-    {
-        // in any case (nested or non-nested) one parallel region is created and
-        // only one thread creates a set of tasks
-        _PSTL_PRAGMA(omp parallel)
-        _PSTL_PRAGMA(omp single nowait) { oneapi::dpl::__omp_backend::__parallel_for_body(__first, __last, __f); }
-    }
-}
-
 template <class _ExecutionPolicy, class _Index, class _Fp>
 void
 __parallel_for(oneapi::dpl::__internal::__omp_backend_tag, _ExecutionPolicy&&, _Index __first, _Index __last, _Fp __f)
diff --git a/include/oneapi/dpl/pstl/parallel_backend_serial.h b/include/oneapi/dpl/pstl/parallel_backend_serial.h
index 264947fe550..9d9f3ea2314 100644
--- a/include/oneapi/dpl/pstl/parallel_backend_serial.h
+++ b/include/oneapi/dpl/pstl/parallel_backend_serial.h
@@ -64,13 +64,6 @@ __cancel_execution()
 {
 }
 
-template <class _ExecutionPolicy, class _Index, class _Fp>
-void
-__parallel_for(_ExecutionPolicy&&, _Index __first, _Index __last, _Fp __f)
-{
-    __f(__first, __last);
-}
-
 template <class _ExecutionPolicy, class _Index, class _Fp>
 void
 __parallel_for(oneapi::dpl::__internal::__serial_backend_tag, _ExecutionPolicy&&, _Index __first, _Index __last,
diff --git a/include/oneapi/dpl/pstl/parallel_backend_tbb.h b/include/oneapi/dpl/pstl/parallel_backend_tbb.h
index 7639551a1c6..74c1eade762 100644
--- a/include/oneapi/dpl/pstl/parallel_backend_tbb.h
+++ b/include/oneapi/dpl/pstl/parallel_backend_tbb.h
@@ -115,17 +115,6 @@ class __parallel_for_body
     _RealBody _M_body;
 };
 
-//! Evaluation of brick f[i,j) for each subrange [i,j) of [first,last)
-// wrapper over tbb::parallel_for
-template <class _ExecutionPolicy, class _Index, class _Fp>
-void
-__parallel_for(_ExecutionPolicy&&, _Index __first, _Index __last, _Fp __f)
-{
-    tbb::this_task_arena::isolate([=]() {
-        tbb::parallel_for(tbb::blocked_range<_Index>(__first, __last), __parallel_for_body<_Index, _Fp>(__f));
-    });
-}
-
 //! Evaluation of brick f[i,j) for each subrange [i,j) of [first,last)
 // wrapper over tbb::parallel_for
 template <class _ExecutionPolicy, class _Index, class _Fp>

From 3c493ea4a56e6a35470277dfd63a26e240e25658 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 09:55:26 +0100
Subject: [PATCH 284/566] Remove __parallel_transform_scan implementations
 without tags

---
 include/oneapi/dpl/pstl/omp/parallel_transform_scan.h |  9 ---------
 include/oneapi/dpl/pstl/parallel_backend_serial.h     |  7 -------
 include/oneapi/dpl/pstl/parallel_backend_tbb.h        | 11 -----------
 3 files changed, 27 deletions(-)

diff --git a/include/oneapi/dpl/pstl/omp/parallel_transform_scan.h b/include/oneapi/dpl/pstl/omp/parallel_transform_scan.h
index f0093ccab98..35c28b4330c 100644
--- a/include/oneapi/dpl/pstl/omp/parallel_transform_scan.h
+++ b/include/oneapi/dpl/pstl/omp/parallel_transform_scan.h
@@ -25,15 +25,6 @@ namespace dpl
 namespace __omp_backend
 {
 
-template <class _ExecutionPolicy, class _Index, class _Up, class _Tp, class _Cp, class _Rp, class _Sp>
-_Tp
-__parallel_transform_scan(_ExecutionPolicy&&, _Index __n, _Up /* __u */, _Tp __init, _Cp /* __combine */,
-                          _Rp /* __brick_reduce */, _Sp __scan)
-{
-    // TODO: parallelize this function.
-    return __scan(_Index(0), __n, __init);
-}
-
 template <class _ExecutionPolicy, class _Index, class _Up, class _Tp, class _Cp, class _Rp, class _Sp>
 _Tp
 __parallel_transform_scan(oneapi::dpl::__internal::__omp_backend_tag, _ExecutionPolicy&&, _Index __n, _Up /* __u */,
diff --git a/include/oneapi/dpl/pstl/parallel_backend_serial.h b/include/oneapi/dpl/pstl/parallel_backend_serial.h
index 9d9f3ea2314..a73a973de32 100644
--- a/include/oneapi/dpl/pstl/parallel_backend_serial.h
+++ b/include/oneapi/dpl/pstl/parallel_backend_serial.h
@@ -144,13 +144,6 @@ __parallel_strict_scan(oneapi::dpl::__internal::__serial_backend_tag, _Execution
         __scan(_Index(0), __n, __initial);
 }
 
-template <class _ExecutionPolicy, class _Index, class _UnaryOp, class _Tp, class _BinaryOp, class _Reduce, class _Scan>
-_Tp
-__parallel_transform_scan(_ExecutionPolicy&&, _Index __n, _UnaryOp, _Tp __init, _BinaryOp, _Reduce, _Scan __scan)
-{
-    return __scan(_Index(0), __n, __init);
-}
-
 template <class _ExecutionPolicy, class _Index, class _UnaryOp, class _Tp, class _BinaryOp, class _Reduce, class _Scan>
 _Tp
 __parallel_transform_scan(oneapi::dpl::__internal::__serial_backend_tag, _ExecutionPolicy&&, _Index __n, _UnaryOp,
diff --git a/include/oneapi/dpl/pstl/parallel_backend_tbb.h b/include/oneapi/dpl/pstl/parallel_backend_tbb.h
index 74c1eade762..2f48fe9cc98 100644
--- a/include/oneapi/dpl/pstl/parallel_backend_tbb.h
+++ b/include/oneapi/dpl/pstl/parallel_backend_tbb.h
@@ -485,17 +485,6 @@ __parallel_strict_scan(oneapi::dpl::__internal::__tbb_backend_tag, _ExecutionPol
     });
 }
 
-template <class _ExecutionPolicy, class _Index, class _Up, class _Tp, class _Cp, class _Rp, class _Sp>
-_Tp
-__parallel_transform_scan(_ExecutionPolicy&&, _Index __n, _Up __u, _Tp __init, _Cp __combine, _Rp __brick_reduce,
-                          _Sp __scan)
-{
-    __trans_scan_body<_Index, _Up, _Tp, _Cp, _Rp, _Sp> __body(__u, __init, __combine, __brick_reduce, __scan);
-    auto __range = tbb::blocked_range<_Index>(0, __n);
-    tbb::this_task_arena::isolate([__range, &__body]() { tbb::parallel_scan(__range, __body); });
-    return __body.sum();
-}
-
 template <class _ExecutionPolicy, class _Index, class _Up, class _Tp, class _Cp, class _Rp, class _Sp>
 _Tp
 __parallel_transform_scan(oneapi::dpl::__internal::__tbb_backend_tag, _ExecutionPolicy&&, _Index __n, _Up __u,

From f8321f6c4f93750c67ae1e6f7183a74e4cc19841 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 15 Feb 2024 19:10:12 +0100
Subject: [PATCH 285/566] Remove extra changes from
 __parallel_for_fpga_submitter

---
 .../hetero/dpcpp/parallel_backend_sycl_fpga.h | 30 +------------------
 1 file changed, 1 insertion(+), 29 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
index 769e5abe4af..8d563fd1cd1 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
@@ -75,33 +75,6 @@ struct __parallel_for_fpga_submitter<__internal::__optional_kernel_name<_Name...
         });
         return __future(__event);
     }
-
-// TODO is this define check really required here?
-#if _ONEDPL_FPGA_DEVICE
-    template <typename _ExecutionPolicy, typename _Fp, typename _Index, typename... _Ranges>
-    auto
-    operator()(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionPolicy&& __exec, _Fp __brick, _Index __count,
-               _Ranges&&... __rngs) const
-    {
-        auto __n = oneapi::dpl::__ranges::__get_first_range_size(__rngs...);
-        assert(__n > 0);
-
-        _PRINT_INFO_IN_DEBUG_MODE(__exec);
-        auto __event = __exec.queue().submit([&__rngs..., &__brick, __count](sycl::handler& __cgh) {
-            //get an access to data under SYCL buffer:
-            oneapi::dpl::__ranges::__require_access(__cgh, __rngs...);
-
-            __cgh.single_task<_Name...>([=]() {
-#    pragma unroll(::std::decay <_ExecutionPolicy>::type::unroll_factor)
-                for (auto __idx = 0; __idx < __count; ++__idx)
-                {
-                    __brick(__idx, __rngs...);
-                }
-            });
-        });
-        return __future(__event);
-    }
-#endif // _ONEDPL_FPGA_DEVICE
 };
 
 template <typename _ExecutionPolicy, typename _Fp, typename _Index, typename... _Ranges>
@@ -112,8 +85,7 @@ __parallel_for(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionPolicy&& _
     using _Policy = ::std::decay_t<_ExecutionPolicy>;
     using __parallel_for_name = __internal::__kernel_name_provider<typename _Policy::kernel_name>;
 
-    return __parallel_for_fpga_submitter<__parallel_for_name>()(oneapi::dpl::__internal::__fpga_backend_tag{},
-                                                                std::forward<_ExecutionPolicy>(__exec), __brick,
+    return __parallel_for_fpga_submitter<__parallel_for_name>()(std::forward<_ExecutionPolicy>(__exec), __brick,
                                                                 __count, std::forward<_Ranges>(__rngs)...);
 }
 

From f2c5a7b2856ba79894d23ac3164a43af28a55761 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 10:39:54 +0100
Subject: [PATCH 286/566] Remove __parallel_stable_sort implementations without
 tags

---
 .../dpl/pstl/omp/parallel_stable_sort.h       | 41 -------------------
 .../oneapi/dpl/pstl/parallel_backend_serial.h |  8 ----
 .../oneapi/dpl/pstl/parallel_backend_tbb.h    | 25 -----------
 3 files changed, 74 deletions(-)

diff --git a/include/oneapi/dpl/pstl/omp/parallel_stable_sort.h b/include/oneapi/dpl/pstl/omp/parallel_stable_sort.h
index 2c5c760f584..4633a3fcade 100644
--- a/include/oneapi/dpl/pstl/omp/parallel_stable_sort.h
+++ b/include/oneapi/dpl/pstl/omp/parallel_stable_sort.h
@@ -121,47 +121,6 @@ __parallel_stable_sort_body(_RandomAccessIterator __xs, _RandomAccessIterator __
     }
 }
 
-template <class _ExecutionPolicy, typename _RandomAccessIterator, typename _Compare, typename _LeafSort>
-void
-__parallel_stable_sort(_ExecutionPolicy&& /*__exec*/, _RandomAccessIterator __xs, _RandomAccessIterator __xe,
-                       _Compare __comp, _LeafSort __leaf_sort, std::size_t __nsort = 0)
-{
-    auto __count = static_cast<std::size_t>(__xe - __xs);
-    if (__count <= __default_chunk_size || __nsort < __count)
-    {
-        __leaf_sort(__xs, __xe, __comp);
-        return;
-    }
-
-    // TODO: the partial sort implementation should
-    // be shared with the other backends.
-
-    if (omp_in_parallel())
-    {
-        if (__count <= __nsort)
-        {
-            oneapi::dpl::__omp_backend::__parallel_stable_sort_body(__xs, __xe, __comp, __leaf_sort);
-        }
-        else
-        {
-            oneapi::dpl::__omp_backend::__parallel_stable_partial_sort(__xs, __xe, __comp, __leaf_sort, __nsort);
-        }
-    }
-    else
-    {
-        _PSTL_PRAGMA(omp parallel)
-        _PSTL_PRAGMA(omp single nowait)
-        if (__count <= __nsort)
-        {
-            oneapi::dpl::__omp_backend::__parallel_stable_sort_body(__xs, __xe, __comp, __leaf_sort);
-        }
-        else
-        {
-            oneapi::dpl::__omp_backend::__parallel_stable_partial_sort(__xs, __xe, __comp, __leaf_sort, __nsort);
-        }
-    }
-}
-
 template <class _ExecutionPolicy, typename _RandomAccessIterator, typename _Compare, typename _LeafSort>
 void
 __parallel_stable_sort(oneapi::dpl::__internal::__omp_backend_tag, _ExecutionPolicy&& /*__exec*/,
diff --git a/include/oneapi/dpl/pstl/parallel_backend_serial.h b/include/oneapi/dpl/pstl/parallel_backend_serial.h
index a73a973de32..cfb799fc73a 100644
--- a/include/oneapi/dpl/pstl/parallel_backend_serial.h
+++ b/include/oneapi/dpl/pstl/parallel_backend_serial.h
@@ -152,14 +152,6 @@ __parallel_transform_scan(oneapi::dpl::__internal::__serial_backend_tag, _Execut
     return __scan(_Index(0), __n, __init);
 }
 
-template <class _ExecutionPolicy, typename _RandomAccessIterator, typename _Compare, typename _LeafSort>
-void
-__parallel_stable_sort(_ExecutionPolicy&&, _RandomAccessIterator __first, _RandomAccessIterator __last, _Compare __comp,
-                       _LeafSort __leaf_sort, ::std::size_t = 0)
-{
-    __leaf_sort(__first, __last, __comp);
-}
-
 template <class _ExecutionPolicy, typename _RandomAccessIterator, typename _Compare, typename _LeafSort>
 void
 __parallel_stable_sort(oneapi::dpl::__internal::__serial_backend_tag, _ExecutionPolicy&&, _RandomAccessIterator __first,
diff --git a/include/oneapi/dpl/pstl/parallel_backend_tbb.h b/include/oneapi/dpl/pstl/parallel_backend_tbb.h
index 2f48fe9cc98..1a4bd50a68e 100644
--- a/include/oneapi/dpl/pstl/parallel_backend_tbb.h
+++ b/include/oneapi/dpl/pstl/parallel_backend_tbb.h
@@ -1248,31 +1248,6 @@ __stable_sort_func<_RandomAccessIterator1, _RandomAccessIterator2, _Compare, _Le
     return __self;
 }
 
-template <class _ExecutionPolicy, typename _RandomAccessIterator, typename _Compare, typename _LeafSort>
-void
-__parallel_stable_sort(_ExecutionPolicy&&, _RandomAccessIterator __xs, _RandomAccessIterator __xe, _Compare __comp,
-                       _LeafSort __leaf_sort, ::std::size_t __nsort)
-{
-    tbb::this_task_arena::isolate([=, &__nsort]() {
-        //sorting based on task tree and parallel merge
-        typedef typename ::std::iterator_traits<_RandomAccessIterator>::value_type _ValueType;
-        typedef typename ::std::iterator_traits<_RandomAccessIterator>::difference_type _DifferenceType;
-        const _DifferenceType __n = __xe - __xs;
-
-        const _DifferenceType __sort_cut_off = _ONEDPL_STABLE_SORT_CUT_OFF;
-        if (__n > __sort_cut_off)
-        {
-            __tbb_backend::__buffer<_ExecutionPolicy, _ValueType> __buf(__n);
-            __root_task<__stable_sort_func<_RandomAccessIterator, _ValueType*, _Compare, _LeafSort>> __root{
-                __xs, __xe, __buf.get(), true, __comp, __leaf_sort, __nsort, __xs, __buf.get()};
-            __task::spawn_root_and_wait(__root);
-            return;
-        }
-        //serial sort
-        __leaf_sort(__xs, __xe, __comp);
-    });
-}
-
 template <class _ExecutionPolicy, typename _RandomAccessIterator, typename _Compare, typename _LeafSort>
 void
 __parallel_stable_sort(oneapi::dpl::__internal::__tbb_backend_tag, _ExecutionPolicy&&, _RandomAccessIterator __xs,

From a97ec6f14c84fecd8521e9e8b529903f2259bc56 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 11:14:56 +0100
Subject: [PATCH 287/566] Remove __parallel_or implementations without tags

---
 include/oneapi/dpl/pstl/parallel_impl.h | 19 -------------------
 1 file changed, 19 deletions(-)

diff --git a/include/oneapi/dpl/pstl/parallel_impl.h b/include/oneapi/dpl/pstl/parallel_impl.h
index 2d3f327bf53..846fc993e8c 100644
--- a/include/oneapi/dpl/pstl/parallel_impl.h
+++ b/include/oneapi/dpl/pstl/parallel_impl.h
@@ -110,25 +110,6 @@ __parallel_find(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Index __f
 // parallel_or
 //------------------------------------------------------------------------
 //! Return true if brick f[i,j) returns true for some subrange [i,j) of [first,last)
-template <class _ExecutionPolicy, class _Index, class _Brick>
-bool
-__parallel_or(_ExecutionPolicy&& __exec, _Index __first, _Index __last, _Brick __f)
-{
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Index>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
-    ::std::atomic<bool> __found(false);
-    __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
-                                  [__f, &__found](_Index __i, _Index __j) {
-                                      if (!__found.load(::std::memory_order_relaxed) && __f(__i, __j))
-                                      {
-                                          __found.store(true, ::std::memory_order_relaxed);
-                                          __par_backend::__cancel_execution();
-                                      }
-                                  });
-    return __found;
-}
-
 template <class _IsVector, class _ExecutionPolicy, class _Index, class _Brick>
 bool
 __parallel_or(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _Index __first, _Index __last, _Brick __f)

From 25efb732527fabe2d2a70b3c5a491eac30ea507e Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 11:23:36 +0100
Subject: [PATCH 288/566] Remove __parallel_find implementations without tags

---
 include/oneapi/dpl/pstl/parallel_impl.h | 37 -------------------------
 1 file changed, 37 deletions(-)

diff --git a/include/oneapi/dpl/pstl/parallel_impl.h b/include/oneapi/dpl/pstl/parallel_impl.h
index 846fc993e8c..fc5507f50d1 100644
--- a/include/oneapi/dpl/pstl/parallel_impl.h
+++ b/include/oneapi/dpl/pstl/parallel_impl.h
@@ -32,43 +32,6 @@ namespace __internal
 //-----------------------------------------------------------------------
 /** Return extremum value returned by brick f[i,j) for subranges [i,j) of [first,last)
 Each f[i,j) must return a value in [i,j). */
-template <class _ExecutionPolicy, class _Index, class _Brick, class _IsFirst>
-_Index
-__parallel_find(_ExecutionPolicy&& __exec, _Index __first, _Index __last, _Brick __f, _IsFirst)
-{
-    typedef typename ::std::iterator_traits<_Index>::difference_type _DifferenceType;
-    const _DifferenceType __n = __last - __first;
-    _DifferenceType __initial_dist = _IsFirst::value ? __n : -1;
-
-    constexpr auto __comp = ::std::conditional_t<_IsFirst::value, __pstl_less, __pstl_greater>{};
-
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Index>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
-    ::std::atomic<_DifferenceType> __extremum(__initial_dist);
-    // TODO: find out what is better here: parallel_for or parallel_reduce
-    __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
-                                  [__comp, __f, __first, &__extremum](_Index __i, _Index __j) {
-                                      // See "Reducing Contention Through Priority Updates", PPoPP '13, for discussion of
-                                      // why using a shared variable scales fairly well in this situation.
-                                      if (__comp(__i - __first, __extremum))
-                                      {
-                                          _Index __res = __f(__i, __j);
-                                          // If not '__last' returned then we found what we want so put this to extremum
-                                          if (__res != __j)
-                                          {
-                                              const _DifferenceType __k = __res - __first;
-                                              for (_DifferenceType __old = __extremum; __comp(__k, __old);
-                                                   __old = __extremum)
-                                              {
-                                                  __extremum.compare_exchange_weak(__old, __k);
-                                              }
-                                          }
-                                      }
-                                  });
-    return __extremum != __initial_dist ? __first + __extremum : __last;
-}
-
 template <class _IsVector, class _ExecutionPolicy, class _Index, class _Brick, class _IsFirst>
 _Index
 __parallel_find(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Index __first, _Index __last, _Brick __f,

From a081f6e44b451e01c9cc57e7644f6a40efd22e68 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 11:27:28 +0100
Subject: [PATCH 289/566] Remove __parallel_merge implementations without tags

---
 .../pstl/hetero/dpcpp/parallel_backend_sycl.h | 27 ----------------
 include/oneapi/dpl/pstl/omp/parallel_merge.h  | 31 -------------------
 .../oneapi/dpl/pstl/parallel_backend_serial.h | 10 ------
 .../oneapi/dpl/pstl/parallel_backend_tbb.h    | 29 -----------------
 4 files changed, 97 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
index af3724db1e5..ada6a261fd3 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
@@ -1841,33 +1841,6 @@ struct __parallel_merge_submitter<_IdType, __internal::__optional_kernel_name<_N
 template <typename... _Name>
 class __merge_kernel_name;
 
-template <typename _ExecutionPolicy, typename _Range1, typename _Range2, typename _Range3, typename _Compare,
-          oneapi::dpl::__internal::__enable_if_device_execution_policy<_ExecutionPolicy, int> = 0>
-auto
-__parallel_merge(_ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&& __rng2, _Range3&& __rng3, _Compare __comp)
-{
-    using _CustomName = oneapi::dpl::__internal::__policy_kernel_name<_ExecutionPolicy>;
-
-    const auto __n = __rng1.size() + __rng2.size();
-    if (__n <= std::numeric_limits<::std::uint32_t>::max())
-    {
-        using _wi_index_type = ::std::uint32_t;
-        using _MergeKernel = oneapi::dpl::__par_backend_hetero::__internal::__kernel_name_provider<__merge_kernel_name<_CustomName, _wi_index_type>>;
-        return __parallel_merge_submitter<_wi_index_type, _MergeKernel>()(::std::forward<_ExecutionPolicy>(__exec),
-                                                      ::std::forward<_Range1>(__rng1), ::std::forward<_Range2>(__rng2),
-                                                      ::std::forward<_Range3>(__rng3), __comp);
-
-    }
-    else
-    {
-        using _wi_index_type = ::std::uint64_t;
-        using _MergeKernel = oneapi::dpl::__par_backend_hetero::__internal::__kernel_name_provider<__merge_kernel_name<_CustomName, _wi_index_type>>;
-        return __parallel_merge_submitter<_wi_index_type, _MergeKernel>()(::std::forward<_ExecutionPolicy>(__exec),
-                                                      ::std::forward<_Range1>(__rng1), ::std::forward<_Range2>(__rng2),
-                                                      ::std::forward<_Range3>(__rng3), __comp);
-    }
-}
-
 template <typename _ExecutionPolicy, typename _Range1, typename _Range2, typename _Range3, typename _Compare>
 auto
 __parallel_merge(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&& __exec, _Range1&& __rng1,
diff --git a/include/oneapi/dpl/pstl/omp/parallel_merge.h b/include/oneapi/dpl/pstl/omp/parallel_merge.h
index 39e8606d4ea..162ef097801 100644
--- a/include/oneapi/dpl/pstl/omp/parallel_merge.h
+++ b/include/oneapi/dpl/pstl/omp/parallel_merge.h
@@ -68,37 +68,6 @@ __parallel_merge_body(std::size_t __size_x, std::size_t __size_y, _RandomAccessI
     _PSTL_PRAGMA(omp taskwait)
 }
 
-template <class _ExecutionPolicy, typename _RandomAccessIterator1, typename _RandomAccessIterator2,
-          typename _RandomAccessIterator3, typename _Compare, typename _LeafMerge>
-void
-__parallel_merge(_ExecutionPolicy&& /*__exec*/, _RandomAccessIterator1 __xs, _RandomAccessIterator1 __xe,
-                 _RandomAccessIterator2 __ys, _RandomAccessIterator2 __ye, _RandomAccessIterator3 __zs, _Compare __comp,
-                 _LeafMerge __leaf_merge)
-{
-    std::size_t __size_x = __xe - __xs;
-    std::size_t __size_y = __ye - __ys;
-
-    /*
-     * Run the merge in parallel by chunking it up. Use the smaller range (if any) as the iteration range, and the
-     * larger range as the search range.
-     */
-
-    if (omp_in_parallel())
-    {
-        oneapi::dpl::__omp_backend::__parallel_merge_body(__size_x, __size_y, __xs, __xe, __ys, __ye, __zs, __comp,
-                                                          __leaf_merge);
-    }
-    else
-    {
-        _PSTL_PRAGMA(omp parallel)
-        {
-            _PSTL_PRAGMA(omp single nowait)
-            oneapi::dpl::__omp_backend::__parallel_merge_body(__size_x, __size_y, __xs, __xe, __ys, __ye, __zs, __comp,
-                                                              __leaf_merge);
-        }
-    }
-}
-
 template <class _ExecutionPolicy, typename _RandomAccessIterator1, typename _RandomAccessIterator2,
           typename _RandomAccessIterator3, typename _Compare, typename _LeafMerge>
 void
diff --git a/include/oneapi/dpl/pstl/parallel_backend_serial.h b/include/oneapi/dpl/pstl/parallel_backend_serial.h
index cfb799fc73a..ff8f5e162e2 100644
--- a/include/oneapi/dpl/pstl/parallel_backend_serial.h
+++ b/include/oneapi/dpl/pstl/parallel_backend_serial.h
@@ -160,16 +160,6 @@ __parallel_stable_sort(oneapi::dpl::__internal::__serial_backend_tag, _Execution
     __leaf_sort(__first, __last, __comp);
 }
 
-template <class _ExecutionPolicy, typename _RandomAccessIterator1, typename _RandomAccessIterator2,
-          typename _RandomAccessIterator3, typename _Compare, typename _LeafMerge>
-void
-__parallel_merge(_ExecutionPolicy&&, _RandomAccessIterator1 __first1, _RandomAccessIterator1 __last1,
-                 _RandomAccessIterator2 __first2, _RandomAccessIterator2 __last2, _RandomAccessIterator3 __outit,
-                 _Compare __comp, _LeafMerge __leaf_merge)
-{
-    __leaf_merge(__first1, __last1, __first2, __last2, __outit, __comp);
-}
-
 template <class _ExecutionPolicy, typename _RandomAccessIterator1, typename _RandomAccessIterator2,
           typename _RandomAccessIterator3, typename _Compare, typename _LeafMerge>
 void
diff --git a/include/oneapi/dpl/pstl/parallel_backend_tbb.h b/include/oneapi/dpl/pstl/parallel_backend_tbb.h
index 1a4bd50a68e..47c43867d54 100644
--- a/include/oneapi/dpl/pstl/parallel_backend_tbb.h
+++ b/include/oneapi/dpl/pstl/parallel_backend_tbb.h
@@ -1339,35 +1339,6 @@ operator()(__task* __self)
     return __self;
 }
 
-template <class _ExecutionPolicy, typename _RandomAccessIterator1, typename _RandomAccessIterator2,
-          typename _RandomAccessIterator3, typename _Compare, typename _LeafMerge>
-void
-__parallel_merge(_ExecutionPolicy&&, _RandomAccessIterator1 __xs, _RandomAccessIterator1 __xe,
-                 _RandomAccessIterator2 __ys, _RandomAccessIterator2 __ye, _RandomAccessIterator3 __zs, _Compare __comp,
-                 _LeafMerge __leaf_merge)
-{
-    typedef typename ::std::iterator_traits<_RandomAccessIterator1>::difference_type _DifferenceType1;
-    typedef typename ::std::iterator_traits<_RandomAccessIterator2>::difference_type _DifferenceType2;
-    typedef typename ::std::common_type_t<_DifferenceType1, _DifferenceType2> _SizeType;
-    const _SizeType __n = (__xe - __xs) + (__ye - __ys);
-    const _SizeType __merge_cut_off = _ONEDPL_MERGE_CUT_OFF;
-    if (__n <= __merge_cut_off)
-    {
-        // Fall back on serial merge
-        __leaf_merge(__xs, __xe, __ys, __ye, __zs, __comp);
-    }
-    else
-    {
-        tbb::this_task_arena::isolate([=]() {
-            typedef __merge_func_static<_RandomAccessIterator1, _RandomAccessIterator2, _RandomAccessIterator3,
-                                        _Compare, _LeafMerge>
-                _TaskType;
-            __root_task<_TaskType> __root{__xs, __xe, __ys, __ye, __zs, __comp, __leaf_merge};
-            __task::spawn_root_and_wait(__root);
-        });
-    }
-}
-
 template <class _ExecutionPolicy, typename _RandomAccessIterator1, typename _RandomAccessIterator2,
           typename _RandomAccessIterator3, typename _Compare, typename _LeafMerge>
 void

From 57119b1fe7f2d2bf0964c6a6e1fea720bfb0eaef Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 11:34:53 +0100
Subject: [PATCH 290/566] Remove __parallel_partial_sort implementations
 without tags

---
 .../pstl/hetero/dpcpp/parallel_backend_sycl.h | 19 -------------------
 1 file changed, 19 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
index ada6a261fd3..ebbe14843c8 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
@@ -2194,25 +2194,6 @@ __parallel_stable_sort(oneapi::dpl::__internal::__device_backend_tag, _Execution
 // parallel_partial_sort - async pattern
 //-----------------------------------------------------------------------
 
-// TODO: check if it makes sense to move these wrappers out of backend to a common place
-// TODO: consider changing __partial_merge_kernel to make it compatible with
-//       __full_merge_kernel in order to use __parallel_sort_impl routine
-template <typename _ExecutionPolicy, typename _Iterator, typename _Compare,
-          oneapi::dpl::__internal::__enable_if_device_execution_policy<_ExecutionPolicy, int> = 0>
-auto
-__parallel_partial_sort(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __mid, _Iterator __last,
-                        _Compare __comp)
-{
-    const auto __mid_idx = __mid - __first;
-
-    auto __keep = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read_write, _Iterator>();
-    auto __buf = __keep(__first, __last);
-
-    return __parallel_partial_sort_impl(oneapi::dpl::__internal::__device_backend_tag{},
-                                        ::std::forward<_ExecutionPolicy>(__exec), __buf.all_view(),
-                                        __partial_merge_kernel<decltype(__mid_idx)>{__mid_idx}, __comp);
-}
-
 // TODO: check if it makes sense to move these wrappers out of backend to a common place
 // TODO: consider changing __partial_merge_kernel to make it compatible with
 //       __full_merge_kernel in order to use __parallel_sort_impl routine

From 3c37fdd31e9b871698b6c97646a533245b4404fa Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 12:40:45 +0100
Subject: [PATCH 291/566] Remove __parallel_for_each implementations without
 tags

---
 .../oneapi/dpl/pstl/omp/parallel_for_each.h   | 19 -------------------
 .../oneapi/dpl/pstl/parallel_backend_serial.h |  8 --------
 .../oneapi/dpl/pstl/parallel_backend_tbb.h    |  6 ------
 3 files changed, 33 deletions(-)

diff --git a/include/oneapi/dpl/pstl/omp/parallel_for_each.h b/include/oneapi/dpl/pstl/omp/parallel_for_each.h
index 9cc4f249f80..32410cbe927 100644
--- a/include/oneapi/dpl/pstl/omp/parallel_for_each.h
+++ b/include/oneapi/dpl/pstl/omp/parallel_for_each.h
@@ -42,25 +42,6 @@ __parallel_for_each_body(_ForwardIterator __first, _ForwardIterator __last, _Fp
     }
 }
 
-template <class _ExecutionPolicy, class _ForwardIterator, class _Fp>
-void
-__parallel_for_each(_ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _Fp __f)
-{
-    if (omp_in_parallel())
-    {
-        // we don't create a nested parallel region in an existing parallel
-        // region: just create tasks
-        oneapi::dpl::__omp_backend::__parallel_for_each_body(__first, __last, __f);
-    }
-    else
-    {
-        // in any case (nested or non-nested) one parallel region is created and
-        // only one thread creates a set of tasks
-        _PSTL_PRAGMA(omp parallel)
-        _PSTL_PRAGMA(omp single nowait) { oneapi::dpl::__omp_backend::__parallel_for_each_body(__first, __last, __f); }
-    }
-}
-
 template <class _ExecutionPolicy, class _ForwardIterator, class _Fp>
 void
 __parallel_for_each(oneapi::dpl::__internal::__omp_backend_tag, _ExecutionPolicy&&, _ForwardIterator __first,
diff --git a/include/oneapi/dpl/pstl/parallel_backend_serial.h b/include/oneapi/dpl/pstl/parallel_backend_serial.h
index ff8f5e162e2..4bdfa18f21c 100644
--- a/include/oneapi/dpl/pstl/parallel_backend_serial.h
+++ b/include/oneapi/dpl/pstl/parallel_backend_serial.h
@@ -186,14 +186,6 @@ __parallel_invoke(oneapi::dpl::__internal::__serial_backend_tag, _ExecutionPolic
     ::std::forward<_F2>(__f2)();
 }
 
-template <class _ExecutionPolicy, class _ForwardIterator, class _Fp>
-void
-__parallel_for_each(_ExecutionPolicy&&, _ForwardIterator __begin, _ForwardIterator __end, _Fp __f)
-{
-    for (auto __iter = __begin; __iter != __end; ++__iter)
-        __f(*__iter);
-}
-
 template <class _ExecutionPolicy, class _ForwardIterator, class _Fp>
 void
 __parallel_for_each(oneapi::dpl::__internal::__serial_backend_tag, _ExecutionPolicy&&, _ForwardIterator __begin,
diff --git a/include/oneapi/dpl/pstl/parallel_backend_tbb.h b/include/oneapi/dpl/pstl/parallel_backend_tbb.h
index 47c43867d54..bddf9b04849 100644
--- a/include/oneapi/dpl/pstl/parallel_backend_tbb.h
+++ b/include/oneapi/dpl/pstl/parallel_backend_tbb.h
@@ -1392,12 +1392,6 @@ __parallel_invoke(oneapi::dpl::__internal::__tbb_backend_tag, _ExecutionPolicy&&
 //------------------------------------------------------------------------
 // parallel_for_each
 //------------------------------------------------------------------------
-template <class _ExecutionPolicy, class _ForwardIterator, class _Fp>
-void
-__parallel_for_each(_ExecutionPolicy&&, _ForwardIterator __begin, _ForwardIterator __end, _Fp __f)
-{
-    tbb::this_task_arena::isolate([&]() { tbb::parallel_for_each(__begin, __end, __f); });
-}
 
 template <class _ExecutionPolicy, class _ForwardIterator, class _Fp>
 void

From 81076d59159a9aef700e3bedf27cbc39e31c0354 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 12:41:29 +0100
Subject: [PATCH 292/566] Remove __parallel_invoke implementations without tags

---
 include/oneapi/dpl/pstl/omp/parallel_invoke.h    | 16 ----------------
 .../oneapi/dpl/pstl/parallel_backend_serial.h    |  8 --------
 include/oneapi/dpl/pstl/parallel_backend_tbb.h   |  8 --------
 3 files changed, 32 deletions(-)

diff --git a/include/oneapi/dpl/pstl/omp/parallel_invoke.h b/include/oneapi/dpl/pstl/omp/parallel_invoke.h
index 2520526a191..3503096add5 100644
--- a/include/oneapi/dpl/pstl/omp/parallel_invoke.h
+++ b/include/oneapi/dpl/pstl/omp/parallel_invoke.h
@@ -36,22 +36,6 @@ __parallel_invoke_body(_F1&& __f1, _F2&& __f2)
     }
 }
 
-template <class _ExecutionPolicy, typename _F1, typename _F2>
-void
-__parallel_invoke(_ExecutionPolicy&&, _F1&& __f1, _F2&& __f2)
-{
-    if (omp_in_parallel())
-    {
-        oneapi::dpl::__omp_backend::__parallel_invoke_body(std::forward<_F1>(__f1), std::forward<_F2>(__f2));
-    }
-    else
-    {
-        _PSTL_PRAGMA(omp parallel)
-        _PSTL_PRAGMA(omp single nowait)
-        oneapi::dpl::__omp_backend::__parallel_invoke_body(std::forward<_F1>(__f1), std::forward<_F2>(__f2));
-    }
-}
-
 template <class _ExecutionPolicy, typename _F1, typename _F2>
 void
 __parallel_invoke(oneapi::dpl::__internal::__omp_backend_tag, _ExecutionPolicy&&, _F1&& __f1, _F2&& __f2)
diff --git a/include/oneapi/dpl/pstl/parallel_backend_serial.h b/include/oneapi/dpl/pstl/parallel_backend_serial.h
index 4bdfa18f21c..b26707adf77 100644
--- a/include/oneapi/dpl/pstl/parallel_backend_serial.h
+++ b/include/oneapi/dpl/pstl/parallel_backend_serial.h
@@ -170,14 +170,6 @@ __parallel_merge(oneapi::dpl::__internal::__serial_backend_tag, _ExecutionPolicy
     __leaf_merge(__first1, __last1, __first2, __last2, __outit, __comp);
 }
 
-template <class _ExecutionPolicy, typename _F1, typename _F2>
-void
-__parallel_invoke(_ExecutionPolicy&&, _F1&& __f1, _F2&& __f2)
-{
-    ::std::forward<_F1>(__f1)();
-    ::std::forward<_F2>(__f2)();
-}
-
 template <class _ExecutionPolicy, typename _F1, typename _F2>
 void
 __parallel_invoke(oneapi::dpl::__internal::__serial_backend_tag, _ExecutionPolicy&&, _F1&& __f1, _F2&& __f2)
diff --git a/include/oneapi/dpl/pstl/parallel_backend_tbb.h b/include/oneapi/dpl/pstl/parallel_backend_tbb.h
index bddf9b04849..a43665f05b6 100644
--- a/include/oneapi/dpl/pstl/parallel_backend_tbb.h
+++ b/include/oneapi/dpl/pstl/parallel_backend_tbb.h
@@ -1371,14 +1371,6 @@ __parallel_merge(oneapi::dpl::__internal::__tbb_backend_tag, _ExecutionPolicy&&,
 //------------------------------------------------------------------------
 // parallel_invoke
 //------------------------------------------------------------------------
-template <class _ExecutionPolicy, typename _F1, typename _F2>
-void
-__parallel_invoke(_ExecutionPolicy&&, _F1&& __f1, _F2&& __f2)
-{
-    //TODO: a version of tbb::this_task_arena::isolate with variadic arguments pack should be added in the future
-    tbb::this_task_arena::isolate(
-        [&]() { tbb::parallel_invoke(::std::forward<_F1>(__f1), ::std::forward<_F2>(__f2)); });
-}
 
 template <class _ExecutionPolicy, typename _F1, typename _F2>
 void

From 9ccfa70c3829ec1fc4f28ac9f3c055f359015bd2 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 12:43:11 +0100
Subject: [PATCH 293/566] Remove __parallel_reduce implementations without tags

---
 include/oneapi/dpl/pstl/omp/parallel_reduce.h | 27 -------------------
 .../oneapi/dpl/pstl/parallel_backend_serial.h | 15 -----------
 .../oneapi/dpl/pstl/parallel_backend_tbb.h    | 15 -----------
 3 files changed, 57 deletions(-)

diff --git a/include/oneapi/dpl/pstl/omp/parallel_reduce.h b/include/oneapi/dpl/pstl/omp/parallel_reduce.h
index a59a56d374f..4fc62cdf3d8 100644
--- a/include/oneapi/dpl/pstl/omp/parallel_reduce.h
+++ b/include/oneapi/dpl/pstl/omp/parallel_reduce.h
@@ -50,33 +50,6 @@ __parallel_reduce_body(_RandomAccessIterator __first, _RandomAccessIterator __la
 //      c(x,y) combines values x and y that were the result of r
 //------------------------------------------------------------------------
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _Value, typename _RealBody, typename _Reduction>
-_Value
-__parallel_reduce(_ExecutionPolicy&&, _RandomAccessIterator __first, _RandomAccessIterator __last, _Value __identity,
-                  _RealBody __real_body, _Reduction __reduction)
-{
-    // We don't create a nested parallel region in an existing parallel region:
-    // just create tasks.
-    if (omp_in_parallel())
-    {
-        return oneapi::dpl::__omp_backend::__parallel_reduce_body(__first, __last, __identity, __real_body,
-                                                                  __reduction);
-    }
-
-    // In any case (nested or non-nested) one parallel region is created and only
-    // one thread creates a set of tasks.
-    _Value __res = __identity;
-
-    _PSTL_PRAGMA(omp parallel)
-    _PSTL_PRAGMA(omp single nowait)
-    {
-        __res =
-            oneapi::dpl::__omp_backend::__parallel_reduce_body(__first, __last, __identity, __real_body, __reduction);
-    }
-
-    return __res;
-}
-
 template <class _ExecutionPolicy, class _RandomAccessIterator, class _Value, typename _RealBody, typename _Reduction>
 _Value
 __parallel_reduce(oneapi::dpl::__internal::__omp_backend_tag, _ExecutionPolicy&&, _RandomAccessIterator __first,
diff --git a/include/oneapi/dpl/pstl/parallel_backend_serial.h b/include/oneapi/dpl/pstl/parallel_backend_serial.h
index b26707adf77..1e597f277c9 100644
--- a/include/oneapi/dpl/pstl/parallel_backend_serial.h
+++ b/include/oneapi/dpl/pstl/parallel_backend_serial.h
@@ -72,21 +72,6 @@ __parallel_for(oneapi::dpl::__internal::__serial_backend_tag, _ExecutionPolicy&&
     __f(__first, __last);
 }
 
-template <class _ExecutionPolicy, class _Value, class _Index, typename _RealBody, typename _Reduction>
-_Value
-__parallel_reduce(_ExecutionPolicy&&, _Index __first, _Index __last, const _Value& __identity,
-                  const _RealBody& __real_body, const _Reduction&)
-{
-    if (__first == __last)
-    {
-        return __identity;
-    }
-    else
-    {
-        return __real_body(__first, __last, __identity);
-    }
-}
-
 template <class _ExecutionPolicy, class _Value, class _Index, typename _RealBody, typename _Reduction>
 _Value
 __parallel_reduce(oneapi::dpl::__internal::__serial_backend_tag, _ExecutionPolicy&&, _Index __first, _Index __last,
diff --git a/include/oneapi/dpl/pstl/parallel_backend_tbb.h b/include/oneapi/dpl/pstl/parallel_backend_tbb.h
index a43665f05b6..e85ce3c4cf6 100644
--- a/include/oneapi/dpl/pstl/parallel_backend_tbb.h
+++ b/include/oneapi/dpl/pstl/parallel_backend_tbb.h
@@ -128,21 +128,6 @@ __parallel_for(oneapi::dpl::__internal::__tbb_backend_tag, _ExecutionPolicy&&, _
 
 //! Evaluation of brick f[i,j) for each subrange [i,j) of [first,last)
 // wrapper over tbb::parallel_reduce
-template <class _ExecutionPolicy, class _Value, class _Index, typename _RealBody, typename _Reduction>
-_Value
-__parallel_reduce(_ExecutionPolicy&&, _Index __first, _Index __last, const _Value& __identity,
-                  const _RealBody& __real_body, const _Reduction& __reduction)
-{
-    return tbb::this_task_arena::isolate([__first, __last, &__identity, &__real_body, &__reduction]() -> _Value {
-        return tbb::parallel_reduce(
-            tbb::blocked_range<_Index>(__first, __last), __identity,
-            [__real_body](const tbb::blocked_range<_Index>& __r, const _Value& __value) -> _Value {
-                return __real_body(__r.begin(), __r.end(), __value);
-            },
-            __reduction);
-    });
-}
-
 template <class _ExecutionPolicy, class _Value, class _Index, typename _RealBody, typename _Reduction>
 _Value
 __parallel_reduce(oneapi::dpl::__internal::__tbb_backend_tag, _ExecutionPolicy&&, _Index __first, _Index __last,

From 2f7f3ecedecaaa166001587439282fc6f08ee52a Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 12:44:13 +0100
Subject: [PATCH 294/566] Remove __parallel_strict_scan implementations without
 tags

---
 include/oneapi/dpl/pstl/omp/parallel_scan.h   | 36 -----------------
 .../oneapi/dpl/pstl/parallel_backend_serial.h | 13 ------
 .../oneapi/dpl/pstl/parallel_backend_tbb.h    | 40 -------------------
 3 files changed, 89 deletions(-)

diff --git a/include/oneapi/dpl/pstl/omp/parallel_scan.h b/include/oneapi/dpl/pstl/omp/parallel_scan.h
index 7001d4267b3..d3609152319 100644
--- a/include/oneapi/dpl/pstl/omp/parallel_scan.h
+++ b/include/oneapi/dpl/pstl/omp/parallel_scan.h
@@ -106,42 +106,6 @@ __parallel_strict_scan_body(_Index __n, _Tp __initial, _Rp __reduce, _Cp __combi
                                             __initial, __combine, __scan);
 }
 
-template <class _ExecutionPolicy, typename _Index, typename _Tp, typename _Rp, typename _Cp, typename _Sp, typename _Ap>
-void
-__parallel_strict_scan(_ExecutionPolicy&&, _Index __n, _Tp __initial, _Rp __reduce, _Cp __combine, _Sp __scan,
-                       _Ap __apex)
-{
-    if (__n <= __default_chunk_size)
-    {
-        _Tp __sum = __initial;
-        if (__n)
-        {
-            __sum = __combine(__sum, __reduce(_Index(0), __n));
-        }
-        __apex(__sum);
-        if (__n)
-        {
-            __scan(_Index(0), __n, __initial);
-        }
-        return;
-    }
-
-    if (omp_in_parallel())
-    {
-        oneapi::dpl::__omp_backend::__parallel_strict_scan_body<_ExecutionPolicy>(__n, __initial, __reduce, __combine,
-                                                                                  __scan, __apex);
-    }
-    else
-    {
-        _PSTL_PRAGMA(omp parallel)
-        _PSTL_PRAGMA(omp single nowait)
-        {
-            oneapi::dpl::__omp_backend::__parallel_strict_scan_body<_ExecutionPolicy>(__n, __initial, __reduce,
-                                                                                      __combine, __scan, __apex);
-        }
-    }
-}
-
 template <class _ExecutionPolicy, typename _Index, typename _Tp, typename _Rp, typename _Cp, typename _Sp, typename _Ap>
 void
 __parallel_strict_scan(oneapi::dpl::__internal::__omp_backend_tag, _ExecutionPolicy&&, _Index __n, _Tp __initial,
diff --git a/include/oneapi/dpl/pstl/parallel_backend_serial.h b/include/oneapi/dpl/pstl/parallel_backend_serial.h
index 1e597f277c9..54704bbcf6c 100644
--- a/include/oneapi/dpl/pstl/parallel_backend_serial.h
+++ b/include/oneapi/dpl/pstl/parallel_backend_serial.h
@@ -103,19 +103,6 @@ __parallel_transform_reduce(oneapi::dpl::__internal::__serial_backend_tag, _Exec
     return __reduce(__first, __last, __init);
 }
 
-template <class _ExecutionPolicy, typename _Index, typename _Tp, typename _Rp, typename _Cp, typename _Sp, typename _Ap>
-void
-__parallel_strict_scan(_ExecutionPolicy&&, _Index __n, _Tp __initial, _Rp __reduce, _Cp __combine, _Sp __scan,
-                       _Ap __apex)
-{
-    _Tp __sum = __initial;
-    if (__n)
-        __sum = __combine(__sum, __reduce(_Index(0), __n));
-    __apex(__sum);
-    if (__n)
-        __scan(_Index(0), __n, __initial);
-}
-
 template <class _ExecutionPolicy, typename _Index, typename _Tp, typename _Rp, typename _Cp, typename _Sp, typename _Ap>
 void
 __parallel_strict_scan(oneapi::dpl::__internal::__serial_backend_tag, _ExecutionPolicy&&, _Index __n, _Tp __initial,
diff --git a/include/oneapi/dpl/pstl/parallel_backend_tbb.h b/include/oneapi/dpl/pstl/parallel_backend_tbb.h
index e85ce3c4cf6..7687094382b 100644
--- a/include/oneapi/dpl/pstl/parallel_backend_tbb.h
+++ b/include/oneapi/dpl/pstl/parallel_backend_tbb.h
@@ -390,46 +390,6 @@ __downsweep(_Index __i, _Index __m, _Index __tilesize, _Tp* __r, _Index __lastsi
 // apex is called exactly once, after all calls to reduce and before all calls to scan.
 // For example, it's useful for allocating a __buffer used by scan but whose size is the sum of all reduction values.
 // T must have a trivial constructor and destructor.
-template <class _ExecutionPolicy, typename _Index, typename _Tp, typename _Rp, typename _Cp, typename _Sp, typename _Ap>
-void
-__parallel_strict_scan(_ExecutionPolicy&&, _Index __n, _Tp __initial, _Rp __reduce, _Cp __combine, _Sp __scan,
-                       _Ap __apex)
-{
-    tbb::this_task_arena::isolate([=, &__combine]() {
-        if (__n > 1)
-        {
-            _Index __p = tbb::this_task_arena::max_concurrency();
-            const _Index __slack = 4;
-            _Index __tilesize = (__n - 1) / (__slack * __p) + 1;
-            _Index __m = (__n - 1) / __tilesize;
-            __tbb_backend::__buffer<_ExecutionPolicy, _Tp> __buf(__m + 1);
-            _Tp* __r = __buf.get();
-            __tbb_backend::__upsweep(_Index(0), _Index(__m + 1), __tilesize, __r, __n - __m * __tilesize, __reduce,
-                                     __combine);
-
-            // When __apex is a no-op and __combine has no side effects, a good optimizer
-            // should be able to eliminate all code between here and __apex.
-            // Alternatively, provide a default value for __apex that can be
-            // recognized by metaprogramming that conditionlly executes the following.
-            size_t __k = __m + 1;
-            _Tp __t = __r[__k - 1];
-            while ((__k &= __k - 1))
-                __t = __combine(__r[__k - 1], __t);
-            __apex(__combine(__initial, __t));
-            __tbb_backend::__downsweep(_Index(0), _Index(__m + 1), __tilesize, __r, __n - __m * __tilesize, __initial,
-                                       __combine, __scan);
-            return;
-        }
-        // Fewer than 2 elements in sequence, or out of memory.  Handle has single block.
-        _Tp __sum = __initial;
-        if (__n)
-            __sum = __combine(__sum, __reduce(_Index(0), __n));
-        __apex(__sum);
-        if (__n)
-            __scan(_Index(0), __n, __initial);
-    });
-}
-
 template <class _ExecutionPolicy, typename _Index, typename _Tp, typename _Rp, typename _Cp, typename _Sp, typename _Ap>
 void
 __parallel_strict_scan(oneapi::dpl::__internal::__tbb_backend_tag, _ExecutionPolicy&&, _Index __n, _Tp __initial,

From a3476cd702fbf6bc44638f1127f28134083b3eeb Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 13:11:59 +0100
Subject: [PATCH 295/566] Remove __parallel_transform_reduce implementations
 without tags

---
 .../dpl/pstl/omp/parallel_transform_reduce.h  | 29 -------------------
 .../oneapi/dpl/pstl/parallel_backend_serial.h |  8 -----
 .../oneapi/dpl/pstl/parallel_backend_tbb.h    | 12 --------
 3 files changed, 49 deletions(-)

diff --git a/include/oneapi/dpl/pstl/omp/parallel_transform_reduce.h b/include/oneapi/dpl/pstl/omp/parallel_transform_reduce.h
index aa400692f51..2c6cf06577b 100644
--- a/include/oneapi/dpl/pstl/omp/parallel_transform_reduce.h
+++ b/include/oneapi/dpl/pstl/omp/parallel_transform_reduce.h
@@ -83,35 +83,6 @@ __transform_reduce_body(_RandomAccessIterator __first, _RandomAccessIterator __l
     return __init;
 }
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _UnaryOp, class _Value, class _Combiner,
-          class _Reduction>
-_Value
-__parallel_transform_reduce(_ExecutionPolicy&&, _RandomAccessIterator __first, _RandomAccessIterator __last,
-                            _UnaryOp __unary_op, _Value __init, _Combiner __combiner, _Reduction __reduction)
-{
-    _Value __result = __init;
-    if (omp_in_parallel())
-    {
-        // We don't create a nested parallel region in an existing parallel
-        // region: just create tasks
-        __result = oneapi::dpl::__omp_backend::__transform_reduce_body(__first, __last, __unary_op, __init, __combiner,
-                                                                       __reduction);
-    }
-    else
-    {
-        // Create a parallel region, and a single thread will create tasks
-        // for the region.
-        _PSTL_PRAGMA(omp parallel)
-        _PSTL_PRAGMA(omp single nowait)
-        {
-            __result = oneapi::dpl::__omp_backend::__transform_reduce_body(__first, __last, __unary_op, __init,
-                                                                           __combiner, __reduction);
-        }
-    }
-
-    return __result;
-}
-
 template <class _ExecutionPolicy, class _RandomAccessIterator, class _UnaryOp, class _Value, class _Combiner,
           class _Reduction>
 _Value
diff --git a/include/oneapi/dpl/pstl/parallel_backend_serial.h b/include/oneapi/dpl/pstl/parallel_backend_serial.h
index 54704bbcf6c..5876e3383d7 100644
--- a/include/oneapi/dpl/pstl/parallel_backend_serial.h
+++ b/include/oneapi/dpl/pstl/parallel_backend_serial.h
@@ -87,14 +87,6 @@ __parallel_reduce(oneapi::dpl::__internal::__serial_backend_tag, _ExecutionPolic
     }
 }
 
-template <class _ExecutionPolicy, class _Index, class _UnaryOp, class _Tp, class _BinaryOp, class _Reduce>
-_Tp
-__parallel_transform_reduce(_ExecutionPolicy&&, _Index __first, _Index __last, _UnaryOp, _Tp __init, _BinaryOp,
-                            _Reduce __reduce)
-{
-    return __reduce(__first, __last, __init);
-}
-
 template <class _ExecutionPolicy, class _Index, class _UnaryOp, class _Tp, class _BinaryOp, class _Reduce>
 _Tp
 __parallel_transform_reduce(oneapi::dpl::__internal::__serial_backend_tag, _ExecutionPolicy&&, _Index __first,
diff --git a/include/oneapi/dpl/pstl/parallel_backend_tbb.h b/include/oneapi/dpl/pstl/parallel_backend_tbb.h
index 7687094382b..b67e7c7931d 100644
--- a/include/oneapi/dpl/pstl/parallel_backend_tbb.h
+++ b/include/oneapi/dpl/pstl/parallel_backend_tbb.h
@@ -209,18 +209,6 @@ struct __par_trans_red_body
     }
 };
 
-template <class _ExecutionPolicy, class _Index, class _Up, class _Tp, class _Cp, class _Rp>
-_Tp
-__parallel_transform_reduce(_ExecutionPolicy&&, _Index __first, _Index __last, _Up __u, _Tp __init, _Cp __combine,
-                            _Rp __brick_reduce)
-{
-    __tbb_backend::__par_trans_red_body<_Index, _Up, _Tp, _Cp, _Rp> __body(__u, __init, __combine, __brick_reduce);
-    // The grain size of 3 is used in order to provide minimum 2 elements for each body
-    tbb::this_task_arena::isolate(
-        [__first, __last, &__body]() { tbb::parallel_reduce(tbb::blocked_range<_Index>(__first, __last, 3), __body); });
-    return __body.sum();
-}
-
 template <class _ExecutionPolicy, class _Index, class _Up, class _Tp, class _Cp, class _Rp>
 _Tp
 __parallel_transform_reduce(oneapi::dpl::__internal::__tbb_backend_tag, _ExecutionPolicy&&, _Index __first,

From 11dc7750c48cde521d6c1726b86f6403faa4515a Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 15 Feb 2024 17:42:20 +0100
Subject: [PATCH 296/566] Remove old implementations with
 __enable_if_host_execution_policy

---
 include/oneapi/dpl/pstl/algorithm_fwd.h  |  630 +-----
 include/oneapi/dpl/pstl/algorithm_impl.h | 2635 +---------------------
 include/oneapi/dpl/pstl/numeric_fwd.h    |   54 -
 include/oneapi/dpl/pstl/numeric_impl.h   |  144 --
 4 files changed, 55 insertions(+), 3408 deletions(-)

diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index e456f7a5093..bcf5fa3b829 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -46,20 +46,10 @@ bool
 __brick_any_of(const _RandomAccessIterator, const _RandomAccessIterator, _Pred,
                /*__is_vector=*/::std::true_type) noexcept;
 
-template <class _ExecutionPolicy, class _ForwardIterator, class _Pred, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, bool>
-__pattern_any_of(_ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _Pred, _IsVector,
-                 /*parallel=*/::std::false_type) noexcept;
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Pred>
 bool
 __pattern_any_of(_Tag, _ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _Pred) noexcept;
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _Pred, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, bool>
-__pattern_any_of(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _Pred, _IsVector,
-                 /*parallel=*/::std::true_type);
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Pred>
 bool
 __pattern_any_of(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _Pred);
@@ -78,11 +68,6 @@ template <class _RandomAccessIterator, class _Function>
 void __brick_walk1(_RandomAccessIterator, _RandomAccessIterator, _Function,
                    /*vector=*/::std::true_type) noexcept;
 
-template <class _ExecutionPolicy, class _ForwardIterator, class _Function, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>
-__pattern_walk1(_ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _Function, _IsVector,
-                /*parallel=*/::std::false_type) noexcept;
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Function, class _IsVector>
 void
 __pattern_walk1(_Tag, _ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _Function) noexcept;
@@ -109,20 +94,10 @@ template <class _IsVector, class _ExecutionPolicy, class _ForwardIterator, class
 void
 __pattern_walk1(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _Function);
 
-template <class _ExecutionPolicy, class _ForwardIterator, class _Brick>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>
-__pattern_walk_brick(_ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _Brick,
-                     /*parallel=*/::std::false_type) noexcept;
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Brick>
 void
 __pattern_walk_brick(_Tag, _ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _Brick) noexcept;
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _Brick>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>
-__pattern_walk_brick(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _Brick,
-                     /*parallel=*/::std::true_type);
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Brick>
 void
 __pattern_walk_brick(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator,
@@ -140,38 +115,18 @@ template <class _RandomAccessIterator, class _DifferenceType, class _Function>
 _RandomAccessIterator __brick_walk1_n(_RandomAccessIterator, _DifferenceType, _Function,
                                       /*vectorTag=*/::std::true_type) noexcept;
 
-template <class _ExecutionPolicy, class _ForwardIterator, class _Size, class _Function, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _ForwardIterator>
-__pattern_walk1_n(_ExecutionPolicy&&, _ForwardIterator, _Size, _Function, _IsVector,
-                  /*is_parallel=*/::std::false_type) noexcept;
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Size, class _Function>
 _ForwardIterator
 __pattern_walk1_n(_Tag, _ExecutionPolicy&&, _ForwardIterator, _Size, _Function) noexcept;
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _Size, class _Function, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator>
-__pattern_walk1_n(_ExecutionPolicy&&, _RandomAccessIterator, _Size, _Function, _IsVector,
-                  /*is_parallel=*/::std::true_type);
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Size, class _Function>
 _RandomAccessIterator
 __pattern_walk1_n(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator, _Size, _Function);
 
-template <class _ExecutionPolicy, class _ForwardIterator, class _Size, class _Brick>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _ForwardIterator>
-__pattern_walk_brick_n(_ExecutionPolicy&&, _ForwardIterator, _Size, _Brick,
-                       /*is_parallel=*/::std::false_type) noexcept;
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Size, class _Brick>
 _ForwardIterator
 __pattern_walk_brick_n(_Tag, _ExecutionPolicy&&, _ForwardIterator, _Size, _Brick) noexcept;
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _Size, class _Brick>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator>
-__pattern_walk_brick_n(_ExecutionPolicy&&, _RandomAccessIterator, _Size, _Brick,
-                       /*is_parallel=*/::std::true_type);
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Size, class _Brick>
 _RandomAccessIterator
 __pattern_walk_brick_n(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator, _Size, _Brick);
@@ -198,11 +153,6 @@ template <class _RandomAccessIterator1, class _Size, class _RandomAccessIterator
 _RandomAccessIterator2 __brick_walk2_n(_RandomAccessIterator1, _Size, _RandomAccessIterator2, _Function,
                                        /*vector=*/::std::true_type) noexcept;
 
-template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Function, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _ForwardIterator2>
-__pattern_walk2(_ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2, _Function, _IsVector,
-                /*parallel=*/::std::false_type) noexcept;
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Function>
 _ForwardIterator2
 __pattern_walk2(_Tag, _ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2, _Function) noexcept;
@@ -232,34 +182,17 @@ _ForwardIterator2
 __pattern_walk2(__parallel_forward_tag, _ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2,
                 _Function);
 
-template <class _ExecutionPolicy, class _ForwardIterator1, class _Size, class _ForwardIterator2, class _Function,
-          class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _ForwardIterator2>
-__pattern_walk2_n(_ExecutionPolicy&&, _ForwardIterator1, _Size, _ForwardIterator2, _Function, _IsVector,
-                  /*parallel=*/::std::false_type) noexcept;
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _Size, class _ForwardIterator2,
           class _Function>
 _ForwardIterator2
 __pattern_walk2_n(_Tag, _ExecutionPolicy&&, _ForwardIterator1, _Size, _ForwardIterator2, _Function) noexcept;
 
-template <class _ExecutionPolicy, class _RandomAccessIterator1, class _Size, class _RandomAccessIterator2,
-          class _Function, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator2>
-__pattern_walk2_n(_ExecutionPolicy&&, _RandomAccessIterator1, _Size, _RandomAccessIterator2, _Function, _IsVector,
-                  /*parallel=*/::std::true_type);
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _Size,
           class _RandomAccessIterator2, class _Function>
 _RandomAccessIterator2
 __pattern_walk2_n(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator1, _Size, _RandomAccessIterator2,
                   _Function);
 
-template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Brick>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _ForwardIterator2>
-__pattern_walk2_brick(_ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2, _Brick,
-                      /*parallel=*/::std::false_type) noexcept;
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Brick>
 _ForwardIterator2
 __pattern_walk2_brick(_Tag, _ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2,
@@ -290,21 +223,11 @@ _ForwardIterator2
 __pattern_walk2_brick(__parallel_forward_tag, _ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1,
                       _ForwardIterator2, _Brick);
 
-template <class _ExecutionPolicy, class _ForwardIterator1, class _Size, class _ForwardIterator2, class _Brick>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _ForwardIterator2>
-__pattern_walk2_brick_n(_ExecutionPolicy&&, _ForwardIterator1, _Size, _ForwardIterator2, _Brick,
-                        /*parallel=*/::std::false_type) noexcept;
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _Size, class _ForwardIterator2,
           class _Brick>
 _ForwardIterator2
 __pattern_walk2_brick_n(_Tag, _ExecutionPolicy&&, _ForwardIterator1, _Size, _ForwardIterator2, _Brick) noexcept;
 
-template <class _ExecutionPolicy, class _RandomAccessIterator1, class _Size, class _RandomAccessIterator2, class _Brick>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator2>
-__pattern_walk2_brick_n(_ExecutionPolicy&&, _RandomAccessIterator1, _Size, _RandomAccessIterator2, _Brick,
-                        /*parallel=*/::std::true_type);
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _Size,
           class _RandomAccessIterator2, class _Brick>
 _RandomAccessIterator2
@@ -326,13 +249,6 @@ _RandomAccessIterator3 __brick_walk3(_RandomAccessIterator1, _RandomAccessIterat
                                      _RandomAccessIterator3, _Function,
                                      /*vector=*/::std::true_type) noexcept;
 
-template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _ForwardIterator3,
-          class _Function, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _ForwardIterator3>
-__pattern_walk3(_ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2, _ForwardIterator3,
-                _Function, _IsVector,
-                /*parallel=*/::std::false_type) noexcept;
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _ForwardIterator3,
           class _Function>
 _ForwardIterator3
@@ -375,26 +291,12 @@ __pattern_walk3(__parallel_forward_tag, _ExecutionPolicy&&, _ForwardIterator1, _
 // transform_if
 //------------------------------------------------------------------------
 
-template <typename _ExecutionPolicy, typename _ForwardIterator1, typename _ForwardIterator2, typename _Function,
-          class _IsVector, class _IsParallel>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _ForwardIterator2>
-__pattern_walk2_transform_if(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
-                             _ForwardIterator2 __first2, _Function __func, _IsVector __is_vector,
-                             _IsParallel __is_parallel) noexcept;
-
 template <class _Tag, typename _ExecutionPolicy, typename _ForwardIterator1, typename _ForwardIterator2,
           typename _Function>
 _ForwardIterator2
 __pattern_walk2_transform_if(_Tag, _ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2,
                              _Function) noexcept;
 
-template <typename _ExecutionPolicy, typename _ForwardIterator1, typename _ForwardIterator2, typename _ForwardIterator3,
-          typename _Function, class _IsVector, class _IsParallel>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _ForwardIterator3>
-__pattern_walk3_transform_if(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
-                             _ForwardIterator2 __first2, _ForwardIterator3 __first3, _Function __func,
-                             _IsVector __is_vector, _IsParallel __is_parallel) noexcept;
-
 template <class _Tag, typename _ExecutionPolicy, typename _ForwardIterator1, typename _ForwardIterator2,
           typename _ForwardIterator3, typename _Function>
 _ForwardIterator3
@@ -413,23 +315,11 @@ template <class _RandomAccessIterator1, class _RandomAccessIterator2, class _Bin
 bool __brick_equal(_RandomAccessIterator1, _RandomAccessIterator1, _RandomAccessIterator2, _BinaryPredicate,
                    /* is_vector = */ ::std::true_type) noexcept;
 
-template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _BinaryPredicate,
-          class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, bool>
-__pattern_equal(_ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2, _BinaryPredicate,
-                _IsVector, /* is_parallel = */ ::std::false_type) noexcept;
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _BinaryPredicate>
 bool
 __pattern_equal(_Tag, _ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2, _ForwardIterator2,
                 _BinaryPredicate) noexcept;
 
-template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _BinaryPredicate,
-          class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, bool>
-__pattern_equal(_ExecutionPolicy&&, _RandomAccessIterator1, _RandomAccessIterator1, _RandomAccessIterator2,
-                _BinaryPredicate, _IsVector, /* is_parallel = */ ::std::true_type);
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
           class _BinaryPredicate>
 bool
@@ -444,23 +334,11 @@ template <class _RandomAccessIterator1, class _RandomAccessIterator2, class _Bin
 bool __brick_equal(_RandomAccessIterator1, _RandomAccessIterator1, _RandomAccessIterator2, _RandomAccessIterator2,
                    _BinaryPredicate, /* is_vector = */ ::std::true_type) noexcept;
 
-template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _BinaryPredicate,
-          class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, bool>
-__pattern_equal(_ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2, _ForwardIterator2,
-                _BinaryPredicate, _IsVector, /* is_parallel = */ ::std::false_type) noexcept;
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _BinaryPredicate>
 bool
 __pattern_equal(_Tag, _ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2,
                 _BinaryPredicate) noexcept;
 
-template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _BinaryPredicate,
-          class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, bool>
-__pattern_equal(_ExecutionPolicy&&, _RandomAccessIterator1, _RandomAccessIterator1, _RandomAccessIterator2,
-                _RandomAccessIterator2, _BinaryPredicate, _IsVector, /* is_parallel = */ ::std::true_type);
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
           class _BinaryPredicate>
 bool
@@ -501,25 +379,11 @@ _RandomAccessIterator1 __brick_find_end(_RandomAccessIterator1, _RandomAccessIte
                                         _RandomAccessIterator2, _BinaryPredicate,
                                         /*__is_vector=*/::std::true_type) noexcept;
 
-template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _BinaryPredicate,
-          class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _ForwardIterator1>
-__pattern_find_end(_ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2, _ForwardIterator2,
-                   _BinaryPredicate, _IsVector,
-                   /*is_parallel=*/::std::false_type) noexcept;
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _BinaryPredicate>
 _ForwardIterator1
 __pattern_find_end(_Tag, _ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2, _ForwardIterator2,
                    _BinaryPredicate) noexcept;
 
-template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _BinaryPredicate,
-          class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator1>
-__pattern_find_end(_ExecutionPolicy&&, _RandomAccessIterator1, _RandomAccessIterator1, _RandomAccessIterator2,
-                   _RandomAccessIterator2, _BinaryPredicate, _IsVector,
-                   /*is_parallel=*/::std::true_type);
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
           class _BinaryPredicate>
 _RandomAccessIterator1
@@ -540,23 +404,11 @@ _RandomAccessIterator1 __brick_find_first_of(_RandomAccessIterator1, _RandomAcce
                                              _RandomAccessIterator2, _BinaryPredicate,
                                              /*__is_vector=*/::std::true_type) noexcept;
 
-template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _BinaryPredicate,
-          class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _ForwardIterator1>
-__pattern_find_first_of(_ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2, _ForwardIterator2,
-                        _BinaryPredicate, _IsVector, /*is_parallel=*/::std::false_type) noexcept;
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _BinaryPredicate>
 _ForwardIterator1
 __pattern_find_first_of(_Tag, _ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2,
                         _ForwardIterator2, _BinaryPredicate) noexcept;
 
-template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _BinaryPredicate,
-          class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator1>
-__pattern_find_first_of(_ExecutionPolicy&&, _RandomAccessIterator1, _RandomAccessIterator1, _RandomAccessIterator2,
-                        _RandomAccessIterator2, _BinaryPredicate, _IsVector, /*is_parallel=*/::std::true_type);
-
 template <class _IsVector, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2,
           class _BinaryPredicate>
 _ForwardIterator1
@@ -577,25 +429,11 @@ _RandomAccessIterator1 __brick_search(_RandomAccessIterator1, _RandomAccessItera
                                       _RandomAccessIterator2, _BinaryPredicate,
                                       /*vector=*/::std::true_type) noexcept;
 
-template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _BinaryPredicate,
-          class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _ForwardIterator1>
-__pattern_search(_ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2, _ForwardIterator2,
-                 _BinaryPredicate, _IsVector,
-                 /*is_parallel=*/::std::false_type) noexcept;
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _BinaryPredicate>
 _ForwardIterator1
 __pattern_search(_Tag, _ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2, _ForwardIterator2,
                  _BinaryPredicate) noexcept;
 
-template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _BinaryPredicate,
-          class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator1>
-__pattern_search(_ExecutionPolicy&&, _RandomAccessIterator1, _RandomAccessIterator1, _RandomAccessIterator2,
-                 _RandomAccessIterator2, _BinaryPredicate, _IsVector,
-                 /*is_parallel=*/::std::true_type);
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
           class _BinaryPredicate>
 _RandomAccessIterator1
@@ -616,25 +454,11 @@ _RandomAccessIterator
 __brick_search_n(_RandomAccessIterator, _RandomAccessIterator, _Size, const _Tp&, _BinaryPredicate,
                  /*vector=*/::std::true_type) noexcept;
 
-template <class _ExecutionPolicy, class _ForwardIterator, class _Size, class _Tp, class _BinaryPredicate,
-          class IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _ForwardIterator>
-__pattern_search_n(_ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _Size, const _Tp&, _BinaryPredicate,
-                   IsVector,
-                   /*is_parallel=*/::std::false_type) noexcept;
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Size, class _Tp, class _BinaryPredicate>
 _ForwardIterator
 __pattern_search_n(_Tag, _ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _Size, const _Tp&,
                    _BinaryPredicate) noexcept;
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _Size, class _Tp, class _BinaryPredicate,
-          class IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator>
-__pattern_search_n(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _Size, const _Tp&,
-                   _BinaryPredicate, IsVector,
-                   /*is_parallel=*/::std::true_type);
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Size, class _Tp,
           class _BinaryPredicate>
 _RandomAccessIterator
@@ -708,22 +532,11 @@ void
 __brick_partition_by_mask(_RandomAccessIterator, _RandomAccessIterator, _OutputIterator1, _OutputIterator2, bool*,
                           /*vector=*/::std::true_type) noexcept;
 
-template <class _ExecutionPolicy, class _ForwardIterator, class _OutputIterator, class _UnaryPredicate, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _OutputIterator>
-__pattern_copy_if(_ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _OutputIterator, _UnaryPredicate, _IsVector,
-                  /*parallel=*/::std::false_type) noexcept;
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _OutputIterator, class _UnaryPredicate>
 _OutputIterator
 __pattern_copy_if(_Tag, _ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _OutputIterator,
                   _UnaryPredicate) noexcept;
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _OutputIterator, class _UnaryPredicate,
-          class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _OutputIterator>
-__pattern_copy_if(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _OutputIterator, _UnaryPredicate,
-                  _IsVector, /*parallel=*/::std::true_type);
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
           class _UnaryPredicate>
 _RandomAccessIterator2
@@ -744,22 +557,10 @@ typename ::std::iterator_traits<_ForwardIterator>::difference_type
     __brick_count(_ForwardIterator, _ForwardIterator, _Predicate,
                   /* is_vector = */ ::std::false_type) noexcept;
 
-template <class _ExecutionPolicy, class _ForwardIterator, class _Predicate, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<
-    _ExecutionPolicy, typename ::std::iterator_traits<_ForwardIterator>::difference_type>
-__pattern_count(_ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _Predicate,
-                /* is_parallel */ ::std::false_type, _IsVector) noexcept;
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Predicate>
 typename ::std::iterator_traits<_ForwardIterator>::difference_type
 __pattern_count(_Tag, _ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _Predicate) noexcept;
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _Predicate, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<
-    _ExecutionPolicy, typename ::std::iterator_traits<_RandomAccessIterator>::difference_type>
-__pattern_count(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _Predicate,
-                /* is_parallel */ ::std::true_type, _IsVector);
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Predicate>
 typename ::std::iterator_traits<_RandomAccessIterator>::difference_type
 __pattern_count(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator,
@@ -777,20 +578,10 @@ template <class _RandomAccessIterator, class _BinaryPredicate>
 _RandomAccessIterator __brick_unique(_RandomAccessIterator, _RandomAccessIterator, _BinaryPredicate,
                                      /*is_vector=*/::std::true_type) noexcept;
 
-template <class _ExecutionPolicy, class _ForwardIterator, class _BinaryPredicate, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _ForwardIterator>
-__pattern_unique(_ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _BinaryPredicate, _IsVector,
-                 /*is_parallel=*/::std::false_type) noexcept;
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _BinaryPredicate>
 _ForwardIterator
 __pattern_unique(_Tag, _ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _BinaryPredicate) noexcept;
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _BinaryPredicate, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator>
-__pattern_unique(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _BinaryPredicate, _IsVector,
-                 /*is_parallel=*/::std::true_type);
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _BinaryPredicate>
 _RandomAccessIterator
 __pattern_unique(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator,
@@ -808,12 +599,6 @@ template <class _RandomAccessIterator, class _OutputIterator, class _BinaryPredi
 _OutputIterator __brick_unique_copy(_RandomAccessIterator, _RandomAccessIterator, _OutputIterator, _BinaryPredicate,
                                     /*vector=*/::std::true_type) noexcept;
 
-template <class _ExecutionPolicy, class _ForwardIterator, class _OutputIterator, class _BinaryPredicate,
-          class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _OutputIterator>
-__pattern_unique_copy(_ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _OutputIterator, _BinaryPredicate,
-                      _IsVector, /*parallel=*/::std::false_type) noexcept;
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _OutputIterator, class _BinaryPredicate>
 _OutputIterator
 __pattern_unique_copy(_Tag, _ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _OutputIterator,
@@ -829,12 +614,6 @@ _DifferenceType
 __brick_calc_mask_2(_RandomAccessIterator, _RandomAccessIterator, bool* __restrict, _BinaryPredicate,
                     /*vector=*/::std::true_type) noexcept;
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _OutputIterator, class _BinaryPredicate,
-          class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _OutputIterator>
-__pattern_unique_copy(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _OutputIterator,
-                      _BinaryPredicate, _IsVector, /*parallel=*/::std::true_type);
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
           class _BinaryPredicate>
 _RandomAccessIterator2
@@ -861,20 +640,10 @@ template <class _RandomAccessIterator>
 void __brick_reverse(_RandomAccessIterator, _RandomAccessIterator, _RandomAccessIterator,
                      /*is_vector=*/::std::true_type) noexcept;
 
-template <class _ExecutionPolicy, class _BidirectionalIterator, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>
-__pattern_reverse(_ExecutionPolicy&&, _BidirectionalIterator, _BidirectionalIterator, _IsVector,
-                  /*is_parallel=*/::std::false_type) noexcept;
-
 template <class _Tag, class _ExecutionPolicy, class _BidirectionalIterator>
 void
 __pattern_reverse(_Tag, _ExecutionPolicy&&, _BidirectionalIterator, _BidirectionalIterator) noexcept;
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>
-__pattern_reverse(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _IsVector,
-                  /*is_parallel=*/::std::true_type);
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator>
 void
 __pattern_reverse(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator);
@@ -891,21 +660,11 @@ template <class _RandomAccessIterator, class _OutputIterator>
 _OutputIterator __brick_reverse_copy(_RandomAccessIterator, _RandomAccessIterator, _OutputIterator,
                                      /*is_vector=*/::std::true_type) noexcept;
 
-template <class _ExecutionPolicy, class _BidirectionalIterator, class _OutputIterator, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _OutputIterator>
-__pattern_reverse_copy(_ExecutionPolicy&&, _BidirectionalIterator, _BidirectionalIterator, _OutputIterator, _IsVector,
-                       /*is_parallel=*/::std::false_type) noexcept;
-
 template <class _Tag, class _ExecutionPolicy, class _BidirectionalIterator, class _OutputIterator>
 _OutputIterator
 __pattern_reverse_copy(_Tag, _ExecutionPolicy&&, _BidirectionalIterator, _BidirectionalIterator,
                        _OutputIterator) noexcept;
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _OutputIterator, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _OutputIterator>
-__pattern_reverse_copy(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _OutputIterator, _IsVector,
-                       /*is_parallel=*/::std::true_type);
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2>
 _RandomAccessIterator2
 __pattern_reverse_copy(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator1, _RandomAccessIterator1,
@@ -923,20 +682,10 @@ template <class _RandomAccessIterator>
 _RandomAccessIterator __brick_rotate(_RandomAccessIterator, _RandomAccessIterator, _RandomAccessIterator,
                                      /*is_vector=*/::std::true_type) noexcept;
 
-template <class _ExecutionPolicy, class _ForwardIterator, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _ForwardIterator>
-__pattern_rotate(_ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _ForwardIterator, _IsVector,
-                 /*is_parallel=*/::std::false_type) noexcept;
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator>
 _ForwardIterator
 __pattern_rotate(_Tag, _ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _ForwardIterator) noexcept;
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator>
-__pattern_rotate(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _RandomAccessIterator, _IsVector,
-                 /*is_parallel=*/::std::true_type);
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator>
 _RandomAccessIterator
 __pattern_rotate(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator,
@@ -955,23 +704,11 @@ _OutputIterator __brick_rotate_copy(_RandomAccessIterator, _RandomAccessIterator
                                     _OutputIterator,
                                     /*__is_vector=*/::std::true_type) noexcept;
 
-template <class _ExecutionPolicy, class _ForwardIterator, class _OutputIterator, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _OutputIterator>
-__pattern_rotate_copy(_ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _ForwardIterator, _OutputIterator,
-                      _IsVector,
-                      /*is_parallel=*/::std::false_type) noexcept;
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _OutputIterator>
 _OutputIterator
 __pattern_rotate_copy(_Tag, _ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _ForwardIterator,
                       _OutputIterator) noexcept;
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _OutputIterator, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _OutputIterator>
-__pattern_rotate_copy(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _RandomAccessIterator,
-                      _OutputIterator, _IsVector,
-                      /*is_parallel=*/::std::true_type);
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2>
 _RandomAccessIterator2
 __pattern_rotate_copy(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator1, _RandomAccessIterator1,
@@ -989,20 +726,10 @@ template <class _RandomAccessIterator, class _UnaryPredicate>
 bool __brick_is_partitioned(_RandomAccessIterator, _RandomAccessIterator, _UnaryPredicate,
                             /*is_vector=*/::std::true_type) noexcept;
 
-template <class _ExecutionPolicy, class _ForwardIterator, class _UnaryPredicate, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, bool>
-__pattern_is_partitioned(_ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _UnaryPredicate, _IsVector,
-                         /*is_parallel=*/::std::false_type) noexcept;
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _UnaryPredicate>
 bool
 __pattern_is_partitioned(_Tag, _ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _UnaryPredicate) noexcept;
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _UnaryPredicate, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, bool>
-__pattern_is_partitioned(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _UnaryPredicate, _IsVector,
-                         /*is_parallel=*/::std::true_type);
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _UnaryPredicate>
 bool
 __pattern_is_partitioned(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator,
@@ -1020,20 +747,10 @@ template <class _RandomAccessIterator, class _UnaryPredicate>
 _RandomAccessIterator __brick_partition(_RandomAccessIterator, _RandomAccessIterator, _UnaryPredicate,
                                         /*is_vector=*/::std::true_type) noexcept;
 
-template <class _ExecutionPolicy, class _ForwardIterator, class _UnaryPredicate, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _ForwardIterator>
-__pattern_partition(_ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _UnaryPredicate, _IsVector,
-                    /*is_parallel=*/::std::false_type) noexcept;
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _UnaryPredicate>
 _ForwardIterator
 __pattern_partition(_Tag, _ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _UnaryPredicate) noexcept;
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _UnaryPredicate, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator>
-__pattern_partition(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _UnaryPredicate, _IsVector,
-                    /*is_parallel=*/::std::true_type);
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _UnaryPredicate>
 _RandomAccessIterator
 __pattern_partition(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator,
@@ -1051,22 +768,11 @@ template <class _RandomAccessIterator, class _UnaryPredicate>
 _RandomAccessIterator __brick_stable_partition(_RandomAccessIterator, _RandomAccessIterator, _UnaryPredicate,
                                                /*__is_vector=*/::std::true_type) noexcept;
 
-template <class _ExecutionPolicy, class _BidirectionalIterator, class _UnaryPredicate, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _BidirectionalIterator>
-__pattern_stable_partition(_ExecutionPolicy&&, _BidirectionalIterator, _BidirectionalIterator, _UnaryPredicate,
-                           _IsVector,
-                           /*is_parallelization=*/::std::false_type) noexcept;
-
 template <class _Tag, class _ExecutionPolicy, class _BidirectionalIterator, class _UnaryPredicate>
 _BidirectionalIterator
 __pattern_stable_partition(_Tag, _ExecutionPolicy&&, _BidirectionalIterator, _BidirectionalIterator,
                            _UnaryPredicate) noexcept;
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _UnaryPredicate, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator>
-__pattern_stable_partition(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _UnaryPredicate, _IsVector,
-                           /*is_parallelization=*/::std::true_type);
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _UnaryPredicate>
 _RandomAccessIterator
 __pattern_stable_partition(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator,
@@ -1087,28 +793,12 @@ ::std::pair<_OutputIterator1, _OutputIterator2> __brick_partition_copy(_RandomAc
                                                                        _UnaryPredicate,
                                                                        /*is_vector=*/::std::true_type) noexcept;
 
-template <class _ExecutionPolicy, class _ForwardIterator, class _OutputIterator1, class _OutputIterator2,
-          class _UnaryPredicate, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy,
-                                                           ::std::pair<_OutputIterator1, _OutputIterator2>>
-__pattern_partition_copy(_ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _OutputIterator1, _OutputIterator2,
-                         _UnaryPredicate, _IsVector,
-                         /*is_parallelization=*/::std::false_type) noexcept;
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _OutputIterator1, class _OutputIterator2,
           class _UnaryPredicate>
 ::std::pair<_OutputIterator1, _OutputIterator2>
 __pattern_partition_copy(_Tag, _ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _OutputIterator1,
                          _OutputIterator2, _UnaryPredicate) noexcept;
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _OutputIterator1, class _OutputIterator2,
-          class _UnaryPredicate, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy,
-                                                           ::std::pair<_OutputIterator1, _OutputIterator2>>
-__pattern_partition_copy(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _OutputIterator1,
-                         _OutputIterator2, _UnaryPredicate, _IsVector,
-                         /*is_parallelization=*/::std::true_type);
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
           class _RandomAccessIterator3, class _UnaryPredicate>
 ::std::pair<_RandomAccessIterator2, _RandomAccessIterator3>
@@ -1119,23 +809,11 @@ __pattern_partition_copy(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomA
 // sort
 //------------------------------------------------------------------------
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _Compare, class _IsVector,
-          class _IsMoveConstructible>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>
-__pattern_sort(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _Compare, _IsVector /*is_vector*/,
-               /*is_parallel=*/::std::false_type, _IsMoveConstructible) noexcept;
-
 template <class _Tag, class _ExecutionPolicy, class _RandomAccessIterator, class _Compare, class _IsMoveConstructible>
 void
 __pattern_sort(_Tag, _ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _Compare,
                _IsMoveConstructible) noexcept;
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _Compare, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>
-__pattern_sort(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _Compare, _IsVector /*is_vector*/,
-               /*is_parallel=*/::std::true_type,
-               /*is_move_constructible=*/::std::true_type);
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Compare>
 void
 __pattern_sort(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _Compare,
@@ -1145,22 +823,10 @@ __pattern_sort(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessItera
 // stable_sort
 //------------------------------------------------------------------------
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _Compare, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>
-__pattern_stable_sort(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _Compare,
-                      _IsVector /*is_vector*/,
-                      /*is_parallel=*/::std::false_type) noexcept;
-
 template <class _Tag, class _ExecutionPolicy, class _RandomAccessIterator, class _Compare>
 void
 __pattern_stable_sort(_Tag, _ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _Compare) noexcept;
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _Compare, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>
-__pattern_stable_sort(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _Compare,
-                      _IsVector /*is_vector*/,
-                      /*is_parallel=*/::std::true_type);
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Compare>
 void
 __pattern_stable_sort(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator,
@@ -1169,12 +835,6 @@ __pattern_stable_sort(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAcce
 //------------------------------------------------------------------------
 // sort_by_key
 //------------------------------------------------------------------------
-template <typename _ExecutionPolicy, typename _RandomAccessIterator1, typename _RandomAccessIterator2,
-          typename _Compare, typename _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>
-__pattern_sort_by_key(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __keys_first,
-                      _RandomAccessIterator1 __keys_last, _RandomAccessIterator2 __values_first, _Compare __comp,
-                      _IsVector /*vector=*/, /*is_parallel=*/::std::false_type) noexcept;
 
 template <class _Tag, typename _ExecutionPolicy, typename _RandomAccessIterator1, typename _RandomAccessIterator2,
           typename _Compare>
@@ -1182,13 +842,6 @@ void
 __pattern_sort_by_key(_Tag, _ExecutionPolicy&&, _RandomAccessIterator1, _RandomAccessIterator1, _RandomAccessIterator2,
                       _Compare) noexcept;
 
-template <typename _ExecutionPolicy, typename _RandomAccessIterator1, typename _RandomAccessIterator2,
-          typename _Compare, typename _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>
-__pattern_sort_by_key(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __keys_first,
-                      _RandomAccessIterator1 __keys_last, _RandomAccessIterator2 __values_first, _Compare __comp,
-                      _IsVector /*vector=*/, /*is_parallel=*/::std::true_type);
-
 template <typename _IsVector, typename _ExecutionPolicy, typename _RandomAccessIterator1,
           typename _RandomAccessIterator2, typename _Compare>
 void
@@ -1199,23 +852,11 @@ __pattern_sort_by_key(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAcce
 // partial_sort
 //------------------------------------------------------------------------
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _Compare, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>
-__pattern_partial_sort(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _RandomAccessIterator,
-                       _Compare, _IsVector,
-                       /*is_parallel=*/::std::false_type) noexcept;
-
 template <class _Tag, class _ExecutionPolicy, class _RandomAccessIterator, class _Compare>
 void
 __pattern_partial_sort(_Tag, _ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _RandomAccessIterator,
                        _Compare) noexcept;
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _Compare, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>
-__pattern_partial_sort(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _RandomAccessIterator,
-                       _Compare, _IsVector,
-                       /*is_parallel=*/::std::true_type);
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Compare>
 void
 __pattern_partial_sort(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator,
@@ -1225,25 +866,11 @@ __pattern_partial_sort(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAcc
 // partial_sort_copy
 //------------------------------------------------------------------------
 
-template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _Compare,
-          class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator2>
-__pattern_partial_sort_copy(_ExecutionPolicy&&, _RandomAccessIterator1, _RandomAccessIterator1, _RandomAccessIterator2,
-                            _RandomAccessIterator2, _Compare, _IsVector,
-                            /*is_parallel=*/::std::false_type) noexcept;
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _RandomAccessIterator, class _Compare>
 _RandomAccessIterator
 __pattern_partial_sort_copy(_Tag, _ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _RandomAccessIterator,
                             _RandomAccessIterator, _Compare) noexcept;
 
-template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _Compare,
-          class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator2>
-__pattern_partial_sort_copy(_ExecutionPolicy&&, _RandomAccessIterator1, _RandomAccessIterator1, _RandomAccessIterator2,
-                            _RandomAccessIterator2, _Compare, _IsVector,
-                            /*is_parallel=*/::std::true_type);
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _RandomAccessIterator, class _Compare>
 _RandomAccessIterator
 __pattern_partial_sort_copy(_Tag, _ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _RandomAccessIterator,
@@ -1263,21 +890,11 @@ _ForwardIterator
 __brick_adjacent_find(_ForwardIterator, _ForwardIterator, _BinaryPredicate,
                       /* IsVector = */ ::std::false_type, bool) noexcept;
 
-template <class _ExecutionPolicy, class _ForwardIterator, class _BinaryPredicate, class _IsVector, class _Semantic>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _ForwardIterator>
-__pattern_adjacent_find(_ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _BinaryPredicate,
-                        /* is_parallel */ ::std::false_type, _IsVector, _Semantic) noexcept;
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _BinaryPredicate, class _Semantic>
 _ForwardIterator
 __pattern_adjacent_find(_Tag, _ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _BinaryPredicate,
                         _Semantic) noexcept;
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _BinaryPredicate, class _IsVector, class _Semantic>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator>
-__pattern_adjacent_find(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _BinaryPredicate,
-                        /* is_parallel */ ::std::true_type, _IsVector, _Semantic);
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _BinaryPredicate, class _Semantic>
 _RandomAccessIterator
 __pattern_adjacent_find(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator,
@@ -1286,23 +903,12 @@ __pattern_adjacent_find(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAc
 //------------------------------------------------------------------------
 // nth_element
 //------------------------------------------------------------------------
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _Compare, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>
-__pattern_nth_element(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _RandomAccessIterator, _Compare,
-                      _IsVector,
-                      /*is_parallel=*/::std::false_type) noexcept;
 
 template <class _Tag, class _ExecutionPolicy, class _RandomAccessIterator, class _Compare>
 void
 __pattern_nth_element(_Tag, _ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _RandomAccessIterator,
                       _Compare) noexcept;
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _Compare, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>
-__pattern_nth_element(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _RandomAccessIterator, _Compare,
-                      _IsVector,
-                      /*is_parallel=*/::std::true_type);
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Compare>
 void
 __pattern_nth_element(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator,
@@ -1314,20 +920,10 @@ __pattern_nth_element(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAcce
 template <typename _Tp, typename _ExecutionPolicy, typename = void>
 struct __brick_fill;
 
-template <class _ExecutionPolicy, class _ForwardIterator, class _Tp, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>
-__pattern_fill(_ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, const _Tp&,
-               /*is_parallel=*/::std::false_type, _IsVector) noexcept;
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Tp>
 void
 __pattern_fill(_Tag, _ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, const _Tp&) noexcept;
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _Tp, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator>
-__pattern_fill(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, const _Tp&,
-               /*is_parallel=*/::std::true_type, _IsVector);
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Tp>
 _RandomAccessIterator
 __pattern_fill(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, const _Tp&);
@@ -1335,20 +931,10 @@ __pattern_fill(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessItera
 template <typename _Tp, typename _ExecutionPolicy, typename = void>
 struct __brick_fill_n;
 
-template <class _ExecutionPolicy, class _OutputIterator, class _Size, class _Tp, class _IsVector>
-_OutputIterator
-__pattern_fill_n(_ExecutionPolicy&&, _OutputIterator, _Size, const _Tp&,
-                 /*is_parallel=*/::std::false_type, _IsVector) noexcept;
-
 template <class _Tag, class _ExecutionPolicy, class _OutputIterator, class _Size, class _Tp>
 _OutputIterator
 __pattern_fill_n(_Tag, _ExecutionPolicy&&, _OutputIterator, _Size, const _Tp&) noexcept;
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _Size, class _Tp, class _IsVector>
-_RandomAccessIterator
-__pattern_fill_n(_ExecutionPolicy&&, _RandomAccessIterator, _Size, const _Tp&,
-                 /*is_parallel=*/::std::true_type, _IsVector);
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Size, class _Tp>
 _RandomAccessIterator
 __pattern_fill_n(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator, _Size, const _Tp&);
@@ -1365,20 +951,10 @@ template <class _ForwardIterator, class _Generator>
 void __brick_generate(_ForwardIterator, _ForwardIterator, _Generator,
                       /* is_vector = */ ::std::false_type) noexcept;
 
-template <class _ExecutionPolicy, class _ForwardIterator, class _Generator, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>
-__pattern_generate(_ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _Generator,
-                   /*is_parallel=*/::std::false_type, _IsVector) noexcept;
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Generator>
 void
 __pattern_generate(_Tag, _ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _Generator) noexcept;
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _Generator, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator>
-__pattern_generate(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _Generator,
-                   /*is_parallel=*/::std::true_type, _IsVector);
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Generator>
 _RandomAccessIterator
 __pattern_generate(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator,
@@ -1392,20 +968,10 @@ template <class OutputIterator, class Size, class _Generator>
 OutputIterator __brick_generate_n(OutputIterator, Size, _Generator,
                                   /* is_vector = */ ::std::false_type) noexcept;
 
-template <class _ExecutionPolicy, class OutputIterator, class Size, class _Generator, class _IsVector>
-OutputIterator
-__pattern_generate_n(_ExecutionPolicy&&, OutputIterator, Size, _Generator,
-                     /*is_parallel=*/::std::false_type, _IsVector) noexcept;
-
 template <class _Tag, class _ExecutionPolicy, class _OutputIterator, class _Size, class _Generator>
 _OutputIterator
 __pattern_generate_n(_Tag, _ExecutionPolicy&&, _OutputIterator, _Size, _Generator) noexcept;
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class Size, class _Generator, class _IsVector>
-_RandomAccessIterator
-__pattern_generate_n(_ExecutionPolicy&&, _RandomAccessIterator, Size, _Generator,
-                     /*is_parallel=*/::std::true_type, _IsVector);
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Size, class _Generator>
 _RandomAccessIterator
 __pattern_generate_n(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator, _Size, _Generator);
@@ -1421,20 +987,10 @@ template <class _RandomAccessIterator, class _UnaryPredicate>
 _RandomAccessIterator __brick_remove_if(_RandomAccessIterator, _RandomAccessIterator, _UnaryPredicate,
                                         /* __is_vector = */ ::std::true_type) noexcept;
 
-template <class _ExecutionPolicy, class _ForwardIterator, class _UnaryPredicate, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _ForwardIterator>
-__pattern_remove_if(_ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _UnaryPredicate, _IsVector,
-                    /*is_parallel*/ ::std::false_type) noexcept;
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _UnaryPredicate>
 _ForwardIterator
 __pattern_remove_if(_Tag, _ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _UnaryPredicate) noexcept;
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _UnaryPredicate, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator>
-__pattern_remove_if(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _UnaryPredicate, _IsVector,
-                    /*is_parallel*/ ::std::true_type);
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _UnaryPredicate>
 _RandomAccessIterator
 __pattern_remove_if(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator,
@@ -1454,25 +1010,12 @@ _OutputIterator __brick_merge(_RandomAccessIterator1, _RandomAccessIterator1, _R
                               _RandomAccessIterator2, _OutputIterator, _Compare,
                               /* __is_vector = */ ::std::true_type) noexcept;
 
-template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _OutputIterator,
-          class _Compare, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _OutputIterator>
-__pattern_merge(_ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2, _ForwardIterator2,
-                _OutputIterator, _Compare, _IsVector, /* is_parallel = */ ::std::false_type) noexcept;
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _OutputIterator,
           class _Compare>
 _OutputIterator
 __pattern_merge(_Tag, _ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2, _ForwardIterator2,
                 _OutputIterator, _Compare) noexcept;
 
-template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _OutputIterator,
-          class _Compare, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _OutputIterator>
-__pattern_merge(_ExecutionPolicy&&, _RandomAccessIterator1, _RandomAccessIterator1, _RandomAccessIterator2,
-                _RandomAccessIterator2, _OutputIterator, _Compare, _IsVector,
-                /* is_parallel = */ ::std::true_type);
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
           class _RandomAccessIterator3, class _Compare>
 _RandomAccessIterator3
@@ -1491,23 +1034,11 @@ template <class _RandomAccessIterator, class _Compare>
 void __brick_inplace_merge(_RandomAccessIterator, _RandomAccessIterator, _RandomAccessIterator, _Compare,
                            /* __is_vector = */ ::std::true_type) noexcept;
 
-template <class _ExecutionPolicy, class _BidirectionalIterator, class _Compare, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>
-__pattern_inplace_merge(_ExecutionPolicy&&, _BidirectionalIterator, _BidirectionalIterator, _BidirectionalIterator,
-                        _Compare, _IsVector,
-                        /* is_parallel = */ ::std::false_type) noexcept;
-
 template <class _Tag, class _ExecutionPolicy, class _BidirectionalIterator, class _Compare>
 void
 __pattern_inplace_merge(_Tag, _ExecutionPolicy&&, _BidirectionalIterator, _BidirectionalIterator,
                         _BidirectionalIterator, _Compare) noexcept;
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _Compare, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>
-__pattern_inplace_merge(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _RandomAccessIterator,
-                        _Compare, _IsVector,
-                        /*is_parallel=*/::std::true_type);
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Compare>
 void
 __pattern_inplace_merge(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator,
@@ -1517,24 +1048,11 @@ __pattern_inplace_merge(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAc
 // includes
 //------------------------------------------------------------------------
 
-template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Compare, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, bool>
-__pattern_includes(_ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2, _ForwardIterator2,
-                   _Compare, _IsVector,
-                   /*is_parallel=*/::std::false_type) noexcept;
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Compare>
 bool
 __pattern_includes(_Tag, _ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2, _ForwardIterator2,
                    _Compare) noexcept;
 
-template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _Compare,
-          class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, bool>
-__pattern_includes(_ExecutionPolicy&&, _RandomAccessIterator1, _RandomAccessIterator1, _RandomAccessIterator2,
-                   _RandomAccessIterator2, _Compare, _IsVector,
-                   /*is_parallel=*/::std::true_type);
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
           class _Compare>
 bool
@@ -1555,24 +1073,12 @@ _OutputIterator __brick_set_union(_RandomAccessIterator1, _RandomAccessIterator1
                                   _RandomAccessIterator2, _OutputIterator, _Compare,
                                   /*__is_vector=*/::std::true_type) noexcept;
 
-template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _OutputIterator,
-          class _Compare, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _OutputIterator>
-__pattern_set_union(_ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2, _ForwardIterator2,
-                    _OutputIterator, _Compare, _IsVector, /*is_parallel=*/::std::false_type) noexcept;
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _OutputIterator,
           class _Compare>
 _OutputIterator
 __pattern_set_union(_Tag, _ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2,
                     _ForwardIterator2, _OutputIterator, _Compare) noexcept;
 
-template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _OutputIterator,
-          class _Compare, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _OutputIterator>
-__pattern_set_union(_ExecutionPolicy&&, _RandomAccessIterator1, _RandomAccessIterator1, _RandomAccessIterator2,
-                    _RandomAccessIterator2, _OutputIterator, _Compare, _IsVector, /*is_parallel=*/::std::true_type);
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
           class _OutputIterator, class _Compare>
 _OutputIterator
@@ -1593,26 +1099,12 @@ _OutputIterator __brick_set_intersection(_RandomAccessIterator1, _RandomAccessIt
                                          _RandomAccessIterator2, _OutputIterator, _Compare,
                                          /*__is_vector=*/::std::true_type) noexcept;
 
-template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _OutputIterator,
-          class _Compare, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _OutputIterator>
-__pattern_set_intersection(_ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2,
-                           _ForwardIterator2, _OutputIterator, _Compare, _IsVector,
-                           /*is_parallel=*/::std::false_type) noexcept;
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _OutputIterator,
           class _Compare>
 _OutputIterator
 __pattern_set_intersection(_Tag, _ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2,
                            _ForwardIterator2, _OutputIterator, _Compare) noexcept;
 
-template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _OutputIterator,
-          class _Compare, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _OutputIterator>
-__pattern_set_intersection(_ExecutionPolicy&&, _RandomAccessIterator1, _RandomAccessIterator1, _RandomAccessIterator2,
-                           _RandomAccessIterator2, _OutputIterator, _Compare, _IsVector,
-                           /*is_parallel=*/::std::true_type);
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
           class _RandomAccessIterator3, class _Compare>
 _RandomAccessIterator3
@@ -1634,25 +1126,12 @@ _OutputIterator __brick_set_difference(_RandomAccessIterator1, _RandomAccessIter
                                        _RandomAccessIterator2, _OutputIterator, _Compare,
                                        /*__is_vector=*/::std::true_type) noexcept;
 
-template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _OutputIterator,
-          class _Compare, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _OutputIterator>
-__pattern_set_difference(_ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2, _ForwardIterator2,
-                         _OutputIterator, _Compare, _IsVector, /*is_parallel=*/::std::false_type) noexcept;
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _OutputIterator,
           class _Compare>
 _OutputIterator
 __pattern_set_difference(_Tag, _ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2,
                          _ForwardIterator2, _OutputIterator, _Compare) noexcept;
 
-template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _OutputIterator,
-          class _Compare, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _OutputIterator>
-__pattern_set_difference(_ExecutionPolicy&&, _RandomAccessIterator1, _RandomAccessIterator1, _RandomAccessIterator2,
-                         _RandomAccessIterator2, _OutputIterator, _Compare, _IsVector,
-                         /*is_parallel=*/::std::true_type);
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
           class _RandomAccessIterator3, class _Compare>
 _RandomAccessIterator3
@@ -1673,26 +1152,12 @@ _OutputIterator __brick_set_symmetric_difference(_RandomAccessIterator1, _Random
                                                  _RandomAccessIterator2, _OutputIterator, _Compare,
                                                  /*__is_vector=*/::std::true_type) noexcept;
 
-template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _OutputIterator,
-          class _Compare, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _OutputIterator>
-__pattern_set_symmetric_difference(_ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2,
-                                   _ForwardIterator2, _OutputIterator, _Compare, _IsVector,
-                                   /*is_parallel=*/::std::false_type) noexcept;
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _OutputIterator,
           class _Compare>
 _OutputIterator
 __pattern_set_symmetric_difference(_Tag, _ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2,
                                    _ForwardIterator2, _OutputIterator, _Compare) noexcept;
 
-template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _OutputIterator,
-          class _Compare, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _OutputIterator>
-__pattern_set_symmetric_difference(_ExecutionPolicy&&, _RandomAccessIterator1, _RandomAccessIterator1,
-                                   _RandomAccessIterator2, _RandomAccessIterator2, _OutputIterator, _Compare, _IsVector,
-                                   /*is_parallel=*/::std::true_type);
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
           class _RandomAccessIterator3, class _Compare>
 _RandomAccessIterator3
@@ -1712,20 +1177,10 @@ template <class _RandomAccessIterator, class _Compare>
 _RandomAccessIterator __brick_is_heap_until(_RandomAccessIterator, _RandomAccessIterator, _Compare,
                                             /* __is_vector = */ ::std::true_type) noexcept;
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _Compare, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator>
-__pattern_is_heap_until(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _Compare, _IsVector,
-                        /* is_parallel = */ ::std::false_type) noexcept;
-
 template <class _Tag, class _ExecutionPolicy, class _RandomAccessIterator, class _Compare>
 _RandomAccessIterator
 __pattern_is_heap_until(_Tag, _ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _Compare) noexcept;
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _Compare, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator>
-__pattern_is_heap_until(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _Compare, _IsVector,
-                        /* is_parallel = */ ::std::true_type);
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Compare>
 _RandomAccessIterator
 __pattern_is_heap_until(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator,
@@ -1743,20 +1198,10 @@ template <class _RandomAccessIterator, class _Compare>
 bool __brick_is_heap(_RandomAccessIterator, _RandomAccessIterator, _Compare,
                      /* __is_vector = */ ::std::true_type) noexcept;
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _Compare, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, bool>
-__pattern_is_heap(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _Compare, _IsVector,
-                  /* is_parallel = */ ::std::false_type) noexcept;
-
 template <class _Tag, class _ExecutionPolicy, class _RandomAccessIterator, class _Compare>
 bool
 __pattern_is_heap(_Tag, _ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _Compare) noexcept;
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _Compare, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, bool>
-__pattern_is_heap(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _Compare, _IsVector,
-                  /* is_parallel = */ ::std::true_type);
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Compare>
 bool
 __pattern_is_heap(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator,
@@ -1774,20 +1219,10 @@ template <typename _RandomAccessIterator, typename _Compare>
 _RandomAccessIterator __brick_min_element(_RandomAccessIterator, _RandomAccessIterator, _Compare,
                                           /* __is_vector = */ ::std::true_type) noexcept;
 
-template <typename _ExecutionPolicy, typename _ForwardIterator, typename _Compare, typename _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _ForwardIterator>
-__pattern_min_element(_ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _Compare, _IsVector,
-                      /* is_parallel = */ ::std::false_type) noexcept;
-
 template <class _Tag, typename _ExecutionPolicy, typename _ForwardIterator, typename _Compare>
 _ForwardIterator
 __pattern_min_element(_Tag, _ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _Compare) noexcept;
 
-template <typename _ExecutionPolicy, typename _RandomAccessIterator, typename _Compare, typename _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator>
-__pattern_min_element(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _Compare, _IsVector,
-                      /* is_parallel = */ ::std::true_type);
-
 template <typename _IsVector, typename _ExecutionPolicy, typename _RandomAccessIterator, typename _Compare>
 _RandomAccessIterator
 __pattern_min_element(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator,
@@ -1806,22 +1241,10 @@ ::std::pair<_RandomAccessIterator, _RandomAccessIterator>
     __brick_minmax_element(_RandomAccessIterator, _RandomAccessIterator, _Compare,
                            /* __is_vector = */ ::std::true_type) noexcept;
 
-template <typename _ExecutionPolicy, typename _ForwardIterator, typename _Compare, typename _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy,
-                                                           ::std::pair<_ForwardIterator, _ForwardIterator>>
-__pattern_minmax_element(_ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _Compare, _IsVector,
-                         /* is_parallel = */ ::std::false_type) noexcept;
-
 template <class _Tag, typename _ExecutionPolicy, typename _ForwardIterator, typename _Compare>
 ::std::pair<_ForwardIterator, _ForwardIterator>
 __pattern_minmax_element(_Tag, _ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _Compare) noexcept;
 
-template <typename _ExecutionPolicy, typename _RandomAccessIterator, typename _Compare, typename _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy,
-                                                           ::std::pair<_RandomAccessIterator, _RandomAccessIterator>>
-__pattern_minmax_element(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _Compare, _IsVector,
-                         /* is_parallel = */ ::std::true_type);
-
 template <typename _IsVector, typename _ExecutionPolicy, typename _RandomAccessIterator, typename _Compare>
 ::std::pair<_RandomAccessIterator, _RandomAccessIterator>
 __pattern_minmax_element(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator,
@@ -1842,25 +1265,11 @@ ::std::pair<_RandomAccessIterator1, _RandomAccessIterator2>
                      _Predicate,
                      /* __is_vector = */ ::std::true_type) noexcept;
 
-template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Predicate, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy,
-                                                           ::std::pair<_ForwardIterator1, _ForwardIterator2>>
-__pattern_mismatch(_ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2, _ForwardIterator2,
-                   _Predicate, _IsVector,
-                   /* is_parallel = */ ::std::false_type) noexcept;
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Predicate>
 ::std::pair<_ForwardIterator1, _ForwardIterator2>
 __pattern_mismatch(_Tag, _ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2, _ForwardIterator2,
                    _Predicate) noexcept;
 
-template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _Predicate,
-          class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy,
-                                                           ::std::pair<_RandomAccessIterator1, _RandomAccessIterator2>>
-__pattern_mismatch(_ExecutionPolicy&&, _RandomAccessIterator1, _RandomAccessIterator1, _RandomAccessIterator2,
-                   _RandomAccessIterator2, _Predicate, _IsVector, /* is_parallel = */ ::std::true_type);
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
           class _Predicate>
 ::std::pair<_RandomAccessIterator1, _RandomAccessIterator2>
@@ -1881,34 +1290,19 @@ bool __brick_lexicographical_compare(_RandomAccessIterator1, _RandomAccessIterat
                                      _RandomAccessIterator2, _Compare,
                                      /* __is_vector = */ ::std::true_type) noexcept;
 
-template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Compare, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, bool>
-__pattern_lexicographical_compare(_ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2,
-                                  _ForwardIterator2, _Compare, _IsVector,
-                                  /* is_parallel = */ ::std::false_type) noexcept;
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Compare>
 bool
 __pattern_lexicographical_compare(_Tag, _ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2,
                                   _ForwardIterator2, _Compare) noexcept;
 
-template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _Compare,
-          class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, bool>
-__pattern_lexicographical_compare(_ExecutionPolicy&&, _RandomAccessIterator1, _RandomAccessIterator1,
-                                  _RandomAccessIterator2, _RandomAccessIterator2, _Compare, _IsVector,
-                                  /* is_parallel = */ ::std::true_type);
-
 template <class _IsVector, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Compare>
 bool
 __pattern_lexicographical_compare(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1,
                                   _ForwardIterator2, _ForwardIterator2, _Compare) noexcept;
 
-template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Function, class _IsVector,
-          class _IsParallel>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _ForwardIterator2>
-__pattern_swap(_ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2, _Function, _IsVector,
-               _IsParallel);
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Function>
+_ForwardIterator2
+__pattern_swap(_Tag, _ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2, _Function);
 
 //------------------------------------------------------------------------
 // shift_left
@@ -1924,34 +1318,16 @@ _ForwardIterator __brick_shift_left(_ForwardIterator, _ForwardIterator,
                                     typename ::std::iterator_traits<_ForwardIterator>::difference_type,
                                     /*__is_vector=*/::std::true_type) noexcept;
 
-template <class _ExecutionPolicy, class _ForwardIterator, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _ForwardIterator>
-__pattern_shift_left(_ExecutionPolicy&&, _ForwardIterator, _ForwardIterator,
-                     typename ::std::iterator_traits<_ForwardIterator>::difference_type, _IsVector,
-                     /*is_parallel=*/::std::false_type) noexcept;
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator>
 _ForwardIterator
 __pattern_shift_left(_Tag, _ExecutionPolicy&&, _ForwardIterator, _ForwardIterator,
                      typename ::std::iterator_traits<_ForwardIterator>::difference_type) noexcept;
 
-template <class _ExecutionPolicy, class _ForwardIterator, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _ForwardIterator>
-__pattern_shift_left(_ExecutionPolicy&&, _ForwardIterator, _ForwardIterator,
-                     typename ::std::iterator_traits<_ForwardIterator>::difference_type, _IsVector,
-                     /*is_parallel=*/::std::true_type);
-
 template <class _IsVector, class _ExecutionPolicy, class _ForwardIterator>
 _ForwardIterator
 __pattern_shift_left(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _ForwardIterator, _ForwardIterator,
                      typename ::std::iterator_traits<_ForwardIterator>::difference_type);
 
-template <class _ExecutionPolicy, class _BidirectionalIterator, class _IsVector, class _IsParallel>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _BidirectionalIterator>
-__pattern_shift_right(_ExecutionPolicy&&, _BidirectionalIterator, _BidirectionalIterator,
-                      typename ::std::iterator_traits<_BidirectionalIterator>::difference_type, _IsVector,
-                      _IsParallel is_parallel);
-
 } // namespace __internal
 } // namespace dpl
 } // namespace oneapi
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 8c4569daf09..11c5d758454 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -63,14 +63,6 @@ __brick_any_of(const _RandomAccessIterator __first, const _RandomAccessIterator
     return __unseq_backend::__simd_or(__first, __last - __first, __pred);
 };
 
-template <class _ExecutionPolicy, class _ForwardIterator, class _Pred, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, bool>
-__pattern_any_of(_ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _Pred __pred,
-                 _IsVector __is_vector, /*parallel=*/::std::false_type) noexcept
-{
-    return __internal::__brick_any_of(__first, __last, __pred, __is_vector);
-}
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Pred>
 bool
 __pattern_any_of(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _Pred __pred) noexcept
@@ -80,22 +72,6 @@ __pattern_any_of(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIte
     return __internal::__brick_any_of(__first, __last, __pred, typename _Tag::__is_vector{});
 }
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _Pred, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, bool>
-__pattern_any_of(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIterator __last, _Pred __pred,
-                 _IsVector __is_vector, /*parallel=*/::std::true_type)
-{
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
-
-    return __internal::__except_handler([&]() {
-        return __internal::__parallel_or(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
-                                         [__pred, __is_vector](_RandomAccessIterator __i, _RandomAccessIterator __j) {
-                                             return __internal::__brick_any_of(__i, __j, __pred, __is_vector);
-                                         });
-    });
-}
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Pred>
 bool
 __pattern_any_of(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
@@ -168,15 +144,6 @@ __pattern_walk1(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIter
     __internal::__brick_walk1(__first, __last, __f, typename _Tag::__is_vector{});
 }
 
-template <class _ExecutionPolicy, class _ForwardIterator, class _Function, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>
-__pattern_walk1(_ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _Function __f,
-                _IsVector __is_vector,
-                /*parallel=*/::std::false_type) noexcept
-{
-    __internal::__brick_walk1(__first, __last, __f, __is_vector);
-}
-
 template <class _ExecutionPolicy, class _ForwardIterator, class _Function>
 void
 __pattern_walk1(__parallel_forward_tag, _ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last,
@@ -221,48 +188,6 @@ __pattern_replace_if(_Tag __tag, _ExecutionPolicy&& __exec, _ForwardIterator __f
             oneapi::dpl::__internal::__ref_or_copy<_ExecutionPolicy, _UnaryPredicate>>(__new_value, __pred));
 }
 
-//template <class _ExecutionPolicy, class _ForwardIterator, class _Function, class _IsVector>
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _Function, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy_conditional<
-    _ExecutionPolicy, __is_random_access_iterator_v<_RandomAccessIterator>>
-__pattern_walk1(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIterator __last, _Function __f,
-                _IsVector __is_vector,
-                /*parallel=*/::std::true_type)
-{
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
-    __internal::__except_handler([&]() {
-        __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
-                                      [__f, __is_vector](_RandomAccessIterator __i, _RandomAccessIterator __j) {
-                                          __internal::__brick_walk1(__i, __j, __f, __is_vector);
-                                      });
-    });
-}
-
-template <class _ExecutionPolicy, class _ForwardIterator, class _Function, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy_conditional<_ExecutionPolicy,
-                                                                       !__is_random_access_iterator_v<_ForwardIterator>>
-__pattern_walk1(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last, _Function __f, _IsVector,
-                /*parallel=*/::std::true_type)
-{
-    typedef typename ::std::iterator_traits<_ForwardIterator>::reference _ReferenceType;
-    auto __func = [&__f](_ReferenceType arg) { __f(arg); };
-    __internal::__except_handler([&]() {
-        __par_backend::__parallel_for_each(::std::forward<_ExecutionPolicy>(__exec), __first, __last, __func);
-    });
-}
-
-template <class _ExecutionPolicy, class _ForwardIterator, class _Brick>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>
-__pattern_walk_brick(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last, _Brick __brick,
-                     /*parallel=*/::std::false_type) noexcept
-{
-    constexpr auto __is_vector = __internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>();
-    __brick(__first, __last, __is_vector);
-}
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Brick>
 void
 __pattern_walk_brick(_Tag, _ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last,
@@ -273,25 +198,6 @@ __pattern_walk_brick(_Tag, _ExecutionPolicy&& __exec, _ForwardIterator __first,
     __brick(__first, __last, typename _Tag::__is_vector{});
 }
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _Brick>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>
-__pattern_walk_brick(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIterator __last,
-                     _Brick __brick,
-                     /*parallel=*/::std::true_type)
-{
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
-    constexpr auto __is_vector = __internal::__is_vectorization_preferred<_ExecutionPolicy, _RandomAccessIterator>();
-    __internal::__except_handler([&]() {
-        __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
-                                      [__brick, __is_vector](_RandomAccessIterator __i, _RandomAccessIterator __j) {
-                                          __brick(__i, __j, __is_vector);
-                                      });
-    });
-}
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Brick>
 void
 __pattern_walk_brick(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
@@ -325,14 +231,6 @@ __brick_walk1_n(_RandomAccessIterator __first, _DifferenceType __n, _Function __
     return __unseq_backend::__simd_walk_1(__first, __n, __f);
 }
 
-template <class _ExecutionPolicy, class _ForwardIterator, class _Size, class _Function, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _ForwardIterator>
-__pattern_walk1_n(_ExecutionPolicy&&, _ForwardIterator __first, _Size __n, _Function __f, _IsVector __is_vector,
-                  /*is_parallel=*/::std::false_type) noexcept
-{
-    return __internal::__brick_walk1_n(__first, __n, __f, __is_vector);
-}
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Size, class _Function>
 _ForwardIterator
 __pattern_walk1_n(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _Size __n, _Function __f) noexcept
@@ -342,20 +240,6 @@ __pattern_walk1_n(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _Size __n,
     return __internal::__brick_walk1_n(__first, __n, __f, typename _Tag::__is_vector{});
 }
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _Size, class _Function, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator>
-__pattern_walk1_n(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _Size __n, _Function __f,
-                  _IsVector __is_vector,
-                  /*is_parallel=*/::std::true_type)
-{
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
-
-    oneapi::dpl::__internal::__pattern_walk1(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first,
-                                             __first + __n, __f);
-    return __first + __n;
-}
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Size, class _Function>
 _RandomAccessIterator
 __pattern_walk1_n(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _RandomAccessIterator __first, _Size __n,
@@ -366,15 +250,6 @@ __pattern_walk1_n(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _R
     return __first + __n;
 }
 
-template <class _ExecutionPolicy, class _ForwardIterator, class _Size, class _Brick>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _ForwardIterator>
-__pattern_walk_brick_n(_ExecutionPolicy&& __exec, _ForwardIterator __first, _Size __n, _Brick __brick,
-                       /*is_parallel=*/::std::false_type) noexcept
-{
-    constexpr auto __is_vector = __internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator>();
-    return __brick(__first, __n, __is_vector);
-}
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Size, class _Brick>
 _ForwardIterator
 __pattern_walk_brick_n(_Tag, _ExecutionPolicy&& __exec, _ForwardIterator __first, _Size __n, _Brick __brick) noexcept
@@ -384,25 +259,6 @@ __pattern_walk_brick_n(_Tag, _ExecutionPolicy&& __exec, _ForwardIterator __first
     return __brick(__first, __n, typename _Tag::__is_vector{});
 }
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _Size, class _Brick>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator>
-__pattern_walk_brick_n(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _Size __n, _Brick __brick,
-                       /*is_parallel=*/::std::true_type)
-{
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
-    constexpr auto __is_vector = __internal::__is_vectorization_preferred<_ExecutionPolicy, _RandomAccessIterator>();
-    return __internal::__except_handler([&]() {
-        __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __first + __n,
-                                      [__brick, __is_vector](_RandomAccessIterator __i, _RandomAccessIterator __j) {
-                                          __brick(__i, __j - __i, __is_vector);
-                                      });
-        return __first + __n;
-    });
-}
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Size, class _Brick>
 _RandomAccessIterator
 __pattern_walk_brick_n(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator __first, _Size __n,
@@ -460,14 +316,6 @@ __brick_walk2_n(_RandomAccessIterator1 __first1, _Size __n, _RandomAccessIterato
     return __unseq_backend::__simd_walk_2(__first1, __n, __first2, __f);
 }
 
-template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Function, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _ForwardIterator2>
-__pattern_walk2(_ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1 __last1, _ForwardIterator2 __first2,
-                _Function __f, _IsVector __is_vector, /*parallel=*/::std::false_type) noexcept
-{
-    return __internal::__brick_walk2(__first1, __last1, __first2, __f, __is_vector);
-}
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Function>
 _ForwardIterator2
 __pattern_walk2(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
@@ -573,15 +421,6 @@ __pattern_walk2(__parallel_forward_tag, _ExecutionPolicy&& __exec, _ForwardItera
     });
 }
 
-template <class _ExecutionPolicy, class _ForwardIterator1, class _Size, class _ForwardIterator2, class _Function,
-          class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _ForwardIterator2>
-__pattern_walk2_n(_ExecutionPolicy&&, _ForwardIterator1 __first1, _Size __n, _ForwardIterator2 __first2, _Function __f,
-                  _IsVector __is_vector, /*parallel=*/::std::false_type) noexcept
-{
-    return __internal::__brick_walk2_n(__first1, __n, __first2, __f, __is_vector);
-}
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _Size, class _ForwardIterator2,
           class _Function>
 _ForwardIterator2
@@ -593,19 +432,6 @@ __pattern_walk2_n(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _Size __
     return __internal::__brick_walk2_n(__first1, __n, __first2, __f, typename _Tag::__is_vector{});
 }
 
-template <class _ExecutionPolicy, class _RandomAccessIterator1, class _Size, class _RandomAccessIterator2,
-          class _Function, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator2>
-__pattern_walk2_n(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1, _Size __n,
-                  _RandomAccessIterator2 __first2, _Function __f, _IsVector __is_vector, /*parallel=*/::std::true_type)
-{
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2>();
-
-    return __internal::__pattern_walk2(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first1,
-                                       __first1 + __n, __first2, __f);
-}
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _Size,
           class _RandomAccessIterator2, class _Function>
 _RandomAccessIterator2
@@ -616,16 +442,6 @@ __pattern_walk2_n(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _R
                                        __first2, __f);
 }
 
-template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Brick>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _ForwardIterator2>
-__pattern_walk2_brick(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
-                      _ForwardIterator2 __first2, _Brick __brick, /*parallel=*/::std::false_type) noexcept
-{
-    constexpr auto __is_vector =
-        __internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
-    return __brick(__first1, __last1, __first2, __is_vector);
-}
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Brick>
 _ForwardIterator2
 __pattern_walk2_brick(_Tag, _ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
@@ -636,30 +452,6 @@ __pattern_walk2_brick(_Tag, _ExecutionPolicy&& __exec, _ForwardIterator1 __first
     return __brick(__first1, __last1, __first2, typename _Tag::__is_vector{});
 }
 
-template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _Brick>
-oneapi::dpl::__internal::__enable_if_host_execution_policy_conditional<
-    _ExecutionPolicy, __is_random_access_iterator_v<_RandomAccessIterator1, _RandomAccessIterator2>,
-    _RandomAccessIterator2>
-__pattern_walk2_brick(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1, _RandomAccessIterator1 __last1,
-                      _RandomAccessIterator2 __first2, _Brick __brick, /*parallel=*/::std::true_type)
-{
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
-    constexpr auto __is_vector =
-        __internal::__is_vectorization_preferred<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2>();
-
-    return __except_handler([&]() {
-        __par_backend::__parallel_for(
-            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1,
-            [&__is_vector, __first1, __first2, __brick](_RandomAccessIterator1 __i, _RandomAccessIterator1 __j) {
-                __brick(__i, __j, __first2 + (__i - __first1), __is_vector);
-            });
-        return __first2 + (__last1 - __first1);
-    });
-}
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
           class _Brick>
 _RandomAccessIterator2
@@ -737,28 +529,6 @@ __pattern_walk2_brick(__parallel_forward_tag, _ExecutionPolicy&& __exec, _Forwar
     });
 }
 
-template <class _ExecutionPolicy, class _RandomAccessIterator1, class _Size, class _RandomAccessIterator2, class _Brick>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator2>
-__pattern_walk2_brick_n(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1, _Size __n,
-                        _RandomAccessIterator2 __first2, _Brick __brick, /*parallel=*/::std::true_type)
-{
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
-    constexpr auto __is_vector =
-        __internal::__is_vectorization_preferred<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2>();
-
-    return __except_handler([&]() {
-        __par_backend::__parallel_for(
-            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first1, __first1 + __n,
-            [&__is_vector, __first1, __first2, __brick](_RandomAccessIterator1 __i, _RandomAccessIterator1 __j) {
-                __brick(__i, __j - __i, __first2 + (__i - __first1), __is_vector);
-            });
-        return __first2 + __n;
-    });
-}
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _Size,
           class _RandomAccessIterator2, class _Brick>
 _RandomAccessIterator2
@@ -777,16 +547,6 @@ __pattern_walk2_brick_n(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _R
     });
 }
 
-template <class _ExecutionPolicy, class _ForwardIterator1, class _Size, class _ForwardIterator2, class _Brick>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _ForwardIterator2>
-__pattern_walk2_brick_n(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _Size __n, _ForwardIterator2 __first2,
-                        _Brick __brick, /*parallel=*/::std::false_type) noexcept
-{
-    constexpr auto __is_vector =
-        __internal::__is_vectorization_preferred<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
-    return __brick(__first1, __n, __first2, __is_vector);
-}
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _Size, class _ForwardIterator2,
           class _Brick>
 _ForwardIterator2
@@ -821,16 +581,6 @@ __brick_walk3(_RandomAccessIterator1 __first1, _RandomAccessIterator1 __last1, _
     return __unseq_backend::__simd_walk_3(__first1, __last1 - __first1, __first2, __first3, __f);
 }
 
-template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _ForwardIterator3,
-          class _Function, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _ForwardIterator3>
-__pattern_walk3(_ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1 __last1, _ForwardIterator2 __first2,
-                _ForwardIterator3 __first3, _Function __f, _IsVector __is_vector,
-                /*parallel=*/::std::false_type) noexcept
-{
-    return __internal::__brick_walk3(__first1, __last1, __first2, __first3, __f, __is_vector);
-}
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _ForwardIterator3,
           class _Function>
 _ForwardIterator3
@@ -955,17 +705,6 @@ __pattern_walk3(__parallel_forward_tag, _ExecutionPolicy&& __exec, _ForwardItera
 // transform_if
 //------------------------------------------------------------------------
 
-template <typename _ExecutionPolicy, typename _ForwardIterator1, typename _ForwardIterator2, typename _Function,
-          class _IsVector, class _IsParallel>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _ForwardIterator2>
-__pattern_walk2_transform_if(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
-                             _ForwardIterator2 __first2, _Function __func, _IsVector __is_vector,
-                             _IsParallel __is_parallel) noexcept
-{
-    return __pattern_walk2(::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __func, __is_vector,
-                           __is_parallel);
-}
-
 template <class _Tag, typename _ExecutionPolicy, typename _ForwardIterator1, typename _ForwardIterator2,
           typename _Function>
 _ForwardIterator2
@@ -977,17 +716,6 @@ __pattern_walk2_transform_if(_Tag __tag, _ExecutionPolicy&& __exec, _ForwardIter
     return __pattern_walk2(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __func);
 }
 
-template <typename _ExecutionPolicy, typename _ForwardIterator1, typename _ForwardIterator2, typename _ForwardIterator3,
-          typename _Function, class _IsVector, class _IsParallel>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _ForwardIterator3>
-__pattern_walk3_transform_if(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
-                             _ForwardIterator2 __first2, _ForwardIterator3 __first3, _Function __func,
-                             _IsVector __is_vector, _IsParallel __is_parallel) noexcept
-{
-    return __pattern_walk3(::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __first3, __func,
-                           __is_vector, __is_parallel);
-}
-
 template <class _Tag, typename _ExecutionPolicy, typename _ForwardIterator1, typename _ForwardIterator2,
           typename _ForwardIterator3, typename _Function>
 _ForwardIterator3
@@ -1025,16 +753,6 @@ __brick_equal(_RandomAccessIterator1 __first1, _RandomAccessIterator1 __last1, _
                .first == __last1;
 }
 
-template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _BinaryPredicate,
-          class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, bool>
-__pattern_equal(_ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1 __last1, _ForwardIterator2 __first2,
-                _ForwardIterator2 __last2, _BinaryPredicate __p, _IsVector __is_vector, /* is_parallel = */
-                ::std::false_type) noexcept
-{
-    return __internal::__brick_equal(__first1, __last1, __first2, __last2, __p, __is_vector);
-}
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _BinaryPredicate>
 bool
 __pattern_equal(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
@@ -1045,29 +763,6 @@ __pattern_equal(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIt
     return __internal::__brick_equal(__first1, __last1, __first2, __last2, __p, typename _Tag::__is_vector{});
 }
 
-template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _BinaryPredicate,
-          class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, bool>
-__pattern_equal(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1, _RandomAccessIterator1 __last1,
-                _RandomAccessIterator2 __first2, _RandomAccessIterator2 __last2, _BinaryPredicate __p,
-                _IsVector __is_vector, /*is_parallel=*/::std::true_type)
-{
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2>();
-
-    if (__last1 - __first1 != __last2 - __first2)
-        return false;
-
-    return __internal::__except_handler([&]() {
-        return !__internal::__parallel_or(
-            __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1,
-            [__first1, __first2, __p, __is_vector](_RandomAccessIterator1 __i, _RandomAccessIterator1 __j) {
-                return !__internal::__brick_equal(__i, __j, __first2 + (__i - __first1), __first2 + (__j - __first1),
-                                                  __p, __is_vector);
-            });
-    });
-}
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
           class _BinaryPredicate>
 bool
@@ -1109,15 +804,6 @@ __brick_equal(_RandomAccessIterator1 __first1, _RandomAccessIterator1 __last1, _
                .first == __last1;
 }
 
-template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _BinaryPredicate,
-          class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, bool>
-__pattern_equal(_ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1 __last1, _ForwardIterator2 __first2,
-                _BinaryPredicate __p, _IsVector __is_vector, /* is_parallel = */ ::std::false_type) noexcept
-{
-    return __internal::__brick_equal(__first1, __last1, __first2, __p, __is_vector);
-}
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _BinaryPredicate>
 bool
 __pattern_equal(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
@@ -1128,25 +814,6 @@ __pattern_equal(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIt
     return __internal::__brick_equal(__first1, __last1, __first2, __p, typename _Tag::__is_vector{});
 }
 
-template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _BinaryPredicate,
-          class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, bool>
-__pattern_equal(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1, _RandomAccessIterator1 __last1,
-                _RandomAccessIterator2 __first2, _BinaryPredicate __p, _IsVector __is_vector,
-                /*is_parallel=*/::std::true_type)
-{
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2>();
-
-    return __internal::__except_handler([&]() {
-        return !__internal::__parallel_or(
-            __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1,
-            [__first1, __first2, __p, __is_vector](_RandomAccessIterator1 __i, _RandomAccessIterator1 __j) {
-                return !__internal::__brick_equal(__i, __j, __first2 + (__i - __first1), __p, __is_vector);
-            });
-    });
-}
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
           class _BinaryPredicate>
 bool
@@ -1327,16 +994,6 @@ __brick_find_end(_RandomAccessIterator1 __first, _RandomAccessIterator1 __last,
     return __find_subrange(__first, __last, __last, __s_first, __s_last, __pred, false, ::std::true_type());
 }
 
-template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _BinaryPredicate,
-          class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _ForwardIterator1>
-__pattern_find_end(_ExecutionPolicy&&, _ForwardIterator1 __first, _ForwardIterator1 __last, _ForwardIterator2 __s_first,
-                   _ForwardIterator2 __s_last, _BinaryPredicate __pred, _IsVector __is_vector,
-                   /*is_parallel=*/::std::false_type) noexcept
-{
-    return __internal::__brick_find_end(__first, __last, __s_first, __s_last, __pred, __is_vector);
-}
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _BinaryPredicate>
 _ForwardIterator1
 __pattern_find_end(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first, _ForwardIterator1 __last,
@@ -1347,38 +1004,6 @@ __pattern_find_end(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first, _Forward
     return __internal::__brick_find_end(__first, __last, __s_first, __s_last, __pred, typename _Tag::__is_vector{});
 }
 
-template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _BinaryPredicate,
-          class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator1>
-__pattern_find_end(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first, _RandomAccessIterator1 __last,
-                   _RandomAccessIterator2 __s_first, _RandomAccessIterator2 __s_last, _BinaryPredicate __pred,
-                   _IsVector __is_vector, /*is_parallel=*/::std::true_type)
-{
-    // TODO is it correct that we check _RandomAccessIterator2 in __select_backend ?
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2>();
-
-    if (__last - __first == __s_last - __s_first)
-    {
-        const bool __res = __internal::__pattern_equal(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
-                                                       __first, __last, __s_first, __pred);
-        return __res ? __first : __last;
-    }
-    else
-    {
-        return __internal::__except_handler([&]() {
-            return __internal::__parallel_find(
-                __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
-                [__last, __s_first, __s_last, __pred, __is_vector](_RandomAccessIterator1 __i,
-                                                                   _RandomAccessIterator1 __j) {
-                    return __internal::__find_subrange(__i, __j, __last, __s_first, __s_last, __pred, false,
-                                                       __is_vector);
-                },
-                ::std::false_type{});
-        });
-    }
-}
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
           class _BinaryPredicate>
 _RandomAccessIterator1
@@ -1425,16 +1050,6 @@ __brick_find_first_of(_ForwardIterator1 __first, _ForwardIterator1 __last, _Forw
     return __unseq_backend::__simd_find_first_of(__first, __last, __s_first, __s_last, __pred);
 }
 
-template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _BinaryPredicate,
-          class _IsVector>
-_ForwardIterator1
-__pattern_find_first_of(_ExecutionPolicy&&, _ForwardIterator1 __first, _ForwardIterator1 __last,
-                        _ForwardIterator2 __s_first, _ForwardIterator2 __s_last, _BinaryPredicate __pred,
-                        _IsVector __is_vector, /*is_parallel=*/::std::false_type) noexcept
-{
-    return __internal::__brick_find_first_of(__first, __last, __s_first, __s_last, __pred, __is_vector);
-}
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _BinaryPredicate>
 _ForwardIterator1
 __pattern_find_first_of(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first, _ForwardIterator1 __last,
@@ -1446,26 +1061,6 @@ __pattern_find_first_of(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first, _Fo
                                              typename _Tag::__is_vector{});
 }
 
-template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _BinaryPredicate,
-          class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _ForwardIterator1>
-__pattern_find_first_of(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _ForwardIterator1 __last,
-                        _ForwardIterator2 __s_first, _ForwardIterator2 __s_last, _BinaryPredicate __pred,
-                        _IsVector __is_vector, /*is_parallel=*/::std::true_type)
-{
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
-
-    return __internal::__except_handler([&]() {
-        return __internal::__parallel_find(
-            __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
-            [__s_first, __s_last, &__pred, __is_vector](_ForwardIterator1 __i, _ForwardIterator1 __j) {
-                return __internal::__brick_find_first_of(__i, __j, __s_first, __s_last, __pred, __is_vector);
-            },
-            ::std::true_type{});
-    });
-}
-
 template <class _IsVector, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2,
           class _BinaryPredicate>
 _ForwardIterator1
@@ -1504,16 +1099,6 @@ __brick_search(_ForwardIterator1 __first, _ForwardIterator1 __last, _ForwardIter
     return __internal::__find_subrange(__first, __last, __last, __s_first, __s_last, __pred, true, ::std::true_type());
 }
 
-template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _BinaryPredicate,
-          class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _ForwardIterator1>
-__pattern_search(_ExecutionPolicy&&, _ForwardIterator1 __first, _ForwardIterator1 __last, _ForwardIterator2 __s_first,
-                 _ForwardIterator2 __s_last, _BinaryPredicate __pred, _IsVector __is_vector,
-                 /*is_parallel=*/::std::false_type) noexcept
-{
-    return __internal::__brick_search(__first, __last, __s_first, __s_last, __pred, __is_vector);
-}
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _BinaryPredicate>
 _ForwardIterator1
 __pattern_search(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first, _ForwardIterator1 __last,
@@ -1524,39 +1109,6 @@ __pattern_search(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first, _ForwardIt
     return __internal::__brick_search(__first, __last, __s_first, __s_last, __pred, typename _Tag::__is_vector{});
 }
 
-template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _BinaryPredicate,
-          class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator1>
-__pattern_search(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first, _RandomAccessIterator1 __last,
-                 _RandomAccessIterator2 __s_first, _RandomAccessIterator2 __s_last, _BinaryPredicate __pred,
-                 _IsVector __is_vector,
-                 /*is_parallel=*/::std::true_type)
-{
-    // TODO is it correct that we check _RandomAccessIterator2 in __select_backend ?
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2>();
-
-    if (__last - __first == __s_last - __s_first)
-    {
-        const bool __res = __internal::__pattern_equal(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
-                                                       __first, __last, __s_first, __pred);
-        return __res ? __first : __last;
-    }
-    else
-    {
-        return __internal::__except_handler([&]() {
-            return __internal::__parallel_find(
-                __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
-                [__last, __s_first, __s_last, __pred, __is_vector](_RandomAccessIterator1 __i,
-                                                                   _RandomAccessIterator1 __j) {
-                    return __internal::__find_subrange(__i, __j, __last, __s_first, __s_last, __pred, true,
-                                                       __is_vector);
-                },
-                ::std::true_type{});
-        });
-    }
-}
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
           class _BinaryPredicate>
 _RandomAccessIterator1
@@ -1603,16 +1155,6 @@ __brick_search_n(_RandomAccessIterator __first, _RandomAccessIterator __last, _S
     return __internal::__find_subrange(__first, __last, __last, __count, __value, __pred, ::std::true_type());
 }
 
-template <class _ExecutionPolicy, class _ForwardIterator, class _Size, class _Tp, class _BinaryPredicate,
-          class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _ForwardIterator>
-__pattern_search_n(_ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _Size __count,
-                   const _Tp& __value, _BinaryPredicate __pred, _IsVector __is_vector,
-                   /*is_parallel=*/::std::false_type) noexcept
-{
-    return __internal::__brick_search_n(__first, __last, __count, __value, __pred, __is_vector);
-}
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Size, class _Tp, class _BinaryPredicate>
 _ForwardIterator
 __pattern_search_n(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _Size __count,
@@ -1623,53 +1165,22 @@ __pattern_search_n(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardI
     return __internal::__brick_search_n(__first, __last, __count, __value, __pred, typename _Tag::__is_vector{});
 }
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _Size, class _Tp, class _BinaryPredicate,
-          class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator>
-__pattern_search_n(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIterator __last,
-                   _Size __count, const _Tp& __value, _BinaryPredicate __pred, _IsVector __is_vector,
-                   /*is_parallel=*/::std::true_type)
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Size, class _Tp,
+          class _BinaryPredicate>
+_RandomAccessIterator
+__pattern_search_n(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
+                   _RandomAccessIterator __last, _Size __count, const _Tp& __value, _BinaryPredicate __pred)
 {
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
-
     if (static_cast<_Size>(__last - __first) == __count)
     {
-        const bool __result = !__internal::__pattern_any_of(
-            ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
-            [&__value, &__pred](const _Tp& __val) { return !__pred(__val, __value); }, __is_vector,
-            /*is_parallel*/ ::std::true_type());
+        const bool __result =
+            !__internal::__pattern_any_of(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+                                          [&__value, &__pred](const _Tp& __val) { return !__pred(__val, __value); });
         return __result ? __first : __last;
     }
     else
     {
-        return __internal::__except_handler([&__exec, __first, __last, __count, &__value, __pred, __is_vector]() {
-            return __internal::__parallel_find(
-                __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
-                [__last, __count, &__value, __pred, __is_vector](_RandomAccessIterator __i, _RandomAccessIterator __j) {
-                    return __internal::__find_subrange(__i, __j, __last, __count, __value, __pred, __is_vector);
-                },
-                ::std::true_type{});
-        });
-    }
-}
-
-template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Size, class _Tp,
-          class _BinaryPredicate>
-_RandomAccessIterator
-__pattern_search_n(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
-                   _RandomAccessIterator __last, _Size __count, const _Tp& __value, _BinaryPredicate __pred)
-{
-    if (static_cast<_Size>(__last - __first) == __count)
-    {
-        const bool __result =
-            !__internal::__pattern_any_of(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
-                                          [&__value, &__pred](const _Tp& __val) { return !__pred(__val, __value); });
-        return __result ? __first : __last;
-    }
-    else
-    {
-        return __internal::__except_handler([&__exec, __first, __last, __count, &__value, __pred, __tag]() {
+        return __internal::__except_handler([&__exec, __first, __last, __count, &__value, __pred, __tag]() {
             return __internal::__parallel_find(
                 __tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
                 [__last, __count, &__value, __pred](_RandomAccessIterator __i, _RandomAccessIterator __j) {
@@ -1947,14 +1458,6 @@ __brick_partition_by_mask(_RandomAccessIterator1 __first, _RandomAccessIterator1
 #endif
 }
 
-template <class _ExecutionPolicy, class _ForwardIterator, class _OutputIterator, class _UnaryPredicate, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _OutputIterator>
-__pattern_copy_if(_ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _OutputIterator __result,
-                  _UnaryPredicate __pred, _IsVector __is_vector, /*parallel=*/::std::false_type) noexcept
-{
-    return __internal::__brick_copy_if(__first, __last, __result, __pred, __is_vector);
-}
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _OutputIterator, class _UnaryPredicate>
 _OutputIterator
 __pattern_copy_if(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _OutputIterator __result,
@@ -1965,46 +1468,6 @@ __pattern_copy_if(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIt
     return __internal::__brick_copy_if(__first, __last, __result, __pred, typename _Tag::__is_vector{});
 }
 
-template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _UnaryPredicate,
-          class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator2>
-__pattern_copy_if(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first, _RandomAccessIterator1 __last,
-                  _RandomAccessIterator2 __result, _UnaryPredicate __pred, _IsVector __is_vector,
-                  /*parallel=*/::std::true_type)
-{
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
-    typedef typename ::std::iterator_traits<_RandomAccessIterator1>::difference_type _DifferenceType;
-    const _DifferenceType __n = __last - __first;
-    if (_DifferenceType(1) < __n)
-    {
-        __par_backend::__buffer<_ExecutionPolicy, bool> __mask_buf(__n);
-        return __internal::__except_handler([&__exec, __n, __first, __result, __is_vector, __pred, &__mask_buf]() {
-            bool* __mask = __mask_buf.get();
-            _DifferenceType __m{};
-            __par_backend::__parallel_strict_scan(
-                __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __n, _DifferenceType(0),
-                [=](_DifferenceType __i, _DifferenceType __len) { // Reduce
-                    return __internal::__brick_calc_mask_1<_DifferenceType>(__first + __i, __first + (__i + __len),
-                                                                            __mask + __i, __pred, __is_vector)
-                        .first;
-                },
-                ::std::plus<_DifferenceType>(),                                              // Combine
-                [=](_DifferenceType __i, _DifferenceType __len, _DifferenceType __initial) { // Scan
-                    __internal::__brick_copy_by_mask(
-                        __first + __i, __first + (__i + __len), __result + __initial, __mask + __i,
-                        [](_RandomAccessIterator1 __x, _RandomAccessIterator2 __z) { *__z = *__x; }, __is_vector);
-                },
-                [&__m](_DifferenceType __total) { __m = __total; });
-            return __result + __m;
-        });
-    }
-    // trivial sequence - use serial algorithm
-    return __internal::__brick_copy_if(__first, __last, __result, __pred, __is_vector);
-}
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
           class _UnaryPredicate>
 _RandomAccessIterator2
@@ -2061,15 +1524,6 @@ __brick_count(_ForwardIterator __first, _ForwardIterator __last, _Predicate __pr
     return ::std::count_if(__first, __last, __pred);
 }
 
-template <class _ExecutionPolicy, class _ForwardIterator, class _Predicate, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<
-    _ExecutionPolicy, typename ::std::iterator_traits<_ForwardIterator>::difference_type>
-__pattern_count(_ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _Predicate __pred,
-                /* is_parallel */ ::std::false_type, _IsVector __is_vector) noexcept
-{
-    return __internal::__brick_count(__first, __last, __pred, __is_vector);
-}
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Predicate>
 typename ::std::iterator_traits<_ForwardIterator>::difference_type
 __pattern_count(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _Predicate __pred) noexcept
@@ -2079,32 +1533,6 @@ __pattern_count(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIter
     return __internal::__brick_count(__first, __last, __pred, typename _Tag::__is_vector{});
 }
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _Predicate, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<
-    _ExecutionPolicy, typename ::std::iterator_traits<_RandomAccessIterator>::difference_type>
-__pattern_count(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIterator __last,
-                _Predicate __pred,
-                /* is_parallel */ ::std::true_type, _IsVector __is_vector)
-{
-    typedef typename ::std::iterator_traits<_RandomAccessIterator>::difference_type _SizeType;
-
-    //trivial pre-checks
-    if (__first == __last)
-        return _SizeType(0);
-
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
-    return __internal::__except_handler([&]() {
-        return __par_backend::__parallel_reduce(
-            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, _SizeType(0),
-            [__pred, __is_vector](_RandomAccessIterator __begin, _RandomAccessIterator __end, _SizeType __value)
-                -> _SizeType { return __value + __internal::__brick_count(__begin, __end, __pred, __is_vector); },
-            ::std::plus<_SizeType>());
-    });
-}
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Predicate>
 typename ::std::iterator_traits<_RandomAccessIterator>::difference_type
 __pattern_count(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
@@ -2149,14 +1577,6 @@ __brick_unique(_RandomAccessIterator __first, _RandomAccessIterator __last, _Bin
     return ::std::unique(__first, __last, __pred);
 }
 
-template <class _ExecutionPolicy, class _ForwardIterator, class _BinaryPredicate, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _ForwardIterator>
-__pattern_unique(_ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _BinaryPredicate __pred,
-                 _IsVector __is_vector, /*is_parallel=*/::std::false_type) noexcept
-{
-    return __internal::__brick_unique(__first, __last, __pred, __is_vector);
-}
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _BinaryPredicate>
 _ForwardIterator
 __pattern_unique(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
@@ -2251,32 +1671,6 @@ __remove_elements(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _F
     });
 }
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _BinaryPredicate, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator>
-__pattern_unique(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIterator __last,
-                 _BinaryPredicate __pred, _IsVector __is_vector, /*is_parallel=*/::std::true_type)
-{
-    typedef typename ::std::iterator_traits<_RandomAccessIterator>::reference _ReferenceType;
-
-    if (__first == __last)
-    {
-        return __last;
-    }
-    if (__first + 1 == __last || __first + 2 == __last)
-    {
-        // Trivial sequence - use serial algorithm
-        return __internal::__brick_unique(__first, __last, __pred, __is_vector);
-    }
-    return __internal::__remove_elements(
-        ::std::forward<_ExecutionPolicy>(__exec), ++__first, __last,
-        [&__pred, __is_vector](bool* __b, bool* __e, _RandomAccessIterator __it) {
-            __internal::__brick_walk3(
-                __b, __e, __it - 1, __it,
-                [&__pred](bool& __x, _ReferenceType __y, _ReferenceType __z) { __x = !__pred(__y, __z); }, __is_vector);
-        },
-        __is_vector);
-}
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _BinaryPredicate>
 _RandomAccessIterator
 __pattern_unique(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
@@ -2328,15 +1722,6 @@ __brick_unique_copy(_RandomAccessIterator1 __first, _RandomAccessIterator1 __las
 #endif
 }
 
-template <class _ExecutionPolicy, class _ForwardIterator, class _OutputIterator, class _BinaryPredicate,
-          class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _OutputIterator>
-__pattern_unique_copy(_ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _OutputIterator __result,
-                      _BinaryPredicate __pred, _IsVector __is_vector, /*parallel=*/::std::false_type) noexcept
-{
-    return __internal::__brick_unique_copy(__first, __last, __result, __pred, __is_vector);
-}
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _OutputIterator, class _BinaryPredicate>
 _OutputIterator
 __pattern_unique_copy(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
@@ -2369,60 +1754,6 @@ __brick_calc_mask_2(_RandomAccessIterator __first, _RandomAccessIterator __last,
     return __unseq_backend::__simd_calc_mask_2(__first, __last - __first, __mask, __pred);
 }
 
-template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _BinaryPredicate,
-          class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator2>
-__pattern_unique_copy(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first, _RandomAccessIterator1 __last,
-                      _RandomAccessIterator2 __result, _BinaryPredicate __pred, _IsVector __is_vector,
-                      /*parallel=*/::std::true_type)
-{
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
-    typedef typename ::std::iterator_traits<_RandomAccessIterator1>::difference_type _DifferenceType;
-    const _DifferenceType __n = __last - __first;
-    if (_DifferenceType(2) < __n)
-    {
-        __par_backend::__buffer<_ExecutionPolicy, bool> __mask_buf(__n);
-        if (_DifferenceType(2) < __n)
-        {
-            return __internal::__except_handler([&__exec, __n, __first, __result, __pred, __is_vector, &__mask_buf]() {
-                bool* __mask = __mask_buf.get();
-                _DifferenceType __m{};
-                __par_backend::__parallel_strict_scan(
-                    __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __n, _DifferenceType(0),
-                    [=](_DifferenceType __i, _DifferenceType __len) -> _DifferenceType { // Reduce
-                        _DifferenceType __extra = 0;
-                        if (__i == 0)
-                        {
-                            // Special boundary case
-                            __mask[__i] = true;
-                            if (--__len == 0)
-                                return 1;
-                            ++__i;
-                            ++__extra;
-                        }
-                        return __internal::__brick_calc_mask_2<_DifferenceType>(__first + __i, __first + (__i + __len),
-                                                                                __mask + __i, __pred, __is_vector) +
-                               __extra;
-                    },
-                    ::std::plus<_DifferenceType>(),                                              // Combine
-                    [=](_DifferenceType __i, _DifferenceType __len, _DifferenceType __initial) { // Scan
-                        // Phase 2 is same as for __pattern_copy_if
-                        __internal::__brick_copy_by_mask(
-                            __first + __i, __first + (__i + __len), __result + __initial, __mask + __i,
-                            [](_RandomAccessIterator1 __x, _RandomAccessIterator2 __z) { *__z = *__x; }, __is_vector);
-                    },
-                    [&__m](_DifferenceType __total) { __m = __total; });
-                return __result + __m;
-            });
-        }
-    }
-    // trivial sequence - use serial algorithm
-    return __internal::__brick_unique_copy(__first, __last, __result, __pred, __is_vector);
-}
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
           class _BinaryPredicate>
 _RandomAccessIterator2
@@ -2528,15 +1859,6 @@ __brick_reverse(_RandomAccessIterator __first, _RandomAccessIterator __last, _Ra
                                    });
 }
 
-template <class _ExecutionPolicy, class _BidirectionalIterator, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>
-__pattern_reverse(_ExecutionPolicy&&, _BidirectionalIterator __first, _BidirectionalIterator __last,
-                  _IsVector _is_vector,
-                  /*is_parallel=*/::std::false_type) noexcept
-{
-    __internal::__brick_reverse(__first, __last, _is_vector);
-}
-
 template <class _Tag, class _ExecutionPolicy, class _BidirectionalIterator>
 void
 __pattern_reverse(_Tag, _ExecutionPolicy&&, _BidirectionalIterator __first, _BidirectionalIterator __last) noexcept
@@ -2546,22 +1868,6 @@ __pattern_reverse(_Tag, _ExecutionPolicy&&, _BidirectionalIterator __first, _Bid
     __internal::__brick_reverse(__first, __last, typename _Tag::__is_vector{});
 }
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>
-__pattern_reverse(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIterator __last,
-                  _IsVector __is_vector, /*is_parallel=*/::std::true_type)
-{
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
-    __par_backend::__parallel_for(
-        __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __first + (__last - __first) / 2,
-        [__is_vector, __first, __last](_RandomAccessIterator __inner_first, _RandomAccessIterator __inner_last) {
-            __internal::__brick_reverse(__inner_first, __inner_last, __last - (__inner_first - __first), __is_vector);
-        });
-}
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator>
 void
 __pattern_reverse(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
@@ -2600,14 +1906,6 @@ __brick_reverse_copy(_RandomAccessIterator1 __first, _RandomAccessIterator1 __la
                                           __d_first, [](_ReferenceType1 __x, _ReferenceType2 __y) { __y = __x; });
 }
 
-template <class _ExecutionPolicy, class _BidirectionalIterator, class _OutputIterator, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _OutputIterator>
-__pattern_reverse_copy(_ExecutionPolicy&&, _BidirectionalIterator __first, _BidirectionalIterator __last,
-                       _OutputIterator __d_first, _IsVector __is_vector, /*is_parallel=*/::std::false_type) noexcept
-{
-    return __internal::__brick_reverse_copy(__first, __last, __d_first, __is_vector);
-}
-
 template <class _Tag, class _ExecutionPolicy, class _BidirectionalIterator, class _OutputIterator>
 _OutputIterator
 __pattern_reverse_copy(_Tag, _ExecutionPolicy&&, _BidirectionalIterator __first, _BidirectionalIterator __last,
@@ -2618,26 +1916,6 @@ __pattern_reverse_copy(_Tag, _ExecutionPolicy&&, _BidirectionalIterator __first,
     return __internal::__brick_reverse_copy(__first, __last, __d_first, typename _Tag::__is_vector{});
 }
 
-template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator2>
-__pattern_reverse_copy(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first, _RandomAccessIterator1 __last,
-                       _RandomAccessIterator2 __d_first, _IsVector __is_vector, /*is_parallel=*/::std::true_type)
-{
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
-    auto __len = __last - __first;
-    __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
-                                  [__is_vector, __first, __len, __d_first](_RandomAccessIterator1 __inner_first,
-                                                                           _RandomAccessIterator1 __inner_last) {
-                                      __internal::__brick_reverse_copy(__inner_first, __inner_last,
-                                                                       __d_first + (__len - (__inner_last - __first)),
-                                                                       __is_vector);
-                                  });
-    return __d_first + __len;
-}
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2>
 _RandomAccessIterator2
 __pattern_reverse_copy(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator1 __first,
@@ -2707,14 +1985,6 @@ __brick_rotate(_RandomAccessIterator __first, _RandomAccessIterator __middle, _R
     return __ret;
 }
 
-template <class _ExecutionPolicy, class _ForwardIterator, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _ForwardIterator>
-__pattern_rotate(_ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __middle, _ForwardIterator __last,
-                 _IsVector __is_vector, /*is_parallel=*/::std::false_type) noexcept
-{
-    return __internal::__brick_rotate(__first, __middle, __last, __is_vector);
-}
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator>
 _ForwardIterator
 __pattern_rotate(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __middle,
@@ -2725,73 +1995,6 @@ __pattern_rotate(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIte
     return __internal::__brick_rotate(__first, __middle, __last, typename _Tag::__is_vector{});
 }
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator>
-__pattern_rotate(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIterator __middle,
-                 _RandomAccessIterator __last, _IsVector __is_vector, /*is_parallel=*/::std::true_type)
-{
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
-    typedef typename ::std::iterator_traits<_RandomAccessIterator>::value_type _Tp;
-    auto __n = __last - __first;
-    auto __m = __middle - __first;
-    if (__m <= __n / 2)
-    {
-        __par_backend::__buffer<_ExecutionPolicy, _Tp> __buf(__n - __m);
-        return __internal::__except_handler([&__exec, __n, __m, __first, __middle, __last, __is_vector, &__buf]() {
-            _Tp* __result = __buf.get();
-            __par_backend::__parallel_for(
-                __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __middle, __last,
-                [__middle, __result, __is_vector](_RandomAccessIterator __b, _RandomAccessIterator __e) {
-                    __internal::__brick_uninitialized_move(__b, __e, __result + (__b - __middle), __is_vector);
-                });
-
-            __par_backend::__parallel_for(
-                __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __middle,
-                [__last, __middle, __is_vector](_RandomAccessIterator __b, _RandomAccessIterator __e) {
-                    __internal::__brick_move<_ExecutionPolicy>{}(__b, __e, __b + (__last - __middle), __is_vector);
-                });
-
-            __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __result,
-                                          __result + (__n - __m), [__first, __result, __is_vector](_Tp* __b, _Tp* __e) {
-                                              __brick_move_destroy<_ExecutionPolicy>{}(
-                                                  __b, __e, __first + (__b - __result), __is_vector);
-                                          });
-
-            return __first + (__last - __middle);
-        });
-    }
-    else
-    {
-        __par_backend::__buffer<_ExecutionPolicy, _Tp> __buf(__m);
-        return __internal::__except_handler([&__exec, __n, __m, __first, __middle, __last, __is_vector, &__buf]() {
-            _Tp* __result = __buf.get();
-            __par_backend::__parallel_for(
-                __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __middle,
-                [__first, __result, __is_vector](_RandomAccessIterator __b, _RandomAccessIterator __e) {
-                    __internal::__brick_uninitialized_move(__b, __e, __result + (__b - __first), __is_vector);
-                });
-
-            __par_backend::__parallel_for(
-                __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __middle, __last,
-                [__first, __middle, __is_vector](_RandomAccessIterator __b, _RandomAccessIterator __e) {
-                    __internal::__brick_move<_ExecutionPolicy>{}(__b, __e, __first + (__b - __middle), __is_vector);
-                });
-
-            __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __result,
-                                          __result + __m,
-                                          [__n, __m, __first, __result, __is_vector](_Tp* __b, _Tp* __e) {
-                                              __brick_move_destroy<_ExecutionPolicy>{}(
-                                                  __b, __e, __first + ((__n - __m) + (__b - __result)), __is_vector);
-                                          });
-
-            return __first + (__last - __middle);
-        });
-    }
-}
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator>
 _RandomAccessIterator
 __pattern_rotate(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
@@ -2878,16 +2081,6 @@ __brick_rotate_copy(_ExecutionPolicy&&, _RandomAccessIterator1 __first, _RandomA
     return __internal::__brick_copy<_ExecutionPolicy>{}(__first, __middle, __res, ::std::true_type());
 }
 
-template <class _ExecutionPolicy, class _ForwardIterator, class _OutputIterator, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _OutputIterator>
-__pattern_rotate_copy(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __middle,
-                      _ForwardIterator __last, _OutputIterator __result, _IsVector __is_vector,
-                      /*is_parallel=*/::std::false_type) noexcept
-{
-    return __internal::__brick_rotate_copy(::std::forward<_ExecutionPolicy>(__exec), __first, __middle, __last,
-                                           __result, __is_vector);
-}
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _OutputIterator>
 _OutputIterator
 __pattern_rotate_copy(_Tag, _ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __middle,
@@ -2899,41 +2092,6 @@ __pattern_rotate_copy(_Tag, _ExecutionPolicy&& __exec, _ForwardIterator __first,
                                            __result, typename _Tag::__is_vector{});
 }
 
-template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator2>
-__pattern_rotate_copy(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first, _RandomAccessIterator1 __middle,
-                      _RandomAccessIterator1 __last, _RandomAccessIterator2 __result, _IsVector __is_vector,
-                      /*is_parallel=*/::std::true_type)
-{
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
-    __par_backend::__parallel_for(
-        __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
-        [__first, __last, __middle, __result, __is_vector](_RandomAccessIterator1 __b, _RandomAccessIterator1 __e) {
-            __internal::__brick_copy<_ExecutionPolicy> __copy{};
-            if (__b > __middle)
-            {
-                __copy(__b, __e, __result + (__b - __middle), __is_vector);
-            }
-            else
-            {
-                _RandomAccessIterator2 __new_result = __result + ((__last - __middle) + (__b - __first));
-                if (__e < __middle)
-                {
-                    __copy(__b, __e, __new_result, __is_vector);
-                }
-                else
-                {
-                    __copy(__b, __middle, __new_result, __is_vector);
-                    __copy(__middle, __e, __result, __is_vector);
-                }
-            }
-        });
-    return __result + (__last - __first);
-}
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2>
 _RandomAccessIterator2
 __pattern_rotate_copy(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator1 __first,
@@ -3005,14 +2163,6 @@ __brick_is_partitioned(_RandomAccessIterator __first, _RandomAccessIterator __la
     }
 }
 
-template <class _ExecutionPolicy, class _ForwardIterator, class _UnaryPredicate, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, bool>
-__pattern_is_partitioned(_ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _UnaryPredicate __pred,
-                         _IsVector __is_vector, /*is_parallel=*/::std::false_type) noexcept
-{
-    return __internal::__brick_is_partitioned(__first, __last, __pred, __is_vector);
-}
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _UnaryPredicate>
 bool
 __pattern_is_partitioned(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
@@ -3023,19 +2173,19 @@ __pattern_is_partitioned(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _Fo
     return __internal::__brick_is_partitioned(__first, __last, __pred, typename _Tag::__is_vector{});
 }
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _UnaryPredicate, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, bool>
-__pattern_is_partitioned(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIterator __last,
-                         _UnaryPredicate __pred, _IsVector __is_vector, /*is_parallel=*/::std::true_type)
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _UnaryPredicate>
+bool
+__pattern_is_partitioned(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
+                         _RandomAccessIterator __last, _UnaryPredicate __pred)
 {
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
     //trivial pre-checks
     if (__first == __last)
         return true;
 
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
+    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
     return __internal::__except_handler([&]() {
         // State of current range:
         // broken     - current range is not partitioned by pred
@@ -3070,8 +2220,8 @@ __pattern_is_partitioned(_ExecutionPolicy&& __exec, _RandomAccessIterator __firs
 
         _ReduceType __result = __par_backend::__parallel_reduce(
             __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __identity,
-            [&__pred, __combine, __is_vector](_RandomAccessIterator __i, _RandomAccessIterator __j,
-                                              _ReduceType __value) -> _ReduceType {
+            [&__pred, __combine](_RandomAccessIterator __i, _RandomAccessIterator __j,
+                                 _ReduceType __value) -> _ReduceType {
                 if (__value.__val == __broken)
                     return _ReduceType{__broken, __i};
 
@@ -3081,112 +2231,11 @@ __pattern_is_partitioned(_ExecutionPolicy&& __exec, _RandomAccessIterator __firs
                 {
                     // find first element that don't satisfy pred
                     _RandomAccessIterator __x =
-                        __internal::__brick_find_if(__i + 1, __j, __not_pred<_UnaryPredicate&>(__pred), __is_vector);
+                        __internal::__brick_find_if(__i + 1, __j, __not_pred<_UnaryPredicate&>(__pred), _IsVector{});
                     if (__x != __j)
                     {
                         // find first element after "x" that satisfy pred
-                        _RandomAccessIterator __y = __internal::__brick_find_if(__x + 1, __j, __pred, __is_vector);
-                        // if it was found then range isn't partitioned by pred
-                        if (__y != __j)
-                            return _ReduceType{__broken, __i};
-
-                        __res = _ReduceType{__true_false, __i};
-                    }
-                    else
-                        __res = _ReduceType{__all_true, __i};
-                }
-                else
-                { // if first element doesn't satisfy pred
-                    // then we should find the first element that satisfy pred.
-                    // If we found it then range isn't partitioned by pred
-                    if (__internal::__brick_find_if(__i + 1, __j, __pred, __is_vector) != __j)
-                        return _ReduceType{__broken, __i};
-
-                    __res = _ReduceType{__all_false, __i};
-                }
-                // if we have value from left range then we should calculate the result
-                return (__value.__val == __not_init) ? __res : __combine(__value, __res);
-            },
-
-            [__combine](_ReduceType __val1, _ReduceType __val2) -> _ReduceType {
-                if (__val1.__val == __not_init)
-                    return __val2;
-                if (__val2.__val == __not_init)
-                    return __val1;
-                assert(__val1.__val != __not_init && __val2.__val != __not_init);
-
-                if (__val1.__val == __broken || __val2.__val == __broken)
-                    return _ReduceType{__broken, __val1.__pos};
-                // calculate the result for new big range
-                return __combine(__val1, __val2);
-            });
-        return __result.__val != __broken;
-    });
-}
-
-template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _UnaryPredicate>
-bool
-__pattern_is_partitioned(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
-                         _RandomAccessIterator __last, _UnaryPredicate __pred)
-{
-    //trivial pre-checks
-    if (__first == __last)
-        return true;
-
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
-    return __internal::__except_handler([&]() {
-        // State of current range:
-        // broken     - current range is not partitioned by pred
-        // all_true   - all elements in current range satisfy pred
-        // all_false  - all elements in current range don't satisfy pred
-        // true_false - elements satisfy pred are placed before elements that don't satisfy pred
-        enum _ReduceRes
-        {
-            __not_init = -1,
-            __broken,
-            __all_true,
-            __all_false,
-            __true_false
-        };
-        // Array with states that we'll have when state from the left branch is merged with state from the right branch.
-        // State is calculated by formula: new_state = table[left_state * 4 + right_state]
-        const _ReduceRes __table[] = {__broken,     __broken,     __broken,     __broken, __broken,    __all_true,
-                                      __true_false, __true_false, __broken,     __broken, __all_false, __broken,
-                                      __broken,     __broken,     __true_false, __broken};
-        struct _ReduceType
-        {
-            _ReduceRes __val;
-            _RandomAccessIterator __pos;
-        };
-        //a commutative combiner
-        auto __combine = [&__table](_ReduceType __x, _ReduceType __y) {
-            return __x.__pos > __y.__pos ? _ReduceType{__table[__y.__val * 4 + __x.__val], __y.__pos}
-                                         : _ReduceType{__table[__x.__val * 4 + __y.__val], __x.__pos};
-        };
-
-        const _ReduceType __identity{__not_init, __last};
-
-        _ReduceType __result = __par_backend::__parallel_reduce(
-            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __identity,
-            [&__pred, __combine](_RandomAccessIterator __i, _RandomAccessIterator __j,
-                                 _ReduceType __value) -> _ReduceType {
-                if (__value.__val == __broken)
-                    return _ReduceType{__broken, __i};
-
-                _ReduceType __res{__not_init, __i};
-                // if first element satisfy pred
-                if (__pred(*__i))
-                {
-                    // find first element that don't satisfy pred
-                    _RandomAccessIterator __x =
-                        __internal::__brick_find_if(__i + 1, __j, __not_pred<_UnaryPredicate&>(__pred), _IsVector{});
-                    if (__x != __j)
-                    {
-                        // find first element after "x" that satisfy pred
-                        _RandomAccessIterator __y = __internal::__brick_find_if(__x + 1, __j, __pred, _IsVector{});
+                        _RandomAccessIterator __y = __internal::__brick_find_if(__x + 1, __j, __pred, _IsVector{});
                         // if it was found then range isn't partitioned by pred
                         if (__y != __j)
                             return _ReduceType{__broken, __i};
@@ -3246,14 +2295,6 @@ __brick_partition(_RandomAccessIterator __first, _RandomAccessIterator __last, _
     return ::std::partition(__first, __last, __pred);
 }
 
-template <class _ExecutionPolicy, class _ForwardIterator, class _UnaryPredicate, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _ForwardIterator>
-__pattern_partition(_ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _UnaryPredicate __pred,
-                    _IsVector __is_vector, /*is_parallel=*/::std::false_type) noexcept
-{
-    return __internal::__brick_partition(__first, __last, __pred, __is_vector);
-}
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _UnaryPredicate>
 _ForwardIterator
 __pattern_partition(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
@@ -3264,77 +2305,6 @@ __pattern_partition(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _Forward
     return __internal::__brick_partition(__first, __last, __pred, typename _Tag::__is_vector{});
 }
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _UnaryPredicate, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator>
-__pattern_partition(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIterator __last,
-                    _UnaryPredicate __pred, _IsVector __is_vector, /*is_parallel=*/::std::true_type)
-{
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
-    // partitioned range: elements before pivot satisfy pred (true part),
-    //                    elements after pivot don't satisfy pred (false part)
-    struct _PartitionRange
-    {
-        _RandomAccessIterator __begin;
-        _RandomAccessIterator __pivot;
-        _RandomAccessIterator __end;
-    };
-
-    return __internal::__except_handler([&]() {
-        _PartitionRange __init{__last, __last, __last};
-
-        // lambda for merging two partitioned ranges to one partitioned range
-        auto __reductor = [&__exec, __is_vector](_PartitionRange __val1, _PartitionRange __val2) -> _PartitionRange {
-            auto __size1 = __val1.__end - __val1.__pivot;
-            auto __size2 = __val2.__pivot - __val2.__begin;
-            auto __new_begin = __val2.__begin - (__val1.__end - __val1.__begin);
-
-            // if all elements in left range satisfy pred then we can move new pivot to pivot of right range
-            if (__val1.__end == __val1.__pivot)
-            {
-                return {__new_begin, __val2.__pivot, __val2.__end};
-            }
-            // if true part of right range greater than false part of left range
-            // then we should swap the false part of left range and last part of true part of right range
-            else if (__size2 > __size1)
-            {
-                __par_backend::__parallel_for(
-                    __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __val1.__pivot, __val1.__pivot + __size1,
-                    [__val1, __val2, __size1, __is_vector](_RandomAccessIterator __i, _RandomAccessIterator __j) {
-                        __internal::__brick_swap_ranges(__i, __j, (__val2.__pivot - __size1) + (__i - __val1.__pivot),
-                                                        __is_vector);
-                    });
-                return {__new_begin, __val2.__pivot - __size1, __val2.__end};
-            }
-            // else we should swap the first part of false part of left range and true part of right range
-            else
-            {
-                __par_backend::__parallel_for(
-                    __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __val1.__pivot, __val1.__pivot + __size2,
-                    [__val1, __val2, __is_vector](_RandomAccessIterator __i, _RandomAccessIterator __j) {
-                        __internal::__brick_swap_ranges(__i, __j, __val2.__begin + (__i - __val1.__pivot), __is_vector);
-                    });
-                return {__new_begin, __val1.__pivot + __size2, __val2.__end};
-            }
-        };
-
-        _PartitionRange __result = __par_backend::__parallel_reduce(
-            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __init,
-            [__pred, __is_vector, __reductor](_RandomAccessIterator __i, _RandomAccessIterator __j,
-                                              _PartitionRange __value) -> _PartitionRange {
-                //1. serial partition
-                _RandomAccessIterator __pivot = __internal::__brick_partition(__i, __j, __pred, __is_vector);
-
-                // 2. merging of two ranges (left and right respectively)
-                return __reductor(__value, {__i, __pivot, __j});
-            },
-            __reductor);
-        return __result.__pivot;
-    });
-}
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _UnaryPredicate>
 _RandomAccessIterator
 __pattern_partition(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
@@ -3437,15 +2407,6 @@ __brick_stable_partition(_RandomAccessIterator __first, _RandomAccessIterator __
     return ::std::stable_partition(__first, __last, __pred);
 }
 
-template <class _ExecutionPolicy, class _BidirectionalIterator, class _UnaryPredicate, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _BidirectionalIterator>
-__pattern_stable_partition(_ExecutionPolicy&&, _BidirectionalIterator __first, _BidirectionalIterator __last,
-                           _UnaryPredicate __pred, _IsVector __is_vector,
-                           /*is_parallelization=*/::std::false_type) noexcept
-{
-    return __internal::__brick_stable_partition(__first, __last, __pred, __is_vector);
-}
-
 template <class _Tag, class _ExecutionPolicy, class _BidirectionalIterator, class _UnaryPredicate>
 _BidirectionalIterator
 __pattern_stable_partition(_Tag, _ExecutionPolicy&&, _BidirectionalIterator __first, _BidirectionalIterator __last,
@@ -3456,62 +2417,6 @@ __pattern_stable_partition(_Tag, _ExecutionPolicy&&, _BidirectionalIterator __fi
     return __internal::__brick_stable_partition(__first, __last, __pred, typename _Tag::__is_vector{});
 }
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _UnaryPredicate, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator>
-__pattern_stable_partition(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIterator __last,
-                           _UnaryPredicate __pred, _IsVector __is_vector,
-                           /*is_parallelization=*/::std::true_type)
-{
-    // partitioned range: elements before pivot satisfy pred (true part),
-    //                    elements after pivot don't satisfy pred (false part)
-    struct _PartitionRange
-    {
-        _RandomAccessIterator __begin;
-        _RandomAccessIterator __pivot;
-        _RandomAccessIterator __end;
-    };
-
-    return __internal::__except_handler([&]() {
-        _PartitionRange __init{__last, __last, __last};
-
-        // lambda for merging two partitioned ranges to one partitioned range
-        auto __reductor = [__is_vector](_PartitionRange __val1, _PartitionRange __val2) -> _PartitionRange {
-            auto __size1 = __val1.__end - __val1.__pivot;
-            auto __new_begin = __val2.__begin - (__val1.__end - __val1.__begin);
-
-            // if all elements in left range satisfy pred then we can move new pivot to pivot of right range
-            if (__val1.__end == __val1.__pivot)
-            {
-                return {__new_begin, __val2.__pivot, __val2.__end};
-            }
-            // if true part of right range greater than false part of left range
-            // then we should swap the false part of left range and last part of true part of right range
-            else
-            {
-                __internal::__brick_rotate(__val1.__pivot, __val2.__begin, __val2.__pivot, __is_vector);
-                return {__new_begin, __val2.__pivot - __size1, __val2.__end};
-            }
-        };
-
-        constexpr auto __dispatch_tag =
-            oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
-        using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
-        _PartitionRange __result = __par_backend::__parallel_reduce(
-            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __init,
-            [&__pred, __is_vector, __reductor](_RandomAccessIterator __i, _RandomAccessIterator __j,
-                                               _PartitionRange __value) -> _PartitionRange {
-                //1. serial stable_partition
-                _RandomAccessIterator __pivot = __internal::__brick_stable_partition(__i, __j, __pred, __is_vector);
-
-                // 2. merging of two ranges (left and right respectively)
-                return __reductor(__value, {__i, __pivot, __j});
-            },
-            __reductor);
-        return __result.__pivot;
-    });
-}
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _UnaryPredicate>
 _RandomAccessIterator
 __pattern_stable_partition(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
@@ -3593,17 +2498,6 @@ __brick_partition_copy(_RandomAccessIterator1 __first, _RandomAccessIterator1 __
 #endif
 }
 
-template <class _ExecutionPolicy, class _ForwardIterator, class _OutputIterator1, class _OutputIterator2,
-          class _UnaryPredicate, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy,
-                                                           ::std::pair<_OutputIterator1, _OutputIterator2>>
-__pattern_partition_copy(_ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
-                         _OutputIterator1 __out_true, _OutputIterator2 __out_false, _UnaryPredicate __pred,
-                         _IsVector __is_vector, /*is_parallelization=*/::std::false_type) noexcept
-{
-    return __internal::__brick_partition_copy(__first, __last, __out_true, __out_false, __pred, __is_vector);
-}
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _OutputIterator1, class _OutputIterator2,
           class _UnaryPredicate>
 ::std::pair<_OutputIterator1, _OutputIterator2>
@@ -3616,52 +2510,6 @@ __pattern_partition_copy(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _Fo
                                               typename _Tag::__is_vector{});
 }
 
-template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
-          class _RandomAccessIterator3, class _UnaryPredicate, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy,
-                                                           ::std::pair<_RandomAccessIterator2, _RandomAccessIterator3>>
-__pattern_partition_copy(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first, _RandomAccessIterator1 __last,
-                         _RandomAccessIterator2 __out_true, _RandomAccessIterator3 __out_false, _UnaryPredicate __pred,
-                         _IsVector __is_vector, /*is_parallelization=*/::std::true_type)
-{
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2,
-                                                  _RandomAccessIterator3>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
-    typedef typename ::std::iterator_traits<_RandomAccessIterator1>::difference_type _DifferenceType;
-    typedef ::std::pair<_DifferenceType, _DifferenceType> _ReturnType;
-    const _DifferenceType __n = __last - __first;
-    if (_DifferenceType(1) < __n)
-    {
-        __par_backend::__buffer<_ExecutionPolicy, bool> __mask_buf(__n);
-        return __internal::__except_handler([&__exec, __n, __first, __out_true, __out_false, __is_vector, __pred,
-                                             &__mask_buf]() {
-            bool* __mask = __mask_buf.get();
-            _ReturnType __m{};
-            __par_backend::__parallel_strict_scan(
-                __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __n,
-                ::std::make_pair(_DifferenceType(0), _DifferenceType(0)),
-                [=](_DifferenceType __i, _DifferenceType __len) { // Reduce
-                    return __internal::__brick_calc_mask_1<_DifferenceType>(__first + __i, __first + (__i + __len),
-                                                                            __mask + __i, __pred, __is_vector);
-                },
-                [](const _ReturnType& __x, const _ReturnType& __y) -> _ReturnType {
-                    return ::std::make_pair(__x.first + __y.first, __x.second + __y.second);
-                },                                                                       // Combine
-                [=](_DifferenceType __i, _DifferenceType __len, _ReturnType __initial) { // Scan
-                    __internal::__brick_partition_by_mask(__first + __i, __first + (__i + __len),
-                                                          __out_true + __initial.first, __out_false + __initial.second,
-                                                          __mask + __i, __is_vector);
-                },
-                [&__m](_ReturnType __total) { __m = __total; });
-            return ::std::make_pair(__out_true + __m.first, __out_false + __m.second);
-        });
-    }
-    // trivial sequence - use serial algorithm
-    return __internal::__brick_partition_copy(__first, __last, __out_true, __out_false, __pred, __is_vector);
-}
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
           class _RandomAccessIterator3, class _UnaryPredicate>
 ::std::pair<_RandomAccessIterator2, _RandomAccessIterator3>
@@ -3707,15 +2555,6 @@ __pattern_partition_copy(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _
 // sort
 //------------------------------------------------------------------------
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _Compare, class _IsVector,
-          class _IsMoveConstructible>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>
-__pattern_sort(_ExecutionPolicy&&, _RandomAccessIterator __first, _RandomAccessIterator __last, _Compare __comp,
-               _IsVector /*is_vector*/, /*is_parallel=*/::std::false_type, _IsMoveConstructible) noexcept
-{
-    ::std::sort(__first, __last, __comp);
-}
-
 template <class _Tag, class _ExecutionPolicy, class _RandomAccessIterator, class _Compare, class _IsMoveConstructible>
 void
 __pattern_sort(_Tag, _ExecutionPolicy&&, _RandomAccessIterator __first, _RandomAccessIterator __last, _Compare __comp,
@@ -3726,25 +2565,6 @@ __pattern_sort(_Tag, _ExecutionPolicy&&, _RandomAccessIterator __first, _RandomA
     ::std::sort(__first, __last, __comp);
 }
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _Compare, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>
-__pattern_sort(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIterator __last, _Compare __comp,
-               _IsVector /*is_vector*/, /*is_parallel=*/::std::true_type, /*is_move_constructible=*/::std::true_type)
-{
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
-    __internal::__except_handler([&]() {
-        __par_backend::__parallel_stable_sort(
-            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __comp,
-            [](_RandomAccessIterator __first, _RandomAccessIterator __last, _Compare __comp) {
-                ::std::sort(__first, __last, __comp);
-            },
-            __last - __first);
-    });
-}
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Compare>
 void
 __pattern_sort(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
@@ -3767,14 +2587,6 @@ __pattern_sort(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAcce
 // stable_sort
 //------------------------------------------------------------------------
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _Compare, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>
-__pattern_stable_sort(_ExecutionPolicy&&, _RandomAccessIterator __first, _RandomAccessIterator __last, _Compare __comp,
-                      _IsVector /*is_vector*/, /*is_parallel=*/::std::false_type) noexcept
-{
-    ::std::stable_sort(__first, __last, __comp);
-}
-
 template <class _Tag, class _ExecutionPolicy, class _RandomAccessIterator, class _Compare>
 void
 __pattern_stable_sort(_Tag, _ExecutionPolicy&&, _RandomAccessIterator __first, _RandomAccessIterator __last,
@@ -3785,25 +2597,6 @@ __pattern_stable_sort(_Tag, _ExecutionPolicy&&, _RandomAccessIterator __first, _
     ::std::stable_sort(__first, __last, __comp);
 }
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _Compare, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>
-__pattern_stable_sort(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIterator __last,
-                      _Compare __comp, _IsVector /*is_vector*/, /*is_parallel=*/::std::true_type)
-{
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
-    __internal::__except_handler([&]() {
-        __par_backend::__parallel_stable_sort(
-            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __comp,
-            [](_RandomAccessIterator __first, _RandomAccessIterator __last, _Compare __comp) {
-                ::std::stable_sort(__first, __last, __comp);
-            },
-            __last - __first);
-    });
-}
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Compare>
 void
 __pattern_stable_sort(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
@@ -3825,22 +2618,6 @@ __pattern_stable_sort(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Ran
 // sort_by_key
 //------------------------------------------------------------------------
 
-template <typename _ExecutionPolicy, typename _RandomAccessIterator1, typename _RandomAccessIterator2,
-          typename _Compare, typename _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>
-__pattern_sort_by_key(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __keys_first,
-                      _RandomAccessIterator1 __keys_last, _RandomAccessIterator2 __values_first, _Compare __comp,
-                      _IsVector /*vector=*/, /*is_parallel=*/::std::false_type) noexcept
-{
-    auto __beg = oneapi::dpl::make_zip_iterator(__keys_first, __values_first);
-    auto __end = __beg + (__keys_last - __keys_first);
-    auto __cmp_f = [__comp](const auto& __a, const auto& __b) {
-        return __comp(::std::get<0>(__a), ::std::get<0>(__b));
-    };
-
-    ::std::sort(__beg, __end, __cmp_f);
-}
-
 template <class _Tag, typename _ExecutionPolicy, typename _RandomAccessIterator1, typename _RandomAccessIterator2,
           typename _Compare>
 void
@@ -3859,35 +2636,6 @@ __pattern_sort_by_key(_Tag, _ExecutionPolicy&& __exec, _RandomAccessIterator1 __
     ::std::sort(__beg, __end, __cmp_f);
 }
 
-template <typename _ExecutionPolicy, typename _RandomAccessIterator1, typename _RandomAccessIterator2,
-          typename _Compare, typename _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>
-__pattern_sort_by_key(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __keys_first,
-                      _RandomAccessIterator1 __keys_last, _RandomAccessIterator2 __values_first, _Compare __comp,
-                      _IsVector /*vector=*/, /*is_parallel=*/::std::true_type)
-{
-    static_assert(
-        ::std::is_move_constructible_v<typename ::std::iterator_traits<_RandomAccessIterator1>::value_type> &&
-            ::std::is_move_constructible_v<typename ::std::iterator_traits<_RandomAccessIterator2>::value_type>,
-        "The keys and values should be move constructible in case of parallel execution.");
-
-    auto __beg = oneapi::dpl::make_zip_iterator(__keys_first, __values_first);
-    auto __end = __beg + (__keys_last - __keys_first);
-    auto __cmp_f = [__comp](const auto& __a, const auto& __b) {
-        return __comp(::std::get<0>(__a), ::std::get<0>(__b));
-    };
-
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
-    __internal::__except_handler([&]() {
-        __par_backend::__parallel_stable_sort(
-            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __beg, __end, __cmp_f,
-            [](auto __first, auto __last, auto __cmp) { ::std::sort(__first, __last, __cmp); }, __end - __beg);
-    });
-}
-
 template <typename _IsVector, typename _ExecutionPolicy, typename _RandomAccessIterator1,
           typename _RandomAccessIterator2, typename _Compare>
 void
@@ -3918,15 +2666,6 @@ __pattern_sort_by_key(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Ran
 // partial_sort
 //------------------------------------------------------------------------
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _Compare, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>
-__pattern_partial_sort(_ExecutionPolicy&&, _RandomAccessIterator __first, _RandomAccessIterator __middle,
-                       _RandomAccessIterator __last, _Compare __comp, _IsVector,
-                       /*is_parallel=*/::std::false_type) noexcept
-{
-    ::std::partial_sort(__first, __middle, __last, __comp);
-}
-
 template <class _Tag, class _ExecutionPolicy, class _RandomAccessIterator, class _Compare>
 void
 __pattern_partial_sort(_Tag, _ExecutionPolicy&&, _RandomAccessIterator __first, _RandomAccessIterator __middle,
@@ -3937,32 +2676,6 @@ __pattern_partial_sort(_Tag, _ExecutionPolicy&&, _RandomAccessIterator __first,
     ::std::partial_sort(__first, __middle, __last, __comp);
 }
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _Compare, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>
-__pattern_partial_sort(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIterator __middle,
-                       _RandomAccessIterator __last, _Compare __comp, _IsVector, /*is_parallel=*/::std::true_type)
-{
-    const auto __n = __middle - __first;
-    if (__n == 0)
-        return;
-
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
-    __except_handler([&]() {
-        __par_backend::__parallel_stable_sort(
-            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __comp,
-            [__n](_RandomAccessIterator __begin, _RandomAccessIterator __end, _Compare __comp) {
-                if (__n < __end - __begin)
-                    ::std::partial_sort(__begin, __begin + __n, __end, __comp);
-                else
-                    ::std::sort(__begin, __end, __comp);
-            },
-            __n);
-    });
-}
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Compare>
 void
 __pattern_partial_sort(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
@@ -3991,15 +2704,6 @@ __pattern_partial_sort(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Ra
 // partial_sort_copy
 //------------------------------------------------------------------------
 
-template <class _ExecutionPolicy, class _ForwardIterator, class _RandomAccessIterator, class _Compare, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator>
-__pattern_partial_sort_copy(_ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
-                            _RandomAccessIterator __d_first, _RandomAccessIterator __d_last, _Compare __comp, _IsVector,
-                            /*is_parallel=*/::std::false_type) noexcept
-{
-    return ::std::partial_sort_copy(__first, __last, __d_first, __d_last, __comp);
-}
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _RandomAccessIterator, class _Compare>
 _RandomAccessIterator
 __pattern_partial_sort_copy(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
@@ -4010,85 +2714,6 @@ __pattern_partial_sort_copy(_Tag, _ExecutionPolicy&&, _ForwardIterator __first,
     return ::std::partial_sort_copy(__first, __last, __d_first, __d_last, __comp);
 }
 
-template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _Compare,
-          class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator2>
-__pattern_partial_sort_copy(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first, _RandomAccessIterator1 __last,
-                            _RandomAccessIterator2 __d_first, _RandomAccessIterator2 __d_last, _Compare __comp,
-                            _IsVector __is_vector, /*is_parallel=*/::std::true_type)
-{
-    if (__last == __first || __d_last == __d_first)
-    {
-        return __d_first;
-    }
-
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
-    auto __n1 = __last - __first;
-    auto __n2 = __d_last - __d_first;
-    return __internal::__except_handler([&]() {
-        if (__n2 >= __n1)
-        {
-            __par_backend::__parallel_stable_sort(
-                __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __d_first, __d_first + __n1, __comp,
-                [__first, __d_first, __is_vector](_RandomAccessIterator2 __i, _RandomAccessIterator2 __j,
-                                                  _Compare __comp) {
-                    _RandomAccessIterator1 __i1 = __first + (__i - __d_first);
-                    _RandomAccessIterator1 __j1 = __first + (__j - __d_first);
-
-                    // 1. Copy elements from input to output
-                    __brick_copy<_ExecutionPolicy>{}(__i1, __j1, __i, __is_vector);
-                    // 2. Sort elements in output sequence
-                    ::std::sort(__i, __j, __comp);
-                },
-                __n1);
-            return __d_first + __n1;
-        }
-        else
-        {
-            typedef typename ::std::iterator_traits<_RandomAccessIterator1>::value_type _T1;
-            typedef typename ::std::iterator_traits<_RandomAccessIterator2>::value_type _T2;
-            __par_backend::__buffer<_ExecutionPolicy, _T1> __buf(__n1);
-            _T1* __r = __buf.get();
-
-            __par_backend::__parallel_stable_sort(
-                __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __r, __r + __n1, __comp,
-                [__n2, __first, __r](_T1* __i, _T1* __j, _Compare __comp) {
-                    _RandomAccessIterator1 __it = __first + (__i - __r);
-
-                    // 1. Copy elements from input to raw memory
-                    for (_T1* __k = __i; __k != __j; ++__k, ++__it)
-                    {
-                        ::new (__k) _T2(*__it);
-                    }
-
-                    // 2. Sort elements in temporary buffer
-                    if (__n2 < __j - __i)
-                        ::std::partial_sort(__i, __i + __n2, __j, __comp);
-                    else
-                        ::std::sort(__i, __j, __comp);
-                },
-                __n2);
-
-            // 3. Move elements from temporary buffer to output
-            __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __r, __r + __n2,
-                                          [__r, __d_first, __is_vector](_T1* __i, _T1* __j) {
-                                              __brick_move_destroy<_ExecutionPolicy>{}(
-                                                  __i, __j, __d_first + (__i - __r), __is_vector);
-                                          });
-
-            if constexpr (!::std::is_trivially_destructible_v<_T1>)
-                __par_backend::__parallel_for(
-                    __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __r + __n2, __r + __n1,
-                    [__is_vector](_T1* __i, _T1* __j) { __brick_destroy(__i, __j, __is_vector); });
-
-            return __d_first + __n2;
-        }
-    });
-}
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
           class _Compare>
 _RandomAccessIterator2
@@ -4183,14 +2808,6 @@ __brick_adjacent_find(_ForwardIterator __first, _ForwardIterator __last, _Binary
     return ::std::adjacent_find(__first, __last, __pred);
 }
 
-template <class _ExecutionPolicy, class _ForwardIterator, class _BinaryPredicate, class _IsVector, class _Semantic>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _ForwardIterator>
-__pattern_adjacent_find(_ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _BinaryPredicate __pred,
-                        /* is_parallel */ ::std::false_type, _IsVector __is_vector, _Semantic) noexcept
-{
-    return __internal::__brick_adjacent_find(__first, __last, __pred, __is_vector, _Semantic::value);
-}
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _BinaryPredicate, class _Semantic>
 _ForwardIterator
 __pattern_adjacent_find(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
@@ -4201,55 +2818,6 @@ __pattern_adjacent_find(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _For
     return __internal::__brick_adjacent_find(__first, __last, __pred, typename _Tag::__is_vector{}, _Semantic::value);
 }
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _BinaryPredicate, class _IsVector, class _Semantic>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator>
-__pattern_adjacent_find(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIterator __last,
-                        _BinaryPredicate __pred, /* is_parallel */ ::std::true_type, _IsVector __is_vector,
-                        _Semantic __or_semantic)
-{
-    if (__last - __first < 2)
-        return __last;
-
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
-    return __internal::__except_handler([&]() {
-        return __par_backend::__parallel_reduce(
-            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __last,
-            [__last, __pred, __is_vector, __or_semantic](_RandomAccessIterator __begin, _RandomAccessIterator __end,
-                                                         _RandomAccessIterator __value) -> _RandomAccessIterator {
-                // TODO: investigate performance benefits from the use of shared variable for the result,
-                // checking (compare_and_swap idiom) its __value at __first.
-                if (__or_semantic && __value < __last)
-                { //found
-                    return __value;
-                }
-
-                if (__value > __begin)
-                {
-                    // modify __end to check the predicate on the boundary __values;
-                    // TODO: to use a custom range with boundaries overlapping
-                    // TODO: investigate what if we remove "if" below and run algorithm on range [__first, __last-1)
-                    // then check the pair [__last-1, __last)
-                    if (__end != __last)
-                        ++__end;
-
-                    //correct the global result iterator if the "brick" returns a local "__last"
-                    const _RandomAccessIterator __res =
-                        __internal::__brick_adjacent_find(__begin, __end, __pred, __is_vector, __or_semantic);
-                    if (__res < __end)
-                        __value = __res;
-                }
-                return __value;
-            },
-            [](_RandomAccessIterator __x, _RandomAccessIterator __y) -> _RandomAccessIterator {
-                return __x < __y ? __x : __y;
-            } //reduce a __value
-        );
-    });
-}
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _BinaryPredicate, class _Semantic>
 _RandomAccessIterator
 __pattern_adjacent_find(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
@@ -4302,15 +2870,6 @@ __pattern_adjacent_find(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _R
 // nth_element
 //------------------------------------------------------------------------
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _Compare, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>
-__pattern_nth_element(_ExecutionPolicy&&, _RandomAccessIterator __first, _RandomAccessIterator __nth,
-                      _RandomAccessIterator __last, _Compare __comp, _IsVector,
-                      /*is_parallel=*/::std::false_type) noexcept
-{
-    ::std::nth_element(__first, __nth, __last, __comp);
-}
-
 template <class _Tag, class _ExecutionPolicy, class _RandomAccessIterator, class _Compare>
 void
 __pattern_nth_element(_Tag, _ExecutionPolicy&&, _RandomAccessIterator __first, _RandomAccessIterator __nth,
@@ -4321,52 +2880,6 @@ __pattern_nth_element(_Tag, _ExecutionPolicy&&, _RandomAccessIterator __first, _
     ::std::nth_element(__first, __nth, __last, __comp);
 }
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _Compare, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>
-__pattern_nth_element(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIterator __nth,
-                      _RandomAccessIterator __last, _Compare __comp, _IsVector __is_vector,
-                      /*is_parallel=*/::std::true_type)
-{
-    if (__first == __last || __nth == __last)
-    {
-        return;
-    }
-
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
-
-    using ::std::iter_swap;
-    typedef typename ::std::iterator_traits<_RandomAccessIterator>::value_type _Tp;
-    _RandomAccessIterator __x;
-    do
-    {
-        __x = __internal::__pattern_partition(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first + 1,
-                                              __last,
-                                              [&__comp, __first](const _Tp& __x) { return __comp(__x, *__first); });
-        --__x;
-        if (__x != __first)
-        {
-            iter_swap(__first, __x);
-        }
-        // if x > nth then our new range for partition is [first, x)
-        if (__x - __nth > 0)
-        {
-            __last = __x;
-        }
-        // if x < nth then our new range for partition is [x, last)
-        else if (__x - __nth < 0)
-        {
-            // if *x == *nth then we start the new partition at the next index where *x != *nth
-            while (!__comp(*__nth, *__x) && !__comp(*__x, *__nth) && __x - __nth < 0)
-            {
-                ++__x;
-            }
-            iter_swap(__nth, __x);
-            __first = __x;
-        }
-    } while (__x != __nth);
-}
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Compare>
 void
 __pattern_nth_element(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
@@ -4433,41 +2946,13 @@ struct __brick_fill<_Tp, _ExecutionPolicy, oneapi::dpl::__internal::__enable_if_
     }
 };
 
-template <class _ExecutionPolicy, class _ForwardIterator, class _Tp, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>
-__pattern_fill(_ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, const _Tp& __value,
-               /*is_parallel=*/::std::false_type, _IsVector __is_vector) noexcept
-{
-    __internal::__brick_fill<_Tp, _ExecutionPolicy>{__value}(__first, __last, __is_vector);
-}
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Tp>
 void
-__pattern_fill(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, const _Tp& __value) noexcept
-{
-    static_assert(__is_backend_tag_v<_Tag>);
-
-    __internal::__brick_fill<_Tp, _ExecutionPolicy>{__value}(__first, __last, typename _Tag::__is_vector{});
-}
-
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _Tp, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator>
-__pattern_fill(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIterator __last,
-               const _Tp& __value,
-               /*is_parallel=*/::std::true_type, _IsVector __is_vector)
-{
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
-    return __internal::__except_handler([&__exec, __first, __last, &__value, __is_vector]() {
-        __par_backend::__parallel_for(
-            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
-            [&__value, __is_vector](_RandomAccessIterator __begin, _RandomAccessIterator __end) {
-                __internal::__brick_fill<_Tp, _ExecutionPolicy>{__value}(__begin, __end, __is_vector);
-            });
-        return __last;
-    });
+__pattern_fill(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, const _Tp& __value) noexcept
+{
+    static_assert(__is_backend_tag_v<_Tag>);
+
+    __internal::__brick_fill<_Tp, _ExecutionPolicy>{__value}(__first, __last, typename _Tag::__is_vector{});
 }
 
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Tp>
@@ -4510,14 +2995,6 @@ struct __brick_fill_n<_Tp, _ExecutionPolicy,
     }
 };
 
-template <class _ExecutionPolicy, class _OutputIterator, class _Size, class _Tp, class _IsVector>
-_OutputIterator
-__pattern_fill_n(_ExecutionPolicy&&, _OutputIterator __first, _Size __count, const _Tp& __value,
-                 /*is_parallel=*/::std::false_type, _IsVector __is_vector) noexcept
-{
-    return __internal::__brick_fill_n<_Tp, _ExecutionPolicy>{__value}(__first, __count, __is_vector);
-}
-
 template <class _Tag, class _ExecutionPolicy, class _OutputIterator, class _Size, class _Tp>
 _OutputIterator
 __pattern_fill_n(_Tag, _ExecutionPolicy&&, _OutputIterator __first, _Size __count, const _Tp& __value) noexcept
@@ -4527,18 +3004,6 @@ __pattern_fill_n(_Tag, _ExecutionPolicy&&, _OutputIterator __first, _Size __coun
     return __internal::__brick_fill_n<_Tp, _ExecutionPolicy>{__value}(__first, __count, typename _Tag::__is_vector{});
 }
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _Size, class _Tp, class _IsVector>
-_RandomAccessIterator
-__pattern_fill_n(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _Size __count, const _Tp& __value,
-                 /*is_parallel=*/::std::true_type, _IsVector /*__is_vector*/)
-{
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
-
-    return __internal::__pattern_fill(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first,
-                                      __first + __count, __value);
-}
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Size, class _Tp>
 _RandomAccessIterator
 __pattern_fill_n(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
@@ -4567,14 +3032,6 @@ __brick_generate(_ForwardIterator __first, _ForwardIterator __last, _Generator _
     ::std::generate(__first, __last, __g);
 }
 
-template <class _ExecutionPolicy, class _ForwardIterator, class _Generator, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>
-__pattern_generate(_ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _Generator __g,
-                   /*is_parallel=*/::std::false_type, _IsVector __is_vector) noexcept
-{
-    __internal::__brick_generate(__first, __last, __g, __is_vector);
-}
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Generator>
 void
 __pattern_generate(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _Generator __g) noexcept
@@ -4584,25 +3041,6 @@ __pattern_generate(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardI
     __internal::__brick_generate(__first, __last, __g, typename _Tag::__is_vector{});
 }
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _Generator, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator>
-__pattern_generate(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIterator __last,
-                   _Generator __g,
-                   /*is_parallel=*/::std::true_type, _IsVector __is_vector)
-{
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
-    return __internal::__except_handler([&]() {
-        __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
-                                      [__g, __is_vector](_RandomAccessIterator __begin, _RandomAccessIterator __end) {
-                                          __internal::__brick_generate(__begin, __end, __g, __is_vector);
-                                      });
-        return __last;
-    });
-}
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Generator>
 _RandomAccessIterator
 __pattern_generate(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
@@ -4634,14 +3072,6 @@ __brick_generate_n(OutputIterator __first, Size __count, _Generator __g, /* is_v
     return ::std::generate_n(__first, __count, __g);
 }
 
-template <class _ExecutionPolicy, class _OutputIterator, class _Size, class _Generator, class _IsVector>
-_OutputIterator
-__pattern_generate_n(_ExecutionPolicy&&, _OutputIterator __first, _Size __count, _Generator __g,
-                     /*is_parallel=*/::std::false_type, _IsVector __is_vector) noexcept
-{
-    return __internal::__brick_generate_n(__first, __count, __g, __is_vector);
-}
-
 template <class _Tag, class _ExecutionPolicy, class _OutputIterator, class _Size, class _Generator>
 _OutputIterator
 __pattern_generate_n(_Tag, _ExecutionPolicy&&, _OutputIterator __first, _Size __count, _Generator __g) noexcept
@@ -4651,17 +3081,6 @@ __pattern_generate_n(_Tag, _ExecutionPolicy&&, _OutputIterator __first, _Size __
     return __internal::__brick_generate_n(__first, __count, __g, typename _Tag::__is_vector{});
 }
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _Size, class _Generator, class _IsVector>
-_RandomAccessIterator
-__pattern_generate_n(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _Size __count, _Generator __g,
-                     /*is_parallel=*/::std::true_type, _IsVector __is_vector)
-{
-    static_assert(__is_random_access_iterator_v<_RandomAccessIterator>,
-                  "Pattern-brick error. Should be a random access iterator.");
-    return __internal::__pattern_generate(::std::forward<_ExecutionPolicy>(__exec), __first, __first + __count, __g,
-                                          ::std::true_type(), __is_vector);
-}
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Size, class _Generator>
 _RandomAccessIterator
 __pattern_generate_n(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
@@ -4697,14 +3116,6 @@ __brick_remove_if(_RandomAccessIterator __first, _RandomAccessIterator __last, _
 #endif
 }
 
-template <class _ExecutionPolicy, class _ForwardIterator, class _UnaryPredicate, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _ForwardIterator>
-__pattern_remove_if(_ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _UnaryPredicate __pred,
-                    _IsVector __is_vector, /*is_parallel*/ ::std::false_type) noexcept
-{
-    return __internal::__brick_remove_if(__first, __last, __pred, __is_vector);
-}
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _UnaryPredicate>
 _ForwardIterator
 __pattern_remove_if(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
@@ -4715,28 +3126,6 @@ __pattern_remove_if(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _Forward
     return __internal::__brick_remove_if(__first, __last, __pred, typename _Tag::__is_vector{});
 }
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _UnaryPredicate, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator>
-__pattern_remove_if(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIterator __last,
-                    _UnaryPredicate __pred, _IsVector __is_vector, /*is_parallel*/ ::std::true_type)
-{
-    typedef typename ::std::iterator_traits<_RandomAccessIterator>::reference _ReferenceType;
-
-    if (__first == __last || __first + 1 == __last)
-    {
-        // Trivial sequence - use serial algorithm
-        return __internal::__brick_remove_if(__first, __last, __pred, __is_vector);
-    }
-
-    return __internal::__remove_elements(
-        ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
-        [&__pred, __is_vector](bool* __b, bool* __e, _RandomAccessIterator __it) {
-            __internal::__brick_walk2(
-                __b, __e, __it, [&__pred](bool& __x, _ReferenceType __y) { __x = !__pred(__y); }, __is_vector);
-        },
-        __is_vector);
-}
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _UnaryPredicate>
 _RandomAccessIterator
 __pattern_remove_if(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
@@ -4783,16 +3172,6 @@ __brick_merge(_RandomAccessIterator1 __first1, _RandomAccessIterator1 __last1, _
     return ::std::merge(__first1, __last1, __first2, __last2, __d_first, __comp);
 }
 
-template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _OutputIterator,
-          class _Compare, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _OutputIterator>
-__pattern_merge(_ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1 __last1, _ForwardIterator2 __first2,
-                _ForwardIterator2 __last2, _OutputIterator __d_first, _Compare __comp, _IsVector __is_vector,
-                /* is_parallel = */ ::std::false_type) noexcept
-{
-    return __internal::__brick_merge(__first1, __last1, __first2, __last2, __d_first, __comp, __is_vector);
-}
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _OutputIterator,
           class _Compare>
 _OutputIterator
@@ -4806,28 +3185,6 @@ __pattern_merge(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIt
                                      typename _Tag::__is_vector{});
 }
 
-template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
-          class _RandomAccessIterator3, class _Compare, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator3>
-__pattern_merge(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1, _RandomAccessIterator1 __last1,
-                _RandomAccessIterator2 __first2, _RandomAccessIterator2 __last2, _RandomAccessIterator3 __d_first,
-                _Compare __comp, _IsVector __is_vector, /* is_parallel = */ ::std::true_type)
-{
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2,
-                                                  _RandomAccessIterator3>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
-    __par_backend::__parallel_merge(
-        __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2, __d_first,
-        __comp,
-        [__is_vector](_RandomAccessIterator1 __f1, _RandomAccessIterator1 __l1, _RandomAccessIterator2 __f2,
-                      _RandomAccessIterator2 __l2, _RandomAccessIterator3 __f3, _Compare __comp) {
-            return __internal::__brick_merge(__f1, __l1, __f2, __l2, __f3, __comp, __is_vector);
-        });
-    return __d_first + (__last1 - __first1) + (__last2 - __first2);
-}
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
           class _RandomAccessIterator3, class _Compare>
 _RandomAccessIterator3
@@ -4866,15 +3223,6 @@ __brick_inplace_merge(_RandomAccessIterator __first, _RandomAccessIterator __mid
     ::std::inplace_merge(__first, __middle, __last, __comp);
 }
 
-template <class _ExecutionPolicy, class _BidirectionalIterator, class _Compare, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>
-__pattern_inplace_merge(_ExecutionPolicy&&, _BidirectionalIterator __first, _BidirectionalIterator __middle,
-                        _BidirectionalIterator __last, _Compare __comp, _IsVector __is_vector,
-                        /* is_parallel = */ ::std::false_type) noexcept
-{
-    __internal::__brick_inplace_merge(__first, __middle, __last, __comp, __is_vector);
-}
-
 template <class _Tag, class _ExecutionPolicy, class _BidirectionalIterator, class _Compare>
 void
 __pattern_inplace_merge(_Tag, _ExecutionPolicy&&, _BidirectionalIterator __first, _BidirectionalIterator __middle,
@@ -4885,54 +3233,6 @@ __pattern_inplace_merge(_Tag, _ExecutionPolicy&&, _BidirectionalIterator __first
     __internal::__brick_inplace_merge(__first, __middle, __last, __comp, typename _Tag::__is_vector{});
 }
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _Compare, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>
-__pattern_inplace_merge(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIterator __middle,
-                        _RandomAccessIterator __last, _Compare __comp, _IsVector __is_vector,
-                        /*is_parallel=*/::std::true_type)
-{
-    if (__first == __last || __first == __middle || __middle == __last)
-    {
-        return;
-    }
-
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
-    typedef typename ::std::iterator_traits<_RandomAccessIterator>::value_type _Tp;
-    auto __n = __last - __first;
-    __par_backend::__buffer<_ExecutionPolicy, _Tp> __buf(__n);
-    _Tp* __r = __buf.get();
-    __internal::__except_handler([&]() {
-        auto __move_values = [](_RandomAccessIterator __x, _Tp* __z) {
-            if constexpr (::std::is_trivial_v<_Tp>)
-                *__z = ::std::move(*__x);
-            else
-                ::new (::std::addressof(*__z)) _Tp(::std::move(*__x));
-        };
-
-        auto __move_sequences = [](_RandomAccessIterator __first1, _RandomAccessIterator __last1, _Tp* __first2) {
-            return __internal::__brick_uninitialized_move(__first1, __last1, __first2, _IsVector());
-        };
-
-        __par_backend::__parallel_merge(
-            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __middle, __middle, __last, __r, __comp,
-            [__n, __move_values, __move_sequences](_RandomAccessIterator __f1, _RandomAccessIterator __l1,
-                                                   _RandomAccessIterator __f2, _RandomAccessIterator __l2, _Tp* __f3,
-                                                   _Compare __comp) {
-                (__utils::__serial_move_merge(__n))(__f1, __l1, __f2, __l2, __f3, __comp, __move_values, __move_values,
-                                                    __move_sequences, __move_sequences);
-                return __f3 + (__l1 - __f1) + (__l2 - __f2);
-            });
-        __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __r, __r + __n,
-                                      [__r, __first, __is_vector](_Tp* __i, _Tp* __j) {
-                                          __brick_move_destroy<_ExecutionPolicy>{}(__i, __j, __first + (__i - __r),
-                                                                                   __is_vector);
-                                      });
-    });
-}
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Compare>
 void
 __pattern_inplace_merge(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
@@ -4982,15 +3282,6 @@ __pattern_inplace_merge(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _R
 // includes
 //------------------------------------------------------------------------
 
-template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Compare, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, bool>
-__pattern_includes(_ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
-                   _ForwardIterator2 __first2, _ForwardIterator2 __last2, _Compare __comp, _IsVector,
-                   /*is_parallel=*/::std::false_type) noexcept
-{
-    return ::std::includes(__first1, __last1, __first2, __last2, __comp);
-}
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Compare>
 bool
 __pattern_includes(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
@@ -5001,71 +3292,6 @@ __pattern_includes(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _Forwar
     return ::std::includes(__first1, __last1, __first2, __last2, __comp);
 }
 
-template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _Compare,
-          class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, bool>
-__pattern_includes(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1, _RandomAccessIterator1 __last1,
-                   _RandomAccessIterator2 __first2, _RandomAccessIterator2 __last2, _Compare __comp, _IsVector,
-                   /*is_parallel=*/::std::true_type)
-{
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2>();
-
-    if (__first2 == __last2)
-        return true;
-
-    //optimization; {1} - the first sequence, {2} - the second sequence
-    //{1} is empty or size_of{2} > size_of{1}
-    if (__first1 == __last1 || __last2 - __first2 > __last1 - __first1 ||
-        // {1}:     [**********]     or   [**********]
-        // {2}: [***********]                   [***********]
-        __comp(*__first2, *__first1) || __comp(*(__last1 - 1), *(__last2 - 1)))
-        return false;
-
-    __first1 = ::std::lower_bound(__first1, __last1, *__first2, __comp);
-    if (__first1 == __last1)
-        return false;
-
-    if (__last2 - __first2 == 1)
-        return !__comp(*__first1, *__first2) && !__comp(*__first2, *__first1);
-
-    return __internal::__except_handler([&]() {
-        return !__internal::__parallel_or(
-            __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first2, __last2,
-            [__first1, __last1, __first2, __last2, &__comp](_RandomAccessIterator2 __i, _RandomAccessIterator2 __j) {
-                assert(__j > __i);
-                //assert(__j - __i > 1);
-
-                //1. moving boundaries to "consume" subsequence of equal elements
-                auto __is_equal_sorted = [&__comp](_RandomAccessIterator2 __a, _RandomAccessIterator2 __b) -> bool {
-                    //enough one call of __comp due to compared couple belongs to one sorted sequience
-                    return !__comp(*__a, *__b);
-                };
-
-                //1.1 left bound, case "aaa[aaaxyz...]" - searching "x"
-                if (__i > __first2 && __is_equal_sorted(__i - 1, __i))
-                {
-                    //whole subrange continues to have equal elements - return "no op"
-                    if (__is_equal_sorted(__i, __j - 1))
-                        return false;
-
-                    __i = ::std::upper_bound(__i, __last2, *__i, __comp);
-                }
-
-                //1.2 right bound, case "[...aaa]aaaxyz" - searching "x"
-                if (__j < __last2 && __is_equal_sorted(__j - 1, __j))
-                    __j = ::std::upper_bound(__j, __last2, *__j, __comp);
-
-                //2. testing is __a subsequence of the second range included into the first range
-                auto __b = ::std::lower_bound(__first1, __last1, *__i, __comp);
-
-                assert(!__comp(*(__last1 - 1), *__b));
-                assert(!__comp(*(__j - 1), *__i));
-                return !::std::includes(__b, __last1, __i, __j, __comp);
-            });
-    });
-}
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
           class _Compare>
 bool
@@ -5121,109 +3347,15 @@ __pattern_includes(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _
                 //2. testing is __a subsequence of the second range included into the first range
                 auto __b = ::std::lower_bound(__first1, __last1, *__i, __comp);
 
-                assert(!__comp(*(__last1 - 1), *__b));
-                assert(!__comp(*(__j - 1), *__i));
-                return !::std::includes(__b, __last1, __i, __j, __comp);
-            });
-    });
-}
-
-inline constexpr auto __set_algo_cut_off = 1000;
-
-template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _OutputIterator,
-          class _Compare, class _IsVector, class _SizeFunction, class _SetOP>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _OutputIterator>
-__parallel_set_op(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
-                  _ForwardIterator2 __first2, _ForwardIterator2 __last2, _OutputIterator __result, _Compare __comp,
-                  _SizeFunction __size_func, _SetOP __set_op, _IsVector __is_vector)
-{
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1,
-                                                                              _ForwardIterator2, _OutputIterator>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
-    typedef typename ::std::iterator_traits<_ForwardIterator1>::difference_type _DifferenceType;
-    typedef typename ::std::iterator_traits<_OutputIterator>::value_type _T;
-
-    struct _SetRange
-    {
-        _DifferenceType __pos, __len, __buf_pos;
-        bool
-        empty() const
-        {
-            return __len == 0;
-        }
-    };
-
-    const _DifferenceType __n1 = __last1 - __first1;
-    const _DifferenceType __n2 = __last2 - __first2;
-
-    __par_backend::__buffer<_ExecutionPolicy, _T> __buf(__size_func(__n1, __n2));
-
-    return __internal::__except_handler([&__exec, __n1, __first1, __last1, __first2, __last2, __result, __is_vector,
-                                         __comp, __size_func, __set_op, &__buf]() {
-        auto __tmp_memory = __buf.get();
-        _DifferenceType __m{};
-        auto __scan = [=](_DifferenceType, _DifferenceType, const _SetRange& __s) { // Scan
-            if (!__s.empty())
-                __brick_move_destroy<_ExecutionPolicy>{}(__tmp_memory + __s.__buf_pos,
-                                                         __tmp_memory + (__s.__buf_pos + __s.__len),
-                                                         __result + __s.__pos, __is_vector);
-        };
-        __par_backend::__parallel_strict_scan(
-            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __n1, _SetRange{0, 0, 0}, //-1, 0},
-            [=](_DifferenceType __i, _DifferenceType __len) {                                    // Reduce
-                //[__b; __e) - a subrange of the first sequence, to reduce
-                _ForwardIterator1 __b = __first1 + __i, __e = __first1 + (__i + __len);
-
-                //try searching for the first element which not equal to *__b
-                if (__b != __first1)
-                    __b = ::std::upper_bound(__b, __last1, *__b, __comp);
-
-                //try searching for the first element which not equal to *__e
-                if (__e != __last1)
-                    __e = ::std::upper_bound(__e, __last1, *__e, __comp);
-
-                //check is [__b; __e) empty
-                if (__e - __b < 1)
-                {
-                    _ForwardIterator2 __bb = __last2;
-                    if (__b != __last1)
-                        __bb = ::std::lower_bound(__first2, __last2, *__b, __comp);
-
-                    const _DifferenceType __buf_pos = __size_func((__b - __first1), (__bb - __first2));
-                    return _SetRange{0, 0, __buf_pos};
-                }
-
-                //try searching for "corresponding" subrange [__bb; __ee) in the second sequence
-                _ForwardIterator2 __bb = __first2;
-                if (__b != __first1)
-                    __bb = ::std::lower_bound(__first2, __last2, *__b, __comp);
-
-                _ForwardIterator2 __ee = __last2;
-                if (__e != __last1)
-                    __ee = ::std::lower_bound(__bb, __last2, *__e, __comp);
-
-                const _DifferenceType __buf_pos = __size_func((__b - __first1), (__bb - __first2));
-                auto __buffer_b = __tmp_memory + __buf_pos;
-                auto __res = __set_op(__b, __e, __bb, __ee, __buffer_b, __comp);
-
-                return _SetRange{0, __res - __buffer_b, __buf_pos};
-            },
-            [](const _SetRange& __a, const _SetRange& __b) { // Combine
-                if (__b.__buf_pos > __a.__buf_pos || ((__b.__buf_pos == __a.__buf_pos) && !__b.empty()))
-                    return _SetRange{__a.__pos + __a.__len + __b.__pos, __b.__len, __b.__buf_pos};
-                return _SetRange{__b.__pos + __b.__len + __a.__pos, __a.__len, __a.__buf_pos};
-            },
-            __scan,                                     // Scan
-            [&__m, &__scan](const _SetRange& __total) { // Apex
-                //final scan
-                __scan(0, 0, __total);
-                __m = __total.__pos + __total.__len;
+                assert(!__comp(*(__last1 - 1), *__b));
+                assert(!__comp(*(__j - 1), *__i));
+                return !::std::includes(__b, __last1, __i, __j, __comp);
             });
-        return __result + __m;
     });
 }
 
+inline constexpr auto __set_algo_cut_off = 1000;
+
 template <class _IsVector, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2,
           class _OutputIterator, class _Compare, class _SizeFunction, class _SetOP>
 _OutputIterator
@@ -5316,121 +3448,6 @@ __parallel_set_op(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _F
     });
 }
 
-//a shared parallel pattern for '__pattern_set_union' and '__pattern_set_symmetric_difference'
-template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _OutputIterator,
-          class _Compare, class _SetUnionOp, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _OutputIterator>
-__parallel_set_union_op(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
-                        _ForwardIterator2 __first2, _ForwardIterator2 __last2, _OutputIterator __result,
-                        _Compare __comp, _SetUnionOp __set_union_op, _IsVector __is_vector)
-{
-    typedef typename ::std::iterator_traits<_ForwardIterator1>::difference_type _DifferenceType;
-
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1,
-                                                                              _ForwardIterator2, _OutputIterator>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
-    const auto __n1 = __last1 - __first1;
-    const auto __n2 = __last2 - __first2;
-
-    __brick_copy<_ExecutionPolicy> __copy_range{};
-
-    // {1} {}: parallel copying just first sequence
-    if (__n2 == 0)
-        return __internal::__pattern_walk2_brick(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first1,
-                                                 __last1, __result, __copy_range);
-
-    // {} {2}: parallel copying justmake  second sequence
-    if (__n1 == 0)
-        return __internal::__pattern_walk2_brick(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first2,
-                                                 __last2, __result, __copy_range);
-
-    // testing  whether the sequences are intersected
-    _ForwardIterator1 __left_bound_seq_1 = ::std::lower_bound(__first1, __last1, *__first2, __comp);
-
-    if (__left_bound_seq_1 == __last1)
-    {
-        //{1} < {2}: seq2 is wholly greater than seq1, so, do parallel copying seq1 and seq2
-        __par_backend::__parallel_invoke(
-            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
-            [=] {
-                __internal::__pattern_walk2_brick(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first1,
-                                                  __last1, __result, __copy_range);
-            },
-            [=] {
-                __internal::__pattern_walk2_brick(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first2,
-                                                  __last2, __result + __n1, __copy_range);
-            });
-        return __result + __n1 + __n2;
-    }
-
-    // testing  whether the sequences are intersected
-    _ForwardIterator2 __left_bound_seq_2 = ::std::lower_bound(__first2, __last2, *__first1, __comp);
-
-    if (__left_bound_seq_2 == __last2)
-    {
-        //{2} < {1}: seq2 is wholly greater than seq1, so, do parallel copying seq1 and seq2
-        __par_backend::__parallel_invoke(
-            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
-            [=] {
-                __internal::__pattern_walk2_brick(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first2,
-                                                  __last2, __result, __copy_range);
-            },
-            [=] {
-                __internal::__pattern_walk2_brick(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first1,
-                                                  __last1, __result + __n2, __copy_range);
-            });
-        return __result + __n1 + __n2;
-    }
-
-    const auto __m1 = __left_bound_seq_1 - __first1;
-    if (__m1 > __set_algo_cut_off)
-    {
-        auto __res_or = __result;
-        __result += __m1; //we know proper offset due to [first1; left_bound_seq_1) < [first2; last2)
-        __par_backend::__parallel_invoke(
-            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
-            //do parallel copying of [first1; left_bound_seq_1)
-            [=] {
-                __internal::__pattern_walk2_brick(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first1,
-                                                  __left_bound_seq_1, __res_or, __copy_range);
-            },
-            [=, &__result] {
-                __result = __internal::__parallel_set_op(
-                    __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __left_bound_seq_1, __last1, __first2,
-                    __last2, __result, __comp, [](_DifferenceType __n, _DifferenceType __m) { return __n + __m; },
-                    __set_union_op);
-            });
-        return __result;
-    }
-
-    const auto __m2 = __left_bound_seq_2 - __first2;
-    assert(__m1 == 0 || __m2 == 0);
-    if (__m2 > __set_algo_cut_off)
-    {
-        auto __res_or = __result;
-        __result += __m2; //we know proper offset due to [first2; left_bound_seq_2) < [first1; last1)
-        __par_backend::__parallel_invoke(
-            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
-            //do parallel copying of [first2; left_bound_seq_2)
-            [=] {
-                __internal::__pattern_walk2_brick(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first2,
-                                                  __left_bound_seq_2, __res_or, __copy_range);
-            },
-            [=, &__result] {
-                __result = __internal::__parallel_set_op(
-                    __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __left_bound_seq_2,
-                    __last2, __result, __comp, [](_DifferenceType __n, _DifferenceType __m) { return __n + __m; },
-                    __set_union_op);
-            });
-        return __result;
-    }
-
-    return __internal::__parallel_set_op(
-        __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2, __result,
-        __comp, [](_DifferenceType __n, _DifferenceType __m) { return __n + __m; }, __set_union_op);
-}
-
 //a shared parallel pattern for '__pattern_set_union' and '__pattern_set_symmetric_difference'
 template <class _IsVector, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2,
           class _OutputIterator, class _Compare, class _SetUnionOp>
@@ -5578,17 +3595,6 @@ __brick_set_union(_RandomAccessIterator1 __first1, _RandomAccessIterator1 __last
     return ::std::set_union(__first1, __last1, __first2, __last2, __result, __comp);
 }
 
-template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _OutputIterator,
-          class _Compare, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _OutputIterator>
-__pattern_set_union(_ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
-                    _ForwardIterator2 __first2, _ForwardIterator2 __last2, _OutputIterator __result, _Compare __comp,
-                    _IsVector __is_vector,
-                    /*is_parallel=*/::std::false_type) noexcept
-{
-    return __internal::__brick_set_union(__first1, __last1, __first2, __last2, __result, __comp, __is_vector);
-}
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _OutputIterator,
           class _Compare>
 _OutputIterator
@@ -5602,35 +3608,6 @@ __pattern_set_union(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _Forwa
                                          typename _Tag::__is_vector{});
 }
 
-template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _OutputIterator,
-          class _Compare, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _OutputIterator>
-__pattern_set_union(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1, _RandomAccessIterator1 __last1,
-                    _RandomAccessIterator2 __first2, _RandomAccessIterator2 __last2, _OutputIterator __result,
-                    _Compare __comp, _IsVector __is_vector, /*__is_parallel=*/::std::true_type)
-{
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2,
-                                                  _OutputIterator>();
-
-    const auto __n1 = __last1 - __first1;
-    const auto __n2 = __last2 - __first2;
-
-    // use serial algorithm
-    if (__n1 + __n2 <= __set_algo_cut_off)
-        return ::std::set_union(__first1, __last1, __first2, __last2, __result, __comp);
-
-    typedef typename ::std::iterator_traits<_OutputIterator>::value_type _Tp;
-    return __parallel_set_union_op(
-        __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2, __result,
-        __comp,
-        [](_RandomAccessIterator1 __first1, _RandomAccessIterator1 __last1, _RandomAccessIterator2 __first2,
-           _RandomAccessIterator2 __last2, _Tp* __result, _Compare __comp) {
-            return oneapi::dpl::__utils::__set_union_construct(__first1, __last1, __first2, __last2, __result, __comp,
-                                                               __BrickCopyConstruct<_IsVector>());
-        });
-}
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
           class _OutputIterator, class _Compare>
 _OutputIterator
@@ -5679,16 +3656,6 @@ __brick_set_intersection(_RandomAccessIterator1 __first1, _RandomAccessIterator1
     return ::std::set_intersection(__first1, __last1, __first2, __last2, __result, __comp);
 }
 
-template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _OutputIterator,
-          class _Compare, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _OutputIterator>
-__pattern_set_intersection(_ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
-                           _ForwardIterator2 __first2, _ForwardIterator2 __last2, _OutputIterator __result,
-                           _Compare __comp, _IsVector __is_vector, /*is_parallel=*/::std::false_type) noexcept
-{
-    return __internal::__brick_set_intersection(__first1, __last1, __first2, __last2, __result, __comp, __is_vector);
-}
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _OutputIterator,
           class _Compare>
 _OutputIterator
@@ -5702,73 +3669,6 @@ __pattern_set_intersection(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1,
                                                 typename _Tag::__is_vector{});
 }
 
-template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
-          class _RandomAccessIterator3, class _Compare, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator3>
-__pattern_set_intersection(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1, _RandomAccessIterator1 __last1,
-                           _RandomAccessIterator2 __first2, _RandomAccessIterator2 __last2,
-                           _RandomAccessIterator3 __result, _Compare __comp, _IsVector __is_vector,
-                           /*is_parallel=*/::std::true_type)
-{
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2,
-                                                  _RandomAccessIterator3>();
-
-    typedef typename ::std::iterator_traits<_RandomAccessIterator3>::value_type _T;
-    typedef typename ::std::iterator_traits<_RandomAccessIterator1>::difference_type _DifferenceType;
-
-    const auto __n1 = __last1 - __first1;
-    const auto __n2 = __last2 - __first2;
-
-    // intersection is empty
-    if (__n1 == 0 || __n2 == 0)
-        return __result;
-
-    // testing  whether the sequences are intersected
-    _RandomAccessIterator1 __left_bound_seq_1 = ::std::lower_bound(__first1, __last1, *__first2, __comp);
-    //{1} < {2}: seq 2 is wholly greater than seq 1, so, the intersection is empty
-    if (__left_bound_seq_1 == __last1)
-        return __result;
-
-    // testing  whether the sequences are intersected
-    _RandomAccessIterator2 __left_bound_seq_2 = ::std::lower_bound(__first2, __last2, *__first1, __comp);
-    //{2} < {1}: seq 1 is wholly greater than seq 2, so, the intersection is empty
-    if (__left_bound_seq_2 == __last2)
-        return __result;
-
-    const auto __m1 = __last1 - __left_bound_seq_1 + __n2;
-    if (__m1 > __set_algo_cut_off)
-    {
-        //we know proper offset due to [first1; left_bound_seq_1) < [first2; last2)
-        return __internal::__parallel_set_op(
-            __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __left_bound_seq_1, __last1, __first2, __last2,
-            __result, __comp, [](_DifferenceType __n, _DifferenceType __m) { return ::std::min(__n, __m); },
-            [](_RandomAccessIterator1 __first1, _RandomAccessIterator1 __last1, _RandomAccessIterator2 __first2,
-               _RandomAccessIterator2 __last2, _T* __result, _Compare __comp) {
-                return oneapi::dpl::__utils::__set_intersection_construct(__first1, __last1, __first2, __last2,
-                                                                          __result, __comp);
-            });
-    }
-
-    const auto __m2 = __last2 - __left_bound_seq_2 + __n1;
-    if (__m2 > __set_algo_cut_off)
-    {
-        //we know proper offset due to [first2; left_bound_seq_2) < [first1; last1)
-        __result = __internal::__parallel_set_op(
-            __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __left_bound_seq_2, __last2,
-            __result, __comp, [](_DifferenceType __n, _DifferenceType __m) { return ::std::min(__n, __m); },
-            [](_RandomAccessIterator1 __first1, _RandomAccessIterator1 __last1, _RandomAccessIterator2 __first2,
-               _RandomAccessIterator2 __last2, _T* __result, _Compare __comp) {
-                return oneapi::dpl::__utils::__set_intersection_construct(__first2, __last2, __first1, __last1,
-                                                                          __result, __comp);
-            });
-        return __result;
-    }
-
-    // [left_bound_seq_1; last1) and [left_bound_seq_2; last2) - use serial algorithm
-    return ::std::set_intersection(__left_bound_seq_1, __last1, __left_bound_seq_2, __last2, __result, __comp);
-}
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
           class _RandomAccessIterator3, class _Compare>
 _RandomAccessIterator3
@@ -5842,96 +3742,31 @@ __brick_set_difference(_ForwardIterator1 __first1, _ForwardIterator1 __last1, _F
                        /*__is_vector=*/::std::false_type) noexcept
 {
     return ::std::set_difference(__first1, __last1, __first2, __last2, __result, __comp);
-}
-
-template <class _RandomAccessIterator1, class _RandomAccessIterator2, class _RandomAccessIterator3, class _Compare>
-_RandomAccessIterator3
-__brick_set_difference(_RandomAccessIterator1 __first1, _RandomAccessIterator1 __last1, _RandomAccessIterator2 __first2,
-                       _RandomAccessIterator2 __last2, _RandomAccessIterator3 __result, _Compare __comp,
-                       /*__is_vector=*/::std::true_type) noexcept
-{
-    _PSTL_PRAGMA_MESSAGE("Vectorized algorithm unimplemented, redirected to serial");
-    return ::std::set_difference(__first1, __last1, __first2, __last2, __result, __comp);
-}
-
-template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _OutputIterator,
-          class _Compare, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _OutputIterator>
-__pattern_set_difference(_ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
-                         _ForwardIterator2 __first2, _ForwardIterator2 __last2, _OutputIterator __result,
-                         _Compare __comp, _IsVector __is_vector, /*is_parallel=*/::std::false_type) noexcept
-{
-    return __internal::__brick_set_difference(__first1, __last1, __first2, __last2, __result, __comp, __is_vector);
-}
-
-template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _OutputIterator,
-          class _Compare>
-_OutputIterator
-__pattern_set_difference(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
-                         _ForwardIterator2 __first2, _ForwardIterator2 __last2, _OutputIterator __result,
-                         _Compare __comp) noexcept
-{
-    static_assert(__is_backend_tag_v<_Tag>);
-
-    return __internal::__brick_set_difference(__first1, __last1, __first2, __last2, __result, __comp,
-                                              typename _Tag::__is_vector{});
-}
-
-template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
-          class _RandomAccessIterator3, class _Compare, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator3>
-__pattern_set_difference(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1, _RandomAccessIterator1 __last1,
-                         _RandomAccessIterator2 __first2, _RandomAccessIterator2 __last2,
-                         _RandomAccessIterator3 __result, _Compare __comp, _IsVector __is_vector,
-                         /*is_parallel=*/::std::true_type)
-{
-    typedef typename ::std::iterator_traits<_RandomAccessIterator3>::value_type _T;
-    typedef typename ::std::iterator_traits<_RandomAccessIterator1>::difference_type _DifferenceType;
-
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2,
-                                                  _RandomAccessIterator3>();
-
-    const auto __n1 = __last1 - __first1;
-    const auto __n2 = __last2 - __first2;
-
-    // {} \ {2}: the difference is empty
-    if (__n1 == 0)
-        return __result;
-
-    // {1} \ {}: parallel copying just first sequence
-    if (__n2 == 0)
-        return __pattern_walk2_brick(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1,
-                                     __result, __internal::__brick_copy<_ExecutionPolicy>{});
-
-    // testing  whether the sequences are intersected
-    _RandomAccessIterator1 __left_bound_seq_1 = ::std::lower_bound(__first1, __last1, *__first2, __comp);
-    //{1} < {2}: seq 2 is wholly greater than seq 1, so, parallel copying just first sequence
-    if (__left_bound_seq_1 == __last1)
-        return __pattern_walk2_brick(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1,
-                                     __result, __internal::__brick_copy<_ExecutionPolicy>{});
-
-    // testing  whether the sequences are intersected
-    _RandomAccessIterator2 __left_bound_seq_2 = ::std::lower_bound(__first2, __last2, *__first1, __comp);
-    //{2} < {1}: seq 1 is wholly greater than seq 2, so, parallel copying just first sequence
-    if (__left_bound_seq_2 == __last2)
-        return __internal::__pattern_walk2_brick(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first1,
-                                                 __last1, __result, __brick_copy<_ExecutionPolicy>{});
-
-    if (__n1 + __n2 > __set_algo_cut_off)
-        return __parallel_set_op(
-            __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2, __result,
-            __comp, [](_DifferenceType __n, _DifferenceType) { return __n; },
-            [](_RandomAccessIterator1 __first1, _RandomAccessIterator1 __last1, _RandomAccessIterator2 __first2,
-               _RandomAccessIterator2 __last2, _T* __result, _Compare __comp) {
-                return oneapi::dpl::__utils::__set_difference_construct(__first1, __last1, __first2, __last2, __result,
-                                                                        __comp, __BrickCopyConstruct<_IsVector>());
-            });
+}
 
-    // use serial algorithm
+template <class _RandomAccessIterator1, class _RandomAccessIterator2, class _RandomAccessIterator3, class _Compare>
+_RandomAccessIterator3
+__brick_set_difference(_RandomAccessIterator1 __first1, _RandomAccessIterator1 __last1, _RandomAccessIterator2 __first2,
+                       _RandomAccessIterator2 __last2, _RandomAccessIterator3 __result, _Compare __comp,
+                       /*__is_vector=*/::std::true_type) noexcept
+{
+    _PSTL_PRAGMA_MESSAGE("Vectorized algorithm unimplemented, redirected to serial");
     return ::std::set_difference(__first1, __last1, __first2, __last2, __result, __comp);
 }
 
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _OutputIterator,
+          class _Compare>
+_OutputIterator
+__pattern_set_difference(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
+                         _ForwardIterator2 __first2, _ForwardIterator2 __last2, _OutputIterator __result,
+                         _Compare __comp) noexcept
+{
+    static_assert(__is_backend_tag_v<_Tag>);
+
+    return __internal::__brick_set_difference(__first1, __last1, __first2, __last2, __result, __comp,
+                                              typename _Tag::__is_vector{});
+}
+
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
           class _RandomAccessIterator3, class _Compare>
 _RandomAccessIterator3
@@ -6006,17 +3841,6 @@ __brick_set_symmetric_difference(_RandomAccessIterator1 __first1, _RandomAccessI
     return ::std::set_symmetric_difference(__first1, __last1, __first2, __last2, __result, __comp);
 }
 
-template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _OutputIterator,
-          class _Compare, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _OutputIterator>
-__pattern_set_symmetric_difference(_ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
-                                   _ForwardIterator2 __first2, _ForwardIterator2 __last2, _OutputIterator __result,
-                                   _Compare __comp, _IsVector __is_vector, /*is_parallel=*/::std::false_type) noexcept
-{
-    return __internal::__brick_set_symmetric_difference(__first1, __last1, __first2, __last2, __result, __comp,
-                                                        __is_vector);
-}
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _OutputIterator,
           class _Compare>
 _OutputIterator
@@ -6030,36 +3854,6 @@ __pattern_set_symmetric_difference(_Tag, _ExecutionPolicy&&, _ForwardIterator1 _
                                                         typename _Tag::__is_vector{});
 }
 
-template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
-          class _RandomAccessIterator3, class _Compare, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator3>
-__pattern_set_symmetric_difference(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1,
-                                   _RandomAccessIterator1 __last1, _RandomAccessIterator2 __first2,
-                                   _RandomAccessIterator2 __last2, _RandomAccessIterator3 __result, _Compare __comp,
-                                   _IsVector __is_vector, /*is_parallel=*/::std::true_type)
-{
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2,
-                                                  _RandomAccessIterator3>();
-
-    const auto __n1 = __last1 - __first1;
-    const auto __n2 = __last2 - __first2;
-
-    // use serial algorithm
-    if (__n1 + __n2 <= __set_algo_cut_off)
-        return ::std::set_symmetric_difference(__first1, __last1, __first2, __last2, __result, __comp);
-
-    typedef typename ::std::iterator_traits<_RandomAccessIterator3>::value_type _T;
-    return __internal::__parallel_set_union_op(
-        __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2, __result,
-        __comp,
-        [](_RandomAccessIterator1 __first1, _RandomAccessIterator1 __last1, _RandomAccessIterator2 __first2,
-           _RandomAccessIterator2 __last2, _T* __result, _Compare __comp) {
-            return oneapi::dpl::__utils::__set_symmetric_difference_construct(
-                __first1, __last1, __first2, __last2, __result, __comp, __BrickCopyConstruct<_IsVector>());
-        });
-}
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
           class _RandomAccessIterator3, class _Compare>
 _RandomAccessIterator3
@@ -6108,14 +3902,6 @@ __brick_is_heap_until(_RandomAccessIterator __first, _RandomAccessIterator __las
         [&__comp](_RandomAccessIterator __it, _SizeType __i) { return __comp(__it[(__i - 1) / 2], __it[__i]); });
 }
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _Compare, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator>
-__pattern_is_heap_until(_ExecutionPolicy&&, _RandomAccessIterator __first, _RandomAccessIterator __last,
-                        _Compare __comp, _IsVector __is_vector, /* is_parallel = */ ::std::false_type) noexcept
-{
-    return __internal::__brick_is_heap_until(__first, __last, __comp, __is_vector);
-}
-
 template <class _Tag, class _ExecutionPolicy, class _RandomAccessIterator, class _Compare>
 _RandomAccessIterator
 __pattern_is_heap_until(_Tag, _ExecutionPolicy&&, _RandomAccessIterator __first, _RandomAccessIterator __last,
@@ -6148,25 +3934,6 @@ __is_heap_until_local(_RandomAccessIterator __first, _DifferenceType __begin, _D
         [&__comp](_RandomAccessIterator __it, _DifferenceType __i) { return __comp(__it[(__i - 1) / 2], __it[__i]); });
 }
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _Compare, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator>
-__pattern_is_heap_until(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIterator __last,
-                        _Compare __comp, _IsVector __is_vector, /* is_parallel = */ ::std::true_type)
-{
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
-    return __internal::__except_handler([&]() {
-        return __parallel_find(
-            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
-            [__first, __comp, __is_vector](_RandomAccessIterator __i, _RandomAccessIterator __j) {
-                return __internal::__is_heap_until_local(__first, __i - __first, __j - __first, __comp, __is_vector);
-            },
-            ::std::true_type{});
-    });
-}
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Compare>
 _RandomAccessIterator
 __pattern_is_heap_until(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
@@ -6223,14 +3990,6 @@ __is_heap_local(_RandomAccessIterator __first, _DifferenceType __begin, _Differe
                                             });
 }
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _Compare, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, bool>
-__pattern_is_heap(_ExecutionPolicy&&, _RandomAccessIterator __first, _RandomAccessIterator __last, _Compare __comp,
-                  _IsVector __is_vector, /* is_parallel = */ ::std::false_type) noexcept
-{
-    return __internal::__brick_is_heap(__first, __last, __comp, __is_vector);
-}
-
 template <class _Tag, class _ExecutionPolicy, class _RandomAccessIterator, class _Compare>
 bool
 __pattern_is_heap(_Tag, _ExecutionPolicy&&, _RandomAccessIterator __first, _RandomAccessIterator __last,
@@ -6241,23 +4000,6 @@ __pattern_is_heap(_Tag, _ExecutionPolicy&&, _RandomAccessIterator __first, _Rand
     return __internal::__brick_is_heap(__first, __last, __comp, typename _Tag::__is_vector{});
 }
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _Compare, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, bool>
-__pattern_is_heap(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIterator __last,
-                  _Compare __comp, _IsVector __is_vector, /* is_parallel = */ ::std::true_type)
-{
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
-
-    return __internal::__except_handler([&]() {
-        return !__parallel_or(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
-                              [__first, __comp, __is_vector](_RandomAccessIterator __i, _RandomAccessIterator __j) {
-                                  return !__internal::__is_heap_local(__first, __i - __first, __j - __first, __comp,
-                                                                      __is_vector);
-                              });
-    });
-}
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Compare>
 bool
 __pattern_is_heap(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
@@ -6296,14 +4038,6 @@ __brick_min_element(_RandomAccessIterator __first, _RandomAccessIterator __last,
 #endif
 }
 
-template <typename _ExecutionPolicy, typename _ForwardIterator, typename _Compare, typename _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _ForwardIterator>
-__pattern_min_element(_ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _Compare __comp,
-                      _IsVector __is_vector, /* is_parallel = */ ::std::false_type) noexcept
-{
-    return __internal::__brick_min_element(__first, __last, __comp, __is_vector);
-}
-
 template <class _Tag, typename _ExecutionPolicy, typename _ForwardIterator, typename _Compare>
 _ForwardIterator
 __pattern_min_element(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
@@ -6314,41 +4048,6 @@ __pattern_min_element(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _Forwa
     return __internal::__brick_min_element(__first, __last, __comp, typename _Tag::__is_vector{});
 }
 
-template <typename _ExecutionPolicy, typename _RandomAccessIterator, typename _Compare, typename _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator>
-__pattern_min_element(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIterator __last,
-                      _Compare __comp, _IsVector __is_vector, /* is_parallel = */ ::std::true_type)
-{
-    // a trivial case pre-check
-    if (__last - __first < 2)
-        return __first;
-
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
-    return __internal::__except_handler([&]() {
-        return __par_backend::__parallel_reduce(
-            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, /*identity*/ __last,
-            [=](_RandomAccessIterator __begin, _RandomAccessIterator __end,
-                _RandomAccessIterator __init) -> _RandomAccessIterator {
-                const _RandomAccessIterator __subresult =
-                    __internal::__brick_min_element(__begin, __end, __comp, __is_vector);
-                return __init == __last ? __subresult
-                                        : __internal::__cmp_iterators_by_values(__init, __subresult, __comp,
-                                                                                oneapi::dpl::__internal::__pstl_less());
-            },
-            [=](_RandomAccessIterator __it1, _RandomAccessIterator __it2) -> _RandomAccessIterator {
-                if (__it1 == __last)
-                    return __it2;
-                if (__it2 == __last)
-                    return __it1;
-                return __internal::__cmp_iterators_by_values(__it1, __it2, __comp,
-                                                             oneapi::dpl::__internal::__pstl_less());
-            });
-    });
-}
-
 template <typename _IsVector, typename _ExecutionPolicy, typename _RandomAccessIterator, typename _Compare>
 _RandomAccessIterator
 __pattern_min_element(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
@@ -6406,15 +4105,6 @@ __brick_minmax_element(_RandomAccessIterator __first, _RandomAccessIterator __la
 #endif
 }
 
-template <typename _ExecutionPolicy, typename _ForwardIterator, typename _Compare, typename _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy,
-                                                           ::std::pair<_ForwardIterator, _ForwardIterator>>
-__pattern_minmax_element(_ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _Compare __comp,
-                         _IsVector __is_vector, /* is_parallel = */ ::std::false_type) noexcept
-{
-    return __internal::__brick_minmax_element(__first, __last, __comp, __is_vector);
-}
-
 template <class _Tag, typename _ExecutionPolicy, typename _ForwardIterator, typename _Compare>
 ::std::pair<_ForwardIterator, _ForwardIterator>
 __pattern_minmax_element(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
@@ -6425,52 +4115,6 @@ __pattern_minmax_element(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _Fo
     return __internal::__brick_minmax_element(__first, __last, __comp, typename _Tag::__is_vector{});
 }
 
-template <typename _ExecutionPolicy, typename _RandomAccessIterator, typename _Compare, typename _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy,
-                                                           ::std::pair<_RandomAccessIterator, _RandomAccessIterator>>
-__pattern_minmax_element(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIterator __last,
-                         _Compare __comp, _IsVector __is_vector, /* is_parallel = */ ::std::true_type)
-{
-    // a trivial case pre-check
-    if (__last - __first < 2)
-        return ::std::make_pair(__first, __first);
-
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
-    return __internal::__except_handler([&]() {
-        typedef ::std::pair<_RandomAccessIterator, _RandomAccessIterator> _Result;
-
-        return __par_backend::__parallel_reduce(
-            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
-            /*identity*/ ::std::make_pair(__last, __last),
-            [=, &__comp](_RandomAccessIterator __begin, _RandomAccessIterator __end, _Result __init) -> _Result {
-                const _Result __subresult = __internal::__brick_minmax_element(__begin, __end, __comp, __is_vector);
-                if (__init.first == __last) // = identity
-                    return __subresult;
-                return ::std::make_pair(
-                    __internal::__cmp_iterators_by_values(__init.first, __subresult.first, __comp,
-                                                          oneapi::dpl::__internal::__pstl_less()),
-                    __internal::__cmp_iterators_by_values(__init.second, __subresult.second,
-                                                          oneapi::dpl::__internal::__reorder_pred<_Compare>(__comp),
-                                                          oneapi::dpl::__internal::__pstl_greater()));
-            },
-            [=, &__comp](_Result __p1, _Result __p2) -> _Result {
-                if (__p1.first == __last)
-                    return __p2;
-                if (__p2.first == __last)
-                    return __p1;
-                return ::std::make_pair(
-                    __internal::__cmp_iterators_by_values(__p1.first, __p2.first, __comp,
-                                                          oneapi::dpl::__internal::__pstl_less()),
-                    __internal::__cmp_iterators_by_values(__p1.second, __p2.second,
-                                                          oneapi::dpl::__internal::__reorder_pred<_Compare>(__comp),
-                                                          oneapi::dpl::__internal::__pstl_greater()));
-            });
-    });
-}
-
 template <typename _IsVector, typename _ExecutionPolicy, typename _RandomAccessIterator, typename _Compare>
 ::std::pair<_RandomAccessIterator, _RandomAccessIterator>
 __pattern_minmax_element(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
@@ -6542,16 +4186,6 @@ __brick_mismatch(_RandomAccessIterator1 __first1, _RandomAccessIterator1 __last1
     return __unseq_backend::__simd_first(__first1, __n, __first2, __not_pred<_Predicate&>(__pred));
 }
 
-template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Predicate, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy,
-                                                           ::std::pair<_ForwardIterator1, _ForwardIterator2>>
-__pattern_mismatch(_ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
-                   _ForwardIterator2 __first2, _ForwardIterator2 __last2, _Predicate __pred, _IsVector __is_vector,
-                   /* is_parallel = */ ::std::false_type) noexcept
-{
-    return __internal::__brick_mismatch(__first1, __last1, __first2, __last2, __pred, __is_vector);
-}
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Predicate>
 ::std::pair<_ForwardIterator1, _ForwardIterator2>
 __pattern_mismatch(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
@@ -6562,31 +4196,6 @@ __pattern_mismatch(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _Forwar
     return __internal::__brick_mismatch(__first1, __last1, __first2, __last2, __pred, typename _Tag::__is_vector{});
 }
 
-template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _Predicate,
-          class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy,
-                                                           ::std::pair<_RandomAccessIterator1, _RandomAccessIterator2>>
-__pattern_mismatch(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1, _RandomAccessIterator1 __last1,
-                   _RandomAccessIterator2 __first2, _RandomAccessIterator2 __last2, _Predicate __pred,
-                   _IsVector __is_vector, /* is_parallel = */ ::std::true_type)
-{
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2>();
-
-    return __internal::__except_handler([&]() {
-        auto __n = ::std::min(__last1 - __first1, __last2 - __first2);
-        auto __result = __internal::__parallel_find(
-            __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __first1 + __n,
-            [__first1, __first2, __pred, __is_vector](_RandomAccessIterator1 __i, _RandomAccessIterator1 __j) {
-                return __internal::__brick_mismatch(__i, __j, __first2 + (__i - __first1), __first2 + (__j - __first1),
-                                                    __pred, __is_vector)
-                    .first;
-            },
-            ::std::true_type{});
-        return ::std::make_pair(__result, __first2 + (__result - __first1));
-    });
-}
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
           class _Predicate>
 ::std::pair<_RandomAccessIterator1, _RandomAccessIterator2>
@@ -6669,15 +4278,6 @@ __pattern_lexicographical_compare(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __
                                                        typename _Tag::__is_vector{});
 }
 
-template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Compare, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, bool>
-__pattern_lexicographical_compare(_ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
-                                  _ForwardIterator2 __first2, _ForwardIterator2 __last2, _Compare __comp,
-                                  _IsVector __is_vector, /* is_parallel = */ ::std::false_type) noexcept
-{
-    return __internal::__brick_lexicographical_compare(__first1, __last1, __first2, __last2, __comp, __is_vector);
-}
-
 template <class _IsVector, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Compare>
 bool
 __pattern_lexicographical_compare(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _ForwardIterator1 __first1,
@@ -6687,71 +4287,10 @@ __pattern_lexicographical_compare(__parallel_tag<_IsVector>, _ExecutionPolicy&&,
     return __internal::__brick_lexicographical_compare(__first1, __last1, __first2, __last2, __comp, _IsVector{});
 }
 
-template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _Compare,
-          class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, bool>
-__pattern_lexicographical_compare(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1,
-                                  _RandomAccessIterator1 __last1, _RandomAccessIterator2 __first2,
-                                  _RandomAccessIterator2 __last2, _Compare __comp, _IsVector __is_vector,
-                                  /* is_parallel = */ ::std::true_type)
-{
-    if (__first2 == __last2)
-    { // if second sequence is empty
-        return false;
-    }
-    else if (__first1 == __last1)
-    { // if first sequence is empty
-        return true;
-    }
-    else
-    {
-        constexpr auto __dispatch_tag =
-            oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1,
-                                                      _RandomAccessIterator2>();
-
-        typedef typename ::std::iterator_traits<_RandomAccessIterator1>::reference _RefType1;
-        typedef typename ::std::iterator_traits<_RandomAccessIterator2>::reference _RefType2;
-        --__last1;
-        --__last2;
-        auto __n = ::std::min(__last1 - __first1, __last2 - __first2);
-        auto __result = __internal::__parallel_find(
-            __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __first1 + __n,
-            [__first1, __first2, &__comp, __is_vector](_RandomAccessIterator1 __i, _RandomAccessIterator1 __j) {
-                return __internal::__brick_mismatch(
-                           __i, __j, __first2 + (__i - __first1), __first2 + (__j - __first1),
-                           [&__comp](const _RefType1 __x, const _RefType2 __y) {
-                               return !__comp(__x, __y) && !__comp(__y, __x);
-                           },
-                           __is_vector)
-                    .first;
-            },
-            ::std::true_type{});
-
-        if (__result == __last1 && __first2 + (__result - __first1) != __last2)
-        { // if first sequence shorter than second
-            return !__comp(*(__first2 + (__result - __first1)), *__result);
-        }
-        else
-        { // if second sequence shorter than first or both have the same number of elements
-            return __comp(*__result, *(__first2 + (__result - __first1)));
-        }
-    }
-}
-
 //------------------------------------------------------------------------
 // swap
 //------------------------------------------------------------------------
 
-template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Function, class _IsVector,
-          class _IsParallel>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _ForwardIterator2>
-__pattern_swap(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
-               _ForwardIterator2 __first2, _Function __f, _IsVector __is_vector, _IsParallel __is_parallel)
-{
-    return __pattern_walk2(::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __f, __is_vector,
-                           __is_parallel);
-}
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Function>
 _ForwardIterator2
 __pattern_swap(_Tag __tag, _ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
@@ -6830,15 +4369,6 @@ __brick_shift_left(_ForwardIterator __first, _ForwardIterator __last,
     return __first + __size_res;
 }
 
-template <class _ExecutionPolicy, class _ForwardIterator, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _ForwardIterator>
-__pattern_shift_left(_ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
-                     typename ::std::iterator_traits<_ForwardIterator>::difference_type __n, _IsVector __is_vector,
-                     /*is_parallel=*/::std::false_type) noexcept
-{
-    return __brick_shift_left(__first, __last, __n, __is_vector);
-}
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator>
 _ForwardIterator
 __pattern_shift_left(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
@@ -6849,53 +4379,6 @@ __pattern_shift_left(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _Forwar
     return __brick_shift_left(__first, __last, __n, typename _Tag::__is_vector{});
 }
 
-template <class _ExecutionPolicy, class _ForwardIterator, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _ForwardIterator>
-__pattern_shift_left(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last,
-                     typename ::std::iterator_traits<_ForwardIterator>::difference_type __n, _IsVector __is_vector,
-                     /*is_parallel=*/::std::true_type)
-{
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
-    //If (n > 0 && n < m), returns first + (m - n). Otherwise, if n  > 0, returns first. Otherwise, returns last.
-    if (__n <= 0)
-        return __last;
-    auto __size = __last - __first;
-    if (__n >= __size)
-        return __first;
-
-    using _DiffType = typename ::std::iterator_traits<_ForwardIterator>::difference_type;
-
-    _DiffType __mid = __size / 2 + __size % 2;
-    _DiffType __size_res = __size - __n;
-
-    //1. n >= size/2; there is enough memory to 'total' parallel copying
-    if (__n >= __mid)
-    {
-        __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __n, __size,
-                                      [__first, __n, __is_vector](_DiffType __i, _DiffType __j) {
-                                          __brick_move<_ExecutionPolicy>{}(__first + __i, __first + __j,
-                                                                           __first + __i - __n, __is_vector);
-                                      });
-    }
-    else //2. n < size/2; there is not enough memory to parallel copying; doing parallel copying by n elements
-    {
-        //TODO: to consider parallel processing by the 'internal' loop (but we may probably get cache locality issues)
-        for (auto __k = __n; __k < __size; __k += __n)
-        {
-            auto __end = ::std::min(__k + __n, __size);
-            __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __k, __end,
-                                          [__first, __n, __is_vector](_DiffType __i, _DiffType __j) {
-                                              __brick_move<_ExecutionPolicy>{}(__first + __i, __first + __j,
-                                                                               __first + __i - __n, __is_vector);
-                                          });
-        }
-    }
-
-    return __first + __size_res;
-}
-
 template <class _IsVector, class _ExecutionPolicy, class _ForwardIterator>
 _ForwardIterator
 __pattern_shift_left(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _ForwardIterator __first,
@@ -6941,20 +4424,6 @@ __pattern_shift_left(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Forw
     return __first + __size_res;
 }
 
-template <class _ExecutionPolicy, class _BidirectionalIterator, class _IsVector, class _IsParallel>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _BidirectionalIterator>
-__pattern_shift_right(_ExecutionPolicy&& __exec, _BidirectionalIterator __first, _BidirectionalIterator __last,
-                      typename ::std::iterator_traits<_BidirectionalIterator>::difference_type __n,
-                      _IsVector __is_vector, _IsParallel is_parallel)
-{
-    using _ReverseIterator = typename ::std::reverse_iterator<_BidirectionalIterator>;
-    auto __res = oneapi::dpl::__internal::__pattern_shift_left(::std::forward<_ExecutionPolicy>(__exec),
-                                                               _ReverseIterator(__last), _ReverseIterator(__first), __n,
-                                                               __is_vector, is_parallel);
-
-    return __res.base();
-}
-
 template <class _Tag, class _ExecutionPolicy, class _BidirectionalIterator>
 _BidirectionalIterator
 __pattern_shift_right(_Tag, _ExecutionPolicy&& __exec, _BidirectionalIterator __first, _BidirectionalIterator __last,
diff --git a/include/oneapi/dpl/pstl/numeric_fwd.h b/include/oneapi/dpl/pstl/numeric_fwd.h
index 0e8ba2a96e7..46826aff9d3 100644
--- a/include/oneapi/dpl/pstl/numeric_fwd.h
+++ b/include/oneapi/dpl/pstl/numeric_fwd.h
@@ -43,26 +43,12 @@ _Tp __brick_transform_reduce(_ForwardIterator1, _ForwardIterator1, _ForwardItera
                              _BinaryOperation2,
                              /*__is_vector=*/::std::false_type) noexcept;
 
-template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Tp, class _BinaryOperation1,
-          class _BinaryOperation2, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _Tp>
-__pattern_transform_reduce(_ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2, _Tp,
-                           _BinaryOperation1, _BinaryOperation2, _IsVector,
-                           /*is_parallel=*/::std::false_type) noexcept;
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Tp,
           class _BinaryOperation1, class _BinaryOperation2>
 _Tp
 __pattern_transform_reduce(_Tag, _ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2, _Tp,
                            _BinaryOperation1, _BinaryOperation2) noexcept;
 
-template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _Tp,
-          class _BinaryOperation1, class _BinaryOperation2, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _Tp>
-__pattern_transform_reduce(_ExecutionPolicy&&, _RandomAccessIterator1, _RandomAccessIterator1, _RandomAccessIterator2,
-                           _Tp, _BinaryOperation1, _BinaryOperation2, _IsVector __is_vector,
-                           /*is_parallel=*/::std::true_type);
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
           class _Tp, class _BinaryOperation1, class _BinaryOperation2>
 _Tp
@@ -81,26 +67,12 @@ template <class _ForwardIterator, class _Tp, class _BinaryOperation, class _Unar
 _Tp __brick_transform_reduce(_ForwardIterator, _ForwardIterator, _Tp, _BinaryOperation, _UnaryOperation,
                              /*is_vector=*/::std::false_type) noexcept;
 
-template <class _ExecutionPolicy, class _ForwardIterator, class _Tp, class _BinaryOperation, class _UnaryOperation,
-          class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _Tp>
-__pattern_transform_reduce(_ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _Tp, _BinaryOperation,
-                           _UnaryOperation, _IsVector,
-                           /*is_parallel=*/::std::false_type) noexcept;
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Tp,
           class _BinaryOperation1, class _BinaryOperation2>
 _Tp
 __pattern_transform_reduce(_Tag, _ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2, _Tp,
                            _BinaryOperation1, _BinaryOperation2 __bnary_op2) noexcept;
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _Tp, class _BinaryOperation, class _UnaryOperation,
-          class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _Tp>
-__pattern_transform_reduce(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _Tp, _BinaryOperation,
-                           _UnaryOperation, _IsVector,
-                           /*is_parallel=*/::std::true_type);
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
           class _Tp, class _BinaryOperation1, class _BinaryOperation2>
 _Tp
@@ -135,13 +107,6 @@ ::std::pair<_OutputIterator, _Tp> __brick_transform_scan(_RandomAccessIterator,
                                                          _UnaryOperation, _Tp, _BinaryOperation,
                                                          /*Inclusive*/ ::std::true_type) noexcept;
 
-template <class _ExecutionPolicy, class _ForwardIterator, class _OutputIterator, class _UnaryOperation, class _Tp,
-          class _BinaryOperation, class _Inclusive, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _OutputIterator>
-__pattern_transform_scan(_ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _OutputIterator, _UnaryOperation, _Tp,
-                         _BinaryOperation, _Inclusive, _IsVector,
-                         /*is_parallel=*/::std::false_type) noexcept;
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _OutputIterator, class _UnaryOperation,
           class _Tp, class _BinaryOperation, class _Inclusive>
 _OutputIterator
@@ -177,13 +142,6 @@ __pattern_transform_scan(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomA
                          _OutputIterator, _UnaryOperation, _Tp, _BinaryOperation, _Inclusive);
 
 // transform_scan without initial element
-template <class _ExecutionPolicy, class _ForwardIterator, class _OutputIterator, class _UnaryOperation,
-          class _BinaryOperation, class _Inclusive, class _IsVector, class _IsParallel>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _OutputIterator>
-__pattern_transform_scan(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last,
-                         _OutputIterator __result, _UnaryOperation __unary_op, _BinaryOperation __binary_op, _Inclusive,
-                         _IsVector __is_vector, _IsParallel __is_parallel);
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _OutputIterator, class _UnaryOperation,
           class _BinaryOperation, class _Inclusive>
 _OutputIterator
@@ -203,23 +161,11 @@ _OutputIterator __brick_adjacent_difference(_RandomAccessIterator, _RandomAccess
                                             _BinaryOperation,
                                             /*is_vector*/ ::std::true_type) noexcept;
 
-template <class _ExecutionPolicy, class _ForwardIterator, class _OutputIterator, class _BinaryOperation,
-          class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _OutputIterator>
-__pattern_adjacent_difference(_ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _OutputIterator, _BinaryOperation,
-                              _IsVector, /*is_parallel*/ ::std::false_type) noexcept;
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _OutputIterator, class _BinaryOperation>
 _OutputIterator
 __pattern_adjacent_difference(_Tag, _ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _OutputIterator,
                               _BinaryOperation) noexcept;
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _OutputIterator, class _BinaryOperation,
-          class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _OutputIterator>
-__pattern_adjacent_difference(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _OutputIterator,
-                              _BinaryOperation, _IsVector, /*is_parallel*/ ::std::true_type);
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
           class _BinaryOperation>
 _RandomAccessIterator2
diff --git a/include/oneapi/dpl/pstl/numeric_impl.h b/include/oneapi/dpl/pstl/numeric_impl.h
index 216f313ffbc..92ca677bc8d 100644
--- a/include/oneapi/dpl/pstl/numeric_impl.h
+++ b/include/oneapi/dpl/pstl/numeric_impl.h
@@ -61,17 +61,6 @@ __brick_transform_reduce(_RandomAccessIterator1 __first1, _RandomAccessIterator1
         [=, &__binary_op2](_DifferenceType __i) { return __binary_op2(__first1[__i], __first2[__i]); });
 }
 
-template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Tp, class _BinaryOperation1,
-          class _BinaryOperation2, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _Tp>
-__pattern_transform_reduce(_ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
-                           _ForwardIterator2 __first2, _Tp __init, _BinaryOperation1 __binary_op1,
-                           _BinaryOperation2 __binary_op2, _IsVector __is_vector,
-                           /*is_parallel=*/::std::false_type) noexcept
-{
-    return __brick_transform_reduce(__first1, __last1, __first2, __init, __binary_op1, __binary_op2, __is_vector);
-}
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Tp,
           class _BinaryOperation1, class _BinaryOperation2>
 _Tp
@@ -85,33 +74,6 @@ __pattern_transform_reduce(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1,
                                     typename _Tag::__is_vector{});
 }
 
-template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _Tp,
-          class _BinaryOperation1, class _BinaryOperation2, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _Tp>
-__pattern_transform_reduce(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1, _RandomAccessIterator1 __last1,
-                           _RandomAccessIterator2 __first2, _Tp __init, _BinaryOperation1 __binary_op1,
-                           _BinaryOperation2 __binary_op2, _IsVector __is_vector, /*is_parallel=*/::std::true_type)
-{
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
-    return __internal::__except_handler([&]() {
-        return __par_backend::__parallel_transform_reduce(
-            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1,
-            [__first1, __first2, __binary_op2](_RandomAccessIterator1 __i) mutable {
-                return __binary_op2(*__i, *(__first2 + (__i - __first1)));
-            },
-            __init,
-            __binary_op1, // Combine
-            [__first1, __first2, __binary_op1, __binary_op2,
-             __is_vector](_RandomAccessIterator1 __i, _RandomAccessIterator1 __j, _Tp __init) -> _Tp {
-                return __internal::__brick_transform_reduce(__i, __j, __first2 + (__i - __first1), __init, __binary_op1,
-                                                            __binary_op2, __is_vector);
-            });
-    });
-}
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
           class _Tp, class _BinaryOperation1, class _BinaryOperation2>
 _Tp
@@ -165,16 +127,6 @@ __brick_transform_reduce(_RandomAccessIterator __first, _RandomAccessIterator __
         [=, &__unary_op](_DifferenceType __i) { return __unary_op(__first[__i]); });
 }
 
-template <class _ExecutionPolicy, class _ForwardIterator, class _Tp, class _BinaryOperation, class _UnaryOperation,
-          class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _Tp>
-__pattern_transform_reduce(_ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _Tp __init,
-                           _BinaryOperation __binary_op, _UnaryOperation __unary_op, _IsVector __is_vector,
-                           /*is_parallel=*/::std::false_type) noexcept
-{
-    return __internal::__brick_transform_reduce(__first, __last, __init, __binary_op, __unary_op, __is_vector);
-}
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Tp, class _BinaryOperation,
           class _UnaryOperation>
 _Tp
@@ -187,27 +139,6 @@ __pattern_transform_reduce(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _
                                                 typename _Tag::__is_vector{});
 }
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _Tp, class _BinaryOperation, class _UnaryOperation,
-          class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _Tp>
-__pattern_transform_reduce(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIterator __last,
-                           _Tp __init, _BinaryOperation __binary_op, _UnaryOperation __unary_op, _IsVector __is_vector,
-                           /*is_parallel=*/::std::true_type)
-{
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
-    return __internal::__except_handler([&]() {
-        return __par_backend::__parallel_transform_reduce(
-            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
-            [__unary_op](_RandomAccessIterator __i) mutable { return __unary_op(*__i); }, __init, __binary_op,
-            [__unary_op, __binary_op, __is_vector](_RandomAccessIterator __i, _RandomAccessIterator __j, _Tp __init) {
-                return __internal::__brick_transform_reduce(__i, __j, __init, __binary_op, __unary_op, __is_vector);
-            });
-    });
-}
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Tp, class _BinaryOperation,
           class _UnaryOperation>
 _Tp
@@ -305,18 +236,6 @@ __brick_transform_scan(_RandomAccessIterator __first, _RandomAccessIterator __la
                                               /*is_vector=*/::std::false_type());
 }
 
-template <class _ExecutionPolicy, class _ForwardIterator, class _OutputIterator, class _UnaryOperation, class _Tp,
-          class _BinaryOperation, class _Inclusive, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _OutputIterator>
-__pattern_transform_scan(_ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
-                         _OutputIterator __result, _UnaryOperation __unary_op, _Tp __init, _BinaryOperation __binary_op,
-                         _Inclusive, _IsVector __is_vector, /*is_parallel=*/::std::false_type) noexcept
-{
-    return __internal::__brick_transform_scan(__first, __last, __result, __unary_op, __init, __binary_op, _Inclusive(),
-                                              __is_vector)
-        .first;
-}
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _OutputIterator, class _UnaryOperation,
           class _Tp, class _BinaryOperation, class _Inclusive>
 _OutputIterator
@@ -475,31 +394,6 @@ __pattern_transform_scan(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _
 }
 
 // transform_scan without initial element
-template <class _ExecutionPolicy, class _ForwardIterator, class _OutputIterator, class _UnaryOperation,
-          class _BinaryOperation, class _Inclusive, class _IsVector, class _IsParallel>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _OutputIterator>
-__pattern_transform_scan(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last,
-                         _OutputIterator __result, _UnaryOperation __unary_op, _BinaryOperation __binary_op, _Inclusive,
-                         _IsVector __is_vector, _IsParallel __is_parallel)
-{
-    typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _ValueType;
-    if (__first != __last)
-    {
-        _ValueType __tmp = __unary_op(*__first);
-        *__result = __tmp;
-
-        constexpr auto __dispatch_tag =
-            oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator, _OutputIterator>();
-
-        return __pattern_transform_scan(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), ++__first, __last,
-                                        ++__result, __unary_op, __tmp, __binary_op, _Inclusive());
-    }
-    else
-    {
-        return __result;
-    }
-}
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _OutputIterator, class _UnaryOperation,
           class _BinaryOperation, class _Inclusive>
 _OutputIterator
@@ -553,16 +447,6 @@ __brick_adjacent_difference(_RandomAccessIterator1 __first, _RandomAccessIterato
         [&__op](_ReferenceType1 __x, _ReferenceType1 __y, _ReferenceType2 __z) { __z = __op(__x, __y); });
 }
 
-template <class _ExecutionPolicy, class _ForwardIterator, class _OutputIterator, class _BinaryOperation,
-          class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _OutputIterator>
-__pattern_adjacent_difference(_ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
-                              _OutputIterator __d_first, _BinaryOperation __op, _IsVector __is_vector,
-                              /*is_parallel*/ ::std::false_type) noexcept
-{
-    return __internal::__brick_adjacent_difference(__first, __last, __d_first, __op, __is_vector);
-}
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _OutputIterator, class _BinaryOperation>
 _OutputIterator
 __pattern_adjacent_difference(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
@@ -573,34 +457,6 @@ __pattern_adjacent_difference(_Tag, _ExecutionPolicy&&, _ForwardIterator __first
     return __internal::__brick_adjacent_difference(__first, __last, __d_first, __op, typename _Tag::__is_vector{});
 }
 
-template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _BinaryOperation,
-          class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy, _RandomAccessIterator2>
-__pattern_adjacent_difference(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first, _RandomAccessIterator1 __last,
-                              _RandomAccessIterator2 __d_first, _BinaryOperation __op, _IsVector __is_vector,
-                              /*is_parallel=*/::std::true_type)
-{
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
-    assert(__first != __last);
-    typedef typename ::std::iterator_traits<_RandomAccessIterator1>::reference _ReferenceType1;
-    typedef typename ::std::iterator_traits<_RandomAccessIterator2>::reference _ReferenceType2;
-
-    *__d_first = *__first;
-    __par_backend::__parallel_for(
-        __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last - 1,
-        [&__op, __is_vector, __d_first, __first](_RandomAccessIterator1 __b, _RandomAccessIterator1 __e) {
-            _RandomAccessIterator2 __d_b = __d_first + (__b - __first);
-            __internal::__brick_walk3(
-                __b, __e, __b + 1, __d_b + 1,
-                [&__op](_ReferenceType1 __x, _ReferenceType1 __y, _ReferenceType2 __z) { __z = __op(__y, __x); },
-                __is_vector);
-        });
-    return __d_first + (__last - __first);
-}
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
           class _BinaryOperation>
 _RandomAccessIterator2

From ba4d8c75fc7b48210e54ea476d923bbd25a5d394 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 10:06:28 +0100
Subject: [PATCH 297/566] Remove old implementations with
 __enable_if_host_execution_policy_conditional

---
 include/oneapi/dpl/pstl/algorithm_fwd.h  |  62 -----------
 include/oneapi/dpl/pstl/algorithm_impl.h | 136 -----------------------
 include/oneapi/dpl/pstl/numeric_fwd.h    |  16 ---
 include/oneapi/dpl/pstl/numeric_impl.h   |  74 ------------
 4 files changed, 288 deletions(-)

diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index bcf5fa3b829..5a1dcba419c 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -72,24 +72,10 @@ template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Fun
 void
 __pattern_walk1(_Tag, _ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _Function) noexcept;
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _Function, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy_conditional<
-    _ExecutionPolicy, __is_random_access_iterator_v<_RandomAccessIterator>>
-__pattern_walk1(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIterator __last, _Function __f,
-                _IsVector __is_vector,
-                /*parallel=*/::std::true_type);
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Function>
 void
 __pattern_walk1(__parallel_forward_tag, _ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _Function);
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _Function, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy_conditional<
-    _ExecutionPolicy, !__is_random_access_iterator_v<_RandomAccessIterator>>
-__pattern_walk1(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIterator __last, _Function __f,
-                _IsVector __is_vector,
-                /*parallel=*/::std::true_type);
-
 template <class _IsVector, class _ExecutionPolicy, class _ForwardIterator, class _Function>
 void
 __pattern_walk1(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _Function);
@@ -157,26 +143,12 @@ template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _Fo
 _ForwardIterator2
 __pattern_walk2(_Tag, _ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2, _Function) noexcept;
 
-template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _Function,
-          class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy_conditional<
-    _ExecutionPolicy, __is_random_access_iterator_v<_RandomAccessIterator1, _RandomAccessIterator2>,
-    _RandomAccessIterator2>
-__pattern_walk2(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1, _RandomAccessIterator1 __last1,
-                _RandomAccessIterator2 __first2, _Function __f, _IsVector __is_vector, /*parallel=*/::std::true_type);
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
           class _Function>
 _RandomAccessIterator2
 __pattern_walk2(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator1, _RandomAccessIterator1,
                 _RandomAccessIterator2, _Function);
 
-template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Function, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy_conditional<
-    _ExecutionPolicy, !__is_random_access_iterator_v<_ForwardIterator1, _ForwardIterator2>, _ForwardIterator2>
-__pattern_walk2(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
-                _ForwardIterator2 __first2, _Function __f, _IsVector __is_vector, /*parallel=*/::std::true_type);
-
 template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Function>
 _ForwardIterator2
 __pattern_walk2(__parallel_forward_tag, _ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2,
@@ -198,26 +170,12 @@ _ForwardIterator2
 __pattern_walk2_brick(_Tag, _ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2,
                       _Brick) noexcept;
 
-template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _Brick>
-oneapi::dpl::__internal::__enable_if_host_execution_policy_conditional<
-    _ExecutionPolicy, __is_random_access_iterator_v<_RandomAccessIterator1, _RandomAccessIterator2>,
-    _RandomAccessIterator2>
-__pattern_walk2_brick(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1, _RandomAccessIterator1 __last1,
-                      _RandomAccessIterator2 __first2, _Brick __brick, /*parallel=*/::std::true_type);
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
           class _Brick>
 _RandomAccessIterator2
 __pattern_walk2_brick(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator1, _RandomAccessIterator1,
                       _RandomAccessIterator2, _Brick);
 
-template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _Brick>
-oneapi::dpl::__internal::__enable_if_host_execution_policy_conditional<
-    _ExecutionPolicy, !__is_random_access_iterator_v<_RandomAccessIterator1, _RandomAccessIterator2>,
-    _RandomAccessIterator2>
-__pattern_walk2_brick(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1, _RandomAccessIterator1 __last1,
-                      _RandomAccessIterator2 __first2, _Brick __brick, /*parallel=*/::std::true_type);
-
 template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Brick>
 _ForwardIterator2
 __pattern_walk2_brick(__parallel_forward_tag, _ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1,
@@ -255,32 +213,12 @@ _ForwardIterator3
 __pattern_walk3(_Tag, _ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2, _ForwardIterator3,
                 _Function) noexcept;
 
-template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
-          class _RandomAccessIterator3, class _Function, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy_conditional<
-    _ExecutionPolicy,
-    __is_random_access_iterator_v<_RandomAccessIterator1, _RandomAccessIterator2, _RandomAccessIterator3>,
-    _RandomAccessIterator3>
-__pattern_walk3(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1, _RandomAccessIterator1 __last1,
-                _RandomAccessIterator2 __first2, _RandomAccessIterator3 __first3, _Function __f, _IsVector __is_vector,
-                /*parallel=*/::std::true_type);
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
           class _RandomAccessIterator3, class _Function>
 _RandomAccessIterator3
 __pattern_walk3(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator1, _RandomAccessIterator1,
                 _RandomAccessIterator2, _RandomAccessIterator3, _Function);
 
-template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
-          class _RandomAccessIterator3, class _Function, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy_conditional<
-    _ExecutionPolicy,
-    !__is_random_access_iterator_v<_RandomAccessIterator1, _RandomAccessIterator2, _RandomAccessIterator3>,
-    _RandomAccessIterator3>
-__pattern_walk3(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1, _RandomAccessIterator1 __last1,
-                _RandomAccessIterator2 __first2, _RandomAccessIterator3 __first3, _Function __f, _IsVector __is_vector,
-                /*parallel=*/::std::true_type);
-
 template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _ForwardIterator3,
           class _Function>
 _ForwardIterator3
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 11c5d758454..c5d96eacaad 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -326,28 +326,6 @@ __pattern_walk2(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIt
     return __internal::__brick_walk2(__first1, __last1, __first2, __f, typename _Tag::__is_vector{});
 }
 
-template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2, class _Function,
-          class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy_conditional<
-    _ExecutionPolicy, __is_random_access_iterator_v<_RandomAccessIterator1, _RandomAccessIterator2>,
-    _RandomAccessIterator2>
-__pattern_walk2(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1, _RandomAccessIterator1 __last1,
-                _RandomAccessIterator2 __first2, _Function __f, _IsVector __is_vector, /*parallel=*/::std::true_type)
-{
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
-    return __internal::__except_handler([&]() {
-        __par_backend::__parallel_for(
-            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1,
-            [__f, __first1, __first2, __is_vector](_RandomAccessIterator1 __i, _RandomAccessIterator1 __j) {
-                __internal::__brick_walk2(__i, __j, __first2 + (__i - __first1), __f, __is_vector);
-            });
-        return __first2 + (__last1 - __first1);
-    });
-}
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
           class _Function>
 _RandomAccessIterator2
@@ -366,33 +344,6 @@ __pattern_walk2(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAcc
     });
 }
 
-template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Function, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy_conditional<
-    _ExecutionPolicy, !__is_random_access_iterator_v<_ForwardIterator1, _ForwardIterator2>, _ForwardIterator2>
-__pattern_walk2(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
-                _ForwardIterator2 __first2, _Function __f, _IsVector, /*parallel=*/::std::true_type)
-{
-    return __internal::__except_handler([&]() {
-        using _iterator_tuple = zip_forward_iterator<_ForwardIterator1, _ForwardIterator2>;
-        auto __begin = _iterator_tuple(__first1, __first2);
-        auto __end = _iterator_tuple(__last1, /*dummy parameter*/ _ForwardIterator2());
-
-        typedef typename ::std::iterator_traits<_ForwardIterator1>::reference _ReferenceType1;
-        typedef typename ::std::iterator_traits<_ForwardIterator2>::reference _ReferenceType2;
-
-        __par_backend::__parallel_for_each(::std::forward<_ExecutionPolicy>(__exec), __begin, __end,
-                                           [&__f](::std::tuple<_ReferenceType1, _ReferenceType2> __val) {
-                                               __f(::std::get<0>(__val), ::std::get<1>(__val));
-                                           });
-
-        //TODO: parallel_for_each does not allow to return correct iterator value according to the ::std::transform
-        // implementation. Therefore, iterator value is calculated separately.
-        for (; __begin != __end; ++__begin)
-            ;
-        return ::std::get<1>(__begin.base());
-    });
-}
-
 template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Function>
 _ForwardIterator2
 __pattern_walk2(__parallel_forward_tag, _ExecutionPolicy&& __exec, _ForwardIterator1 __first1,
@@ -470,35 +421,6 @@ __pattern_walk2_brick(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Ran
     });
 }
 
-//TODO: it postponed till adding more or less effective parallel implementation
-template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Brick>
-oneapi::dpl::__internal::__enable_if_host_execution_policy_conditional<
-    _ExecutionPolicy, !__is_random_access_iterator_v<_ForwardIterator1, _ForwardIterator2>, _ForwardIterator2>
-__pattern_walk2_brick(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
-                      _ForwardIterator2 __first2, _Brick __brick, /*parallel=*/::std::true_type)
-{
-    using _iterator_tuple = zip_forward_iterator<_ForwardIterator1, _ForwardIterator2>;
-    auto __begin = _iterator_tuple(__first1, __first2);
-    auto __end = _iterator_tuple(__last1, /*dummy parameter*/ _ForwardIterator2());
-
-    typedef typename ::std::iterator_traits<_ForwardIterator1>::reference _ReferenceType1;
-    typedef typename ::std::iterator_traits<_ForwardIterator2>::reference _ReferenceType2;
-
-    return __except_handler([&]() {
-        __par_backend::__parallel_for_each(::std::forward<_ExecutionPolicy>(__exec), __begin, __end,
-                                           [__brick](::std::tuple<_ReferenceType1, _ReferenceType2> __val) {
-                                               __brick(::std::get<0>(__val),
-                                                       ::std::forward<_ReferenceType2>(::std::get<1>(__val)));
-                                           });
-
-        //TODO: parallel_for_each does not allow to return correct iterator value according to the ::std::transform
-        // implementation. Therefore, iterator value is calculated separately.
-        for (; __begin != __end; ++__begin)
-            ;
-        return ::std::get<1>(__begin.base());
-    });
-}
-
 //TODO: it postponed till adding more or less effective parallel implementation
 template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Brick>
 _ForwardIterator2
@@ -592,32 +514,6 @@ __pattern_walk3(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIt
     return __internal::__brick_walk3(__first1, __last1, __first2, __first3, __f, typename _Tag::__is_vector{});
 }
 
-template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
-          class _RandomAccessIterator3, class _Function, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy_conditional<
-    _ExecutionPolicy,
-    __is_random_access_iterator_v<_RandomAccessIterator1, _RandomAccessIterator2, _RandomAccessIterator3>,
-    _RandomAccessIterator3>
-__pattern_walk3(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1, _RandomAccessIterator1 __last1,
-                _RandomAccessIterator2 __first2, _RandomAccessIterator3 __first3, _Function __f, _IsVector __is_vector,
-                /*parallel=*/::std::true_type)
-{
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2,
-                                                  _RandomAccessIterator3>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
-    return __internal::__except_handler([&]() {
-        __par_backend::__parallel_for(
-            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1,
-            [__f, __first1, __first2, __first3, __is_vector](_RandomAccessIterator1 __i, _RandomAccessIterator1 __j) {
-                __internal::__brick_walk3(__i, __j, __first2 + (__i - __first1), __first3 + (__i - __first1), __f,
-                                          __is_vector);
-            });
-        return __first3 + (__last1 - __first1);
-    });
-}
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
           class _RandomAccessIterator3, class _Function>
 _RandomAccessIterator3
@@ -638,38 +534,6 @@ __pattern_walk3(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAcc
     });
 }
 
-template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _ForwardIterator3,
-          class _Function, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy_conditional<
-    _ExecutionPolicy, !__is_random_access_iterator_v<_ForwardIterator1, _ForwardIterator2, _ForwardIterator3>,
-    _ForwardIterator3>
-__pattern_walk3(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
-                _ForwardIterator2 __first2, _ForwardIterator3 __first3, _Function __f, _IsVector,
-                /*parallel=*/::std::true_type)
-{
-    return __internal::__except_handler([&]() {
-        using _iterator_tuple = zip_forward_iterator<_ForwardIterator1, _ForwardIterator2, _ForwardIterator3>;
-        auto __begin = _iterator_tuple(__first1, __first2, __first3);
-        auto __end = _iterator_tuple(__last1, /*dummy parameter*/ _ForwardIterator2(),
-                                     /*dummy parameter*/ _ForwardIterator3());
-
-        typedef typename ::std::iterator_traits<_ForwardIterator1>::reference _ReferenceType1;
-        typedef typename ::std::iterator_traits<_ForwardIterator2>::reference _ReferenceType2;
-        typedef typename ::std::iterator_traits<_ForwardIterator3>::reference _ReferenceType3;
-
-        __par_backend::__parallel_for_each(::std::forward<_ExecutionPolicy>(__exec), __begin, __end,
-                                           [&](::std::tuple<_ReferenceType1, _ReferenceType2, _ReferenceType3> __val) {
-                                               __f(::std::get<0>(__val), ::std::get<1>(__val), ::std::get<2>(__val));
-                                           });
-
-        //TODO: parallel_for_each does not allow to return correct iterator value according to the ::std::transform
-        // implementation. Therefore, iterator value is calculated separately.
-        for (; __begin != __end; ++__begin)
-            ;
-        return ::std::get<2>(__begin.base());
-    });
-}
-
 template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _ForwardIterator3,
           class _Function>
 _ForwardIterator3
diff --git a/include/oneapi/dpl/pstl/numeric_fwd.h b/include/oneapi/dpl/pstl/numeric_fwd.h
index 46826aff9d3..9ea0ebb1de4 100644
--- a/include/oneapi/dpl/pstl/numeric_fwd.h
+++ b/include/oneapi/dpl/pstl/numeric_fwd.h
@@ -113,28 +113,12 @@ _OutputIterator
 __pattern_transform_scan(_Tag, _ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _OutputIterator, _UnaryOperation,
                          _Tp, _BinaryOperation, _Inclusive) noexcept;
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _OutputIterator, class _UnaryOperation, class _Tp,
-          class _BinaryOperation, class _Inclusive, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy_conditional<
-    _ExecutionPolicy, !::std::is_floating_point_v<_Tp>, _OutputIterator>
-__pattern_transform_scan(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _OutputIterator,
-                         _UnaryOperation, _Tp, _BinaryOperation, _Inclusive, _IsVector,
-                         /*is_parallel=*/::std::true_type);
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _OutputIterator,
           class _UnaryOperation, class _Tp, class _BinaryOperation, class _Inclusive>
 ::std::enable_if_t<!::std::is_floating_point_v<_Tp>, _OutputIterator>
 __pattern_transform_scan(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator,
                          _OutputIterator, _UnaryOperation, _Tp, _BinaryOperation, _Inclusive);
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _OutputIterator, class _UnaryOperation, class _Tp,
-          class _BinaryOperation, class _Inclusive, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy_conditional<_ExecutionPolicy,
-                                                                       ::std::is_floating_point_v<_Tp>, _OutputIterator>
-__pattern_transform_scan(_ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _OutputIterator,
-                         _UnaryOperation, _Tp, _BinaryOperation, _Inclusive, _IsVector,
-                         /*is_parallel=*/::std::true_type);
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _OutputIterator,
           class _UnaryOperation, class _Tp, class _BinaryOperation, class _Inclusive>
 ::std::enable_if_t<::std::is_floating_point_v<_Tp>, _OutputIterator>
diff --git a/include/oneapi/dpl/pstl/numeric_impl.h b/include/oneapi/dpl/pstl/numeric_impl.h
index 92ca677bc8d..b875883a598 100644
--- a/include/oneapi/dpl/pstl/numeric_impl.h
+++ b/include/oneapi/dpl/pstl/numeric_impl.h
@@ -250,38 +250,6 @@ __pattern_transform_scan(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _Fo
         .first;
 }
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _OutputIterator, class _UnaryOperation, class _Tp,
-          class _BinaryOperation, class _Inclusive, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy_conditional<
-    _ExecutionPolicy, !::std::is_floating_point_v<_Tp>, _OutputIterator>
-__pattern_transform_scan(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIterator __last,
-                         _OutputIterator __result, _UnaryOperation __unary_op, _Tp __init, _BinaryOperation __binary_op,
-                         _Inclusive, _IsVector __is_vector, /*is_parallel=*/::std::true_type)
-{
-    typedef typename ::std::iterator_traits<_RandomAccessIterator>::difference_type _DifferenceType;
-
-    return __internal::__except_handler([&]() {
-        __par_backend::__parallel_transform_scan(
-            oneapi::dpl::__internal::__serial_backend_tag{},
-            ::std::forward<_ExecutionPolicy>(__exec), __last - __first,
-            [__first, __unary_op](_DifferenceType __i) mutable { return __unary_op(__first[__i]); }, __init,
-            __binary_op,
-            [__first, __unary_op, __binary_op](_DifferenceType __i, _DifferenceType __j, _Tp __init) {
-                // Execute serial __brick_transform_reduce, due to the explicit SIMD vectorization (reduction) requires a commutative operation for the guarantee of correct scan.
-                return __internal::__brick_transform_reduce(__first + __i, __first + __j, __init, __binary_op,
-                                                            __unary_op,
-                                                            /*__is_vector*/ ::std::false_type());
-            },
-            [__first, __unary_op, __binary_op, __result, __is_vector](_DifferenceType __i, _DifferenceType __j,
-                                                                      _Tp __init) {
-                return __internal::__brick_transform_scan(__first + __i, __first + __j, __result + __i, __unary_op,
-                                                          __init, __binary_op, _Inclusive(), __is_vector)
-                    .second;
-            });
-        return __result + (__last - __first);
-    });
-}
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _OutputIterator,
           class _UnaryOperation, class _Tp, class _BinaryOperation, class _Inclusive>
 ::std::enable_if_t<!::std::is_floating_point_v<_Tp>, _OutputIterator>
@@ -312,48 +280,6 @@ __pattern_transform_scan(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __e
     });
 }
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _OutputIterator, class _UnaryOperation, class _Tp,
-          class _BinaryOperation, class _Inclusive, class _IsVector>
-oneapi::dpl::__internal::__enable_if_host_execution_policy_conditional<_ExecutionPolicy,
-                                                                       ::std::is_floating_point_v<_Tp>, _OutputIterator>
-__pattern_transform_scan(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIterator __last,
-                         _OutputIterator __result, _UnaryOperation __unary_op, _Tp __init, _BinaryOperation __binary_op,
-                         _Inclusive, _IsVector __is_vector, /*is_parallel=*/::std::true_type)
-{
-    typedef typename ::std::iterator_traits<_RandomAccessIterator>::difference_type _DifferenceType;
-    _DifferenceType __n = __last - __first;
-
-    if (__n <= 0)
-    {
-        return __result;
-    }
-
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator, _OutputIterator>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
-    return __internal::__except_handler([&]() {
-        __par_backend::__parallel_strict_scan(
-            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __n, __init,
-            [__first, __unary_op, __binary_op, __result, __is_vector](_DifferenceType __i, _DifferenceType __len) {
-                return __internal::__brick_transform_scan(__first + __i, __first + (__i + __len), __result + __i,
-                                                          __unary_op, _Tp{}, __binary_op, _Inclusive(), __is_vector)
-                    .second;
-            },
-            __binary_op,
-            [__result, &__binary_op](_DifferenceType __i, _DifferenceType __len, _Tp __initial) {
-                return *(::std::transform(__result + __i, __result + __i + __len, __result + __i,
-                                          [&__initial, &__binary_op](const _Tp& __x) {
-                                              _ONEDPL_PRAGMA_FORCEINLINE
-                                              return __binary_op(__initial, __x);
-                                          }) -
-                         1);
-            },
-            [](_Tp) {});
-        return __result + (__last - __first);
-    });
-}
-
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _OutputIterator,
           class _UnaryOperation, class _Tp, class _BinaryOperation, class _Inclusive>
 ::std::enable_if_t<::std::is_floating_point_v<_Tp>, _OutputIterator>

From 9aeec7bbfb06ed327c0a3fca6e8c01aa97227f55 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 10:12:57 +0100
Subject: [PATCH 298/566] include/oneapi/dpl/pstl/execution_defs.h - remove
 __enable_if_host_execution_policy_conditional as not required anymore

---
 include/oneapi/dpl/pstl/execution_defs.h | 4 ----
 1 file changed, 4 deletions(-)

diff --git a/include/oneapi/dpl/pstl/execution_defs.h b/include/oneapi/dpl/pstl/execution_defs.h
index 5001ec9fd0f..6d1d0efd569 100644
--- a/include/oneapi/dpl/pstl/execution_defs.h
+++ b/include/oneapi/dpl/pstl/execution_defs.h
@@ -184,10 +184,6 @@ template <class _ExecPolicy, class _T = void>
 using __enable_if_host_execution_policy =
     ::std::enable_if_t<__is_host_execution_policy<::std::decay_t<_ExecPolicy>>::value, _T>;
 
-template <class _ExecPolicy, const bool __condition, class _T = void>
-using __enable_if_host_execution_policy_conditional =
-    ::std::enable_if_t<__is_host_execution_policy<::std::decay_t<_ExecPolicy>>::value && __condition, _T>;
-
 template <typename _ExecPolicy, typename _T>
 struct __ref_or_copy_impl
 {

From b8eb46bb9e743cd5cb91cf4251edc64b31334fd6 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 15 Feb 2024 18:45:41 +0100
Subject: [PATCH 299/566] Remove old implementations with
 __enable_if_hetero_execution_policy

---
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 1776 +----------------
 .../dpl/pstl/hetero/numeric_impl_hetero.h     |  206 --
 2 files changed, 38 insertions(+), 1944 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 2b7108c7ba3..c451be81c52 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -39,28 +39,6 @@ namespace __internal
 // walk1
 //------------------------------------------------------------------------
 
-template <typename _ExecutionPolicy, typename _ForwardIterator, typename _Function>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy>
-__pattern_walk1(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last, _Function __f,
-                /*vector=*/::std::true_type, /*parallel=*/::std::true_type)
-{
-    auto __n = __last - __first;
-    if (__n <= 0)
-        return;
-
-    auto __keep =
-        oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read_write, _ForwardIterator>();
-    auto __buf = __keep(__first, __last);
-
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
-    oneapi::dpl::__par_backend_hetero::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
-                                                      unseq_backend::walk_n<_ExecutionPolicy, _Function>{__f}, __n,
-                                                      __buf.all_view())
-        .wait();
-}
-
 template <typename _BackendTag, typename _ExecutionPolicy, typename _ForwardIterator, typename _Function>
 void
 __pattern_walk1(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last,
@@ -83,17 +61,6 @@ __pattern_walk1(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _ForwardIt
 // walk1_n
 //------------------------------------------------------------------------
 
-template <typename _ExecutionPolicy, typename _ForwardIterator, typename _Size, typename _Function>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, _ForwardIterator>
-__pattern_walk1_n(_ExecutionPolicy&& __exec, _ForwardIterator __first, _Size __n, _Function __f,
-                  /*vector=*/::std::true_type, /*parallel=*/::std::true_type)
-{
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
-
-    __pattern_walk1(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __first + __n, __f);
-    return __first + __n;
-}
-
 template <typename _BackendTag, typename _ExecutionPolicy, typename _ForwardIterator, typename _Size,
           typename _Function>
 _ForwardIterator
@@ -111,38 +78,6 @@ __pattern_walk1_n(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _F
 // TODO: A tag _IsSync is used for provide a patterns call pipeline, where the last one should be synchronous
 // Probably it should be re-designed by a pipeline approach, when a pattern returns some sync obejects
 // and ones are combined into a "pipeline" (probably like Range pipeline)
-template <typename _IsSync = ::std::true_type,
-          __par_backend_hetero::access_mode __acc_mode1 = __par_backend_hetero::access_mode::read,
-          __par_backend_hetero::access_mode __acc_mode2 = __par_backend_hetero::access_mode::write,
-          typename _ExecutionPolicy, typename _ForwardIterator1, typename _ForwardIterator2, typename _Function>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, _ForwardIterator2>
-__pattern_walk2(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
-                _ForwardIterator2 __first2, _Function __f, /*vector=*/::std::true_type, /*parallel=*/::std::true_type)
-{
-    auto __n = __last1 - __first1;
-    if (__n <= 0)
-        return __first2;
-
-    auto __keep1 = oneapi::dpl::__ranges::__get_sycl_range<__acc_mode1, _ForwardIterator1>();
-    auto __buf1 = __keep1(__first1, __last1);
-
-    auto __keep2 = oneapi::dpl::__ranges::__get_sycl_range<__acc_mode2, _ForwardIterator2>();
-    auto __buf2 = __keep2(__first2, __first2 + __n);
-
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
-    auto __future_obj = oneapi::dpl::__par_backend_hetero::__parallel_for(
-        __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
-        unseq_backend::walk_n<_ExecutionPolicy, _Function>{__f}, __n, __buf1.all_view(), __buf2.all_view());
-
-    if constexpr (_IsSync())
-        __future_obj.wait();
-
-    return __first2 + __n;
-}
-
 template <typename _BackendTag, typename _IsSync = ::std::true_type,
           __par_backend_hetero::access_mode __acc_mode1 = __par_backend_hetero::access_mode::read,
           __par_backend_hetero::access_mode __acc_mode2 = __par_backend_hetero::access_mode::write,
@@ -171,16 +106,6 @@ __pattern_walk2(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _ForwardIt
     return __first2 + __n;
 }
 
-template <typename _ExecutionPolicy, typename _ForwardIterator1, typename _Size, typename _ForwardIterator2,
-          typename _Function>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, _ForwardIterator2>
-__pattern_walk2_n(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _Size __n, _ForwardIterator2 __first2,
-                  _Function __f, /*vector=*/::std::true_type, /*parallel=*/::std::true_type)
-{
-    return __pattern_walk2(::std::forward<_ExecutionPolicy>(__exec), __first1, __first1 + __n, __first2, __f,
-                           ::std::true_type(), ::std::true_type());
-}
-
 template <typename _BackendTag, typename _ExecutionPolicy, typename _ForwardIterator1, typename _Size,
           typename _ForwardIterator2, typename _Function>
 _ForwardIterator2
@@ -194,22 +119,6 @@ __pattern_walk2_n(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _F
 // swap
 //------------------------------------------------------------------------
 
-template <typename _ExecutionPolicy, typename _ForwardIterator1, typename _ForwardIterator2, typename _Function>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, _ForwardIterator2>
-__pattern_swap(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
-               _ForwardIterator2 __first2, _Function __f, /*is_vector=*/::std::true_type,
-               /*is_parallel=*/::std::true_type)
-{
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
-
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
-    return __pattern_walk2<__backend_tag, /*_IsSync=*/::std::true_type, __par_backend_hetero::access_mode::read_write,
-                           __par_backend_hetero::access_mode::read_write>(
-        __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __f);
-}
-
 template <typename _BackendTag, typename _ExecutionPolicy, typename _ForwardIterator1, typename _ForwardIterator2,
           typename _Function>
 _ForwardIterator2
@@ -225,39 +134,6 @@ __pattern_swap(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Forw
 // walk3
 //------------------------------------------------------------------------
 
-template <__par_backend_hetero::access_mode __acc_mode1 = __par_backend_hetero::access_mode::read,
-          __par_backend_hetero::access_mode __acc_mode2 = __par_backend_hetero::access_mode::read,
-          __par_backend_hetero::access_mode __acc_mode3 = __par_backend_hetero::access_mode::write,
-          typename _ExecutionPolicy, typename _ForwardIterator1, typename _ForwardIterator2, typename _ForwardIterator3,
-          typename _Function>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, _ForwardIterator3>
-__pattern_walk3(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
-                _ForwardIterator2 __first2, _ForwardIterator3 __first3, _Function __f, /*vector=*/::std::true_type,
-                /*parallel=*/::std::true_type)
-{
-    auto __n = __last1 - __first1;
-    if (__n <= 0)
-        return __first3;
-
-    auto __keep1 = oneapi::dpl::__ranges::__get_sycl_range<__acc_mode1, _ForwardIterator1>();
-    auto __buf1 = __keep1(__first1, __last1);
-    auto __keep2 = oneapi::dpl::__ranges::__get_sycl_range<__acc_mode2, _ForwardIterator2>();
-    auto __buf2 = __keep2(__first2, __first2 + __n);
-    auto __keep3 = oneapi::dpl::__ranges::__get_sycl_range<__acc_mode3, _ForwardIterator3>();
-    auto __buf3 = __keep3(__first3, __first3 + __n);
-
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1,
-                                                                              _ForwardIterator2, _ForwardIterator3>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
-    oneapi::dpl::__par_backend_hetero::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
-                                                      unseq_backend::walk_n<_ExecutionPolicy, _Function>{__f}, __n,
-                                                      __buf1.all_view(), __buf2.all_view(), __buf3.all_view())
-        .wait();
-
-    return __first3 + __n;
-}
-
 template <typename _BackendTag, __par_backend_hetero::access_mode __acc_mode1 = __par_backend_hetero::access_mode::read,
           __par_backend_hetero::access_mode __acc_mode2 = __par_backend_hetero::access_mode::read,
           __par_backend_hetero::access_mode __acc_mode3 = __par_backend_hetero::access_mode::write,
@@ -295,22 +171,6 @@ struct __walk_brick_wrapper
 {
 };
 
-template <typename _ExecutionPolicy, typename _ForwardIterator, typename _Function>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy>
-__pattern_walk_brick(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last, _Function __f,
-                     /*parallel=*/::std::true_type)
-{
-    if (__last - __first <= 0)
-        return;
-
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
-
-    __pattern_walk1(
-        __dispatch_tag,
-        __par_backend_hetero::make_wrapped_policy<__walk_brick_wrapper>(::std::forward<_ExecutionPolicy>(__exec)),
-        __first, __last, __f);
-}
-
 template <typename _BackendTag, typename _ExecutionPolicy, typename _ForwardIterator, typename _Function>
 void
 __pattern_walk_brick(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _ForwardIterator __first,
@@ -330,20 +190,6 @@ struct __walk_brick_n_wrapper
 {
 };
 
-template <typename _ExecutionPolicy, typename _ForwardIterator, typename _Size, typename _Function>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, _ForwardIterator>
-__pattern_walk_brick_n(_ExecutionPolicy&& __exec, _ForwardIterator __first, _Size __n, _Function __f,
-                       /*parallel=*/::std::true_type)
-{
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
-
-    __pattern_walk1(
-        __dispatch_tag,
-        __par_backend_hetero::make_wrapped_policy<__walk_brick_n_wrapper>(::std::forward<_ExecutionPolicy>(__exec)),
-        __first, __first + __n, __f);
-    return __first + __n;
-}
-
 template <typename _BackendTag, typename _ExecutionPolicy, typename _ForwardIterator, typename _Size,
           typename _Function>
 _ForwardIterator
@@ -366,17 +212,6 @@ struct __walk2_brick_wrapper
 {
 };
 
-template <typename _ExecutionPolicy, typename _ForwardIterator1, typename _ForwardIterator2, typename _Brick>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, _ForwardIterator2>
-__pattern_walk2_brick(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
-                      _ForwardIterator2 __first2, _Brick __brick, /*parallel*/ ::std::true_type)
-{
-    return __pattern_walk2(
-        __par_backend_hetero::make_wrapped_policy<__walk2_brick_wrapper>(::std::forward<_ExecutionPolicy>(__exec)),
-        __first1, __last1, __first2, __brick,
-        /*vector=*/::std::true_type{}, /*parallel*/ ::std::true_type{});
-}
-
 template <typename _BackendTag, typename _ExecutionPolicy, typename _ForwardIterator1, typename _ForwardIterator2,
           typename _Brick>
 _ForwardIterator2
@@ -394,19 +229,6 @@ struct __walk2_brick_n_wrapper
 {
 };
 
-template <typename _ExecutionPolicy, typename _ForwardIterator1, typename _Size, typename _ForwardIterator2,
-          typename _Brick>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, _ForwardIterator2>
-__pattern_walk2_brick_n(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _Size __n, _ForwardIterator2 __first2,
-                        _Brick __brick, /*parallel*/ ::std::true_type)
-{
-
-    return __pattern_walk2(
-        __par_backend_hetero::make_wrapped_policy<__walk2_brick_n_wrapper>(::std::forward<_ExecutionPolicy>(__exec)),
-        __first1, __first1 + __n, __first2, __brick,
-        /*vector=*/::std::true_type{}, /*parallel*/ ::std::true_type{});
-}
-
 template <typename _BackendTag, typename _ExecutionPolicy, typename _ForwardIterator1, typename _Size,
           typename _ForwardIterator2, typename _Brick>
 _ForwardIterator2
@@ -428,27 +250,6 @@ struct __walk2_transform_if_wrapper
 {
 };
 
-template <typename _ExecutionPolicy, typename _ForwardIterator1, typename _ForwardIterator2, typename _Function>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, _ForwardIterator2>
-__pattern_walk2_transform_if(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
-                             _ForwardIterator2 __first2, _Function __func,
-                             /*vector=*/::std::true_type,
-                             /*parallel=*/::std::true_type)
-{
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
-
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
-    // Require `read_write` access mode for output sequence to force a copy in for host iterators to capture incoming
-    // values of the output sequence for elements where the predicate is false.
-    return __pattern_walk2<__backend_tag, /*_IsSync=*/::std::true_type, __par_backend_hetero::access_mode::read,
-                           __par_backend_hetero::access_mode::read_write>(
-        __par_backend_hetero::make_wrapped_policy<__walk2_transform_if_wrapper>(
-            __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec)),
-        __first1, __last1, __first2, __func);
-}
-
 template <typename _BackendTag, typename _ExecutionPolicy, typename _ForwardIterator1, typename _ForwardIterator2,
           typename _Function>
 _ForwardIterator2
@@ -470,24 +271,6 @@ struct __walk3_transform_if_wrapper
 {
 };
 
-template <typename _ExecutionPolicy, typename _ForwardIterator1, typename _ForwardIterator2, typename _ForwardIterator3,
-          typename _Function>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, _ForwardIterator3>
-__pattern_walk3_transform_if(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
-                             _ForwardIterator2 __first2, _ForwardIterator3 __first3, _Function __func,
-                             /*vector=*/::std::true_type,
-                             /*parallel=*/::std::true_type)
-{
-    // Require `read_write` access mode for output sequence to force a copy in for host iterators to capture incoming
-    // values of the output sequence for elements where the predicate is false.
-    return __pattern_walk3<__par_backend_hetero::access_mode::read, __par_backend_hetero::access_mode::read,
-                           __par_backend_hetero::access_mode::read_write>(
-        __par_backend_hetero::make_wrapped_policy<__walk3_transform_if_wrapper>(
-            ::std::forward<_ExecutionPolicy>(__exec)),
-        __first1, __last1, __first2, __first3, __func,
-        /*vector=*/::std::true_type{}, /*parallel*/ ::std::true_type{});
-}
-
 template <typename _BackendTag, typename _ExecutionPolicy, typename _ForwardIterator1, typename _ForwardIterator2,
           typename _ForwardIterator3, typename _Function>
 _ForwardIterator3
@@ -521,20 +304,6 @@ struct fill_functor
     }
 };
 
-template <typename _ExecutionPolicy, typename _ForwardIterator, typename _T>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, _ForwardIterator>
-__pattern_fill(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last, const _T& __value,
-               /*vector=*/::std::true_type, /*parallel=*/::std::true_type)
-{
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
-
-    __pattern_walk1(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
-                    __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::write>(__first),
-                    __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::write>(__last),
-                    fill_functor<_T>{__value});
-    return __last;
-}
-
 template <typename _BackendTag, typename _ExecutionPolicy, typename _ForwardIterator, typename _T>
 _ForwardIterator
 __pattern_fill(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _ForwardIterator __first,
@@ -577,20 +346,6 @@ struct generate_functor
     }
 };
 
-template <typename _ExecutionPolicy, typename _ForwardIterator, typename _Generator>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, _ForwardIterator>
-__pattern_generate(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last, _Generator __g,
-                   /*vector=*/::std::true_type, /*parallel=*/::std::true_type)
-{
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
-
-    __pattern_walk1(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
-                    __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::write>(__first),
-                    __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::write>(__last),
-                    generate_functor<_Generator>{__g});
-    return __last;
-}
-
 template <typename _BackendTag, typename _ExecutionPolicy, typename _ForwardIterator, typename _Generator>
 _ForwardIterator
 __pattern_generate(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _ForwardIterator __first,
@@ -683,64 +438,6 @@ struct __brick_fill_n<_SourceT, _ExecutionPolicy,
 // min_element, max_element
 //------------------------------------------------------------------------
 
-template <typename _ExecutionPolicy, typename _Iterator, typename _Compare>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, _Iterator>
-__pattern_min_element(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last, _Compare __comp,
-                      /*vector*/ ::std::true_type, /*parallel*/ ::std::true_type)
-{
-    if (__first == __last)
-        return __last;
-
-    using _IteratorValueType = typename ::std::iterator_traits<_Iterator>::value_type;
-    using _IndexValueType = ::std::make_unsigned_t<typename ::std::iterator_traits<_Iterator>::difference_type>;
-    using _ReduceValueType = tuple<_IndexValueType, _IteratorValueType>;
-    // Commutativity of the reduction operator depends on the compilation target (see __reduce_fn below);
-    // __spirv_target_conditional postpones deciding on commutativity to the device code where the
-    // target can be correctly tested.
-    using _Commutative = oneapi::dpl::__internal::__spirv_target_conditional</*_SpirvT*/ ::std::false_type,
-                                                                             /*_NonSpirvT*/ ::std::true_type>;
-    auto __reduce_fn = [__comp](_ReduceValueType __a, _ReduceValueType __b) {
-        using ::std::get;
-        // TODO: Consider removing the non-commutative operator for SPIR-V targets when we see improved performance with the
-        // non-sequential load path in transform_reduce.
-        if constexpr (oneapi::dpl::__internal::__is_spirv_target_v)
-        {
-            // This operator doesn't track the lowest found index in case of equal min. or max. values. Thus, this operator is
-            // not commutative.
-            if (__comp(get<1>(__b), get<1>(__a)))
-            {
-                return __b;
-            }
-            return __a;
-        }
-        else
-        {
-            // This operator keeps track of the lowest found index in case of equal min. or max. values. Thus, this operator is
-            // commutative.
-            bool _is_a_lt_b = __comp(get<1>(__a), get<1>(__b));
-            bool _is_b_lt_a = __comp(get<1>(__b), get<1>(__a));
-
-            if (_is_b_lt_a || (!_is_a_lt_b && get<0>(__b) < get<0>(__a)))
-            {
-                return __b;
-            }
-            return __a;
-        }
-    };
-    auto __transform_fn = [](auto __gidx, auto __acc) { return _ReduceValueType{__gidx, __acc[__gidx]}; };
-
-    auto __keep = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _Iterator>();
-    auto __buf = __keep(__first, __last);
-
-    auto __ret_idx = oneapi::dpl::__par_backend_hetero::__parallel_transform_reduce<_ReduceValueType, _Commutative>(
-                         ::std::forward<_ExecutionPolicy>(__exec), __reduce_fn, __transform_fn,
-                         unseq_backend::__no_init_value{}, // no initial value
-                         __buf.all_view())
-                         .get();
-
-    return __first + ::std::get<0>(__ret_idx);
-}
-
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator, typename _Compare>
 _Iterator
 __pattern_min_element(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last,
@@ -817,56 +514,6 @@ __pattern_min_element(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec
 //   However the solution requires use of custom pattern or substantial redesign of existing parallel_transform_reduce.
 //
 
-template <typename _ExecutionPolicy, typename _Iterator, typename _Compare>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, ::std::pair<_Iterator, _Iterator>>
-__pattern_minmax_element(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last, _Compare __comp,
-                         /*vector*/ ::std::true_type, /*parallel*/ ::std::true_type)
-{
-    if (__first == __last)
-        return ::std::make_pair(__first, __first);
-
-    using _IteratorValueType = typename ::std::iterator_traits<_Iterator>::value_type;
-    using _IndexValueType = ::std::make_unsigned_t<typename ::std::iterator_traits<_Iterator>::difference_type>;
-    using _ReduceValueType = ::std::tuple<_IndexValueType, _IndexValueType, _IteratorValueType, _IteratorValueType>;
-
-    // This operator doesn't track the lowest found index in case of equal min. values and the highest found index in
-    // case of equal max. values. Thus, this operator is not commutative.
-    auto __reduce_fn = [__comp](_ReduceValueType __a, _ReduceValueType __b) {
-        using ::std::get;
-        auto __chosen_for_min = __a;
-        auto __chosen_for_max = __b;
-
-        assert(get<0>(__a) < get<0>(__b));
-        assert(get<1>(__a) < get<1>(__b));
-
-        if (__comp(get<2>(__b), get<2>(__a)))
-            __chosen_for_min = ::std::move(__b);
-        if (__comp(get<3>(__b), get<3>(__a)))
-            __chosen_for_max = ::std::move(__a);
-        return _ReduceValueType{get<0>(__chosen_for_min), get<1>(__chosen_for_max), get<2>(__chosen_for_min),
-                                get<3>(__chosen_for_max)};
-    };
-
-    // TODO: Doesn't work with `zip_iterator`.
-    //       In that case the first and the second arguments of `_ReduceValueType` will be
-    //       a `tuple` of `difference_type`, not the `difference_type` itself.
-    auto __transform_fn = [](auto __gidx, auto __acc) {
-        return _ReduceValueType{__gidx, __gidx, __acc[__gidx], __acc[__gidx]};
-    };
-
-    auto __keep = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _Iterator>();
-    auto __buf = __keep(__first, __last);
-
-    auto __ret = oneapi::dpl::__par_backend_hetero::__parallel_transform_reduce<_ReduceValueType,
-                                                                                ::std::false_type /*is_commutative*/>(
-                     ::std::forward<_ExecutionPolicy>(__exec), __reduce_fn, __transform_fn,
-                     unseq_backend::__no_init_value{}, // no initial value
-                     __buf.all_view())
-                     .get();
-
-    return ::std::make_pair<_Iterator, _Iterator>(__first + ::std::get<0>(__ret), __first + ::std::get<1>(__ret));
-}
-
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator, typename _Compare>
 ::std::pair<_Iterator, _Iterator>
 __pattern_minmax_element(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator __first,
@@ -921,15 +568,11 @@ __pattern_minmax_element(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __e
 // adjacent_find
 //------------------------------------------------------------------------
 
-template <typename _ExecutionPolicy, typename _Iterator, typename _BinaryPredicate>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, _Iterator>
-__pattern_adjacent_find(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last, _BinaryPredicate __predicate,
-                        /*parallel*/ ::std::true_type, /*vector*/ ::std::true_type,
-                        oneapi::dpl::__internal::__or_semantic)
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator, typename _BinaryPredicate>
+_Iterator
+__pattern_adjacent_find(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last,
+                        _BinaryPredicate __predicate, oneapi::dpl::__internal::__or_semantic)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
     if (__last - __first < 2)
         return __last;
 
@@ -944,7 +587,7 @@ __pattern_adjacent_find(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator
     // TODO: in case of confilicting names
     // __par_backend_hetero::make_wrapped_policy<__par_backend_hetero::__or_policy_wrapper>()
     bool result = __par_backend_hetero::__parallel_find_or(
-        __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
+        _BackendTag{}, ::std::forward<_ExecutionPolicy>(__exec),
         _Predicate{adjacent_find_fn<_BinaryPredicate>{__predicate}}, __par_backend_hetero::__parallel_or_tag{},
         oneapi::dpl::__ranges::make_zip_view(__buf1.all_view(), __buf2.all_view()));
 
@@ -953,66 +596,6 @@ __pattern_adjacent_find(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator
     return result ? __first : __last;
 }
 
-template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator, typename _BinaryPredicate>
-_Iterator
-__pattern_adjacent_find(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last,
-                        _BinaryPredicate __predicate, oneapi::dpl::__internal::__or_semantic)
-{
-    if (__last - __first < 2)
-        return __last;
-
-    using _Predicate =
-        oneapi::dpl::unseq_backend::single_match_pred<_ExecutionPolicy, adjacent_find_fn<_BinaryPredicate>>;
-
-    auto __keep1 = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _Iterator>();
-    auto __buf1 = __keep1(__first, __last - 1);
-    auto __keep2 = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _Iterator>();
-    auto __buf2 = __keep2(__first + 1, __last);
-
-    // TODO: in case of confilicting names
-    // __par_backend_hetero::make_wrapped_policy<__par_backend_hetero::__or_policy_wrapper>()
-    bool result = __par_backend_hetero::__parallel_find_or(
-        _BackendTag{}, ::std::forward<_ExecutionPolicy>(__exec),
-        _Predicate{adjacent_find_fn<_BinaryPredicate>{__predicate}}, __par_backend_hetero::__parallel_or_tag{},
-        oneapi::dpl::__ranges::make_zip_view(__buf1.all_view(), __buf2.all_view()));
-
-    // inverted conditional because of
-    // reorder_predicate in glue_algorithm_impl.h
-    return result ? __first : __last;
-}
-
-template <typename _ExecutionPolicy, typename _Iterator, typename _BinaryPredicate>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, _Iterator>
-__pattern_adjacent_find(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last, _BinaryPredicate __predicate,
-                        /*parallel*/ ::std::true_type, /*vector*/ ::std::true_type,
-                        oneapi::dpl::__internal::__first_semantic)
-{
-    if (__last - __first < 2)
-        return __last;
-
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
-    using _Predicate =
-        oneapi::dpl::unseq_backend::single_match_pred<_ExecutionPolicy, adjacent_find_fn<_BinaryPredicate>>;
-
-    auto __result = __par_backend_hetero::__parallel_find(
-        __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
-        __par_backend_hetero::zip(
-            __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__first),
-            __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__first + 1)),
-        __par_backend_hetero::zip(
-            __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__last - 1),
-            __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__last)),
-        _Predicate{adjacent_find_fn<_BinaryPredicate>{__predicate}}, ::std::true_type{});
-
-    auto __zip_at_first = __par_backend_hetero::zip(
-        __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__first),
-        __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__first + 1));
-    _Iterator __result_iterator = __first + (__result - __zip_at_first);
-    return (__result_iterator == __last - 1) ? __last : __result_iterator;
-}
-
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator, typename _BinaryPredicate>
 _Iterator
 __pattern_adjacent_find(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last,
@@ -1045,35 +628,6 @@ __pattern_adjacent_find(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __ex
 // count, count_if
 //------------------------------------------------------------------------
 
-template <typename _ExecutionPolicy, typename _Iterator, typename _Predicate>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<
-    _ExecutionPolicy, typename ::std::iterator_traits<_Iterator>::difference_type>
-__pattern_count(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last, _Predicate __predicate,
-                /*parallel*/ ::std::true_type, /*vector*/ ::std::true_type)
-{
-    if (__first == __last)
-        return 0;
-
-    using _ReduceValueType = typename ::std::iterator_traits<_Iterator>::difference_type;
-
-    auto __reduce_fn = ::std::plus<_ReduceValueType>{};
-    // int is being implicitly casted to difference_type
-    // otherwise we can only pass the difference_type as a functor template parameter
-    auto __transform_fn = [__predicate](auto __gidx, auto __acc) -> int {
-        return (__predicate(__acc[__gidx]) ? 1 : 0);
-    };
-
-    auto __keep = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _Iterator>();
-    auto __buf = __keep(__first, __last);
-
-    return oneapi::dpl::__par_backend_hetero::__parallel_transform_reduce<_ReduceValueType,
-                                                                          ::std::true_type /*is_commutative*/>(
-               ::std::forward<_ExecutionPolicy>(__exec), __reduce_fn, __transform_fn,
-               unseq_backend::__no_init_value{}, // no initial value
-               __buf.all_view())
-        .get();
-}
-
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator, typename _Predicate>
 typename ::std::iterator_traits<_Iterator>::difference_type
 __pattern_count(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last,
@@ -1106,29 +660,6 @@ __pattern_count(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Ite
 // any_of
 //------------------------------------------------------------------------
 
-template <typename _ExecutionPolicy, typename _Iterator, typename _Pred>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, bool>
-__pattern_any_of(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last, _Pred __pred,
-                 /*vector=*/::std::true_type, /*parallel=*/::std::true_type)
-{
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
-    if (__first == __last)
-        return false;
-
-    using _Predicate = oneapi::dpl::unseq_backend::single_match_pred<_ExecutionPolicy, _Pred>;
-
-    auto __keep = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _Iterator>();
-    auto __buf = __keep(__first, __last);
-
-    return oneapi::dpl::__par_backend_hetero::__parallel_find_or(
-        __backend_tag{},
-        __par_backend_hetero::make_wrapped_policy<__par_backend_hetero::__or_policy_wrapper>(
-            ::std::forward<_ExecutionPolicy>(__exec)),
-        _Predicate{__pred}, __par_backend_hetero::__parallel_or_tag{}, __buf.all_view());
-}
-
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator, typename _Pred>
 bool
 __pattern_any_of(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last,
@@ -1153,34 +684,6 @@ __pattern_any_of(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _It
 // equal
 //------------------------------------------------------------------------
 
-template <typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2, typename _Pred>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, bool>
-__pattern_equal(_ExecutionPolicy&& __exec, _Iterator1 __first1, _Iterator1 __last1, _Iterator2 __first2,
-                _Iterator2 __last2, _Pred __pred,
-                /*vector=*/::std::true_type, /*parallel=*/::std::true_type)
-{
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator1, _Iterator2>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
-    if (__last1 == __first1 || __last2 == __first2 || __last1 - __first1 != __last2 - __first2)
-        return false;
-
-    using _Predicate = oneapi::dpl::unseq_backend::single_match_pred<_ExecutionPolicy, equal_predicate<_Pred>>;
-
-    auto __keep1 = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _Iterator1>();
-    auto __buf1 = __keep1(__first1, __last1);
-    auto __keep2 = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _Iterator2>();
-    auto __buf2 = __keep2(__first2, __last2);
-
-    // TODO: in case of confilicting names
-    // __par_backend_hetero::make_wrapped_policy<__par_backend_hetero::__or_policy_wrapper>()
-    return !__par_backend_hetero::__parallel_find_or(
-        __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), _Predicate{equal_predicate<_Pred>{__pred}},
-        __par_backend_hetero::__parallel_or_tag{},
-        oneapi::dpl::__ranges::make_zip_view(__buf1.all_view(), __buf2.all_view()));
-}
-
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2, typename _Pred>
 bool
 __pattern_equal(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator1 __first1, _Iterator1 __last1,
@@ -1208,19 +711,6 @@ __pattern_equal(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Ite
 // equal version for sequences with equal length
 //------------------------------------------------------------------------
 
-template <typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2, typename _Pred>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, bool>
-__pattern_equal(_ExecutionPolicy&& __exec, _Iterator1 __first1, _Iterator1 __last1, _Iterator2 __first2, _Pred __pred,
-                /*vector=*/::std::true_type, /*parallel=*/::std::true_type)
-{
-    // TODO is it correct that we check _Iterator2 in __select_backend ?
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator1, _Iterator2>();
-
-    return oneapi::dpl::__internal::__pattern_equal(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first1,
-                                                    __last1, __first2, __first2 + (__last1 - __first1), __pred);
-}
-
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2, typename _Pred>
 bool
 __pattern_equal(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator1 __first1, _Iterator1 __last1,
@@ -1255,39 +745,6 @@ __pattern_find_if(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _I
 // find_end
 //------------------------------------------------------------------------
 
-template <typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2, typename _Pred>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, _Iterator1>
-__pattern_find_end(_ExecutionPolicy&& __exec, _Iterator1 __first, _Iterator1 __last, _Iterator2 __s_first,
-                   _Iterator2 __s_last, _Pred __pred, /*vector=*/::std::true_type, /*parallel=*/::std::true_type)
-{
-    if (__first == __last || __s_last == __s_first || __last - __first < __s_last - __s_first)
-        return __last;
-
-    // TODO is it correct that we check _Iterator2 in __select_backend ?
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator1, _Iterator2>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
-    if (__last - __first == __s_last - __s_first)
-    {
-        const bool __res = __pattern_equal(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
-                                           __s_first, __pred);
-        return __res ? __first : __last;
-    }
-    else
-    {
-        using _Predicate = unseq_backend::multiple_match_pred<_ExecutionPolicy, _Pred>;
-
-        return __par_backend_hetero::__parallel_find(
-            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
-            __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__first),
-            __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__last),
-            __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__s_first),
-            __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__s_last), _Predicate{__pred},
-            ::std::false_type{});
-    }
-}
-
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2, typename _Pred>
 _Iterator1
 __pattern_find_end(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator1 __first, _Iterator1 __last,
@@ -1320,31 +777,6 @@ __pattern_find_end(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _
 // find_first_of
 //------------------------------------------------------------------------
 
-template <typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2, typename _Pred>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, _Iterator1>
-__pattern_find_first_of(_ExecutionPolicy&& __exec, _Iterator1 __first, _Iterator1 __last, _Iterator2 __s_first,
-                        _Iterator2 __s_last, _Pred __pred, /*vector=*/::std::true_type, /*parallel=*/::std::true_type)
-{
-    if (__first == __last || __s_last == __s_first)
-        return __last;
-
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator1, _Iterator2>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
-    using _Predicate = unseq_backend::first_match_pred<_ExecutionPolicy, _Pred>;
-
-    // TODO: To check whether it makes sense to iterate over the second sequence in case of
-    // distance(__first, __last) < distance(__s_first, __s_last).
-    return __par_backend_hetero::__parallel_find(
-        __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
-        __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__first),
-        __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__last),
-        __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__s_first),
-        __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__s_last), _Predicate{__pred},
-        ::std::true_type{});
-}
-
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2, typename _Pred>
 _Iterator1
 __pattern_find_first_of(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator1 __first,
@@ -1375,41 +807,6 @@ class equal_wrapper
 {
 };
 
-template <typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2, typename _Pred>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, _Iterator1>
-__pattern_search(_ExecutionPolicy&& __exec, _Iterator1 __first, _Iterator1 __last, _Iterator2 __s_first,
-                 _Iterator2 __s_last, _Pred __pred, /*vector=*/::std::true_type, /*parallel=*/::std::true_type)
-{
-    if (__s_last == __s_first)
-        return __first;
-
-    if (__last - __first < __s_last - __s_first)
-        return __last;
-
-    // TODO is it correct that we check _Iterator2 in __select_backend ?
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator1, _Iterator2>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
-    if (__last - __first == __s_last - __s_first)
-    {
-        const bool __res = __pattern_equal(
-            __dispatch_tag,
-            __par_backend_hetero::make_wrapped_policy<equal_wrapper>(::std::forward<_ExecutionPolicy>(__exec)), __first,
-            __last, __s_first, __pred);
-        return __res ? __first : __last;
-    }
-
-    using _Predicate = unseq_backend::multiple_match_pred<_ExecutionPolicy, _Pred>;
-    return __par_backend_hetero::__parallel_find(
-        __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
-        __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__first),
-        __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__last),
-        __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__s_first),
-        __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__s_last), _Predicate{__pred},
-        ::std::true_type{});
-}
-
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2, typename _Pred>
 _Iterator1
 __pattern_search(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator1 __first, _Iterator1 __last,
@@ -1457,39 +854,6 @@ struct __search_n_unary_predicate
     }
 };
 
-template <typename _ExecutionPolicy, typename _Iterator, typename _Size, typename _Tp, typename _BinaryPredicate>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, _Iterator>
-__pattern_search_n(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last, _Size __count, const _Tp& __value,
-                   _BinaryPredicate __pred, /*vector=*/::std::true_type, /*parallel=*/::std::true_type)
-{
-    if (__count <= 0)
-        return __first;
-
-    if (__last - __first < __count)
-        return __last;
-
-    if (__last - __first == __count)
-    {
-        constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator>();
-
-        return (!__internal::__pattern_any_of(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
-                                              __search_n_unary_predicate<_Tp, _BinaryPredicate>{__value, __pred}))
-                   ? __first
-                   : __last;
-    }
-
-    using _Predicate = unseq_backend::n_elem_match_pred<_ExecutionPolicy, _BinaryPredicate, _Tp, _Size>;
-
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
-    return __par_backend_hetero::__parallel_find(
-        __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
-        __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__first),
-        __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__last),
-        _Predicate{__pred, __value, __count}, ::std::true_type{});
-}
-
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator, typename _Size, typename _Tp,
           typename _BinaryPredicate>
 _Iterator
@@ -1522,31 +886,6 @@ __pattern_search_n(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _
 // mismatch
 //------------------------------------------------------------------------
 
-template <typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2, typename _Pred>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, ::std::pair<_Iterator1, _Iterator2>>
-__pattern_mismatch(_ExecutionPolicy&& __exec, _Iterator1 __first1, _Iterator1 __last1, _Iterator2 __first2,
-                   _Iterator2 __last2, _Pred __pred, /*vector=*/::std::true_type, /*parallel=*/::std::true_type)
-{
-    auto __n = ::std::min(__last1 - __first1, __last2 - __first2);
-    if (__n <= 0)
-        return ::std::make_pair(__first1, __first2);
-
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator1, _Iterator2>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
-    using _Predicate = oneapi::dpl::unseq_backend::single_match_pred<_ExecutionPolicy, equal_predicate<_Pred>>;
-
-    auto __first_zip = __par_backend_hetero::zip(
-        __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__first1),
-        __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__first2));
-    auto __result = __par_backend_hetero::__parallel_find(
-        __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first_zip, __first_zip + __n,
-        _Predicate{equal_predicate<_Pred>{__pred}}, ::std::true_type{});
-    __n = __result - __first_zip;
-    return ::std::make_pair(__first1 + __n, __first2 + __n);
-}
-
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2, typename _Pred>
 ::std::pair<_Iterator1, _Iterator2>
 __pattern_mismatch(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator1 __first1, _Iterator1 __last1,
@@ -1572,17 +911,12 @@ __pattern_mismatch(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _
 // copy_if
 //------------------------------------------------------------------------
 
-template <typename _ExecutionPolicy, typename _Iterator1, typename _IteratorOrTuple, typename _CreateMaskOp,
-          typename _CopyByMaskOp>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<
-    _ExecutionPolicy, ::std::pair<_IteratorOrTuple, typename ::std::iterator_traits<_Iterator1>::difference_type>>
-__pattern_scan_copy(_ExecutionPolicy&& __exec, _Iterator1 __first, _Iterator1 __last, _IteratorOrTuple __output_first,
-                    _CreateMaskOp __create_mask_op, _CopyByMaskOp __copy_by_mask_op)
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator1, typename _IteratorOrTuple,
+          typename _CreateMaskOp, typename _CopyByMaskOp>
+::std::pair<_IteratorOrTuple, typename ::std::iterator_traits<_Iterator1>::difference_type>
+__pattern_scan_copy(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Iterator1 __first, _Iterator1 __last,
+                    _IteratorOrTuple __output_first, _CreateMaskOp __create_mask_op, _CopyByMaskOp __copy_by_mask_op)
 {
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator1, _IteratorOrTuple>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
     using _It1DifferenceType = typename ::std::iterator_traits<_Iterator1>::difference_type;
 
     if (__first == __last)
@@ -1596,7 +930,7 @@ __pattern_scan_copy(_ExecutionPolicy&& __exec, _Iterator1 __first, _Iterator1 __
         oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::write, _IteratorOrTuple>();
     auto __buf2 = __keep2(__output_first, __output_first + __n);
 
-    auto __res = __par_backend_hetero::__parallel_scan_copy(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
+    auto __res = __par_backend_hetero::__parallel_scan_copy(_BackendTag{}, ::std::forward<_ExecutionPolicy>(__exec),
                                                             __buf1.all_view(), __buf2.all_view(), __n, __create_mask_op,
                                                             __copy_by_mask_op);
 
@@ -1604,78 +938,23 @@ __pattern_scan_copy(_ExecutionPolicy&& __exec, _Iterator1 __first, _Iterator1 __
     return ::std::make_pair(__output_first + __n, __num_copied);
 }
 
-template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator1, typename _IteratorOrTuple,
-          typename _CreateMaskOp, typename _CopyByMaskOp>
-::std::pair<_IteratorOrTuple, typename ::std::iterator_traits<_Iterator1>::difference_type>
-__pattern_scan_copy(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Iterator1 __first, _Iterator1 __last,
-                    _IteratorOrTuple __output_first, _CreateMaskOp __create_mask_op, _CopyByMaskOp __copy_by_mask_op)
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2,
+          typename _Predicate>
+_Iterator2
+__pattern_copy_if(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator1 __first, _Iterator1 __last,
+                  _Iterator2 __result_first, _Predicate __pred)
 {
     using _It1DifferenceType = typename ::std::iterator_traits<_Iterator1>::difference_type;
 
     if (__first == __last)
-        return ::std::make_pair(__output_first, _It1DifferenceType{0});
+        return __result_first;
 
     _It1DifferenceType __n = __last - __first;
 
     auto __keep1 = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _Iterator1>();
     auto __buf1 = __keep1(__first, __last);
-    auto __keep2 =
-        oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::write, _IteratorOrTuple>();
-    auto __buf2 = __keep2(__output_first, __output_first + __n);
-
-    auto __res = __par_backend_hetero::__parallel_scan_copy(_BackendTag{}, ::std::forward<_ExecutionPolicy>(__exec),
-                                                            __buf1.all_view(), __buf2.all_view(), __n, __create_mask_op,
-                                                            __copy_by_mask_op);
-
-    ::std::size_t __num_copied = __res.get();
-    return ::std::make_pair(__output_first + __n, __num_copied);
-}
-
-template <typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2, typename _Predicate>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, _Iterator2>
-__pattern_copy_if(_ExecutionPolicy&& __exec, _Iterator1 __first, _Iterator1 __last, _Iterator2 __result_first,
-                  _Predicate __pred, /*vector=*/::std::true_type, /*parallel=*/::std::true_type)
-{
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator1, _Iterator2>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
-    using _It1DifferenceType = typename ::std::iterator_traits<_Iterator1>::difference_type;
-
-    if (__first == __last)
-        return __result_first;
-
-    _It1DifferenceType __n = __last - __first;
-
-    auto __keep1 = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _Iterator1>();
-    auto __buf1 = __keep1(__first, __last);
-    auto __keep2 = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::write, _Iterator2>();
-    auto __buf2 = __keep2(__result_first, __result_first + __n);
-
-    auto __res = __par_backend_hetero::__parallel_copy_if(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
-                                                          __buf1.all_view(), __buf2.all_view(), __n, __pred);
-
-    ::std::size_t __num_copied = __res.get();
-    return __result_first + __num_copied;
-}
-
-template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2,
-          typename _Predicate>
-_Iterator2
-__pattern_copy_if(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator1 __first, _Iterator1 __last,
-                  _Iterator2 __result_first, _Predicate __pred)
-{
-    using _It1DifferenceType = typename ::std::iterator_traits<_Iterator1>::difference_type;
-
-    if (__first == __last)
-        return __result_first;
-
-    _It1DifferenceType __n = __last - __first;
-
-    auto __keep1 = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _Iterator1>();
-    auto __buf1 = __keep1(__first, __last);
-    auto __keep2 = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::write, _Iterator2>();
-    auto __buf2 = __keep2(__result_first, __result_first + __n);
+    auto __keep2 = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::write, _Iterator2>();
+    auto __buf2 = __keep2(__result_first, __result_first + __n);
 
     auto __res = __par_backend_hetero::__parallel_copy_if(_BackendTag{}, ::std::forward<_ExecutionPolicy>(__exec),
                                                           __buf1.all_view(), __buf2.all_view(), __n, __pred);
@@ -1688,32 +967,6 @@ __pattern_copy_if(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _I
 // partition_copy
 //------------------------------------------------------------------------
 
-template <typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2, typename _Iterator3,
-          typename _UnaryPredicate>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, ::std::pair<_Iterator2, _Iterator3>>
-__pattern_partition_copy(_ExecutionPolicy&& __exec, _Iterator1 __first, _Iterator1 __last, _Iterator2 __result1,
-                         _Iterator3 __result2, _UnaryPredicate __pred, /*vector*/ ::std::true_type,
-                         /*parallel*/ ::std::true_type)
-{
-    if (__first == __last)
-        return ::std::make_pair(__result1, __result2);
-
-    using _It1DifferenceType = typename ::std::iterator_traits<_Iterator1>::difference_type;
-    using _ReduceOp = ::std::plus<_It1DifferenceType>;
-
-    unseq_backend::__create_mask<_UnaryPredicate, _It1DifferenceType> __create_mask_op{__pred};
-    unseq_backend::__partition_by_mask<_ReduceOp, /*inclusive*/ ::std::true_type> __copy_by_mask_op{_ReduceOp{}};
-
-    auto __result = __pattern_scan_copy(
-        ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
-        __par_backend_hetero::zip(
-            __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::write>(__result1),
-            __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::write>(__result2)),
-        __create_mask_op, __copy_by_mask_op);
-
-    return ::std::make_pair(__result1 + __result.second, __result2 + (__last - __first - __result.second));
-}
-
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2,
           typename _Iterator3, typename _UnaryPredicate>
 ::std::pair<_Iterator2, _Iterator3>
@@ -1743,24 +996,6 @@ __pattern_partition_copy(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __e
 // unique_copy
 //------------------------------------------------------------------------
 
-template <typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2, typename _BinaryPredicate>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, _Iterator2>
-__pattern_unique_copy(_ExecutionPolicy&& __exec, _Iterator1 __first, _Iterator1 __last, _Iterator2 __result_first,
-                      _BinaryPredicate __pred, /*vector*/ ::std::true_type, /*parallel*/ ::std::true_type)
-{
-    using _It1DifferenceType = typename ::std::iterator_traits<_Iterator1>::difference_type;
-    unseq_backend::__copy_by_mask<::std::plus<_It1DifferenceType>, oneapi::dpl::__internal::__pstl_assign,
-                                  /*inclusive*/ ::std::true_type, 1>
-        __copy_by_mask_op;
-    __create_mask_unique_copy<__not_pred<_BinaryPredicate>, _It1DifferenceType> __create_mask_op{
-        __not_pred<_BinaryPredicate>{__pred}};
-
-    auto __result = __pattern_scan_copy(::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result_first,
-                                        __create_mask_op, __copy_by_mask_op);
-
-    return __result_first + __result.second;
-}
-
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2,
           typename _BinaryPredicate>
 _Iterator2
@@ -1789,29 +1024,6 @@ class copy_back_wrapper2
 {
 };
 
-template <typename _ExecutionPolicy, typename _Iterator, typename _Predicate>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, _Iterator>
-__pattern_remove_if(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last, _Predicate __pred,
-                    /*vector*/ ::std::true_type, /*parallel*/ ::std::true_type)
-{
-    if (__last == __first)
-        return __last;
-
-    using _ValueType = typename ::std::iterator_traits<_Iterator>::value_type;
-
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator>();
-
-    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __buf(__exec, __last - __first);
-    auto __copy_first = __buf.get();
-    auto __copy_last =
-        __pattern_copy_if(__dispatch_tag, __exec, __first, __last, __copy_first, __not_pred<_Predicate>{__pred});
-
-    //TODO: optimize copy back depending on Iterator, i.e. set_final_data for host iterator/pointer
-    return __pattern_walk2(
-        __par_backend_hetero::make_wrapped_policy<copy_back_wrapper>(::std::forward<_ExecutionPolicy>(__exec)),
-        __copy_first, __copy_last, __first, __brick_copy<_ExecutionPolicy>{}, ::std::true_type{}, ::std::true_type{});
-}
-
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator, typename _Predicate>
 _Iterator
 __pattern_remove_if(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last,
@@ -1843,36 +1055,6 @@ __pattern_remove_if(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec,
         __copy_first, __copy_last, __first, __brick_copy<_ExecutionPolicy>{});
 }
 
-template <typename _ExecutionPolicy, typename _Iterator, typename _BinaryPredicate>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, _Iterator>
-__pattern_unique(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last, _BinaryPredicate __pred,
-                 /*vector*/ ::std::true_type, /*parallel*/ ::std::true_type)
-{
-    if (__last - __first < 2)
-        return __last;
-
-    using _ValueType = typename ::std::iterator_traits<_Iterator>::value_type;
-
-    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __buf(__exec, __last - __first);
-    auto __copy_first = __buf.get();
-    auto __copy_last =
-        __pattern_unique_copy(oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, decltype(__first),
-                                                                        decltype(__last), decltype(__copy_first)>(),
-                              __exec, __first, __last, __copy_first, __pred);
-
-    constexpr auto __dispatch_tag1 =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, decltype(__copy_first), decltype(__copy_last),
-                                                  decltype(__first)>();
-    using __backend_tag1 = typename decltype(__dispatch_tag1)::__backend_tag;
-
-    //TODO: optimize copy back depending on Iterator, i.e. set_final_data for host iterator/pointer
-    return __pattern_walk2<__backend_tag1, /*_IsSync=*/::std::true_type, __par_backend_hetero::access_mode::read_write,
-                           __par_backend_hetero::access_mode::read_write>(
-        __dispatch_tag1,
-        __par_backend_hetero::make_wrapped_policy<copy_back_wrapper>(::std::forward<_ExecutionPolicy>(__exec)),
-        __copy_first, __copy_last, __first, __brick_copy<_ExecutionPolicy>{});
-}
-
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator, typename _BinaryPredicate>
 _Iterator
 __pattern_unique(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last,
@@ -1915,38 +1097,6 @@ enum _IsPartitionedReduceType : signed char
     __true_false
 };
 
-template <typename _ExecutionPolicy, typename _Iterator, typename _Predicate>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, bool>
-__pattern_is_partitioned(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last, _Predicate __predicate,
-                         /*parallel*/ ::std::true_type, /*vector*/ ::std::true_type)
-{
-    if (__last - __first < 2)
-        return true;
-
-    using _ReduceValueType = _IsPartitionedReduceType;
-    auto __reduce_fn = [](_ReduceValueType __a, _ReduceValueType __b) {
-        _ReduceValueType __table[] = {__broken,     __broken,     __broken,     __broken, __broken,    __all_true,
-                                      __true_false, __true_false, __broken,     __broken, __all_false, __broken,
-                                      __broken,     __broken,     __true_false, __broken};
-        return __table[__a * 4 + __b];
-    };
-    auto __transform_fn = [__predicate](auto __gidx, auto __acc) {
-        return (__predicate(__acc[__gidx]) ? __all_true : __all_false);
-    };
-
-    auto __keep = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _Iterator>();
-    auto __buf = __keep(__first, __last);
-
-    auto __res = oneapi::dpl::__par_backend_hetero::__parallel_transform_reduce<_ReduceValueType,
-                                                                                ::std::false_type /*is_commutative*/>(
-                     ::std::forward<_ExecutionPolicy>(__exec), __reduce_fn, __transform_fn,
-                     unseq_backend::__no_init_value{}, // no initial value
-                     __buf.all_view())
-                     .get();
-
-    return __broken != __reduce_fn(_ReduceValueType{__all_true}, __res);
-}
-
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator, typename _Predicate>
 bool
 __pattern_is_partitioned(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator __first,
@@ -1998,28 +1148,6 @@ struct __is_heap_check
     }
 };
 
-template <typename _ExecutionPolicy, typename _RandomAccessIterator, typename _Compare>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, _RandomAccessIterator>
-__pattern_is_heap_until(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIterator __last,
-                        _Compare __comp, /* vector */ ::std::true_type, /* parallel = */ ::std::true_type)
-{
-    if (__last - __first < 2)
-        return __last;
-
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
-    using _Predicate =
-        oneapi::dpl::unseq_backend::single_match_pred_by_idx<_ExecutionPolicy, __is_heap_check<_Compare>>;
-
-    return __par_backend_hetero::__parallel_find(
-        __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
-        __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__first),
-        __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__last), _Predicate{__comp},
-        ::std::true_type{});
-}
-
 template <typename _BackendTag, typename _ExecutionPolicy, typename _RandomAccessIterator, typename _Compare>
 _RandomAccessIterator
 __pattern_is_heap_until(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
@@ -2038,27 +1166,6 @@ __pattern_is_heap_until(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __ex
         ::std::true_type{});
 }
 
-template <typename _ExecutionPolicy, typename _RandomAccessIterator, typename _Compare>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, bool>
-__pattern_is_heap(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIterator __last,
-                  _Compare __comp, /* vector */ ::std::true_type, /* parallel = */ ::std::true_type)
-{
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
-    if (__last - __first < 2)
-        return true;
-
-    using _Predicate =
-        oneapi::dpl::unseq_backend::single_match_pred_by_idx<_ExecutionPolicy, __is_heap_check<_Compare>>;
-
-    return !__par_backend_hetero::__parallel_or(
-        __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
-        __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__first),
-        __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__last), _Predicate{__comp});
-}
-
 template <typename _BackendTag, typename _ExecutionPolicy, typename _RandomAccessIterator, typename _Compare>
 bool
 __pattern_is_heap(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
@@ -2079,51 +1186,6 @@ __pattern_is_heap(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _R
 //------------------------------------------------------------------------
 // merge
 //------------------------------------------------------------------------
-template <typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2, typename _Iterator3, typename _Compare>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, _Iterator3>
-__pattern_merge(_ExecutionPolicy&& __exec, _Iterator1 __first1, _Iterator1 __last1, _Iterator2 __first2,
-                _Iterator2 __last2, _Iterator3 __d_first, _Compare __comp, /*vector=*/::std::true_type,
-                /*parallel=*/::std::true_type)
-{
-    auto __n1 = __last1 - __first1;
-    auto __n2 = __last2 - __first2;
-    auto __n = __n1 + __n2;
-    if (__n == 0)
-        return __d_first;
-
-    const auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator1, _Iterator2, _Iterator3>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
-    //To consider the direct copying pattern call in case just one of sequences is empty.
-    if (__n1 == 0)
-        oneapi::dpl::__internal::__pattern_walk2_brick(
-            __dispatch_tag,
-            oneapi::dpl::__par_backend_hetero::make_wrapped_policy<copy_back_wrapper>(
-                ::std::forward<_ExecutionPolicy>(__exec)),
-            __first2, __last2, __d_first, oneapi::dpl::__internal::__brick_copy<_ExecutionPolicy>{});
-    else if (__n2 == 0)
-        oneapi::dpl::__internal::__pattern_walk2_brick(
-            __dispatch_tag,
-            oneapi::dpl::__par_backend_hetero::make_wrapped_policy<copy_back_wrapper2>(
-                ::std::forward<_ExecutionPolicy>(__exec)),
-            __first1, __last1, __d_first, oneapi::dpl::__internal::__brick_copy<_ExecutionPolicy>{});
-    else
-    {
-        auto __keep1 = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _Iterator1>();
-        auto __buf1 = __keep1(__first1, __last1);
-        auto __keep2 = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _Iterator2>();
-        auto __buf2 = __keep2(__first2, __last2);
-
-        auto __keep3 = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::write, _Iterator3>();
-        auto __buf3 = __keep3(__d_first, __d_first + __n);
-
-        __par_backend_hetero::__parallel_merge(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
-                                               __buf1.all_view(), __buf2.all_view(), __buf3.all_view(), __comp)
-            .wait();
-    }
-    return __d_first + __n;
-}
 
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2,
           typename _Iterator3, typename _Compare>
@@ -2170,35 +1232,6 @@ __pattern_merge(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Ite
 //------------------------------------------------------------------------
 // inplace_merge
 //------------------------------------------------------------------------
-template <typename _ExecutionPolicy, typename _Iterator, typename _Compare>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy>
-__pattern_inplace_merge(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __middle, _Iterator __last,
-                        _Compare __comp, /*vector=*/::std::true_type, /*parallel=*/::std::true_type)
-{
-    using _ValueType = typename ::std::iterator_traits<_Iterator>::value_type;
-
-    if (__first == __middle || __middle == __last || __first == __last)
-        return;
-
-    assert(__first < __middle && __middle < __last);
-
-    auto __n = __last - __first;
-    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __buf(__exec, __n);
-    auto __copy_first = __buf.get();
-    auto __copy_last = __copy_first + __n;
-
-    __pattern_merge(__exec, __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__first),
-                    __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__middle),
-                    __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__middle),
-                    __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__last),
-                    __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::write>(__copy_first),
-                    __comp, ::std::true_type{}, ::std::true_type{});
-
-    //TODO: optimize copy back depending on Iterator, i.e. set_final_data for host iterator/pointer
-    __pattern_walk2(
-        __par_backend_hetero::make_wrapped_policy<copy_back_wrapper>(::std::forward<_ExecutionPolicy>(__exec)),
-        __copy_first, __copy_last, __first, __brick_move<_ExecutionPolicy>{}, ::std::true_type{}, ::std::true_type{});
-}
 
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator, typename _Compare>
 void
@@ -2249,15 +1282,6 @@ __stable_sort_with_projection(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&
         .wait();
 }
 
-template <typename _ExecutionPolicy, typename _Iterator, typename _Compare>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy>
-__pattern_sort(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last, _Compare __comp,
-               /*vector=*/::std::true_type, /*parallel=*/::std::true_type, /*is_move_constructible=*/::std::true_type)
-{
-    __stable_sort_with_projection(::std::forward<_ExecutionPolicy>(__exec), __first, __last, __comp,
-                                  oneapi::dpl::identity{});
-}
-
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator, typename _Compare>
 void
 __pattern_sort(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last,
@@ -2270,14 +1294,6 @@ __pattern_sort(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iter
 //------------------------------------------------------------------------
 // stable_sort
 //------------------------------------------------------------------------
-template <typename _ExecutionPolicy, typename _Iterator, typename _Compare>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy>
-__pattern_stable_sort(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last, _Compare __comp,
-                      /*vector=*/::std::true_type, /*parallel=*/::std::true_type)
-{
-    __stable_sort_with_projection(::std::forward<_ExecutionPolicy>(__exec), __first, __last, __comp,
-                                  oneapi::dpl::identity{});
-}
 
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator, typename _Compare>
 void
@@ -2288,22 +1304,6 @@ __pattern_stable_sort(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec
                                   oneapi::dpl::identity{});
 }
 
-template <typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2, typename _Compare>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy>
-__pattern_sort_by_key(_ExecutionPolicy&& __exec, _Iterator1 __keys_first, _Iterator1 __keys_last,
-                      _Iterator2 __values_first, _Compare __comp, /*vector=*/::std::true_type,
-                      /*parallel=*/::std::true_type)
-{
-    static_assert(::std::is_move_constructible_v<typename ::std::iterator_traits<_Iterator1>::value_type>
-        && ::std::is_move_constructible_v<typename ::std::iterator_traits<_Iterator2>::value_type>,
-        "The keys and values should be move constructible in case of parallel execution.");
-
-    auto __beg = oneapi::dpl::make_zip_iterator(__keys_first, __values_first);
-    auto __end = __beg + (__keys_last - __keys_first);
-    __stable_sort_with_projection(::std::forward<_ExecutionPolicy>(__exec), __beg, __end, __comp,
-                                  [](const auto& __a) { return ::std::get<0>(__a); });
-}
-
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2, typename _Compare>
 void
 __pattern_sort_by_key(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator1 __keys_first,
@@ -2319,57 +1319,6 @@ __pattern_sort_by_key(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec
                                   [](const auto& __a) { return ::std::get<0>(__a); });
 }
 
-template <typename _ExecutionPolicy, typename _Iterator, typename _UnaryPredicate>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, _Iterator>
-__pattern_stable_partition(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last, _UnaryPredicate __pred,
-                           /*vector*/ ::std::true_type, /*parallel*/ ::std::true_type)
-{
-    if (__last == __first)
-        return __last;
-    else if (__last - __first < 2)
-    {
-        constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator>();
-
-        return __pattern_any_of(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __pred)
-                   ? __last
-                   : __first;
-    }
-
-    using _ValueType = typename ::std::iterator_traits<_Iterator>::value_type;
-
-    auto __n = __last - __first;
-
-    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __true_buf(__exec, __n);
-    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __false_buf(__exec, __n);
-    auto __true_result = __true_buf.get();
-    auto __false_result = __false_buf.get();
-
-    constexpr auto __dispatch_tag = __select_backend<_ExecutionPolicy, decltype(__first), decltype(__last),
-                                                     decltype(__true_result), decltype(__false_result)>();
-    auto copy_result =
-        __pattern_partition_copy(__dispatch_tag, __exec, __first, __last, __true_result, __false_result, __pred);
-    auto true_count = copy_result.first - __true_result;
-
-    //TODO: optimize copy back if possible (inplace, decrease number of submits)
-    constexpr auto __dispatch_tag1 =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, decltype(__true_result),
-                                                  decltype(copy_result.first), decltype(__first)>();
-    using __backend_tag1 = typename decltype(__dispatch_tag1)::__backend_tag;
-    __pattern_walk2<__backend_tag1, /*_IsSync=*/::std::false_type>(
-        __dispatch_tag1, __par_backend_hetero::make_wrapped_policy<copy_back_wrapper>(__exec), __true_result,
-        copy_result.first, __first, __brick_move<_ExecutionPolicy>{});
-
-    constexpr auto __dispatch_tag2 =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, decltype(__false_result),
-                                                  decltype(copy_result.second), decltype(__first + true_count)>();
-    __pattern_walk2(
-        __dispatch_tag2,
-        __par_backend_hetero::make_wrapped_policy<copy_back_wrapper2>(::std::forward<_ExecutionPolicy>(__exec)),
-        __false_result, copy_result.second, __first + true_count, __brick_move<_ExecutionPolicy>{});
-
-    return __first + true_count;
-}
-
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator, typename _UnaryPredicate>
 _Iterator
 __pattern_stable_partition(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator __first,
@@ -2416,18 +1365,6 @@ __pattern_stable_partition(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& _
     return __first + true_count;
 }
 
-template <typename _ExecutionPolicy, typename _Iterator, typename _UnaryPredicate>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, _Iterator>
-__pattern_partition(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last, _UnaryPredicate __pred,
-                    /*vector*/ ::std::true_type, /*parallel*/ ::std::true_type)
-{
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator>();
-
-    //TODO: consider nonstable approaches
-    return __pattern_stable_partition(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
-                                      __pred);
-}
-
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator, typename _UnaryPredicate>
 _Iterator
 __pattern_partition(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last,
@@ -2444,55 +1381,6 @@ __pattern_partition(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec,
 // lexicographical_compare
 //------------------------------------------------------------------------
 
-template <typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2, typename _Compare>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, bool>
-__pattern_lexicographical_compare(_ExecutionPolicy&& __exec, _Iterator1 __first1, _Iterator1 __last1,
-                                  _Iterator2 __first2, _Iterator2 __last2, _Compare __comp, /*vector*/ ::std::true_type,
-                                  /*parallel*/ ::std::true_type)
-{
-    //trivial pre-checks
-    if (__first2 == __last2)
-        return false;
-    if (__first1 == __last1)
-        return true;
-
-    using _Iterator1DifferenceType = typename ::std::iterator_traits<_Iterator1>::difference_type;
-    using _ReduceValueType = int32_t;
-
-    auto __reduce_fn = [](_ReduceValueType __a, _ReduceValueType __b) {
-        bool __is_mismatched = __a != 0;
-        return __a * __is_mismatched + __b * !__is_mismatched;
-    };
-    auto __transform_fn = [__comp](auto __gidx, auto __acc1, auto __acc2) {
-        auto const& __s1_val = __acc1[__gidx];
-        auto const& __s2_val = __acc2[__gidx];
-
-        ::std::int32_t __is_s1_val_less = __comp(__s1_val, __s2_val);
-        ::std::int32_t __is_s1_val_greater = __comp(__s2_val, __s1_val);
-
-        // 1 if __s1_val <  __s2_val, -1 if __s1_val <  __s2_val, 0 if __s1_val == __s2_val
-        return _ReduceValueType{1 * __is_s1_val_less - 1 * __is_s1_val_greater};
-    };
-
-    auto __shared_size = ::std::min(__last1 - __first1, (_Iterator1DifferenceType)(__last2 - __first2));
-
-    auto __keep1 = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _Iterator1>();
-    auto __buf1 = __keep1(__first1, __first1 + __shared_size);
-
-    auto __keep2 = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _Iterator2>();
-    auto __buf2 = __keep2(__first2, __first2 + __shared_size);
-
-    auto __ret_idx =
-        oneapi::dpl::__par_backend_hetero::__parallel_transform_reduce<_ReduceValueType,
-                                                                       ::std::false_type /*is_commutative*/>(
-            ::std::forward<_ExecutionPolicy>(__exec), __reduce_fn, __transform_fn,
-            unseq_backend::__no_init_value{}, // no initial value
-            __buf1.all_view(), __buf2.all_view())
-            .get();
-
-    return __ret_idx ? __ret_idx == 1 : (__last1 - __first1) < (__last2 - __first2);
-}
-
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2, typename _Compare>
 bool
 __pattern_lexicographical_compare(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator1 __first1,
@@ -2541,38 +1429,6 @@ __pattern_lexicographical_compare(__hetero_tag<_BackendTag> __tag, _ExecutionPol
     return __ret_idx ? __ret_idx == 1 : (__last1 - __first1) < (__last2 - __first2);
 }
 
-template <typename _ExecutionPolicy, typename _ForwardIterator1, typename _ForwardIterator2, typename _Compare>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, bool>
-__pattern_includes(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
-                   _ForwardIterator2 __first2, _ForwardIterator2 __last2, _Compare __comp, /*vector=*/::std::true_type,
-                   /*parallel=*/::std::true_type)
-{
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
-    //according to the spec
-    if (__first2 == __last2)
-        return true;
-
-    //optimization; {1} - the first sequence, {2} - the second sequence
-    //{1} is empty or size_of{2} > size_of{1}
-    if (__first1 == __last1 || __last2 - __first2 > __last1 - __first1)
-        return false;
-
-    typedef typename ::std::iterator_traits<_ForwardIterator1>::difference_type _Size1;
-    typedef typename ::std::iterator_traits<_ForwardIterator2>::difference_type _Size2;
-
-    using __brick_include_type = unseq_backend::__brick_includes<_ExecutionPolicy, _Compare, _Size1, _Size2>;
-    return !__par_backend_hetero::__parallel_or(
-        __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
-        __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__first2),
-        __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__last2),
-        __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__first1),
-        __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__last1),
-        __brick_include_type(__comp, __last1 - __first1, __last2 - __first2));
-}
-
 template <typename _BackendTag, typename _ExecutionPolicy, typename _ForwardIterator1, typename _ForwardIterator2,
           typename _Compare>
 bool
@@ -2604,24 +1460,6 @@ __pattern_includes(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _
 //------------------------------------------------------------------------
 // partial_sort
 //------------------------------------------------------------------------
-template <typename _ExecutionPolicy, typename _Iterator, typename _Compare>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy>
-__pattern_partial_sort(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __mid, _Iterator __last, _Compare __comp,
-                       /*vector=*/::std::true_type, /*parallel=*/::std::true_type)
-{
-    if (__last - __first < 2)
-        return;
-
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
-    __par_backend_hetero::__parallel_partial_sort(
-        __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
-        __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read_write>(__first),
-        __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read_write>(__mid),
-        __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read_write>(__last), __comp)
-        .wait();
-}
 
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator, typename _Compare>
 void
@@ -2668,91 +1506,6 @@ struct __partial_sort_2
 {
 };
 
-template <typename _ExecutionPolicy, typename _InIterator, typename _OutIterator, typename _Compare>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, _OutIterator>
-__pattern_partial_sort_copy(_ExecutionPolicy&& __exec, _InIterator __first, _InIterator __last,
-                            _OutIterator __out_first, _OutIterator __out_last, _Compare __comp,
-                            /*vector=*/::std::true_type, /*parallel=*/::std::true_type)
-{
-    using _ValueType = typename ::std::iterator_traits<_InIterator>::value_type;
-
-    auto __in_size = __last - __first;
-    auto __out_size = __out_last - __out_first;
-
-    if (__in_size == 0 || __out_size == 0)
-        return __out_first;
-
-    // TODO: we can avoid a separate __pattern_walk2 for initial copy: it can be done during sort itself
-    // like it's done for CPU version, but it's better to be done together with merge cutoff implementation
-    // as it uses a similar mechanism.
-    if (__in_size <= __out_size)
-    {
-        constexpr auto __dispatch_tag1 =
-            oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, decltype(__first), decltype(__last),
-                                                      decltype(__out_first)>();
-        using __backend_tag1 = typename decltype(__dispatch_tag1)::__backend_tag;
-
-        // If our output buffer is larger than the input buffer, simply copy elements to the output and use
-        // full sort on them.
-        auto __out_end = __pattern_walk2<__backend_tag1, /*_IsSync=*/::std::false_type>(
-            __dispatch_tag1, __par_backend_hetero::make_wrapped_policy<__initial_copy_1>(__exec), __first, __last,
-            __out_first, __brick_copy<_ExecutionPolicy>{});
-
-        constexpr auto __dispatch_tag2 =
-            oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, decltype(__out_first), decltype(__out_end),
-                                                      decltype(__out_first)>();
-
-        // Use reqular sort as partial_sort isn't required to be stable
-        __pattern_sort(
-            __dispatch_tag2,
-            __par_backend_hetero::make_wrapped_policy<__partial_sort_1>(::std::forward<_ExecutionPolicy>(__exec)),
-            __out_first, __out_end, __comp, ::std::true_type{});
-
-        return __out_end;
-    }
-    else
-    {
-        // If our input buffer is smaller than the input buffer do the following:
-        // - create a temporary buffer and copy all the elements from the input buffer there
-        // - run partial sort on the temporary buffer
-        // - copy k elements from the temporary buffer to the output buffer.
-        oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __buf(__exec, __in_size);
-
-        auto __buf_first = __buf.get();
-
-        constexpr auto __dispatch_tag1 =
-            oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, decltype(__first), decltype(__last),
-                                                      decltype(__buf_first)>();
-        using __backend_tag1 = typename decltype(__dispatch_tag1)::__backend_tag;
-
-        auto __buf_last = __pattern_walk2<__backend_tag1, /*_IsSync=*/::std::false_type>(
-            __par_backend_hetero::make_wrapped_policy<__initial_copy_2>(__exec), __first, __last, __buf_first,
-            __brick_copy<_ExecutionPolicy>{});
-
-        auto __buf_mid = __buf_first + __out_size;
-
-        constexpr auto __dispatch_tag11 =
-            oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, decltype(__buf_first), decltype(__buf_mid),
-                                                      decltype(__buf_last)>();
-        using __backend_tag11 = typename decltype(__dispatch_tag11)::__backend_tag;
-
-        __par_backend_hetero::__parallel_partial_sort(
-            __backend_tag11{}, __par_backend_hetero::make_wrapped_policy<__partial_sort_2>(__exec),
-            __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read_write>(__buf_first),
-            __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read_write>(__buf_mid),
-            __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read_write>(__buf_last), __comp);
-
-        constexpr auto __dispatch_tag2 =
-            oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, decltype(__buf_first), decltype(__buf_mid),
-                                                      decltype(__out_first)>();
-
-        return __pattern_walk2(
-            __dispatch_tag2,
-            __par_backend_hetero::make_wrapped_policy<__copy_back>(::std::forward<_ExecutionPolicy>(__exec)),
-            __buf_first, __buf_mid, __out_first, __brick_copy<_ExecutionPolicy>{});
-    }
-}
-
 template <typename _BackendTag, typename _ExecutionPolicy, typename _InIterator, typename _OutIterator,
           typename _Compare>
 _OutIterator
@@ -2833,23 +1586,6 @@ __pattern_partial_sort_copy(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&&
 // nth_element
 //------------------------------------------------------------------------
 
-template <typename _ExecutionPolicy, typename _Iterator, typename _Compare>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy>
-__pattern_nth_element(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __nth, _Iterator __last, _Compare __comp,
-                      /*vector*/ ::std::true_type, /*parallel*/ ::std::true_type)
-{
-    if (__first == __last || __nth == __last)
-        return;
-
-    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator>();
-
-    // TODO: check partition-based implementation
-    // - try to avoid host dereference issue
-    // - measure performance of the issue-free implementation
-    __pattern_partial_sort(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __nth + 1, __last,
-                           __comp);
-}
-
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator, typename _Compare>
 void
 __pattern_nth_element(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __nth,
@@ -2867,26 +1603,6 @@ __pattern_nth_element(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec
 //------------------------------------------------------------------------
 // reverse
 //------------------------------------------------------------------------
-template <typename _ExecutionPolicy, typename _Iterator>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy>
-__pattern_reverse(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last, /*vector=*/::std::true_type,
-                  /*parallel=*/::std::true_type)
-{
-    auto __n = __last - __first;
-    if (__n <= 0)
-        return;
-
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
-    auto __keep = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read_write, _Iterator>();
-    auto __buf = __keep(__first, __last);
-    oneapi::dpl::__par_backend_hetero::__parallel_for(
-        __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
-        unseq_backend::__reverse_functor<typename ::std::iterator_traits<_Iterator>::difference_type>{__n}, __n / 2,
-        __buf.all_view())
-        .wait();
-}
 
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator>
 void
@@ -2910,34 +1626,6 @@ __pattern_reverse(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _I
 //------------------------------------------------------------------------
 // reverse_copy
 //------------------------------------------------------------------------
-template <typename _ExecutionPolicy, typename _BidirectionalIterator, typename _ForwardIterator>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, _ForwardIterator>
-__pattern_reverse_copy(_ExecutionPolicy&& __exec, _BidirectionalIterator __first, _BidirectionalIterator __last,
-                       _ForwardIterator __result, /*vector=*/::std::true_type, /*parallel=*/::std::true_type)
-{
-    auto __n = __last - __first;
-    if (__n <= 0)
-        return __result;
-
-    auto __keep1 =
-        oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _BidirectionalIterator>();
-    auto __buf1 = __keep1(__first, __last);
-    auto __keep2 =
-        oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::write, _ForwardIterator>();
-
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _BidirectionalIterator>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
-    auto __buf2 = __keep2(__result, __result + __n);
-    oneapi::dpl::__par_backend_hetero::__parallel_for(
-        __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
-        unseq_backend::__reverse_copy<typename ::std::iterator_traits<_BidirectionalIterator>::difference_type>{__n},
-        __n, __buf1.all_view(), __buf2.all_view())
-        .wait();
-
-    return __result + __n;
-}
 
 template <typename _BackendTag, typename _ExecutionPolicy, typename _BidirectionalIterator, typename _ForwardIterator>
 _ForwardIterator
@@ -2976,44 +1664,6 @@ class __rotate_wrapper
 {
 };
 
-template <typename _ExecutionPolicy, typename _Iterator>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, _Iterator>
-__pattern_rotate(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __new_first, _Iterator __last,
-                 /*vector=*/::std::true_type, /*parallel=*/::std::true_type)
-{
-    auto __n = __last - __first;
-    if (__n <= 0)
-        return __first;
-
-    using _Tp = typename ::std::iterator_traits<_Iterator>::value_type;
-
-    auto __keep = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read_write, _Iterator>();
-    auto __buf = __keep(__first, __last);
-    auto __temp_buf = oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _Tp>(__exec, __n);
-
-    auto __temp_rng =
-        oneapi::dpl::__ranges::all_view<_Tp, __par_backend_hetero::access_mode::write>(__temp_buf.get_buffer());
-
-    const auto __shift = __new_first - __first;
-
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
-    oneapi::dpl::__par_backend_hetero::__parallel_for(
-        __backend_tag{}, oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__rotate_wrapper>(__exec),
-        unseq_backend::__rotate_copy<typename ::std::iterator_traits<_Iterator>::difference_type>{__n, __shift}, __n,
-        __buf.all_view(), __temp_rng);
-
-    using _Function = __brick_move<_ExecutionPolicy>;
-    auto __brick = unseq_backend::walk_n<_ExecutionPolicy, _Function>{_Function{}};
-
-    oneapi::dpl::__par_backend_hetero::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
-                                                      __brick, __n, __temp_rng, __buf.all_view())
-        .wait();
-
-    return __first + (__last - __new_first);
-}
-
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator>
 _Iterator
 __pattern_rotate(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __new_first,
@@ -3030,60 +1680,28 @@ __pattern_rotate(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Iterator
     auto __temp_buf = oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _Tp>(__exec, __n);
 
     auto __temp_rng =
-        oneapi::dpl::__ranges::all_view<_Tp, __par_backend_hetero::access_mode::write>(__temp_buf.get_buffer());
-
-    const auto __shift = __new_first - __first;
-    oneapi::dpl::__par_backend_hetero::__parallel_for(
-        _BackendTag{}, oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__rotate_wrapper>(__exec),
-        unseq_backend::__rotate_copy<typename ::std::iterator_traits<_Iterator>::difference_type>{__n, __shift}, __n,
-        __buf.all_view(), __temp_rng);
-
-    using _Function = __brick_move<_ExecutionPolicy>;
-    auto __brick = unseq_backend::walk_n<_ExecutionPolicy, _Function>{_Function{}};
-
-    oneapi::dpl::__par_backend_hetero::__parallel_for(_BackendTag{}, ::std::forward<_ExecutionPolicy>(__exec), __brick,
-                                                      __n, __temp_rng, __buf.all_view())
-        .wait();
-
-    return __first + (__last - __new_first);
-}
-
-//------------------------------------------------------------------------
-// rotate_copy
-//------------------------------------------------------------------------
-template <typename _ExecutionPolicy, typename _BidirectionalIterator, typename _ForwardIterator>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, _ForwardIterator>
-__pattern_rotate_copy(_ExecutionPolicy&& __exec, _BidirectionalIterator __first, _BidirectionalIterator __new_first,
-                      _BidirectionalIterator __last, _ForwardIterator __result, /*vector=*/::std::true_type,
-                      /*parallel=*/::std::true_type)
-{
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _BidirectionalIterator, _ForwardIterator>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
-    auto __n = __last - __first;
-    if (__n <= 0)
-        return __result;
-
-    auto __keep1 =
-        oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _BidirectionalIterator>();
-    auto __buf1 = __keep1(__first, __last);
-    auto __keep2 =
-        oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::write, _ForwardIterator>();
-    auto __buf2 = __keep2(__result, __result + __n);
+        oneapi::dpl::__ranges::all_view<_Tp, __par_backend_hetero::access_mode::write>(__temp_buf.get_buffer());
 
     const auto __shift = __new_first - __first;
-
     oneapi::dpl::__par_backend_hetero::__parallel_for(
-        __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
-        unseq_backend::__rotate_copy<typename ::std::iterator_traits<_BidirectionalIterator>::difference_type>{__n,
-                                                                                                               __shift},
-        __n, __buf1.all_view(), __buf2.all_view())
+        _BackendTag{}, oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__rotate_wrapper>(__exec),
+        unseq_backend::__rotate_copy<typename ::std::iterator_traits<_Iterator>::difference_type>{__n, __shift}, __n,
+        __buf.all_view(), __temp_rng);
+
+    using _Function = __brick_move<_ExecutionPolicy>;
+    auto __brick = unseq_backend::walk_n<_ExecutionPolicy, _Function>{_Function{}};
+
+    oneapi::dpl::__par_backend_hetero::__parallel_for(_BackendTag{}, ::std::forward<_ExecutionPolicy>(__exec), __brick,
+                                                      __n, __temp_rng, __buf.all_view())
         .wait();
 
-    return __result + __n;
+    return __first + (__last - __new_first);
 }
 
+//------------------------------------------------------------------------
+// rotate_copy
+//------------------------------------------------------------------------
+
 template <typename _BackendTag, typename _ExecutionPolicy, typename _BidirectionalIterator, typename _ForwardIterator>
 _ForwardIterator
 __pattern_rotate_copy(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _BidirectionalIterator __first,
@@ -3112,86 +1730,6 @@ __pattern_rotate_copy(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Bid
     return __result + __n;
 }
 
-template <typename _ExecutionPolicy, typename _ForwardIterator1, typename _ForwardIterator2, typename _OutputIterator,
-          typename _Compare, typename _IsOpDifference>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, _OutputIterator>
-__pattern_hetero_set_op(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
-                        _ForwardIterator2 __first2, _ForwardIterator2 __last2, _OutputIterator __result,
-                        _Compare __comp, _IsOpDifference)
-{
-    typedef typename ::std::iterator_traits<_ForwardIterator1>::difference_type _Size1;
-    typedef typename ::std::iterator_traits<_ForwardIterator2>::difference_type _Size2;
-
-    const _Size1 __n1 = __last1 - __first1;
-    const _Size2 __n2 = __last2 - __first2;
-
-    //Algo is based on the recommended approach of set_intersection algo for GPU: binary search + scan (copying by mask).
-    using _ReduceOp = ::std::plus<_Size1>;
-    using _Assigner = unseq_backend::__scan_assigner;
-    using _NoAssign = unseq_backend::__scan_no_assign;
-    using _MaskAssigner = unseq_backend::__mask_assigner<2>;
-    using _InitType = unseq_backend::__no_init_value<_Size1>;
-    using _DataAcc = unseq_backend::walk_n<_ExecutionPolicy, oneapi::dpl::__internal::__no_op>;
-
-    _ReduceOp __reduce_op;
-    _Assigner __assign_op;
-    _DataAcc __get_data_op;
-    unseq_backend::__copy_by_mask<_ReduceOp, oneapi::dpl::__internal::__pstl_assign, /*inclusive*/ ::std::true_type, 2>
-        __copy_by_mask_op;
-    unseq_backend::__brick_set_op<_ExecutionPolicy, _Compare, _Size1, _Size2, _IsOpDifference> __create_mask_op{
-        __comp, __n1, __n2};
-
-    // temporary buffer to store boolean mask
-    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, int32_t> __mask_buf(__exec, __n1);
-
-    auto __keep1 =
-        oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _ForwardIterator1>();
-    auto __buf1 = __keep1(__first1, __last1);
-    auto __keep2 =
-        oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _ForwardIterator2>();
-    auto __buf2 = __keep2(__first2, __last2);
-
-    auto __keep3 = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::write, _OutputIterator>();
-    auto __buf3 = __keep3(__result, __result + __n1);
-
-    auto __result_size =
-        __par_backend_hetero::__parallel_transform_scan_base(
-            ::std::forward<_ExecutionPolicy>(__exec),
-            oneapi::dpl::__ranges::make_zip_view(
-                __buf1.all_view(), __buf2.all_view(),
-                oneapi::dpl::__ranges::all_view<int32_t, __par_backend_hetero::access_mode::read_write>(
-                    __mask_buf.get_buffer())),
-            __buf3.all_view(), __reduce_op, _InitType{},
-            // local scan
-            unseq_backend::__scan</*inclusive*/ ::std::true_type, _ExecutionPolicy, _ReduceOp, _DataAcc, _Assigner,
-                                  _MaskAssigner, decltype(__create_mask_op), _InitType>{
-                __reduce_op, __get_data_op, __assign_op, _MaskAssigner{}, __create_mask_op},
-            // scan between groups
-            unseq_backend::__scan</*inclusive=*/::std::true_type, _ExecutionPolicy, _ReduceOp, _DataAcc, _NoAssign,
-                                  _Assigner, _DataAcc, _InitType>{__reduce_op, __get_data_op, _NoAssign{}, __assign_op,
-                                                                  __get_data_op},
-            // global scan
-            __copy_by_mask_op)
-            .get();
-
-    return __result + __result_size;
-}
-
-template <typename _ExecutionPolicy, typename _ForwardIterator1, typename _ForwardIterator2, typename _OutputIterator,
-          typename _Compare>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, _OutputIterator>
-__pattern_set_intersection(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
-                           _ForwardIterator2 __first2, _ForwardIterator2 __last2, _OutputIterator __result,
-                           _Compare __comp, /*vector=*/::std::true_type, /*parallel=*/::std::true_type)
-{
-    // intersection is empty
-    if (__first1 == __last1 || __first2 == __last2)
-        return __result;
-
-    return __pattern_hetero_set_op(::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2,
-                                   __result, __comp, unseq_backend::_IntersectionTag());
-}
-
 template <typename _BackendTag, typename _ExecutionPolicy, typename _ForwardIterator1, typename _ForwardIterator2,
           typename _OutputIterator, typename _Compare, typename _IsOpDifference>
 _OutputIterator
@@ -3279,35 +1817,6 @@ class __set_difference_copy_case_1
 {
 };
 
-template <typename _ExecutionPolicy, typename _ForwardIterator1, typename _ForwardIterator2, typename _OutputIterator,
-          typename _Compare>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, _OutputIterator>
-__pattern_set_difference(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
-                         _ForwardIterator2 __first2, _ForwardIterator2 __last2, _OutputIterator __result,
-                         _Compare __comp, /*vector=*/::std::true_type,
-                         /*parallel=*/::std::true_type)
-{
-    // {} \ {2}: the difference is empty
-    if (__first1 == __last1)
-        return __result;
-
-    // {1} \ {}: the difference is {1}
-    if (__first2 == __last2)
-    {
-        const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1,
-                                                                              _ForwardIterator2, _OutputIterator>();
-
-        return oneapi::dpl::__internal::__pattern_walk2_brick(
-            __dispatch_tag,
-            oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__set_difference_copy_case_1>(
-                ::std::forward<_ExecutionPolicy>(__exec)),
-            __first1, __last1, __result, oneapi::dpl::__internal::__brick_copy<_ExecutionPolicy>{});
-    }
-
-    return __pattern_hetero_set_op(::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2,
-                                   __result, __comp, unseq_backend::_DifferenceTag());
-}
-
 template <typename _BackendTag, typename _ExecutionPolicy, typename _ForwardIterator1, typename _ForwardIterator2,
           typename _OutputIterator, typename _Compare>
 _OutputIterator
@@ -3344,60 +1853,6 @@ class __set_union_copy_case_2
 {
 };
 
-template <typename _ExecutionPolicy, typename _ForwardIterator1, typename _ForwardIterator2, typename _OutputIterator,
-          typename _Compare>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, _OutputIterator>
-__pattern_set_union(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
-                    _ForwardIterator2 __first2, _ForwardIterator2 __last2, _OutputIterator __result, _Compare __comp,
-                    /*vector=*/::std::true_type, /*parallel=*/::std::true_type)
-{
-    if (__first1 == __last1 && __first2 == __last2)
-        return __result;
-
-    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1,
-                                                                          _ForwardIterator2, _OutputIterator>();
-
-    //{1} is empty
-    if (__first1 == __last1)
-    {
-        return oneapi::dpl::__internal::__pattern_walk2_brick(
-            __dispatch_tag,
-            oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__set_union_copy_case_1>(
-                ::std::forward<_ExecutionPolicy>(__exec)),
-            __first2, __last2, __result, oneapi::dpl::__internal::__brick_copy<_ExecutionPolicy>{});
-    }
-
-    //{2} is empty
-    if (__first2 == __last2)
-    {
-        return oneapi::dpl::__internal::__pattern_walk2_brick(
-            __dispatch_tag,
-            oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__set_union_copy_case_2>(
-                ::std::forward<_ExecutionPolicy>(__exec)),
-            __first1, __last1, __result, oneapi::dpl::__internal::__brick_copy<_ExecutionPolicy>{});
-    }
-
-    typedef typename ::std::iterator_traits<_OutputIterator>::value_type _ValueType;
-
-    // temporary buffer to store intermediate result
-    const auto __n2 = __last2 - __first2;
-    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __diff(__exec, __n2);
-    auto __buf = __diff.get();
-
-    //1. Calc difference {2} \ {1}
-    const auto __n_diff = oneapi::dpl::__internal::__pattern_hetero_set_op(__exec,__first2, __last2, __first1, __last1,
-                                                                           __buf,__comp, unseq_backend::_DifferenceTag()
-                                                                          ) - __buf;
-    //2. Merge {1} and the difference
-    const auto __dispatch_tag1 = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, decltype(__first1),
-                                                                           decltype(__buf), decltype(__result)>();
-    return oneapi::dpl::__internal::__pattern_merge(
-        __dispatch_tag1,
-        oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__set_union_copy_case_2>(
-            ::std::forward<_ExecutionPolicy>(__exec)),
-        __first1, __last1, __buf, __buf + __n_diff, __result, __comp);
-}
-
 template <typename _BackendTag, typename _ExecutionPolicy, typename _ForwardIterator1, typename _ForwardIterator2,
           typename _OutputIterator, typename _Compare>
 _OutputIterator
@@ -3481,72 +1936,6 @@ class __set_symmetric_difference_phase_2
 // 1. Calc difference {1} \ {2}
 // 2. Calc difference {2} \ {1}
 // 3. Merge the differences
-template <typename _ExecutionPolicy, typename _ForwardIterator1, typename _ForwardIterator2, typename _OutputIterator,
-          typename _Compare>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, _OutputIterator>
-__pattern_set_symmetric_difference(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
-                                   _ForwardIterator2 __first2, _ForwardIterator2 __last2, _OutputIterator __result,
-                                   _Compare __comp, /*vector=*/::std::true_type,
-                                   /*parallel=*/::std::true_type)
-{
-    if (__first1 == __last1 && __first2 == __last2)
-        return __result;
-
-    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1,
-                                                                          _ForwardIterator2, _OutputIterator>();
-
-    //{1} is empty
-    if (__first1 == __last1)
-    {
-        return oneapi::dpl::__internal::__pattern_walk2_brick(
-            __dispatch_tag,
-            oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__set_symmetric_difference_copy_case_1>(
-                ::std::forward<_ExecutionPolicy>(__exec)),
-            __first2, __last2, __result, oneapi::dpl::__internal::__brick_copy<_ExecutionPolicy>{});
-    }
-
-    //{2} is empty
-    if (__first2 == __last2)
-    {
-        return oneapi::dpl::__internal::__pattern_walk2_brick(
-            __dispatch_tag,
-            oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__set_symmetric_difference_copy_case_2>(
-                ::std::forward<_ExecutionPolicy>(__exec)),
-            __first1, __last1, __result, oneapi::dpl::__internal::__brick_copy<_ExecutionPolicy>{});
-    }
-
-    typedef typename ::std::iterator_traits<_OutputIterator>::value_type _ValueType;
-
-    // temporary buffers to store intermediate result
-    const auto __n1 = __last1 - __first1;
-    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __diff_1(__exec, __n1);
-    auto __buf_1 = __diff_1.get();
-    const auto __n2 = __last2 - __first2;
-    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __diff_2(__exec, __n2);
-    auto __buf_2 = __diff_2.get();
-
-    //1. Calc difference {1} \ {2}
-    const auto __n_diff_1 =
-        oneapi::dpl::__internal::__pattern_hetero_set_op(
-            oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__set_symmetric_difference_phase_1>(__exec),
-            __first1, __last1, __first2, __last2, __buf_1, __comp, unseq_backend::_DifferenceTag()) -
-        __buf_1;
-
-    //2. Calc difference {2} \ {1}
-    const auto __n_diff_2 =
-        oneapi::dpl::__internal::__pattern_hetero_set_op(
-            oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__set_symmetric_difference_phase_2>(__exec),
-            __first2, __last2, __first1, __last1, __buf_2, __comp, unseq_backend::_DifferenceTag()) -
-        __buf_2;
-
-    //3. Merge the differences
-    constexpr auto __dispatch_tag1 = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, decltype(__buf_1),
-                                                                               decltype(__buf_2), decltype(__result)>();
-    return oneapi::dpl::__internal::__pattern_merge(__dispatch_tag1, ::std::forward<_ExecutionPolicy>(__exec), __buf_1,
-                                                    __buf_1 + __n_diff_1, __buf_2, __buf_2 + __n_diff_2, __result,
-                                                    __comp);
-}
-
 template <typename _BackendTag, typename _ExecutionPolicy, typename _ForwardIterator1, typename _ForwardIterator2,
           typename _OutputIterator, typename _Compare>
 _OutputIterator
@@ -3618,56 +2007,10 @@ class __shift_left_right
 {
 };
 
-template <typename _ExecutionPolicy, typename _Range>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy,
-                                                             oneapi::dpl::__internal::__difference_t<_Range>>
-__pattern_shift_left(_ExecutionPolicy&& __exec, _Range __rng, oneapi::dpl::__internal::__difference_t<_Range> __n)
-{
-    //If (n > 0 && n < m), returns first + (m - n). Otherwise, if n  > 0, returns first. Otherwise, returns last.
-    using _DiffType = oneapi::dpl::__internal::__difference_t<_Range>;
-    _DiffType __size = __rng.size();
-
-    assert(__n > 0 && __n < __size);
-
-    _DiffType __mid = __size / 2 + __size % 2;
-    _DiffType __size_res = __size - __n;
-
-    //TODO: required to implement correct tag selection here
-    using __backend_tag = oneapi::dpl::__internal::__device_backend_tag;
-
-    //1. n >= size/2; 'size - _n' parallel copying
-    if (__n >= __mid)
-    {
-        using _Function = __brick_move<_ExecutionPolicy>;
-        auto __brick = oneapi::dpl::unseq_backend::walk_n<_ExecutionPolicy, _Function>{_Function{}};
-
-        //TODO: to consider use just "read" access mode for a source range and just "write" - for a destination range.
-        auto __src = oneapi::dpl::__ranges::drop_view_simple<_Range, _DiffType>(__rng, __n);
-        auto __dst = oneapi::dpl::__ranges::take_view_simple<_Range, _DiffType>(__rng, __size_res);
-
-        oneapi::dpl::__par_backend_hetero::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
-                                                          __brick, __size_res, __src, __dst)
-            .wait();
-    }
-    else //2. n < size/2; 'n' parallel copying
-    {
-        auto __brick = unseq_backend::__brick_shift_left<_ExecutionPolicy, _DiffType>{__size, __n};
-        oneapi::dpl::__par_backend_hetero::__parallel_for(
-            __backend_tag{},
-            oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__shift_left_right>(
-                ::std::forward<_ExecutionPolicy>(__exec)),
-            __brick, __n, __rng)
-            .wait();
-    }
-
-    return __size_res;
-}
-
-template <typename _ExecutionPolicy, typename _Iterator>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, _Iterator>
-__pattern_shift_left(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last,
-                     typename ::std::iterator_traits<_Iterator>::difference_type __n, /*vector=*/::std::true_type,
-                     /*is_parallel=*/::std::true_type)
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator>
+_Iterator
+__pattern_shift_left(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last,
+                     typename ::std::iterator_traits<_Iterator>::difference_type __n)
 {
     //If (n > 0 && n < m), returns first + (m - n). Otherwise, if n  > 0, returns first. Otherwise, returns last.
     auto __size = __last - __first;
@@ -3684,29 +2027,6 @@ __pattern_shift_left(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __l
     return __first + __res;
 }
 
-template <typename _ExecutionPolicy, typename _Iterator>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, _Iterator>
-__pattern_shift_right(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last,
-                      typename ::std::iterator_traits<_Iterator>::difference_type __n, /*vector=*/::std::true_type,
-                      /*is_parallel=*/::std::true_type)
-{
-    //If (n > 0 && n < m), returns first + n. Otherwise, if n  > 0, returns last. Otherwise, returns first.
-    auto __size = __last - __first;
-    if (__n <= 0)
-        return __first;
-    if (__n >= __size)
-        return __last;
-
-    auto __keep = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read_write, _Iterator>();
-    auto __buf = __keep(__first, __last);
-
-    //A shift right is the shift left with a reverse logic.
-    auto __rng = oneapi::dpl::__ranges::reverse_view_simple<decltype(__buf.all_view())>{__buf.all_view()};
-    auto __res = oneapi::dpl::__internal::__pattern_shift_left(::std::forward<_ExecutionPolicy>(__exec), __rng, __n);
-
-    return __last - __res;
-}
-
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator>
 _Iterator
 __pattern_shift_right(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last,
@@ -3729,26 +2049,6 @@ __pattern_shift_right(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec
     return __last - __res;
 }
 
-template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator>
-_Iterator
-__pattern_shift_left(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last,
-                     typename ::std::iterator_traits<_Iterator>::difference_type __n)
-{
-    //If (n > 0 && n < m), returns first + (m - n). Otherwise, if n  > 0, returns first. Otherwise, returns last.
-    auto __size = __last - __first;
-    if (__n <= 0)
-        return __last;
-    if (__n >= __size)
-        return __first;
-
-    auto __keep = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read_write, _Iterator>();
-    auto __buf = __keep(__first, __last);
-
-    auto __res =
-        oneapi::dpl::__internal::__pattern_shift_left(::std::forward<_ExecutionPolicy>(__exec), __buf.all_view(), __n);
-    return __first + __res;
-}
-
 } // namespace __internal
 } // namespace dpl
 } // namespace oneapi
diff --git a/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
index dde4eb10f58..0a79297e7b8 100644
--- a/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
@@ -37,35 +37,6 @@ namespace __internal
 // transform_reduce (version with two binary functions)
 //------------------------------------------------------------------------
 
-template <typename _ExecutionPolicy, typename _RandomAccessIterator1, typename _RandomAccessIterator2, typename _Tp,
-          typename _BinaryOperation1, typename _BinaryOperation2>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, _Tp>
-__pattern_transform_reduce(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1, _RandomAccessIterator1 __last1,
-                           _RandomAccessIterator2 __first2, _Tp __init, _BinaryOperation1 __binary_op1,
-                           _BinaryOperation2 __binary_op2, /*vector=*/::std::true_type, /*parallel=*/::std::true_type)
-{
-    if (__first1 == __last1)
-        return __init;
-
-    using _Functor = unseq_backend::walk_n<_ExecutionPolicy, _BinaryOperation2>;
-    using _RepackedTp = __par_backend_hetero::__repacked_tuple_t<_Tp>;
-
-    auto __n = __last1 - __first1;
-    auto __keep1 =
-        oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _RandomAccessIterator1>();
-    auto __buf1 = __keep1(__first1, __last1);
-    auto __keep2 =
-        oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _RandomAccessIterator2>();
-    auto __buf2 = __keep2(__first2, __first2 + __n);
-
-    return oneapi::dpl::__par_backend_hetero::__parallel_transform_reduce<_RepackedTp,
-                                                                          ::std::true_type /*is_commutative*/>(
-               ::std::forward<_ExecutionPolicy>(__exec), __binary_op1, _Functor{__binary_op2},
-               unseq_backend::__init_value<_RepackedTp>{__init}, // initial value
-               __buf1.all_view(), __buf2.all_view())
-        .get();
-}
-
 template <typename _BackendTag, typename _ExecutionPolicy, typename _RandomAccessIterator1,
           typename _RandomAccessIterator2, typename _Tp, typename _BinaryOperation1, typename _BinaryOperation2>
 _Tp
@@ -99,30 +70,6 @@ __pattern_transform_reduce(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& _
 // transform_reduce (with unary and binary functions)
 //------------------------------------------------------------------------
 
-template <typename _ExecutionPolicy, typename _ForwardIterator, typename _Tp, typename _BinaryOperation,
-          typename _UnaryOperation>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, _Tp>
-__pattern_transform_reduce(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last, _Tp __init,
-                           _BinaryOperation __binary_op, _UnaryOperation __unary_op, /*vector=*/::std::true_type,
-                           /*parallel=*/::std::true_type)
-{
-    if (__first == __last)
-        return __init;
-
-    using _Functor = unseq_backend::walk_n<_ExecutionPolicy, _UnaryOperation>;
-    using _RepackedTp = __par_backend_hetero::__repacked_tuple_t<_Tp>;
-
-    auto __keep = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _ForwardIterator>();
-    auto __buf = __keep(__first, __last);
-
-    return oneapi::dpl::__par_backend_hetero::__parallel_transform_reduce<_RepackedTp,
-                                                                          ::std::true_type /*is_commutative*/>(
-               ::std::forward<_ExecutionPolicy>(__exec), __binary_op, _Functor{__unary_op},
-               unseq_backend::__init_value<_RepackedTp>{__init}, // initial value
-               __buf.all_view())
-        .get();
-}
-
 template <typename _BackendTag, typename _ExecutionPolicy, typename _ForwardIterator, typename _Tp,
           typename _BinaryOperation, typename _UnaryOperation>
 _Tp
@@ -175,85 +122,6 @@ __iterators_possibly_equal(const sycl_iterator<_Mode1, _T, _Allocator>& __it1,
 }
 #endif // _ONEDPL_BACKEND_SYCL
 
-template <typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2, typename _UnaryOperation,
-          typename _InitType, typename _BinaryOperation, typename _Inclusive>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, _Iterator2>
-__pattern_transform_scan_base(_ExecutionPolicy&& __exec, _Iterator1 __first, _Iterator1 __last, _Iterator2 __result,
-                              _UnaryOperation __unary_op, _InitType __init, _BinaryOperation __binary_op, _Inclusive)
-{
-    if (__first == __last)
-        return __result;
-
-    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator1, _Iterator2>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
-    const auto __n = __last - __first;
-
-    auto __keep1 = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _Iterator1>();
-    auto __buf1 = __keep1(__first, __last);
-
-    // This is a temporary workaround for an in-place exclusive scan while the SYCL backend scan pattern is not fixed.
-    const bool __is_scan_inplace_exclusive = __n > 1 && !_Inclusive{} && __iterators_possibly_equal(__first, __result);
-    if (!__is_scan_inplace_exclusive)
-    {
-        auto __keep2 = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::write, _Iterator2>();
-        auto __buf2 = __keep2(__result, __result + __n);
-
-        oneapi::dpl::__par_backend_hetero::__parallel_transform_scan(
-            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __buf1.all_view(), __buf2.all_view(), __n,
-            __unary_op, __init, __binary_op, _Inclusive{})
-            .wait();
-    }
-    else
-    {
-        assert(__n > 1);
-        assert(!_Inclusive{});
-        assert(__iterators_possibly_equal(__first, __result));
-
-        using _Type = typename _InitType::__value_type;
-
-        auto __policy =
-            __par_backend_hetero::make_wrapped_policy<ExecutionPolicyWrapper>(::std::forward<_ExecutionPolicy>(__exec));
-        using _NewExecutionPolicy = decltype(__policy);
-
-        // Create temporary buffer
-        oneapi::dpl::__par_backend_hetero::__buffer<_NewExecutionPolicy, _Type> __tmp_buf(__policy, __n);
-        auto __first_tmp = __tmp_buf.get();
-        auto __last_tmp = __first_tmp + __n;
-        auto __keep2 = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::write, _Iterator2>();
-        auto __buf2 = __keep2(__first_tmp, __last_tmp);
-
-        // Run main algorithm and save data into temporary buffer
-        oneapi::dpl::__par_backend_hetero::__parallel_transform_scan(__backend_tag{}, __policy, __buf1.all_view(),
-                                                                     __buf2.all_view(), __n, __unary_op, __init,
-                                                                     __binary_op, _Inclusive{})
-            .wait();
-
-        // Move data from temporary buffer into results
-        oneapi::dpl::__internal::__pattern_walk2_brick(__dispatch_tag, ::std::move(__policy), __first_tmp, __last_tmp,
-                                                       __result,
-                                                       oneapi::dpl::__internal::__brick_move<_NewExecutionPolicy>{});
-
-        //TODO: optimize copy back depending on Iterator, i.e. set_final_data for host iterator/pointer
-    }
-
-    return __result + __n;
-}
-
-template <typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2, typename _UnaryOperation, typename _Type,
-          typename _BinaryOperation, typename _Inclusive>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, _Iterator2>
-__pattern_transform_scan(_ExecutionPolicy&& __exec, _Iterator1 __first, _Iterator1 __last, _Iterator2 __result,
-                         _UnaryOperation __unary_op, _Type __init, _BinaryOperation __binary_op, _Inclusive,
-                         /*vector=*/::std::true_type, /*parallel=*/::std::true_type)
-{
-    using _RepackedType = __par_backend_hetero::__repacked_tuple_t<_Type>;
-    using _InitType = unseq_backend::__init_value<_RepackedType>;
-
-    return __pattern_transform_scan_base(::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result,
-                                         __unary_op, _InitType{__init}, __binary_op, _Inclusive{});
-}
-
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2,
           typename _UnaryOperation, typename _InitType, typename _BinaryOperation, typename _Inclusive>
 _Iterator2
@@ -334,21 +202,6 @@ __pattern_transform_scan(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __e
 }
 
 // scan without initial element
-template <typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2, typename _UnaryOperation,
-          typename _BinaryOperation, typename _Inclusive>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, _Iterator2>
-__pattern_transform_scan(_ExecutionPolicy&& __exec, _Iterator1 __first, _Iterator1 __last, _Iterator2 __result,
-                         _UnaryOperation __unary_op, _BinaryOperation __binary_op, _Inclusive,
-                         /*vector=*/::std::true_type, /*parallel=*/::std::true_type)
-{
-    using _Type = typename ::std::iterator_traits<_Iterator1>::value_type;
-    using _RepackedType = __par_backend_hetero::__repacked_tuple_t<_Type>;
-    using _InitType = unseq_backend::__no_init_value<_RepackedType>;
-
-    return __pattern_transform_scan_base(::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result,
-                                         __unary_op, _InitType{}, __binary_op, _Inclusive{});
-}
-
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2,
           typename _UnaryOperation, typename _BinaryOperation, typename _Inclusive>
 _Iterator2
@@ -374,65 +227,6 @@ struct adjacent_difference_wrapper
 {
 };
 
-template <typename _ExecutionPolicy, typename _ForwardIterator1, typename _ForwardIterator2, typename _BinaryOperation>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, _ForwardIterator2>
-__pattern_adjacent_difference(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _ForwardIterator1 __last,
-                              _ForwardIterator2 __d_first, _BinaryOperation __op, /*vector*/ ::std::true_type,
-                              /*parallel*/ ::std::true_type)
-{
-    auto __n = __last - __first;
-    if (__n <= 0)
-        return __d_first;
-
-    const auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
-    using _It1ValueT = typename ::std::iterator_traits<_ForwardIterator1>::value_type;
-    using _It2ValueTRef = typename ::std::iterator_traits<_ForwardIterator2>::reference;
-
-    _ForwardIterator2 __d_last = __d_first + __n;
-
-#if !__SYCL_UNNAMED_LAMBDA__
-    // if we have the only element, just copy it according to the specification
-    if (__n == 1)
-    {
-        return __internal::__except_handler([&__exec, __first, __last, __d_first, __d_last, &__op]() {
-            auto __wrapped_policy = __par_backend_hetero::make_wrapped_policy<adjacent_difference_wrapper>(
-                ::std::forward<_ExecutionPolicy>(__exec));
-
-            __internal::__pattern_walk2_brick(__dispatch_tag, __wrapped_policy, __first, __last, __d_first,
-                                              __internal::__brick_copy<decltype(__wrapped_policy)>{});
-
-            return __d_last;
-        });
-    }
-    else
-#endif
-    {
-        return __internal::__except_handler([&__exec, __first, __last, __d_first, __d_last, &__op, __n]() {
-            auto __fn = [__op](_It1ValueT __in1, _It1ValueT __in2, _It2ValueTRef __out1) {
-                __out1 = __op(__in2, __in1); // This move assignment is allowed by the C++ standard draft N4810
-            };
-
-            auto __keep1 =
-                oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _ForwardIterator1>();
-            auto __buf1 = __keep1(__first, __last);
-            auto __keep2 =
-                oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::write, _ForwardIterator2>();
-            auto __buf2 = __keep2(__d_first, __d_last);
-
-            using _Function = unseq_backend::walk_adjacent_difference<_ExecutionPolicy, decltype(__fn)>;
-
-            oneapi::dpl::__par_backend_hetero::__parallel_for(__backend_tag{}, __exec, _Function{__fn}, __n,
-                                                              __buf1.all_view(), __buf2.all_view())
-                .wait();
-
-            return __d_last;
-        });
-    }
-}
-
 template <typename _BackendTag, typename _ExecutionPolicy, typename _ForwardIterator1, typename _ForwardIterator2,
           typename _BinaryOperation>
 _ForwardIterator2

From c3bbe584729397edbba7e1eaf371dee092296569 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 09:45:46 +0100
Subject: [PATCH 300/566] Remove old implementations with
 __enable_if_device_execution_policy

---
 .../pstl/hetero/dpcpp/parallel_backend_sycl.h | 355 ------------------
 1 file changed, 355 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
index ebbe14843c8..60de6356b37 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
@@ -773,65 +773,6 @@ __parallel_transform_scan_base(oneapi::dpl::__internal::__device_backend_tag, _E
         __binary_op, __init, __local_scan, __group_scan, __global_scan);
 }
 
-template <typename _ExecutionPolicy, typename _Range1, typename _Range2, typename _UnaryOperation, typename _InitType,
-          typename _BinaryOperation, typename _Inclusive,
-          oneapi::dpl::__internal::__enable_if_device_execution_policy<_ExecutionPolicy, int> = 0>
-auto
-__parallel_transform_scan(_ExecutionPolicy&& __exec, _Range1&& __in_rng, _Range2&& __out_rng, ::std::size_t __n,
-                          _UnaryOperation __unary_op, _InitType __init, _BinaryOperation __binary_op, _Inclusive)
-{
-    using _Type = typename _InitType::__value_type;
-
-    // Next power of 2 greater than or equal to __n
-    auto __n_uniform = __n;
-    if ((__n_uniform & (__n_uniform - 1)) != 0)
-        __n_uniform = oneapi::dpl::__internal::__dpl_bit_floor(__n) << 1;
-
-    // Pessimistically only use half of the memory to take into account memory used by compiled kernel
-    const ::std::size_t __max_slm_size =
-        __exec.queue().get_device().template get_info<sycl::info::device::local_mem_size>() / 2;
-    const auto __req_slm_size = sizeof(_Type) * __n_uniform;
-
-    constexpr int __single_group_upper_limit = 16384;
-
-    constexpr bool __can_use_group_scan = unseq_backend::__has_known_identity<_BinaryOperation, _Type>::value;
-    if constexpr (__can_use_group_scan)
-    {
-        if (__n <= __single_group_upper_limit && __max_slm_size >= __req_slm_size)
-        {
-            return __parallel_transform_scan_single_group(
-                std::forward<_ExecutionPolicy>(__exec), ::std::forward<_Range1>(__in_rng),
-                ::std::forward<_Range2>(__out_rng), __n, __unary_op, __init, __binary_op, _Inclusive{});
-        }
-    }
-
-    // Either we can't use group scan or this input is too big for one workgroup
-    using _Assigner = unseq_backend::__scan_assigner;
-    using _NoAssign = unseq_backend::__scan_no_assign;
-    using _UnaryFunctor = unseq_backend::walk_n<_ExecutionPolicy, _UnaryOperation>;
-    using _NoOpFunctor = unseq_backend::walk_n<_ExecutionPolicy, oneapi::dpl::__internal::__no_op>;
-
-    _Assigner __assign_op;
-    _NoAssign __no_assign_op;
-    _NoOpFunctor __get_data_op;
-
-    return __future(
-        __parallel_transform_scan_base(
-            ::std::forward<_ExecutionPolicy>(__exec), ::std::forward<_Range1>(__in_rng),
-            ::std::forward<_Range2>(__out_rng), __binary_op, __init,
-            // local scan
-            unseq_backend::__scan<_Inclusive, _ExecutionPolicy, _BinaryOperation, _UnaryFunctor, _Assigner, _Assigner,
-                                  _NoOpFunctor, _InitType>{__binary_op, _UnaryFunctor{__unary_op}, __assign_op,
-                                                           __assign_op, __get_data_op},
-            // scan between groups
-            unseq_backend::__scan</*inclusive=*/::std::true_type, _ExecutionPolicy, _BinaryOperation, _NoOpFunctor,
-                                  _NoAssign, _Assigner, _NoOpFunctor, unseq_backend::__no_init_value<_Type>>{
-                __binary_op, _NoOpFunctor{}, __no_assign_op, __assign_op, __get_data_op},
-            // global scan
-            unseq_backend::__global_scan_functor<_Inclusive, _BinaryOperation, _InitType>{__binary_op, __init})
-            .event());
-}
-
 template <typename _ExecutionPolicy, typename _Range1, typename _Range2, typename _UnaryOperation, typename _InitType,
           typename _BinaryOperation, typename _Inclusive>
 auto
@@ -933,48 +874,6 @@ struct __invoke_single_group_copy_if
     }
 };
 
-template <typename _ExecutionPolicy, typename _InRng, typename _OutRng, typename _Size, typename _CreateMaskOp,
-          typename _CopyByMaskOp,
-          oneapi::dpl::__internal::__enable_if_device_execution_policy<_ExecutionPolicy, int> = 0>
-auto
-__parallel_scan_copy(_ExecutionPolicy&& __exec, _InRng&& __in_rng, _OutRng&& __out_rng, _Size __n,
-                     _CreateMaskOp __create_mask_op, _CopyByMaskOp __copy_by_mask_op)
-{
-    using _ReduceOp = ::std::plus<_Size>;
-    using _Assigner = unseq_backend::__scan_assigner;
-    using _NoAssign = unseq_backend::__scan_no_assign;
-    using _MaskAssigner = unseq_backend::__mask_assigner<1>;
-    using _DataAcc = unseq_backend::walk_n<_ExecutionPolicy, oneapi::dpl::__internal::__no_op>;
-    using _InitType = unseq_backend::__no_init_value<_Size>;
-
-    _Assigner __assign_op;
-    _ReduceOp __reduce_op;
-    _DataAcc __get_data_op;
-    _MaskAssigner __add_mask_op;
-
-    // temporary buffer to store boolean mask
-    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, int32_t> __mask_buf(__exec, __n);
-
-    return __parallel_transform_scan_base(
-        oneapi::dpl::__internal::__device_backend_tag{},
-        ::std::forward<_ExecutionPolicy>(__exec),
-        oneapi::dpl::__ranges::make_zip_view(
-            ::std::forward<_InRng>(__in_rng),
-            oneapi::dpl::__ranges::all_view<int32_t, __par_backend_hetero::access_mode::read_write>(
-                __mask_buf.get_buffer())),
-        ::std::forward<_OutRng>(__out_rng), __reduce_op, _InitType{},
-        // local scan
-        unseq_backend::__scan</*inclusive*/ ::std::true_type, _ExecutionPolicy, _ReduceOp, _DataAcc, _Assigner,
-                              _MaskAssigner, _CreateMaskOp, _InitType>{__reduce_op, __get_data_op, __assign_op,
-                                                                       __add_mask_op, __create_mask_op},
-        // scan between groups
-        unseq_backend::__scan</*inclusive*/ ::std::true_type, _ExecutionPolicy, _ReduceOp, _DataAcc, _NoAssign,
-                              _Assigner, _DataAcc, _InitType>{__reduce_op, __get_data_op, _NoAssign{}, __assign_op,
-                                                              __get_data_op},
-        // global scan
-        __copy_by_mask_op);
-}
-
 template <typename _ExecutionPolicy, typename _InRng, typename _OutRng, typename _Size, typename _CreateMaskOp,
           typename _CopyByMaskOp>
 auto
@@ -1016,50 +915,6 @@ __parallel_scan_copy(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPo
         __copy_by_mask_op);
 }
 
-template <typename _ExecutionPolicy, typename _InRng, typename _OutRng, typename _Size, typename _Pred,
-          oneapi::dpl::__internal::__enable_if_device_execution_policy<_ExecutionPolicy, int> = 0>
-auto
-__parallel_copy_if(_ExecutionPolicy&& __exec, _InRng&& __in_rng, _OutRng&& __out_rng, _Size __n, _Pred __pred)
-{
-    using _SingleGroupInvoker = __invoke_single_group_copy_if<_Size>;
-
-    // Next power of 2 greater than or equal to __n
-    auto __n_uniform = ::oneapi::dpl::__internal::__dpl_bit_ceil(static_cast<::std::make_unsigned_t<_Size>>(__n));
-
-    // Pessimistically only use half of the memory to take into account memory used by compiled kernel
-    const ::std::size_t __max_slm_size =
-        __exec.queue().get_device().template get_info<sycl::info::device::local_mem_size>() / 2;
-
-    // The kernel stores n integers for the predicate and another n integers for the offsets
-    const auto __req_slm_size = sizeof(::std::uint16_t) * __n_uniform * 2;
-
-    constexpr ::std::uint16_t __single_group_upper_limit = 16384;
-
-    ::std::size_t __max_wg_size = oneapi::dpl::__internal::__max_work_group_size(__exec);
-
-    if (__n <= __single_group_upper_limit && __max_slm_size >= __req_slm_size &&
-        __max_wg_size >= _SingleGroupInvoker::__targeted_wg_size)
-    {
-        using _SizeBreakpoints =
-            ::std::integer_sequence<::std::uint16_t, 16, 32, 64, 128, 256, 512, 1024, 2048, 4096, 8192, 16384>;
-
-        return __par_backend_hetero::__static_monotonic_dispatcher<_SizeBreakpoints>::__dispatch(
-            _SingleGroupInvoker{}, __n, ::std::forward<_ExecutionPolicy>(__exec), __n, ::std::forward<_InRng>(__in_rng),
-            ::std::forward<_OutRng>(__out_rng), __pred);
-    }
-    else
-    {
-        using _ReduceOp = ::std::plus<_Size>;
-        using CreateOp = unseq_backend::__create_mask<_Pred, _Size>;
-        using CopyOp = unseq_backend::__copy_by_mask<_ReduceOp, oneapi::dpl::__internal::__pstl_assign,
-                                                     /*inclusive*/ ::std::true_type, 1>;
-
-        return __parallel_scan_copy(oneapi::dpl::__internal::__device_backend_tag{},
-                                    ::std::forward<_ExecutionPolicy>(__exec), ::std::forward<_InRng>(__in_rng),
-                                    ::std::forward<_OutRng>(__out_rng), __n, CreateOp{__pred}, CopyOp{});
-    }
-}
-
 template <typename _ExecutionPolicy, typename _InRng, typename _OutRng, typename _Size, typename _Pred>
 auto
 __parallel_copy_if(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&& __exec, _InRng&& __in_rng,
@@ -1238,109 +1093,6 @@ struct __early_exit_find_or
 // parallel_find_or - sync pattern
 //------------------------------------------------------------------------
 
-// Base pattern for __parallel_or and __parallel_find. The execution depends on tag type _BrickTag.
-template <typename _ExecutionPolicy, typename _Brick, typename _BrickTag, typename... _Ranges>
-oneapi::dpl::__internal::__enable_if_device_execution_policy<
-    _ExecutionPolicy,
-    ::std::conditional_t<::std::is_same_v<_BrickTag, __parallel_or_tag>, bool,
-                         oneapi::dpl::__internal::__difference_t<
-                             typename oneapi::dpl::__ranges::__get_first_range_type<_Ranges...>::type>>>
-__parallel_find_or(_ExecutionPolicy&& __exec, _Brick __f, _BrickTag __brick_tag, _Ranges&&... __rngs)
-{
-    using _CustomName = oneapi::dpl::__internal::__policy_kernel_name<_ExecutionPolicy>;
-    using _AtomicType = typename _BrickTag::_AtomicType;
-    using _FindOrKernel =
-        oneapi::dpl::__par_backend_hetero::__internal::__kernel_name_generator<__find_or_kernel, _CustomName, _Brick,
-                                                                               _BrickTag, _Ranges...>;
-
-    constexpr bool __or_tag_check = ::std::is_same_v<_BrickTag, __parallel_or_tag>;
-    auto __rng_n = oneapi::dpl::__ranges::__get_first_range_size(__rngs...);
-    assert(__rng_n > 0);
-
-    // TODO: find a way to generalize getting of reliable work-group size
-    auto __wgroup_size = oneapi::dpl::__internal::__max_work_group_size(__exec);
-#if _ONEDPL_COMPILE_KERNEL
-    auto __kernel = __internal::__kernel_compiler<_FindOrKernel>::__compile(__exec);
-    __wgroup_size = ::std::min(__wgroup_size, oneapi::dpl::__internal::__kernel_work_group_size(__exec, __kernel));
-#endif
-    auto __max_cu = oneapi::dpl::__internal::__max_compute_units(__exec);
-
-    auto __n_groups = (__rng_n - 1) / __wgroup_size + 1;
-    // TODO: try to change __n_groups with another formula for more perfect load balancing
-    __n_groups = ::std::min(__n_groups, decltype(__n_groups)(__max_cu));
-
-    auto __n_iter = (__rng_n - 1) / (__n_groups * __wgroup_size) + 1;
-
-    _PRINT_INFO_IN_DEBUG_MODE(__exec, __wgroup_size, __max_cu);
-
-    _AtomicType __init_value = _BrickTag::__init_value(__rng_n);
-    auto __result = __init_value;
-
-    auto __pred = oneapi::dpl::__par_backend_hetero::__early_exit_find_or<_ExecutionPolicy, _Brick>{__f};
-
-    // scope is to copy data back to __result after destruction of temporary sycl:buffer
-    {
-        auto __temp = sycl::buffer<_AtomicType, 1>(&__result, 1); // temporary storage for global atomic
-
-        // main parallel_for
-        __exec.queue().submit([&](sycl::handler& __cgh) {
-            oneapi::dpl::__ranges::__require_access(__cgh, __rngs...);
-            auto __temp_acc = __temp.template get_access<access_mode::read_write>(__cgh);
-
-            // create local accessor to connect atomic with
-            __dpl_sycl::__local_accessor<_AtomicType> __temp_local(1, __cgh);
-#if _ONEDPL_COMPILE_KERNEL && _ONEDPL_KERNEL_BUNDLE_PRESENT
-            __cgh.use_kernel_bundle(__kernel.get_kernel_bundle());
-#endif
-            __cgh.parallel_for<_FindOrKernel>(
-#if _ONEDPL_COMPILE_KERNEL && !_ONEDPL_KERNEL_BUNDLE_PRESENT
-                __kernel,
-#endif
-                sycl::nd_range</*dim=*/1>(sycl::range</*dim=*/1>(__n_groups * __wgroup_size),
-                                          sycl::range</*dim=*/1>(__wgroup_size)),
-                [=](sycl::nd_item</*dim=*/1> __item_id) {
-                    auto __local_idx = __item_id.get_local_id(0);
-
-                    __dpl_sycl::__atomic_ref<_AtomicType, sycl::access::address_space::global_space> __found(
-                        *__dpl_sycl::__get_accessor_ptr(__temp_acc));
-                    __dpl_sycl::__atomic_ref<_AtomicType, sycl::access::address_space::local_space> __found_local(
-                        *__dpl_sycl::__get_accessor_ptr(__temp_local));
-
-                    // 1. Set initial value to local atomic
-                    if (__local_idx == 0)
-                        __found_local.store(__init_value);
-                    __dpl_sycl::__group_barrier(__item_id);
-
-                    // 2. Find any element that satisfies pred and set local atomic value to global atomic
-                    constexpr auto __comp = typename _BrickTag::_Compare{};
-                    __pred(__item_id, __n_iter, __wgroup_size, __comp, __found_local, __brick_tag, __rngs...);
-                    __dpl_sycl::__group_barrier(__item_id);
-
-                    // Set local atomic value to global atomic
-                    if (__local_idx == 0 && __comp(__found_local.load(), __found.load()))
-                    {
-                        if constexpr (__or_tag_check)
-                            __found.store(1);
-                        else
-                        {
-                            for (auto __old = __found.load(); __comp(__found_local.load(), __old);
-                                 __old = __found.load())
-                            {
-                                __found.compare_exchange_strong(__old, __found_local.load());
-                            }
-                        }
-                    }
-                });
-        });
-        //The end of the scope  -  a point of synchronization (on temporary sycl buffer destruction)
-    }
-
-    if constexpr (__or_tag_check)
-        return __result;
-    else
-        return __result != __init_value ? __result : __rng_n;
-}
-
 // Base pattern for __parallel_or and __parallel_find. The execution depends on tag type _BrickTag.
 template <typename _ExecutionPolicy, typename _Brick, typename _BrickTag, typename... _Ranges>
 ::std::conditional_t<
@@ -1452,26 +1204,6 @@ class __or_policy_wrapper
 {
 };
 
-template <typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2, typename _Brick>
-oneapi::dpl::__internal::__enable_if_device_execution_policy<_ExecutionPolicy, bool>
-__parallel_or(_ExecutionPolicy&& __exec, _Iterator1 __first, _Iterator1 __last, _Iterator2 __s_first,
-              _Iterator2 __s_last, _Brick __f)
-{
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator1, _Iterator2>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
-    auto __keep = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _Iterator1>();
-    auto __buf = __keep(__first, __last);
-    auto __s_keep = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _Iterator2>();
-    auto __s_buf = __s_keep(__s_first, __s_last);
-
-    return oneapi::dpl::__par_backend_hetero::__parallel_find_or(
-        __backend_tag{},
-        __par_backend_hetero::make_wrapped_policy<__or_policy_wrapper>(::std::forward<_ExecutionPolicy>(__exec)), __f,
-        __parallel_or_tag{}, __buf.all_view(), __s_buf.all_view());
-}
-
 template <typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2, typename _Brick>
 bool
 __parallel_or(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&& __exec, _Iterator1 __first,
@@ -1488,25 +1220,6 @@ __parallel_or(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&&
         __parallel_or_tag{}, __buf.all_view(), __s_buf.all_view());
 }
 
-// Special overload for single sequence cases.
-// TODO: check if similar pattern may apply to other algorithms. If so, these overloads should be moved out of
-// backend code.
-template <typename _ExecutionPolicy, typename _Iterator, typename _Brick>
-oneapi::dpl::__internal::__enable_if_device_execution_policy<_ExecutionPolicy, bool>
-__parallel_or(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last, _Brick __f)
-{
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
-    auto __keep = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _Iterator>();
-    auto __buf = __keep(__first, __last);
-
-    return oneapi::dpl::__par_backend_hetero::__parallel_find_or(
-        __backend_tag{},
-        __par_backend_hetero::make_wrapped_policy<__or_policy_wrapper>(::std::forward<_ExecutionPolicy>(__exec)), __f,
-        __parallel_or_tag{}, __buf.all_view());
-}
-
 // Special overload for single sequence cases.
 // TODO: check if similar pattern may apply to other algorithms. If so, these overloads should be moved out of
 // backend code.
@@ -1533,29 +1246,6 @@ class __find_policy_wrapper
 {
 };
 
-template <typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2, typename _Brick, typename _IsFirst>
-oneapi::dpl::__internal::__enable_if_device_execution_policy<_ExecutionPolicy, _Iterator1>
-__parallel_find(_ExecutionPolicy&& __exec, _Iterator1 __first, _Iterator1 __last, _Iterator2 __s_first,
-                _Iterator2 __s_last, _Brick __f, _IsFirst)
-{
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator1, _Iterator2>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
-    auto __keep = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _Iterator1>();
-    auto __buf = __keep(__first, __last);
-    auto __s_keep = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _Iterator2>();
-    auto __s_buf = __s_keep(__s_first, __s_last);
-
-    using _TagType = ::std::conditional_t<_IsFirst::value, __parallel_find_forward_tag<decltype(__buf.all_view())>,
-                                          __parallel_find_backward_tag<decltype(__buf.all_view())>>;
-    return __first + oneapi::dpl::__par_backend_hetero::__parallel_find_or(
-                         __backend_tag{},
-                         __par_backend_hetero::make_wrapped_policy<__find_policy_wrapper>(
-                             ::std::forward<_ExecutionPolicy>(__exec)),
-                         __f, _TagType{}, __buf.all_view(), __s_buf.all_view());
-}
-
 template <typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2, typename _Brick, typename _IsFirst>
 _Iterator1
 __parallel_find(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&& __exec, _Iterator1 __first,
@@ -1578,25 +1268,6 @@ __parallel_find(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&
 // Special overload for single sequence cases.
 // TODO: check if similar pattern may apply to other algorithms. If so, these overloads should be moved out of
 // backend code.
-template <typename _ExecutionPolicy, typename _Iterator, typename _Brick, typename _IsFirst>
-oneapi::dpl::__internal::__enable_if_device_execution_policy<_ExecutionPolicy, _Iterator>
-__parallel_find(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last, _Brick __f, _IsFirst)
-{
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
-    auto __keep = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _Iterator>();
-    auto __buf = __keep(__first, __last);
-
-    using _TagType = ::std::conditional_t<_IsFirst::value, __parallel_find_forward_tag<decltype(__buf.all_view())>,
-                                          __parallel_find_backward_tag<decltype(__buf.all_view())>>;
-    return __first + oneapi::dpl::__par_backend_hetero::__parallel_find_or(
-                         __backend_tag{},
-                         __par_backend_hetero::make_wrapped_policy<__find_policy_wrapper>(
-                             ::std::forward<_ExecutionPolicy>(__exec)),
-                         __f, _TagType{}, __buf.all_view());
-}
-
 template <typename _ExecutionPolicy, typename _Iterator, typename _Brick, typename _IsFirst>
 _Iterator
 __parallel_find(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&& __exec, _Iterator __first,
@@ -2135,18 +1806,6 @@ struct __is_radix_sort_usable_for_type
 };
 
 #if _USE_RADIX_SORT
-template <typename _ExecutionPolicy, typename _Range, typename _Compare, typename _Proj,
-          ::std::enable_if_t<
-              oneapi::dpl::__internal::__is_device_execution_policy_v<::std::decay_t<_ExecutionPolicy>> &&
-                  __is_radix_sort_usable_for_type<oneapi::dpl::__internal::__key_t<_Proj, _Range>, _Compare>::value,
-              int> = 0>
-auto
-__parallel_stable_sort(_ExecutionPolicy&& __exec, _Range&& __rng, _Compare, _Proj __proj)
-{
-    return __parallel_radix_sort<__internal::__is_comp_ascending<::std::decay_t<_Compare>>::value>(
-        ::std::forward<_ExecutionPolicy>(__exec), ::std::forward<_Range>(__rng), __proj);
-}
-
 template <
     typename _ExecutionPolicy, typename _Range, typename _Compare, typename _Proj,
     ::std::enable_if_t<
@@ -2161,20 +1820,6 @@ __parallel_stable_sort(oneapi::dpl::__internal::__device_backend_tag, _Execution
 }
 #endif
 
-template <typename _ExecutionPolicy, typename _Range, typename _Compare, typename _Proj,
-          ::std::enable_if_t<
-              oneapi::dpl::__internal::__is_device_execution_policy_v<::std::decay_t<_ExecutionPolicy>> &&
-                  !__is_radix_sort_usable_for_type<oneapi::dpl::__internal::__key_t<_Proj, _Range>, _Compare>::value,
-              int> = 0>
-auto
-__parallel_stable_sort(_ExecutionPolicy&& __exec, _Range&& __rng, _Compare __comp, _Proj __proj)
-{
-    auto __cmp_f = [__comp, __proj](const auto& __a, const auto& __b) mutable {
-        return __comp(__proj(__a), __proj(__b));
-    };
-    return __parallel_sort_impl(::std::forward<_ExecutionPolicy>(__exec), ::std::forward<_Range>(__rng), __cmp_f);
-}
-
 template <
     typename _ExecutionPolicy, typename _Range, typename _Compare, typename _Proj,
     ::std::enable_if_t<

From 0a070ed5f7d19d52d005048996c96d166b007a60 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 15 Feb 2024 19:12:26 +0100
Subject: [PATCH 301/566] Remove old implementations with
 __enable_if_fpga_execution_policy

---
 .../hetero/dpcpp/parallel_backend_sycl_fpga.h | 126 ------------------
 .../dpcpp/parallel_backend_sycl_reduce.h      | 101 --------------
 2 files changed, 227 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
index 8d563fd1cd1..f0a69e9f0c1 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
@@ -93,18 +93,6 @@ __parallel_for(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionPolicy&& _
 // parallel_transform_reduce
 //------------------------------------------------------------------------
 
-template <typename _Tp, typename _Commutative, typename _ExecutionPolicy, typename _ReduceOp, typename _TransformOp,
-          typename _InitType, oneapi::dpl::__internal::__enable_if_fpga_execution_policy<_ExecutionPolicy, int> = 0,
-          typename... _Ranges>
-auto
-__parallel_transform_reduce(_ExecutionPolicy&& __exec, _ReduceOp __reduce_op, _TransformOp __transform_op,
-                            _InitType __init, _Ranges&&... __rngs)
-{
-    // workaround until we implement more performant version for patterns
-    return oneapi::dpl::__par_backend_hetero::__parallel_transform_reduce<_Tp, _Commutative>(
-        __exec.__device_policy(), __reduce_op, __transform_op, __init, ::std::forward<_Ranges>(__rngs)...);
-}
-
 template <typename _Tp, typename _Commutative, typename _ExecutionPolicy, typename _ReduceOp, typename _TransformOp,
           typename _InitType, typename... _Ranges>
 auto
@@ -120,19 +108,6 @@ __parallel_transform_reduce(oneapi::dpl::__internal::__fpga_backend_tag, _Execut
 // parallel_transform_scan
 //------------------------------------------------------------------------
 
-template <typename _ExecutionPolicy, typename _Range1, typename _Range2, typename _UnaryOperation, typename _InitType,
-          typename _BinaryOperation, typename _Inclusive,
-          oneapi::dpl::__internal::__enable_if_fpga_execution_policy<_ExecutionPolicy, int> = 0>
-auto
-__parallel_transform_scan(_ExecutionPolicy&& __exec, _Range1&& __in_rng, _Range2&& __out_rng, ::std::size_t __n,
-                          _UnaryOperation __unary_op, _InitType __init, _BinaryOperation __binary_op, _Inclusive)
-{
-    // workaround until we implement more performant version for patterns
-    return oneapi::dpl::__par_backend_hetero::__parallel_transform_scan(
-        oneapi::dpl::__internal::__device_backend_tag{}, __exec.__device_policy(), ::std::forward<_Range1>(__in_rng),
-        ::std::forward<_Range2>(__out_rng), __n, __unary_op, __init, __binary_op, _Inclusive{});
-}
-
 template <typename _ExecutionPolicy, typename _Range1, typename _Range2, typename _UnaryOperation, typename _InitType,
           typename _BinaryOperation, typename _Inclusive>
 auto
@@ -159,17 +134,6 @@ __parallel_transform_scan_base(oneapi::dpl::__internal::__fpga_backend_tag, _Exe
         ::std::forward<_Range2>(__rng2), __binary_op, __init, __local_scan, __group_scan, __global_scan);
 }
 
-template <typename _ExecutionPolicy, typename _InRng, typename _OutRng, typename _Size, typename _Pred,
-          oneapi::dpl::__internal::__enable_if_fpga_execution_policy<_ExecutionPolicy, int> = 0>
-auto
-__parallel_copy_if(_ExecutionPolicy&& __exec, _InRng&& __in_rng, _OutRng&& __out_rng, _Size __n, _Pred __pred)
-{
-    // workaround until we implement more performant version for patterns
-    return oneapi::dpl::__par_backend_hetero::__parallel_copy_if(
-        oneapi::dpl::__internal::__device_backend_tag{}, __exec.__device_policy(), ::std::forward<_InRng>(__in_rng),
-        ::std::forward<_OutRng>(__out_rng), __n, __pred);
-}
-
 template <typename _ExecutionPolicy, typename _InRng, typename _OutRng, typename _Size, typename _Pred>
 auto
 __parallel_copy_if(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionPolicy&& __exec, _InRng&& __in_rng,
@@ -181,19 +145,6 @@ __parallel_copy_if(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionPolicy
         ::std::forward<_OutRng>(__out_rng), __n, __pred);
 }
 
-template <typename _ExecutionPolicy, typename _InRng, typename _OutRng, typename _Size, typename _CreateMaskOp,
-          typename _CopyByMaskOp,
-          oneapi::dpl::__internal::__enable_if_fpga_execution_policy<_ExecutionPolicy, int> = 0>
-auto
-__parallel_scan_copy(_ExecutionPolicy&& __exec, _InRng&& __in_rng, _OutRng&& __out_rng, _Size __n,
-                     _CreateMaskOp __create_mask_op, _CopyByMaskOp __copy_by_mask_op)
-{
-    // workaround until we implement more performant version for patterns
-    return oneapi::dpl::__par_backend_hetero::__parallel_scan_copy(
-        oneapi::dpl::__internal::__device_backend_tag{}, __exec.__device_policy(), ::std::forward<_InRng>(__in_rng),
-        ::std::forward<_OutRng>(__out_rng), __n, __create_mask_op, __copy_by_mask_op);
-}
-
 template <typename _ExecutionPolicy, typename _InRng, typename _OutRng, typename _Size, typename _CreateMaskOp,
           typename _CopyByMaskOp>
 auto
@@ -209,18 +160,6 @@ __parallel_scan_copy(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionPoli
 //------------------------------------------------------------------------
 // __parallel_find_or
 //-----------------------------------------------------------------------
-template <typename _ExecutionPolicy, typename _Brick, typename _BrickTag, typename... _Ranges>
-oneapi::dpl::__internal::__enable_if_fpga_execution_policy<
-    _ExecutionPolicy,
-    ::std::conditional_t<::std::is_same_v<_BrickTag, __parallel_or_tag>, bool,
-                         oneapi::dpl::__internal::__difference_t<
-                             typename oneapi::dpl::__ranges::__get_first_range_type<_Ranges...>::type>>>
-__parallel_find_or(_ExecutionPolicy&& __exec, _Brick __f, _BrickTag __brick_tag, _Ranges&&... __rngs)
-{
-    return oneapi::dpl::__par_backend_hetero::__parallel_find_or(oneapi::dpl::__internal::__device_backend_tag{},
-                                                                 __exec.__device_policy(), __f, __brick_tag,
-                                                                 ::std::forward<_Ranges>(__rngs)...);
-}
 
 template <typename _ExecutionPolicy, typename _Brick, typename _BrickTag, typename... _Ranges>
 ::std::conditional_t<
@@ -259,14 +198,6 @@ __parallel_or(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionPolicy&& __
                                                             __s_last, __f);
 }
 
-template <typename _ExecutionPolicy, typename _Iterator, typename _Brick>
-oneapi::dpl::__internal::__enable_if_fpga_execution_policy<_ExecutionPolicy, bool>
-__parallel_or(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last, _Brick __f)
-{
-    // workaround until we implement more performant version for patterns
-    return oneapi::dpl::__par_backend_hetero::__parallel_or(__exec.__device_policy(), __first, __last, __f);
-}
-
 template <typename _ExecutionPolicy, typename _Iterator, typename _Brick>
 bool
 __parallel_or(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionPolicy&& __exec, _Iterator __first,
@@ -281,18 +212,6 @@ __parallel_or(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionPolicy&& __
 // parallel_find
 //-----------------------------------------------------------------------
 
-template <typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2, typename _Brick, typename _IsFirst>
-oneapi::dpl::__internal::__enable_if_fpga_execution_policy<_ExecutionPolicy, _Iterator1>
-__parallel_find(_ExecutionPolicy&& __exec, _Iterator1 __first, _Iterator1 __last, _Iterator2 __s_first,
-                _Iterator2 __s_last, _Brick __f, _IsFirst __is_first)
-{
-
-    // workaround until we implement more performant version for patterns
-    return oneapi::dpl::__par_backend_hetero::__parallel_find(oneapi::dpl::__internal::__device_backend_tag{},
-                                                              __exec.__device_policy(), __first, __last, __s_first,
-                                                              __s_last, __f, __is_first);
-}
-
 template <typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2, typename _Brick, typename _IsFirst>
 _Iterator1
 __parallel_find(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionPolicy&& __exec, _Iterator1 __first,
@@ -304,15 +223,6 @@ __parallel_find(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionPolicy&&
                                                               __s_last, __f, __is_first);
 }
 
-template <typename _ExecutionPolicy, typename _Iterator, typename _Brick, typename _IsFirst>
-oneapi::dpl::__internal::__enable_if_fpga_execution_policy<_ExecutionPolicy, _Iterator>
-__parallel_find(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last, _Brick __f, _IsFirst __is_first)
-{
-    // workaround until we implement more performant version for patterns
-    return oneapi::dpl::__par_backend_hetero::__parallel_find(__exec.__device_policy(), __first, __last, __f,
-                                                              __is_first);
-}
-
 template <typename _ExecutionPolicy, typename _Iterator, typename _Brick, typename _IsFirst>
 _Iterator
 __parallel_find(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionPolicy&& __exec, _Iterator __first,
@@ -327,20 +237,6 @@ __parallel_find(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionPolicy&&
 // parallel_merge
 //-----------------------------------------------------------------------
 
-template <typename _ExecutionPolicy, typename _Range1, typename _Range2, typename _Range3, typename _Compare>
-auto
-__parallel_merge(_ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&& __rng2, _Range3&& __rng3, _Compare __comp)
-    -> oneapi::dpl::__internal::__enable_if_fpga_execution_policy<
-        _ExecutionPolicy, decltype(oneapi::dpl::__par_backend_hetero::__parallel_merge(
-                              __exec.__device_policy(), ::std::forward<_Range1>(__rng1),
-                              ::std::forward<_Range2>(__rng2), ::std::forward<_Range3>(__rng3), __comp))>
-{
-    // workaround until we implement more performant version for patterns
-    return oneapi::dpl::__par_backend_hetero::__parallel_merge(
-        __exec.__device_policy(), ::std::forward<_Range1>(__rng1), ::std::forward<_Range2>(__rng2),
-        ::std::forward<_Range3>(__rng3), __comp);
-}
-
 template <typename _ExecutionPolicy, typename _Range1, typename _Range2, typename _Range3, typename _Compare>
 auto
 __parallel_merge(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionPolicy&& __exec, _Range1&& __rng1,
@@ -359,17 +255,6 @@ __parallel_merge(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionPolicy&&
 // parallel_stable_sort
 //-----------------------------------------------------------------------
 
-template <typename _ExecutionPolicy, typename _Range, typename _Compare, typename _Proj,
-          oneapi::dpl::__internal::__enable_if_fpga_execution_policy<_ExecutionPolicy, int> = 0>
-auto
-__parallel_stable_sort(_ExecutionPolicy&& __exec, _Range&& __rng, _Compare __comp, _Proj __proj)
-{
-    // workaround until we implement more performant version for patterns
-    return oneapi::dpl::__par_backend_hetero::__parallel_stable_sort(oneapi::dpl::__internal::__device_backend_tag{},
-                                                                     __exec.__device_policy(),
-                                                                     ::std::forward<_Range>(__rng), __comp, __proj);
-}
-
 template <typename _ExecutionPolicy, typename _Range, typename _Compare, typename _Proj>
 auto
 __parallel_stable_sort(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionPolicy&& __exec, _Range&& __rng,
@@ -386,17 +271,6 @@ __parallel_stable_sort(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionPo
 //-----------------------------------------------------------------------
 
 // TODO: check if it makes sense to move these wrappers out of backend to a common place
-template <typename _ExecutionPolicy, typename _Iterator, typename _Compare,
-          oneapi::dpl::__internal::__enable_if_fpga_execution_policy<_ExecutionPolicy, int> = 0>
-auto
-__parallel_partial_sort(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __mid, _Iterator __last,
-                        _Compare __comp)
-{
-    // workaround until we implement more performant version for patterns
-    return oneapi::dpl::__par_backend_hetero::__parallel_partial_sort(
-        oneapi::dpl::__internal::__device_backend_tag{}, __exec.__device_policy(), __first, __mid, __last, __comp);
-}
-
 template <typename _ExecutionPolicy, typename _Iterator, typename _Compare>
 auto
 __parallel_partial_sort(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionPolicy&& __exec, _Iterator __first,
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h
index f1f8c04cac2..28b48333d52 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h
@@ -411,107 +411,6 @@ struct __parallel_transform_reduce_impl
 // Mid-sized arrays use two tree reductions with independent __iters_per_work_item.
 // Big arrays are processed with a recursive tree reduction. __work_group_size * __iters_per_work_item elements are
 // reduced in each step.
-template <typename _Tp, typename _Commutative, typename _ExecutionPolicy, typename _ReduceOp, typename _TransformOp,
-          typename _InitType, oneapi::dpl::__internal::__enable_if_device_execution_policy<_ExecutionPolicy, int> = 0,
-          typename... _Ranges>
-auto
-__parallel_transform_reduce(_ExecutionPolicy&& __exec, _ReduceOp __reduce_op, _TransformOp __transform_op,
-                            _InitType __init, _Ranges&&... __rngs)
-{
-    auto __n = oneapi::dpl::__ranges::__get_first_range_size(__rngs...);
-    assert(__n > 0);
-
-    // Get the work group size adjusted to the local memory limit.
-    // Pessimistically double the memory requirement to take into account memory used by compiled kernel.
-    // TODO: find a way to generalize getting of reliable work-group size.
-    ::std::size_t __work_group_size = oneapi::dpl::__internal::__slm_adjusted_work_group_size(__exec, sizeof(_Tp) * 2);
-
-    // Use single work group implementation if array < __work_group_size * __iters_per_work_item.
-    if (__work_group_size >= 256)
-    {
-        if (__n <= 256)
-        {
-            return __parallel_transform_reduce_small_impl<_Tp, 256, 1, _Commutative>(
-                ::std::forward<_ExecutionPolicy>(__exec), __n, __reduce_op, __transform_op, __init,
-                ::std::forward<_Ranges>(__rngs)...);
-        }
-        else if (__n <= 512)
-        {
-            return __parallel_transform_reduce_small_impl<_Tp, 256, 2, _Commutative>(
-                ::std::forward<_ExecutionPolicy>(__exec), __n, __reduce_op, __transform_op, __init,
-                ::std::forward<_Ranges>(__rngs)...);
-        }
-        else if (__n <= 1024)
-        {
-            return __parallel_transform_reduce_small_impl<_Tp, 256, 4, _Commutative>(
-                ::std::forward<_ExecutionPolicy>(__exec), __n, __reduce_op, __transform_op, __init,
-                ::std::forward<_Ranges>(__rngs)...);
-        }
-        else if (__n <= 2048)
-        {
-            return __parallel_transform_reduce_small_impl<_Tp, 256, 8, _Commutative>(
-                ::std::forward<_ExecutionPolicy>(__exec), __n, __reduce_op, __transform_op, __init,
-                ::std::forward<_Ranges>(__rngs)...);
-        }
-        else if (__n <= 4096)
-        {
-            return __parallel_transform_reduce_small_impl<_Tp, 256, 16, _Commutative>(
-                ::std::forward<_ExecutionPolicy>(__exec), __n, __reduce_op, __transform_op, __init,
-                ::std::forward<_Ranges>(__rngs)...);
-        }
-        else if (__n <= 8192)
-        {
-            return __parallel_transform_reduce_small_impl<_Tp, 256, 32, _Commutative>(
-                ::std::forward<_ExecutionPolicy>(__exec), __n, __reduce_op, __transform_op, __init,
-                ::std::forward<_Ranges>(__rngs)...);
-        }
-
-        // Use two-step tree reduction.
-        // First step reduces __work_group_size * __iters_per_work_item_device_kernel elements.
-        // Second step reduces __work_group_size * __iters_per_work_item_work_group_kernel elements.
-        else if (__n <= 2097152)
-        {
-            return __parallel_transform_reduce_mid_impl<_Tp, 256, 32, 1, _Commutative>(
-                ::std::forward<_ExecutionPolicy>(__exec), __n, __reduce_op, __transform_op, __init,
-                ::std::forward<_Ranges>(__rngs)...);
-        }
-        else if (__n <= 4194304)
-        {
-            return __parallel_transform_reduce_mid_impl<_Tp, 256, 32, 2, _Commutative>(
-                ::std::forward<_ExecutionPolicy>(__exec), __n, __reduce_op, __transform_op, __init,
-                ::std::forward<_Ranges>(__rngs)...);
-        }
-        else if (__n <= 8388608)
-        {
-            return __parallel_transform_reduce_mid_impl<_Tp, 256, 32, 4, _Commutative>(
-                ::std::forward<_ExecutionPolicy>(__exec), __n, __reduce_op, __transform_op, __init,
-                ::std::forward<_Ranges>(__rngs)...);
-        }
-        else if (__n <= 16777216)
-        {
-            return __parallel_transform_reduce_mid_impl<_Tp, 256, 32, 8, _Commutative>(
-                ::std::forward<_ExecutionPolicy>(__exec), __n, __reduce_op, __transform_op, __init,
-                ::std::forward<_Ranges>(__rngs)...);
-        }
-        else if (__n <= 33554432)
-        {
-            return __parallel_transform_reduce_mid_impl<_Tp, 256, 32, 16, _Commutative>(
-                ::std::forward<_ExecutionPolicy>(__exec), __n, __reduce_op, __transform_op, __init,
-                ::std::forward<_Ranges>(__rngs)...);
-        }
-        else if (__n <= 67108864)
-        {
-            return __parallel_transform_reduce_mid_impl<_Tp, 256, 32, 32, _Commutative>(
-                ::std::forward<_ExecutionPolicy>(__exec), __n, __reduce_op, __transform_op, __init,
-                ::std::forward<_Ranges>(__rngs)...);
-        }
-    }
-    // Otherwise use a recursive tree reduction.
-    return __parallel_transform_reduce_impl<_Tp, 32, _Commutative>::submit(
-        ::std::forward<_ExecutionPolicy>(__exec), __n, __work_group_size, __reduce_op, __transform_op, __init,
-        ::std::forward<_Ranges>(__rngs)...);
-}
-
 template <typename _Tp, typename _Commutative, typename _ExecutionPolicy, typename _ReduceOp, typename _TransformOp,
           typename _InitType, typename... _Ranges>
 auto

From 9f4ec47c26648f5aff6178f4e663b70748283881 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 15:03:53 +0100
Subject: [PATCH 302/566] Fix error: remove extra __select_backend calls

---
 .../oneapi/dpl/internal/binary_search_impl.h  |  4 +-
 include/oneapi/dpl/pstl/algorithm_impl.h      | 51 +++++--------------
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 12 +----
 .../dpl/pstl/hetero/numeric_impl_hetero.h     |  8 ++-
 include/oneapi/dpl/pstl/numeric_impl.h        |  4 +-
 5 files changed, 21 insertions(+), 58 deletions(-)

diff --git a/include/oneapi/dpl/internal/binary_search_impl.h b/include/oneapi/dpl/internal/binary_search_impl.h
index 8a6f6b2a910..4e4a17b7f21 100644
--- a/include/oneapi/dpl/internal/binary_search_impl.h
+++ b/include/oneapi/dpl/internal/binary_search_impl.h
@@ -117,9 +117,7 @@ OutputIterator
 lower_bound_impl(__hetero_tag<_BackendTag> __tag, Policy&& policy, InputIterator1 start, InputIterator1 end,
                  InputIterator2 value_start, InputIterator2 value_end, OutputIterator result, StrictWeakOrdering comp)
 {
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<Policy, InputIterator1, InputIterator2, OutputIterator>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+    using __backend_tag = typename decltype(__tag)::__backend_tag;
 
     namespace __bknd = __par_backend_hetero;
     const auto size = ::std::distance(start, end);
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index c5d96eacaad..d266c94f232 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -2039,16 +2039,14 @@ __pattern_is_partitioned(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _Fo
 
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _UnaryPredicate>
 bool
-__pattern_is_partitioned(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
+__pattern_is_partitioned(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
                          _RandomAccessIterator __last, _UnaryPredicate __pred)
 {
     //trivial pre-checks
     if (__first == __last)
         return true;
 
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+    using __backend_tag = typename decltype(__tag)::__backend_tag;
 
     return __internal::__except_handler([&]() {
         // State of current range:
@@ -2171,9 +2169,11 @@ __pattern_partition(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _Forward
 
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _UnaryPredicate>
 _RandomAccessIterator
-__pattern_partition(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
+__pattern_partition(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
                     _RandomAccessIterator __last, _UnaryPredicate __pred)
 {
+    using __backend_tag = typename decltype(__tag)::__backend_tag;
+
     // partitioned range: elements before pivot satisfy pred (true part),
     //                    elements after pivot don't satisfy pred (false part)
     struct _PartitionRange
@@ -2201,11 +2201,6 @@ __pattern_partition(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Rando
             // then we should swap the false part of left range and last part of true part of right range
             else if (__size2 > __size1)
             {
-                constexpr auto __dispatch_tag =
-                    oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, decltype(__val1.__pivot),
-                                                              decltype(__val1.__pivot + __size1)>();
-                using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
                 __par_backend::__parallel_for(
                     __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __val1.__pivot, __val1.__pivot + __size1,
                     [__val1, __val2, __size1](_RandomAccessIterator __i, _RandomAccessIterator __j) {
@@ -2217,11 +2212,6 @@ __pattern_partition(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Rando
             // else we should swap the first part of false part of left range and true part of right range
             else
             {
-                constexpr auto __dispatch_tag =
-                    oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, decltype(__val1.__pivot),
-                                                              decltype(__val1.__pivot + __size2)>();
-                using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
                 __par_backend::__parallel_for(
                     __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __val1.__pivot, __val1.__pivot + __size2,
                     [__val1, __val2](_RandomAccessIterator __i, _RandomAccessIterator __j) {
@@ -2231,10 +2221,6 @@ __pattern_partition(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Rando
             }
         };
 
-        constexpr auto __dispatch_tag =
-            oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
-        using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
         _PartitionRange __result = __par_backend::__parallel_reduce(
             __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __init,
             [__pred, __reductor](_RandomAccessIterator __i, _RandomAccessIterator __j,
@@ -2283,9 +2269,11 @@ __pattern_stable_partition(_Tag, _ExecutionPolicy&&, _BidirectionalIterator __fi
 
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _UnaryPredicate>
 _RandomAccessIterator
-__pattern_stable_partition(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
+__pattern_stable_partition(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
                            _RandomAccessIterator __last, _UnaryPredicate __pred)
 {
+    using __backend_tag = typename decltype(__tag)::__backend_tag;
+
     // partitioned range: elements before pivot satisfy pred (true part),
     //                    elements after pivot don't satisfy pred (false part)
     struct _PartitionRange
@@ -2317,10 +2305,6 @@ __pattern_stable_partition(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec,
             }
         };
 
-        constexpr auto __dispatch_tag =
-            oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
-        using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
         _PartitionRange __result = __par_backend::__parallel_reduce(
             __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __init,
             [&__pred, __reductor](_RandomAccessIterator __i, _RandomAccessIterator __j,
@@ -2684,16 +2668,14 @@ __pattern_adjacent_find(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _For
 
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _BinaryPredicate, class _Semantic>
 _RandomAccessIterator
-__pattern_adjacent_find(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
+__pattern_adjacent_find(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
                         _RandomAccessIterator __last, _BinaryPredicate __pred, _Semantic __or_semantic)
 {
+    using __backend_tag = typename decltype(__tag)::__backend_tag;
+
     if (__last - __first < 2)
         return __last;
 
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
     return __internal::__except_handler([&]() {
         return __par_backend::__parallel_reduce(
             __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __last,
@@ -4290,20 +4272,15 @@ __pattern_shift_left(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Forw
 
 template <class _Tag, class _ExecutionPolicy, class _BidirectionalIterator>
 _BidirectionalIterator
-__pattern_shift_right(_Tag, _ExecutionPolicy&& __exec, _BidirectionalIterator __first, _BidirectionalIterator __last,
+__pattern_shift_right(_Tag __tag, _ExecutionPolicy&& __exec, _BidirectionalIterator __first, _BidirectionalIterator __last,
                       typename ::std::iterator_traits<_BidirectionalIterator>::difference_type __n)
 {
     static_assert(__is_backend_tag_v<_Tag>);
 
     using _ReverseIterator = typename ::std::reverse_iterator<_BidirectionalIterator>;
 
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, decltype(_ReverseIterator(__last)),
-                                                  decltype(_ReverseIterator(__first))>();
-
-    auto __res =
-        oneapi::dpl::__internal::__pattern_shift_left(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
-                                                      _ReverseIterator(__last), _ReverseIterator(__first), __n);
+    auto __res = oneapi::dpl::__internal::__pattern_shift_left(
+        __tag, ::std::forward<_ExecutionPolicy>(__exec), _ReverseIterator(__last), _ReverseIterator(__first), __n);
 
     return __res.base();
 }
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index c451be81c52..500a6dcbaac 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -1037,12 +1037,7 @@ __pattern_remove_if(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec,
     oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __buf(__exec, __last - __first);
     auto __copy_first = __buf.get();
 
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, decltype(__first), decltype(__last),
-                                                  decltype(__copy_first)>();
-
-    auto __copy_last =
-        __pattern_copy_if(__dispatch_tag, __exec, __first, __last, __copy_first, __not_pred<_Predicate>{__pred});
+    auto __copy_last = __pattern_copy_if(__tag, __exec, __first, __last, __copy_first, __not_pred<_Predicate>{__pred});
 
     constexpr auto __dispatch_tag1 =
         oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, decltype(__copy_first), decltype(__copy_last),
@@ -1370,11 +1365,8 @@ _Iterator
 __pattern_partition(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last,
                     _UnaryPredicate __pred)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator>();
-
     //TODO: consider nonstable approaches
-    return __pattern_stable_partition(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
-                                      __pred);
+    return __pattern_stable_partition(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __pred);
 }
 
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
index 0a79297e7b8..914c9de7012 100644
--- a/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
@@ -125,15 +125,14 @@ __iterators_possibly_equal(const sycl_iterator<_Mode1, _T, _Allocator>& __it1,
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2,
           typename _UnaryOperation, typename _InitType, typename _BinaryOperation, typename _Inclusive>
 _Iterator2
-__pattern_transform_scan_base(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Iterator1 __first,
+__pattern_transform_scan_base(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator1 __first,
                               _Iterator1 __last, _Iterator2 __result, _UnaryOperation __unary_op, _InitType __init,
                               _BinaryOperation __binary_op, _Inclusive)
 {
     if (__first == __last)
         return __result;
 
-    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator1, _Iterator2>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+    using __backend_tag = typename decltype(__tag)::__backend_tag;
 
     const auto __n = __last - __first;
 
@@ -178,8 +177,7 @@ __pattern_transform_scan_base(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __ex
             .wait();
 
         // Move data from temporary buffer into results
-        oneapi::dpl::__internal::__pattern_walk2_brick(__dispatch_tag, ::std::move(__policy), __first_tmp, __last_tmp,
-                                                       __result,
+        oneapi::dpl::__internal::__pattern_walk2_brick(__tag, ::std::move(__policy), __first_tmp, __last_tmp, __result,
                                                        oneapi::dpl::__internal::__brick_move<_NewExecutionPolicy>{});
 
         //TODO: optimize copy back depending on Iterator, i.e. set_final_data for host iterator/pointer
diff --git a/include/oneapi/dpl/pstl/numeric_impl.h b/include/oneapi/dpl/pstl/numeric_impl.h
index b875883a598..bb64049ace8 100644
--- a/include/oneapi/dpl/pstl/numeric_impl.h
+++ b/include/oneapi/dpl/pstl/numeric_impl.h
@@ -390,9 +390,7 @@ __pattern_adjacent_difference(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&
                               _RandomAccessIterator1 __first, _RandomAccessIterator1 __last,
                               _RandomAccessIterator2 __d_first, _BinaryOperation __op)
 {
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+    using __backend_tag = typename decltype(__tag)::__backend_tag;
 
     assert(__first != __last);
     typedef typename ::std::iterator_traits<_RandomAccessIterator1>::reference _ReferenceType1;

From ff94f67e8b03dcadb892d0edd990cca667b03ee5 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 15:07:25 +0100
Subject: [PATCH 303/566] Fix self-review comment: move 'using __backend_tag =
 ' at first line inside function

---
 include/oneapi/dpl/pstl/algorithm_impl.h      | 20 +++++++++----------
 .../dpl/pstl/hetero/numeric_impl_hetero.h     |  4 ++--
 include/oneapi/dpl/pstl/numeric_impl.h        |  4 ++--
 3 files changed, 14 insertions(+), 14 deletions(-)

diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index d266c94f232..15e065e0065 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -2042,12 +2042,12 @@ bool
 __pattern_is_partitioned(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
                          _RandomAccessIterator __last, _UnaryPredicate __pred)
 {
+    using __backend_tag = typename decltype(__tag)::__backend_tag;
+
     //trivial pre-checks
     if (__first == __last)
         return true;
 
-    using __backend_tag = typename decltype(__tag)::__backend_tag;
-
     return __internal::__except_handler([&]() {
         // State of current range:
         // broken     - current range is not partitioned by pred
@@ -2529,12 +2529,12 @@ void
 __pattern_partial_sort(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
                        _RandomAccessIterator __middle, _RandomAccessIterator __last, _Compare __comp)
 {
+    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
+
     const auto __n = __middle - __first;
     if (__n == 0)
         return;
 
-    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
-
     __except_handler([&]() {
         __par_backend::__parallel_stable_sort(
             __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __comp,
@@ -3302,10 +3302,10 @@ __parallel_set_union_op(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __ex
                         _ForwardIterator1 __last1, _ForwardIterator2 __first2, _ForwardIterator2 __last2,
                         _OutputIterator __result, _Compare __comp, _SetUnionOp __set_union_op)
 {
-    typedef typename ::std::iterator_traits<_ForwardIterator1>::difference_type _DifferenceType;
-
     using __backend_tag = typename decltype(__tag)::__backend_tag;
 
+    typedef typename ::std::iterator_traits<_ForwardIterator1>::difference_type _DifferenceType;
+
     const auto __n1 = __last1 - __first1;
     const auto __n2 = __last2 - __first2;
 
@@ -3899,12 +3899,12 @@ _RandomAccessIterator
 __pattern_min_element(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
                       _RandomAccessIterator __last, _Compare __comp)
 {
+    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
+
     // a trivial case pre-check
     if (__last - __first < 2)
         return __first;
 
-    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
-
     return __internal::__except_handler([&]() {
         return __par_backend::__parallel_reduce(
             __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, /*identity*/ __last,
@@ -3966,12 +3966,12 @@ ::std::pair<_RandomAccessIterator, _RandomAccessIterator>
 __pattern_minmax_element(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
                          _RandomAccessIterator __last, _Compare __comp)
 {
+    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
+
     // a trivial case pre-check
     if (__last - __first < 2)
         return ::std::make_pair(__first, __first);
 
-    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
-
     return __internal::__except_handler([&]() {
         typedef ::std::pair<_RandomAccessIterator, _RandomAccessIterator> _Result;
 
diff --git a/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
index 914c9de7012..cd38e03652e 100644
--- a/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
@@ -129,11 +129,11 @@ __pattern_transform_scan_base(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&
                               _Iterator1 __last, _Iterator2 __result, _UnaryOperation __unary_op, _InitType __init,
                               _BinaryOperation __binary_op, _Inclusive)
 {
+    using __backend_tag = typename decltype(__tag)::__backend_tag;
+
     if (__first == __last)
         return __result;
 
-    using __backend_tag = typename decltype(__tag)::__backend_tag;
-
     const auto __n = __last - __first;
 
     auto __keep1 = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _Iterator1>();
diff --git a/include/oneapi/dpl/pstl/numeric_impl.h b/include/oneapi/dpl/pstl/numeric_impl.h
index bb64049ace8..ea6cf5e4601 100644
--- a/include/oneapi/dpl/pstl/numeric_impl.h
+++ b/include/oneapi/dpl/pstl/numeric_impl.h
@@ -287,6 +287,8 @@ __pattern_transform_scan(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _
                          _RandomAccessIterator __last, _OutputIterator __result, _UnaryOperation __unary_op, _Tp __init,
                          _BinaryOperation __binary_op, _Inclusive)
 {
+    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
+
     typedef typename ::std::iterator_traits<_RandomAccessIterator>::difference_type _DifferenceType;
     _DifferenceType __n = __last - __first;
 
@@ -295,8 +297,6 @@ __pattern_transform_scan(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _
         return __result;
     }
 
-    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
-
     return __internal::__except_handler([&]() {
         __par_backend::__parallel_strict_scan(
             __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __n, __init,

From 446a48bc874ebdbe1bccd7ce2fdb7daa44de1da3 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 15:14:44 +0100
Subject: [PATCH 304/566] Fix self-review comment: define 'using __backend_tag
 = ' by the same way in all code

---
 .../oneapi/dpl/internal/binary_search_impl.h  | 12 ++++-----
 include/oneapi/dpl/pstl/algorithm_impl.h      | 26 +++++++++----------
 .../experimental/internal/for_loop_impl.h     |  8 +++---
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   |  4 +--
 .../dpl/pstl/hetero/numeric_impl_hetero.h     |  2 +-
 include/oneapi/dpl/pstl/numeric_impl.h        |  4 +--
 6 files changed, 28 insertions(+), 28 deletions(-)

diff --git a/include/oneapi/dpl/internal/binary_search_impl.h b/include/oneapi/dpl/internal/binary_search_impl.h
index 4e4a17b7f21..0c48edcd468 100644
--- a/include/oneapi/dpl/internal/binary_search_impl.h
+++ b/include/oneapi/dpl/internal/binary_search_impl.h
@@ -114,10 +114,10 @@ binary_search_impl(_Tag, Policy&& policy, InputIterator1 start, InputIterator1 e
 template <typename _BackendTag, typename Policy, typename InputIterator1, typename InputIterator2,
           typename OutputIterator, typename StrictWeakOrdering>
 OutputIterator
-lower_bound_impl(__hetero_tag<_BackendTag> __tag, Policy&& policy, InputIterator1 start, InputIterator1 end,
+lower_bound_impl(__hetero_tag<_BackendTag>, Policy&& policy, InputIterator1 start, InputIterator1 end,
                  InputIterator2 value_start, InputIterator2 value_end, OutputIterator result, StrictWeakOrdering comp)
 {
-    using __backend_tag = typename decltype(__tag)::__backend_tag;
+    using __backend_tag = typename __hetero_tag<_BackendTag>::__backend_tag;
 
     namespace __bknd = __par_backend_hetero;
     const auto size = ::std::distance(start, end);
@@ -146,10 +146,10 @@ lower_bound_impl(__hetero_tag<_BackendTag> __tag, Policy&& policy, InputIterator
 template <typename _BackendTag, typename Policy, typename InputIterator1, typename InputIterator2,
           typename OutputIterator, typename StrictWeakOrdering>
 OutputIterator
-upper_bound_impl(__hetero_tag<_BackendTag> __tag, Policy&& policy, InputIterator1 start, InputIterator1 end,
+upper_bound_impl(__hetero_tag<_BackendTag>, Policy&& policy, InputIterator1 start, InputIterator1 end,
                  InputIterator2 value_start, InputIterator2 value_end, OutputIterator result, StrictWeakOrdering comp)
 {
-    using __backend_tag = typename decltype(__tag)::__backend_tag;
+    using __backend_tag = typename __hetero_tag<_BackendTag>::__backend_tag;
 
     namespace __bknd = __par_backend_hetero;
     const auto size = ::std::distance(start, end);
@@ -178,10 +178,10 @@ upper_bound_impl(__hetero_tag<_BackendTag> __tag, Policy&& policy, InputIterator
 template <typename _BackendTag, typename Policy, typename InputIterator1, typename InputIterator2,
           typename OutputIterator, typename StrictWeakOrdering>
 OutputIterator
-binary_search_impl(__hetero_tag<_BackendTag> __tag, Policy&& policy, InputIterator1 start, InputIterator1 end,
+binary_search_impl(__hetero_tag<_BackendTag>, Policy&& policy, InputIterator1 start, InputIterator1 end,
                    InputIterator2 value_start, InputIterator2 value_end, OutputIterator result, StrictWeakOrdering comp)
 {
-    using __backend_tag = typename decltype(__tag)::__backend_tag;
+    using __backend_tag = typename __hetero_tag<_BackendTag>::__backend_tag;
 
     namespace __bknd = __par_backend_hetero;
     const auto size = ::std::distance(start, end);
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 15e065e0065..2d6569afc2c 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -1455,10 +1455,10 @@ __pattern_unique(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIte
 // So, a caller passes _CalcMask brick into remove_elements.
 template <class _IsVector, class _ExecutionPolicy, class _ForwardIterator, class _CalcMask>
 _ForwardIterator
-__remove_elements(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _ForwardIterator __first,
+__remove_elements(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _ForwardIterator __first,
                   _ForwardIterator __last, _CalcMask __calc_mask)
 {
-    using __backend_tag = typename decltype(__tag)::__backend_tag;
+    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
 
     typedef typename ::std::iterator_traits<_ForwardIterator>::difference_type _DifferenceType;
     typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _Tp;
@@ -2039,10 +2039,10 @@ __pattern_is_partitioned(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _Fo
 
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _UnaryPredicate>
 bool
-__pattern_is_partitioned(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
+__pattern_is_partitioned(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
                          _RandomAccessIterator __last, _UnaryPredicate __pred)
 {
-    using __backend_tag = typename decltype(__tag)::__backend_tag;
+    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
 
     //trivial pre-checks
     if (__first == __last)
@@ -2169,10 +2169,10 @@ __pattern_partition(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _Forward
 
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _UnaryPredicate>
 _RandomAccessIterator
-__pattern_partition(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
+__pattern_partition(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
                     _RandomAccessIterator __last, _UnaryPredicate __pred)
 {
-    using __backend_tag = typename decltype(__tag)::__backend_tag;
+    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
 
     // partitioned range: elements before pivot satisfy pred (true part),
     //                    elements after pivot don't satisfy pred (false part)
@@ -2269,10 +2269,10 @@ __pattern_stable_partition(_Tag, _ExecutionPolicy&&, _BidirectionalIterator __fi
 
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _UnaryPredicate>
 _RandomAccessIterator
-__pattern_stable_partition(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
+__pattern_stable_partition(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
                            _RandomAccessIterator __last, _UnaryPredicate __pred)
 {
-    using __backend_tag = typename decltype(__tag)::__backend_tag;
+    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
 
     // partitioned range: elements before pivot satisfy pred (true part),
     //                    elements after pivot don't satisfy pred (false part)
@@ -2668,10 +2668,10 @@ __pattern_adjacent_find(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _For
 
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _BinaryPredicate, class _Semantic>
 _RandomAccessIterator
-__pattern_adjacent_find(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
+__pattern_adjacent_find(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
                         _RandomAccessIterator __last, _BinaryPredicate __pred, _Semantic __or_semantic)
 {
-    using __backend_tag = typename decltype(__tag)::__backend_tag;
+    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
 
     if (__last - __first < 2)
         return __last;
@@ -3205,11 +3205,11 @@ inline constexpr auto __set_algo_cut_off = 1000;
 template <class _IsVector, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2,
           class _OutputIterator, class _Compare, class _SizeFunction, class _SetOP>
 _OutputIterator
-__parallel_set_op(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _ForwardIterator1 __first1,
+__parallel_set_op(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _ForwardIterator1 __first1,
                   _ForwardIterator1 __last1, _ForwardIterator2 __first2, _ForwardIterator2 __last2,
                   _OutputIterator __result, _Compare __comp, _SizeFunction __size_func, _SetOP __set_op)
 {
-    using __backend_tag = typename decltype(__tag)::__backend_tag;
+    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
 
     typedef typename ::std::iterator_traits<_ForwardIterator1>::difference_type _DifferenceType;
     typedef typename ::std::iterator_traits<_OutputIterator>::value_type _T;
@@ -3302,7 +3302,7 @@ __parallel_set_union_op(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __ex
                         _ForwardIterator1 __last1, _ForwardIterator2 __first2, _ForwardIterator2 __last2,
                         _OutputIterator __result, _Compare __comp, _SetUnionOp __set_union_op)
 {
-    using __backend_tag = typename decltype(__tag)::__backend_tag;
+    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
 
     typedef typename ::std::iterator_traits<_ForwardIterator1>::difference_type _DifferenceType;
 
diff --git a/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h b/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
index e86beeb276d..13d72c4fab4 100644
--- a/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
+++ b/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
@@ -390,10 +390,10 @@ __pattern_for_loop(_Tag, _ExecutionPolicy&&, _Ip __first, _Ip __last, _Function
 template <typename _IsVector, typename _ExecutionPolicy, typename _Ip, typename _Size, typename _Function,
           typename... _Rest>
 void
-__pattern_for_loop_n(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _Ip __first, _Size __n, _Function __f,
+__pattern_for_loop_n(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Ip __first, _Size __n, _Function __f,
                      __single_stride_type, _Rest&&... __rest)
 {
-    using __backend_tag = typename decltype(__tag)::__backend_tag;
+    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
 
     using __pack_type = __reduction_pack<_Rest...>;
 
@@ -427,10 +427,10 @@ __pattern_for_loop_n(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec,
 template <typename _IsVector, typename _ExecutionPolicy, typename _Ip, typename _Size, typename _Function, typename _Sp,
           typename... _Rest>
 void
-__pattern_for_loop_n(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _Ip __first, _Size __n, _Function __f,
+__pattern_for_loop_n(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Ip __first, _Size __n, _Function __f,
                      _Sp __stride, _Rest&&... __rest)
 {
-    using __backend_tag = typename decltype(__tag)::__backend_tag;
+    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
 
     using __pack_type = __reduction_pack<_Rest...>;
 
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 500a6dcbaac..173e7bf4891 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -1598,9 +1598,9 @@ __pattern_nth_element(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec
 
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator>
 void
-__pattern_reverse(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last)
+__pattern_reverse(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last)
 {
-    using __backend_tag = typename decltype(__tag)::__backend_tag;
+    using __backend_tag = typename __hetero_tag<_BackendTag>::__backend_tag;
 
     auto __n = __last - __first;
     if (__n <= 0)
diff --git a/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
index cd38e03652e..cabcc10eecd 100644
--- a/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
@@ -129,7 +129,7 @@ __pattern_transform_scan_base(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&
                               _Iterator1 __last, _Iterator2 __result, _UnaryOperation __unary_op, _InitType __init,
                               _BinaryOperation __binary_op, _Inclusive)
 {
-    using __backend_tag = typename decltype(__tag)::__backend_tag;
+    using __backend_tag = typename __hetero_tag<_BackendTag>::__backend_tag;
 
     if (__first == __last)
         return __result;
diff --git a/include/oneapi/dpl/pstl/numeric_impl.h b/include/oneapi/dpl/pstl/numeric_impl.h
index ea6cf5e4601..e5b7180e510 100644
--- a/include/oneapi/dpl/pstl/numeric_impl.h
+++ b/include/oneapi/dpl/pstl/numeric_impl.h
@@ -386,11 +386,11 @@ __pattern_adjacent_difference(_Tag, _ExecutionPolicy&&, _ForwardIterator __first
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
           class _BinaryOperation>
 _RandomAccessIterator2
-__pattern_adjacent_difference(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec,
+__pattern_adjacent_difference(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec,
                               _RandomAccessIterator1 __first, _RandomAccessIterator1 __last,
                               _RandomAccessIterator2 __d_first, _BinaryOperation __op)
 {
-    using __backend_tag = typename decltype(__tag)::__backend_tag;
+    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
 
     assert(__first != __last);
     typedef typename ::std::iterator_traits<_RandomAccessIterator1>::reference _ReferenceType1;

From b808d267dfd08088564d567adec7900e5f0f23db Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 15:19:35 +0100
Subject: [PATCH 305/566] Fix self-review comment: revert extra changes and
 incorrectly merged code

---
 include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
index 60de6356b37..aba0a6ada17 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
@@ -268,8 +268,7 @@ auto
 __parallel_for(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&& __exec, _Fp __brick, _Index __count,
                _Ranges&&... __rngs)
 {
-    using _Policy = ::std::decay_t<_ExecutionPolicy>;
-    using _CustomName = typename _Policy::kernel_name;
+    using _CustomName = oneapi::dpl::__internal::__policy_kernel_name<_ExecutionPolicy>;
     using _ForKernel = oneapi::dpl::__par_backend_hetero::__internal::__kernel_name_provider<_CustomName>;
 
     return __parallel_for_submitter<_ForKernel>()(oneapi::dpl::__internal::__device_backend_tag{},

From ccfde944fc75c4d0efc9ec6ca051fa3ac1f41f84 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 15:28:41 +0100
Subject: [PATCH 306/566] 
 include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h - fix error: restore
 __pattern_shift_left(_ExecutionPolicy&& __exec, _Range __rng,
 oneapi::dpl::__internal::__difference_t<_Range> __n)

---
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 42 +++++++++++++++++++
 1 file changed, 42 insertions(+)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 173e7bf4891..47432beca9d 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -1999,6 +1999,48 @@ class __shift_left_right
 {
 };
 
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Range>
+oneapi::dpl::__internal::__difference_t<_Range>
+__pattern_shift_left(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Range __rng,
+                     oneapi::dpl::__internal::__difference_t<_Range> __n)
+{
+    //If (n > 0 && n < m), returns first + (m - n). Otherwise, if n  > 0, returns first. Otherwise, returns last.
+    using _DiffType = oneapi::dpl::__internal::__difference_t<_Range>;
+    _DiffType __size = __rng.size();
+
+    assert(__n > 0 && __n < __size);
+
+    _DiffType __mid = __size / 2 + __size % 2;
+    _DiffType __size_res = __size - __n;
+
+    //1. n >= size/2; 'size - _n' parallel copying
+    if (__n >= __mid)
+    {
+        using _Function = __brick_move<_ExecutionPolicy>;
+        auto __brick = oneapi::dpl::unseq_backend::walk_n<_ExecutionPolicy, _Function>{_Function{}};
+
+        //TODO: to consider use just "read" access mode for a source range and just "write" - for a destination range.
+        auto __src = oneapi::dpl::__ranges::drop_view_simple<_Range, _DiffType>(__rng, __n);
+        auto __dst = oneapi::dpl::__ranges::take_view_simple<_Range, _DiffType>(__rng, __size_res);
+
+        oneapi::dpl::__par_backend_hetero::__parallel_for(_BackendTag{}, ::std::forward<_ExecutionPolicy>(__exec),
+                                                          __brick, __size_res, __src, __dst)
+            .wait();
+    }
+    else //2. n < size/2; 'n' parallel copying
+    {
+        auto __brick = unseq_backend::__brick_shift_left<_ExecutionPolicy, _DiffType>{__size, __n};
+        oneapi::dpl::__par_backend_hetero::__parallel_for(
+            _BackendTag{},
+            oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__shift_left_right>(
+                ::std::forward<_ExecutionPolicy>(__exec)),
+            __brick, __n, __rng)
+            .wait();
+    }
+
+    return __size_res;
+}
+
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator>
 _Iterator
 __pattern_shift_left(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last,

From 0ec6d87576729aef14ea7ad06607b083240f9306 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 15:46:06 +0100
Subject: [PATCH 307/566] Revert "Fix review comment: I would propose to add a
 parameter _BackendType to this template definition..."

This reverts commit 56406dfe64df2f20ce038f78d9032bb952728a07.
---
 include/oneapi/dpl/pstl/algorithm_fwd.h  | 3 +--
 include/oneapi/dpl/pstl/execution_impl.h | 7 +++----
 include/oneapi/dpl/pstl/numeric_fwd.h    | 2 +-
 3 files changed, 5 insertions(+), 7 deletions(-)

diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index 5a1dcba419c..fd89678dd3b 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -26,10 +26,9 @@ namespace dpl
 namespace __internal
 {
 
-template <class _IsVector, typename _BackendType = __par_backend_tag>
+template <class _IsVector>
 struct __parallel_tag;
 
-template <typename _BackendType = __par_backend_tag>
 struct __parallel_forward_tag;
 
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/execution_impl.h b/include/oneapi/dpl/pstl/execution_impl.h
index cb2b71225b1..dae807121cf 100644
--- a/include/oneapi/dpl/pstl/execution_impl.h
+++ b/include/oneapi/dpl/pstl/execution_impl.h
@@ -113,22 +113,21 @@ struct __serial_tag
     using __is_vector = _IsVector;
 };
 
-template <class _IsVector, typename _BackendType = __par_backend_tag>
+template <class _IsVector>
 struct __parallel_tag
 {
     using __is_vector = _IsVector;
     // backend tag can be change depending on
     // TBB availability in the environment
-    using __backend_tag = _BackendType;
+    using __backend_tag = __par_backend_tag;
 };
 
-template <typename _BackendType = __par_backend_tag>
 struct __parallel_forward_tag
 {
     using __is_vector = ::std::false_type;
     // backend tag can be change depending on
     // TBB availability in the environment
-    using __backend_tag = _BackendType;
+    using __backend_tag = __par_backend_tag;
 };
 
 template <class _IsVector, class... _IteratorTypes>
diff --git a/include/oneapi/dpl/pstl/numeric_fwd.h b/include/oneapi/dpl/pstl/numeric_fwd.h
index 9ea0ebb1de4..1c835443e78 100644
--- a/include/oneapi/dpl/pstl/numeric_fwd.h
+++ b/include/oneapi/dpl/pstl/numeric_fwd.h
@@ -25,7 +25,7 @@ namespace dpl
 {
 namespace __internal
 {
-template <class _IsVector, typename _BackendType = __par_backend_tag>
+template <class _IsVector>
 struct __parallel_tag;
 
 //------------------------------------------------------------------------

From 4e7bf459892935ffaf116ad2593c6f08ce115744 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 15:48:11 +0100
Subject: [PATCH 308/566] include/oneapi/dpl/pstl/algorithm_impl.h - fix
 compile error: use of undeclared identifier '__is_vector'

---
 include/oneapi/dpl/pstl/algorithm_impl.h | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 2d6569afc2c..acef0bd9cbc 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -1521,15 +1521,15 @@ __remove_elements(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Forward
                         else
                             ::new (::std::addressof(*__z)) _Tp(::std::move(*__x));
                     },
-                    __is_vector);
+                    _IsVector{});
             },
             [&__m](_DifferenceType __total) { __m = __total; });
 
         // 3. Elements from result are moved to [first, last)
         __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __result,
-                                      __result + __m, [__result, __first, __is_vector](_Tp* __i, _Tp* __j) {
+                                      __result + __m, [__result, __first](_Tp* __i, _Tp* __j) {
                                           __brick_move_destroy<_ExecutionPolicy>{}(__i, __j, __first + (__i - __result),
-                                                                                   __is_vector);
+                                                                                   _IsVector{});
                                       });
         return __first + __m;
     });

From d6722694c41f6830b6e9203a3e011f4b81d8214c Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 15:51:38 +0100
Subject: [PATCH 309/566] Fix compile error: no template named '__hetero_tag';
 did you mean '__internal::__hetero_tag'?

---
 include/oneapi/dpl/internal/binary_search_impl.h     | 12 ++++++------
 .../dpl/internal/inclusive_scan_by_segment_impl.h    |  8 ++++----
 include/oneapi/dpl/internal/reduce_by_segment_impl.h |  6 +++---
 include/oneapi/dpl/internal/scan_by_segment_impl.h   |  6 +++---
 4 files changed, 16 insertions(+), 16 deletions(-)

diff --git a/include/oneapi/dpl/internal/binary_search_impl.h b/include/oneapi/dpl/internal/binary_search_impl.h
index 0c48edcd468..1881ed344f8 100644
--- a/include/oneapi/dpl/internal/binary_search_impl.h
+++ b/include/oneapi/dpl/internal/binary_search_impl.h
@@ -114,10 +114,10 @@ binary_search_impl(_Tag, Policy&& policy, InputIterator1 start, InputIterator1 e
 template <typename _BackendTag, typename Policy, typename InputIterator1, typename InputIterator2,
           typename OutputIterator, typename StrictWeakOrdering>
 OutputIterator
-lower_bound_impl(__hetero_tag<_BackendTag>, Policy&& policy, InputIterator1 start, InputIterator1 end,
+lower_bound_impl(__internal::__hetero_tag<_BackendTag>, Policy&& policy, InputIterator1 start, InputIterator1 end,
                  InputIterator2 value_start, InputIterator2 value_end, OutputIterator result, StrictWeakOrdering comp)
 {
-    using __backend_tag = typename __hetero_tag<_BackendTag>::__backend_tag;
+    using __backend_tag = typename __internal::__hetero_tag<_BackendTag>::__backend_tag;
 
     namespace __bknd = __par_backend_hetero;
     const auto size = ::std::distance(start, end);
@@ -146,10 +146,10 @@ lower_bound_impl(__hetero_tag<_BackendTag>, Policy&& policy, InputIterator1 star
 template <typename _BackendTag, typename Policy, typename InputIterator1, typename InputIterator2,
           typename OutputIterator, typename StrictWeakOrdering>
 OutputIterator
-upper_bound_impl(__hetero_tag<_BackendTag>, Policy&& policy, InputIterator1 start, InputIterator1 end,
+upper_bound_impl(__internal::__hetero_tag<_BackendTag>, Policy&& policy, InputIterator1 start, InputIterator1 end,
                  InputIterator2 value_start, InputIterator2 value_end, OutputIterator result, StrictWeakOrdering comp)
 {
-    using __backend_tag = typename __hetero_tag<_BackendTag>::__backend_tag;
+    using __backend_tag = typename __internal::__hetero_tag<_BackendTag>::__backend_tag;
 
     namespace __bknd = __par_backend_hetero;
     const auto size = ::std::distance(start, end);
@@ -178,10 +178,10 @@ upper_bound_impl(__hetero_tag<_BackendTag>, Policy&& policy, InputIterator1 star
 template <typename _BackendTag, typename Policy, typename InputIterator1, typename InputIterator2,
           typename OutputIterator, typename StrictWeakOrdering>
 OutputIterator
-binary_search_impl(__hetero_tag<_BackendTag>, Policy&& policy, InputIterator1 start, InputIterator1 end,
+binary_search_impl(__internal::__hetero_tag<_BackendTag>, Policy&& policy, InputIterator1 start, InputIterator1 end,
                    InputIterator2 value_start, InputIterator2 value_end, OutputIterator result, StrictWeakOrdering comp)
 {
-    using __backend_tag = typename __hetero_tag<_BackendTag>::__backend_tag;
+    using __backend_tag = typename __internal::__hetero_tag<_BackendTag>::__backend_tag;
 
     namespace __bknd = __par_backend_hetero;
     const auto size = ::std::distance(start, end);
diff --git a/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h b/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h
index 43310cd063b..f211f995b23 100644
--- a/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h
@@ -43,7 +43,7 @@ pattern_inclusive_scan_by_segment(_Tag, Policy&& policy, InputIterator1 first1,
                                   InputIterator2 first2, OutputIterator result, BinaryPredicate binary_pred,
                                   BinaryOperator binary_op)
 {
-    static_assert(__is_backend_tag_v<_Tag>);
+    static_assert(__internal::__is_backend_tag_v<_Tag>);
 
     const auto n = ::std::distance(first1, last1);
 
@@ -78,7 +78,7 @@ pattern_inclusive_scan_by_segment(_Tag, Policy&& policy, InputIterator1 first1,
 template <typename _BackendTag, typename Policy, typename InputIterator1, typename InputIterator2,
           typename OutputIterator, typename BinaryPredicate, typename BinaryOperator>
 OutputIterator
-inclusive_scan_by_segment_impl(__hetero_tag<_BackendTag> __tag, Policy&& policy, InputIterator1 first1, InputIterator1 last1,
+inclusive_scan_by_segment_impl(__internal::__hetero_tag<_BackendTag> __tag, Policy&& policy, InputIterator1 first1, InputIterator1 last1,
                                InputIterator2 first2, OutputIterator result, BinaryPredicate binary_pred,
                                BinaryOperator binary_op, ::std::true_type /* has_known_identity */)
 {
@@ -91,7 +91,7 @@ inclusive_scan_by_segment_impl(__hetero_tag<_BackendTag> __tag, Policy&& policy,
 template <typename _BackendTag, typename Policy, typename InputIterator1, typename InputIterator2,
           typename OutputIterator, typename BinaryPredicate, typename BinaryOperator>
 OutputIterator
-inclusive_scan_by_segment_impl(__hetero_tag<_BackendTag>, Policy&& policy, InputIterator1 first1, InputIterator1 last1,
+inclusive_scan_by_segment_impl(__internal::__hetero_tag<_BackendTag>, Policy&& policy, InputIterator1 first1, InputIterator1 last1,
                                InputIterator2 first2, OutputIterator result, BinaryPredicate binary_pred,
                                BinaryOperator binary_op, ::std::false_type /* has_known_identity */)
 {
@@ -129,7 +129,7 @@ inclusive_scan_by_segment_impl(__hetero_tag<_BackendTag>, Policy&& policy, Input
 template <typename _BackendTag, typename Policy, typename InputIterator1, typename InputIterator2,
           typename OutputIterator, typename BinaryPredicate, typename BinaryOperator>
 OutputIterator
-pattern_inclusive_scan_by_segment(__hetero_tag<_BackendTag> __tag, Policy&& policy, InputIterator1 first1,
+pattern_inclusive_scan_by_segment(__internal::__hetero_tag<_BackendTag> __tag, Policy&& policy, InputIterator1 first1,
                                   InputIterator1 last1, InputIterator2 first2, OutputIterator result,
                                   BinaryPredicate binary_pred, BinaryOperator binary_op)
 {
diff --git a/include/oneapi/dpl/internal/reduce_by_segment_impl.h b/include/oneapi/dpl/internal/reduce_by_segment_impl.h
index 860664bde67..d65f7ee831c 100644
--- a/include/oneapi/dpl/internal/reduce_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/reduce_by_segment_impl.h
@@ -193,7 +193,7 @@ using _SegReducePrefixPhase = __seg_reduce_prefix_kernel<_Name...>;
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Range1, typename _Range2, typename _Range3,
           typename _Range4, typename _BinaryPredicate, typename _BinaryOperator>
 oneapi::dpl::__internal::__difference_t<_Range3>
-__sycl_reduce_by_segment(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Range1&& __keys, _Range2&& __values,
+__sycl_reduce_by_segment(__internal::__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Range1&& __keys, _Range2&& __values,
                          _Range3&& __out_keys, _Range4&& __out_values, _BinaryPredicate __binary_pred,
                          _BinaryOperator __binary_op, ::std::false_type /* has_known_identity */)
 {
@@ -205,7 +205,7 @@ __sycl_reduce_by_segment(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Range1, typename _Range2, typename _Range3,
           typename _Range4, typename _BinaryPredicate, typename _BinaryOperator>
 oneapi::dpl::__internal::__difference_t<_Range3>
-__sycl_reduce_by_segment(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Range1&& __keys, _Range2&& __values,
+__sycl_reduce_by_segment(__internal::__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Range1&& __keys, _Range2&& __values,
                          _Range3&& __out_keys, _Range4&& __out_values, _BinaryPredicate __binary_pred,
                          _BinaryOperator __binary_op, ::std::true_type /* has_known_identity */)
 {
@@ -573,7 +573,7 @@ __sycl_reduce_by_segment(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _
 template <typename _BackendTag, typename Policy, typename InputIterator1, typename InputIterator2,
           typename OutputIterator1, typename OutputIterator2, typename BinaryPred, typename BinaryOperator>
 ::std::pair<OutputIterator1, OutputIterator2>
-reduce_by_segment_impl(__hetero_tag<_BackendTag> __tag, Policy&& policy, InputIterator1 first1, InputIterator1 last1,
+reduce_by_segment_impl(__internal::__hetero_tag<_BackendTag> __tag, Policy&& policy, InputIterator1 first1, InputIterator1 last1,
                        InputIterator2 first2, OutputIterator1 result1, OutputIterator2 result2, BinaryPred binary_pred,
                        BinaryOperator binary_op)
 {
diff --git a/include/oneapi/dpl/internal/scan_by_segment_impl.h b/include/oneapi/dpl/internal/scan_by_segment_impl.h
index 98cc9e1d93a..7007e875861 100644
--- a/include/oneapi/dpl/internal/scan_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/scan_by_segment_impl.h
@@ -367,9 +367,9 @@ struct __sycl_scan_by_segment_impl
 template <typename _BackendTag, typename Policy, typename InputIterator1, typename InputIterator2,
           typename OutputIterator, typename T, typename BinaryPredicate, typename Operator, typename Inclusive>
 OutputIterator
-__scan_by_segment_impl_common(__hetero_tag<_BackendTag>, Policy&& policy, InputIterator1 first1, InputIterator1 last1,
-                              InputIterator2 first2, OutputIterator result, T init, BinaryPredicate binary_pred,
-                              Operator binary_op, Inclusive)
+__scan_by_segment_impl_common(__internal::__hetero_tag<_BackendTag>, Policy&& policy, InputIterator1 first1,
+                              InputIterator1 last1, InputIterator2 first2, OutputIterator result, T init,
+                              BinaryPredicate binary_pred, Operator binary_op, Inclusive)
 {
     const auto n = ::std::distance(first1, last1);
 

From 9df00f22c791d60f354ff4d7dfba04648200da6b Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 15:54:12 +0100
Subject: [PATCH 310/566] Fix compile error: use of undeclared identifier
 '__is_backend_tag_v'

---
 include/oneapi/dpl/internal/binary_search_impl.h          | 6 +++---
 .../oneapi/dpl/internal/exclusive_scan_by_segment_impl.h  | 8 ++++----
 include/oneapi/dpl/internal/reduce_by_segment_impl.h      | 2 +-
 3 files changed, 8 insertions(+), 8 deletions(-)

diff --git a/include/oneapi/dpl/internal/binary_search_impl.h b/include/oneapi/dpl/internal/binary_search_impl.h
index 1881ed344f8..11e47d01417 100644
--- a/include/oneapi/dpl/internal/binary_search_impl.h
+++ b/include/oneapi/dpl/internal/binary_search_impl.h
@@ -74,7 +74,7 @@ OutputIterator
 lower_bound_impl(_Tag, Policy&& policy, InputIterator1 start, InputIterator1 end, InputIterator2 value_start,
                  InputIterator2 value_end, OutputIterator result, StrictWeakOrdering comp)
 {
-    static_assert(__is_backend_tag_v<_Tag>);
+    static_assert(__internal::__is_backend_tag_v<_Tag>);
 
     return oneapi::dpl::transform(policy, value_start, value_end, result,
                                   [=](typename ::std::iterator_traits<InputIterator2>::reference val) {
@@ -88,7 +88,7 @@ OutputIterator
 upper_bound_impl(_Tag, Policy&& policy, InputIterator1 start, InputIterator1 end, InputIterator2 value_start,
                  InputIterator2 value_end, OutputIterator result, StrictWeakOrdering comp)
 {
-    static_assert(__is_backend_tag_v<_Tag>);
+    static_assert(__internal::__is_backend_tag_v<_Tag>);
 
     return oneapi::dpl::transform(policy, value_start, value_end, result,
                                   [=](typename ::std::iterator_traits<InputIterator2>::reference val) {
@@ -102,7 +102,7 @@ OutputIterator
 binary_search_impl(_Tag, Policy&& policy, InputIterator1 start, InputIterator1 end, InputIterator2 value_start,
                    InputIterator2 value_end, OutputIterator result, StrictWeakOrdering comp)
 {
-    static_assert(__is_backend_tag_v<_Tag>);
+    static_assert(__internal::__is_backend_tag_v<_Tag>);
 
     return oneapi::dpl::transform(policy, value_start, value_end, result,
                                   [=](typename ::std::iterator_traits<InputIterator2>::reference val) {
diff --git a/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h b/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h
index 07d5b930c2f..24d5db6b630 100644
--- a/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h
@@ -44,7 +44,7 @@ pattern_exclusive_scan_by_segment(_Tag, Policy&& policy, InputIterator1 first1,
                                   InputIterator2 first2, OutputIterator result, T init, BinaryPredicate binary_pred,
                                   Operator binary_op)
 {
-    static_assert(__is_backend_tag_v<_Tag>);
+    static_assert(__internal::__is_backend_tag_v<_Tag>);
 
     const auto n = ::std::distance(first1, last1);
 
@@ -97,7 +97,7 @@ pattern_exclusive_scan_by_segment(_Tag, Policy&& policy, InputIterator1 first1,
 template <typename _BackendTag, typename Policy, typename InputIterator1, typename InputIterator2,
           typename OutputIterator, typename T, typename BinaryPredicate, typename Operator>
 OutputIterator
-exclusive_scan_by_segment_impl(__hetero_tag<_BackendTag> __tag, Policy&& policy, InputIterator1 first1, InputIterator1 last1,
+exclusive_scan_by_segment_impl(__internal::__hetero_tag<_BackendTag> __tag, Policy&& policy, InputIterator1 first1, InputIterator1 last1,
                                InputIterator2 first2, OutputIterator result, T init, BinaryPredicate binary_pred,
                                Operator binary_op, ::std::true_type /* has_known_identity*/)
 {
@@ -108,7 +108,7 @@ exclusive_scan_by_segment_impl(__hetero_tag<_BackendTag> __tag, Policy&& policy,
 template <typename _BackendTag, typename Policy, typename InputIterator1, typename InputIterator2,
           typename OutputIterator, typename T, typename BinaryPredicate, typename Operator>
 OutputIterator
-exclusive_scan_by_segment_impl(__hetero_tag<_BackendTag>, Policy&& policy, InputIterator1 first1, InputIterator1 last1,
+exclusive_scan_by_segment_impl(__internal::__hetero_tag<_BackendTag>, Policy&& policy, InputIterator1 first1, InputIterator1 last1,
                                InputIterator2 first2, OutputIterator result, T init, BinaryPredicate binary_pred,
                                Operator binary_op, ::std::false_type /* has_known_identity*/)
 {
@@ -166,7 +166,7 @@ exclusive_scan_by_segment_impl(__hetero_tag<_BackendTag>, Policy&& policy, Input
 template <typename _BackendTag, typename Policy, typename InputIterator1, typename InputIterator2,
           typename OutputIterator, typename T, typename BinaryPredicate, typename Operator>
 OutputIterator
-pattern_exclusive_scan_by_segment(__hetero_tag<_BackendTag> __tag, Policy&& policy, InputIterator1 first1,
+pattern_exclusive_scan_by_segment(__internal::__hetero_tag<_BackendTag> __tag, Policy&& policy, InputIterator1 first1,
                                   InputIterator1 last1, InputIterator2 first2, OutputIterator result, T init,
                                   BinaryPredicate binary_pred, Operator binary_op)
 {
diff --git a/include/oneapi/dpl/internal/reduce_by_segment_impl.h b/include/oneapi/dpl/internal/reduce_by_segment_impl.h
index d65f7ee831c..60ec2eb9d42 100644
--- a/include/oneapi/dpl/internal/reduce_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/reduce_by_segment_impl.h
@@ -85,7 +85,7 @@ reduce_by_segment_impl(_Tag, Policy&& policy, InputIterator1 first1, InputIterat
                        OutputIterator1 result1, OutputIterator2 result2, BinaryPred binary_pred,
                        BinaryOperator binary_op)
 {
-    static_assert(__is_backend_tag_v<_Tag>);
+    static_assert(__internal::__is_backend_tag_v<_Tag>);
 
     // The algorithm reduces values in [first2, first2 + (last1-first1)) where the associated
     // keys for the values are equal to the adjacent key. This function's implementation is a derivative work

From eec6ebe0352a7a220ebc08fb4a238f464cc15905 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 16:03:00 +0100
Subject: [PATCH 311/566] include/oneapi/dpl/internal/reduce_by_segment_impl.h
 - fix compile error: use of undeclared identifier '_ExecutionPolicy'

---
 include/oneapi/dpl/internal/reduce_by_segment_impl.h | 5 ++---
 1 file changed, 2 insertions(+), 3 deletions(-)

diff --git a/include/oneapi/dpl/internal/reduce_by_segment_impl.h b/include/oneapi/dpl/internal/reduce_by_segment_impl.h
index 60ec2eb9d42..e5ea2180cdb 100644
--- a/include/oneapi/dpl/internal/reduce_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/reduce_by_segment_impl.h
@@ -624,9 +624,8 @@ oneapi::dpl::__internal::__enable_if_execution_policy<Policy, ::std::pair<Output
 reduce_by_segment(Policy&& policy, InputIterator1 first1, InputIterator1 last1, InputIterator2 first2,
                   OutputIterator1 result1, OutputIterator2 result2, BinaryPred binary_pred, BinaryOperator binary_op)
 {
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, InputIterator1, InputIterator2, OutputIterator1,
-                                                  OutputIterator2>();
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<Policy, InputIterator1, InputIterator2,
+                                                                              OutputIterator1, OutputIterator2>();
 
     return internal::reduce_by_segment_impl(__dispatch_tag, ::std::forward<Policy>(policy), first1, last1, first2,
                                             result1, result2, binary_pred, binary_op);

From dd281bcfbced5201bb98dace8743fd7f28e3a5c4 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 16:10:33 +0100
Subject: [PATCH 312/566] 
 include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h - fix
 compile error: redefinition of '__parallel_or'

---
 .../hetero/dpcpp/parallel_backend_sycl_fpga.h | 32 +++++--------------
 1 file changed, 8 insertions(+), 24 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
index f0a69e9f0c1..9aabb51117c 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
@@ -187,17 +187,6 @@ __parallel_or(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionPolicy&& __
                                                             __s_last, __f);
 }
 
-template <typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2, typename _Brick>
-bool
-__parallel_or(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionPolicy&& __exec, _Iterator1 __first,
-              _Iterator1 __last, _Iterator2 __s_first, _Iterator2 __s_last, _Brick __f)
-{
-    // workaround until we implement more performant version for patterns
-    return oneapi::dpl::__par_backend_hetero::__parallel_or(oneapi::dpl::__internal::__device_backend_tag{},
-                                                            __exec.__device_policy(), __first, __last, __s_first,
-                                                            __s_last, __f);
-}
-
 template <typename _ExecutionPolicy, typename _Iterator, typename _Brick>
 bool
 __parallel_or(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionPolicy&& __exec, _Iterator __first,
@@ -292,19 +281,14 @@ __parallel_histogram(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionPoli
                      _Range1&& __input, _Range2&& __bins,
                      const _BinHashMgr& __binhash_manager)
 {
-    if constexpr (sizeof(oneapi::dpl::__internal::__value_t<_Range2>) <= sizeof(::std::uint32_t))
-    {
-        // workaround until we implement more performant version for patterns
-        return oneapi::dpl::__par_backend_hetero::__parallel_histogram(
-            oneapi::dpl::__internal::__device_backend_tag{},
-            __exec.__device_policy(), __init_event, ::std::forward<_Range1>(__input), ::std::forward<_Range2>(__bins),
-            __binhash_manager);
-    }
-    else
-    {
-        static_assert(false, "histogram is not supported on FPGA devices with output types greater than 32 bits");
-        return __future(sycl::event{});
-    }
+    static_assert(sizeof(oneapi::dpl::__internal::__value_t<_Range2>) <= sizeof(::std::uint32_t),
+                  "histogram is not supported on FPGA devices with output types greater than 32 bits");
+
+    // workaround until we implement more performant version for patterns
+    return oneapi::dpl::__par_backend_hetero::__parallel_histogram(
+        oneapi::dpl::__internal::__device_backend_tag{},
+        __exec.__device_policy(), __init_event, ::std::forward<_Range1>(__input), ::std::forward<_Range2>(__bins),
+        __binhash_manager);
 }
 
 } // namespace __par_backend_hetero

From 228fb95e8c7bbda20ef16d8f3e4ab3c8b05a2153 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 16:24:03 +0100
Subject: [PATCH 313/566] 
 include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h - fix compile
 error: no matching function for call to '__parallel_find_or'

---
 .../oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h  | 7 +++++++
 1 file changed, 7 insertions(+)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
index bee1d1f1f69..74189556494 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
@@ -103,6 +103,7 @@ __pattern_equal(_ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&& __rng2, _
     // TODO: in case of confilicting names
     // __par_backend_hetero::make_wrapped_policy<__par_backend_hetero::__or_policy_wrapper>()
     return !oneapi::dpl::__par_backend_hetero::__parallel_find_or(
+        oneapi::dpl::__internal::__device_backend_tag{}, // TODO required to fix backend tag eval
         ::std::forward<_ExecutionPolicy>(__exec), _Predicate{equal_predicate<_Pred>{__pred}},
         oneapi::dpl::__par_backend_hetero::__parallel_or_tag{},
         oneapi::dpl::__ranges::zip_view(::std::forward<_Range1>(__rng1), ::std::forward<_Range2>(__rng2)));
@@ -125,6 +126,7 @@ __pattern_find_if(_ExecutionPolicy&& __exec, _Range&& __rng, _Pred __pred)
     using _TagType = oneapi::dpl::__par_backend_hetero::__parallel_find_forward_tag<_Range>;
 
     return oneapi::dpl::__par_backend_hetero::__parallel_find_or(
+        oneapi::dpl::__internal::__device_backend_tag{}, // TODO required to fix backend tag eval
         __par_backend_hetero::make_wrapped_policy<__par_backend_hetero::__find_policy_wrapper>(
             ::std::forward<_ExecutionPolicy>(__exec)),
         _Predicate{__pred}, _TagType{}, ::std::forward<_Range>(__rng));
@@ -154,6 +156,7 @@ __pattern_find_end(_ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&& __rng2
     using _TagType = __par_backend_hetero::__parallel_find_backward_tag<_Range1>;
 
     return oneapi::dpl::__par_backend_hetero::__parallel_find_or(
+        oneapi::dpl::__internal::__device_backend_tag{}, // TODO required to fix backend tag eval
         __par_backend_hetero::make_wrapped_policy<__par_backend_hetero::__find_policy_wrapper>(
             ::std::forward<_ExecutionPolicy>(__exec)),
         _Predicate{__pred}, _TagType{}, ::std::forward<_Range1>(__rng1), ::std::forward<_Range2>(__rng2));
@@ -177,6 +180,7 @@ __pattern_find_first_of(_ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&& _
 
     //TODO: To check whether it makes sense to iterate over the second sequence in case of __rng1.size() < __rng2.size()
     return oneapi::dpl::__par_backend_hetero::__parallel_find_or(
+        oneapi::dpl::__internal::__device_backend_tag{}, // TODO required to fix backend tag eval
         __par_backend_hetero::make_wrapped_policy<__par_backend_hetero::__find_policy_wrapper>(
             ::std::forward<_ExecutionPolicy>(__exec)),
         _Predicate{__pred}, _TagType{}, ::std::forward<_Range1>(__rng1), ::std::forward<_Range2>(__rng2));
@@ -195,6 +199,7 @@ __pattern_any_of(_ExecutionPolicy&& __exec, _Range&& __rng, _Pred __pred)
 
     using _Predicate = oneapi::dpl::unseq_backend::single_match_pred<_ExecutionPolicy, _Pred>;
     return oneapi::dpl::__par_backend_hetero::__parallel_find_or(
+        oneapi::dpl::__internal::__device_backend_tag{}, // TODO required to fix backend tag eval
         __par_backend_hetero::make_wrapped_policy<oneapi::dpl::__par_backend_hetero::__or_policy_wrapper>(
             ::std::forward<_ExecutionPolicy>(__exec)),
         _Predicate{__pred}, oneapi::dpl::__par_backend_hetero::__parallel_or_tag{}, ::std::forward<_Range>(__rng));
@@ -232,6 +237,7 @@ __pattern_search(_ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&& __rng2,
     using _TagType = oneapi::dpl::__par_backend_hetero::__parallel_find_forward_tag<_Range1>;
 
     return oneapi::dpl::__par_backend_hetero::__parallel_find_or(
+        oneapi::dpl::__internal::__device_backend_tag{}, // TODO required to fix backend tag eval
         oneapi::dpl::__par_backend_hetero::make_wrapped_policy<oneapi::dpl::__par_backend_hetero::__find_policy_wrapper>
             (::std::forward<_ExecutionPolicy>(__exec)),
         _Predicate{__pred}, _TagType{}, ::std::forward<_Range1>(__rng1), ::std::forward<_Range2>(__rng2));
@@ -293,6 +299,7 @@ __pattern_adjacent_find(_ExecutionPolicy&& __exec, _Range&& __rng, _BinaryPredic
     // TODO: in case of confilicting names
     // __par_backend_hetero::make_wrapped_policy<__par_backend_hetero::__or_policy_wrapper>()
     auto result = oneapi::dpl::__par_backend_hetero::__parallel_find_or(
+        oneapi::dpl::__internal::__device_backend_tag{}, // TODO required to fix backend tag eval
         ::std::forward<_ExecutionPolicy>(__exec), _Predicate{adjacent_find_fn<_BinaryPredicate>{__predicate}},
         _TagType{}, oneapi::dpl::__ranges::zip_view(__rng1, __rng2));
 

From 9cd710c30937c9c4883fa3b82dbe13c6e8967a25 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 16:32:36 +0100
Subject: [PATCH 314/566] include/oneapi/dpl/pstl/numeric_impl.h - fix compile
 error: no matching function for call to '__parallel_transform_scan'

---
 include/oneapi/dpl/pstl/numeric_impl.h | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)

diff --git a/include/oneapi/dpl/pstl/numeric_impl.h b/include/oneapi/dpl/pstl/numeric_impl.h
index e5b7180e510..d5dfe6a09bb 100644
--- a/include/oneapi/dpl/pstl/numeric_impl.h
+++ b/include/oneapi/dpl/pstl/numeric_impl.h
@@ -253,15 +253,17 @@ __pattern_transform_scan(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _Fo
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _OutputIterator,
           class _UnaryOperation, class _Tp, class _BinaryOperation, class _Inclusive>
 ::std::enable_if_t<!::std::is_floating_point_v<_Tp>, _OutputIterator>
-__pattern_transform_scan(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
+__pattern_transform_scan(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
                          _RandomAccessIterator __last, _OutputIterator __result, _UnaryOperation __unary_op, _Tp __init,
                          _BinaryOperation __binary_op, _Inclusive)
 {
+    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
+
     typedef typename ::std::iterator_traits<_RandomAccessIterator>::difference_type _DifferenceType;
 
     return __internal::__except_handler([&]() {
         __par_backend::__parallel_transform_scan(
-            oneapi::dpl::__internal::__serial_backend_tag{},
+            __backend_tag{},
             ::std::forward<_ExecutionPolicy>(__exec), __last - __first,
             [__first, __unary_op](_DifferenceType __i) mutable { return __unary_op(__first[__i]); }, __init,
             __binary_op,

From f3593bbcb6f0467e2a2f56dd8e3cb491f8c11336 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 16:42:09 +0100
Subject: [PATCH 315/566] 
 include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h - fix compile error:
 no matching function for call to '__pattern_merge'

---
 .../oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h   | 12 ++++++++----
 1 file changed, 8 insertions(+), 4 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 47432beca9d..415f0d567c5 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -1230,7 +1230,7 @@ __pattern_merge(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Ite
 
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator, typename _Compare>
 void
-__pattern_inplace_merge(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator __first,
+__pattern_inplace_merge(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Iterator __first,
                         _Iterator __middle, _Iterator __last, _Compare __comp)
 {
     using _ValueType = typename ::std::iterator_traits<_Iterator>::value_type;
@@ -1245,16 +1245,20 @@ __pattern_inplace_merge(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __ex
     auto __copy_first = __buf.get();
     auto __copy_last = __copy_first + __n;
 
-    __pattern_merge(__exec, __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__first),
+    __pattern_merge(oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, decltype(__first), decltype(__middle),
+                                                              decltype(__last), decltype(__copy_first)>(),
+                    __exec, __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__first),
                     __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__middle),
                     __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__middle),
                     __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__last),
                     __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::write>(__copy_first),
-                    __comp, ::std::true_type{}, ::std::true_type{});
+                    __comp);
 
     //TODO: optimize copy back depending on Iterator, i.e. set_final_data for host iterator/pointer
     __pattern_walk2(
-        __tag, __par_backend_hetero::make_wrapped_policy<copy_back_wrapper>(::std::forward<_ExecutionPolicy>(__exec)),
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, decltype(__copy_first), decltype(__copy_last),
+                                                  decltype(__first)>(),
+        __par_backend_hetero::make_wrapped_policy<copy_back_wrapper>(::std::forward<_ExecutionPolicy>(__exec)),
         __copy_first, __copy_last, __first, __brick_move<_ExecutionPolicy>{});
 }
 

From fff76e82870f9d4b1b0a25cb9d8dddcea9b9c842 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 16:49:56 +0100
Subject: [PATCH 316/566] Fix compile error: no matching function for call to
 '__parallel_transform_reduce'

---
 include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h       | 5 +++++
 .../dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h       | 1 +
 include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h         | 2 ++
 3 files changed, 8 insertions(+)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 415f0d567c5..659fcc99155 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -488,6 +488,7 @@ __pattern_min_element(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec
     auto __buf = __keep(__first, __last);
 
     auto __ret_idx = oneapi::dpl::__par_backend_hetero::__parallel_transform_reduce<_ReduceValueType, _Commutative>(
+                         _BackendTag{},
                          ::std::forward<_ExecutionPolicy>(__exec), __reduce_fn, __transform_fn,
                          unseq_backend::__no_init_value{}, // no initial value
                          __buf.all_view())
@@ -556,6 +557,7 @@ __pattern_minmax_element(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __e
 
     auto __ret = oneapi::dpl::__par_backend_hetero::__parallel_transform_reduce<_ReduceValueType,
                                                                                 ::std::false_type /*is_commutative*/>(
+                     _BackendTag{},
                      ::std::forward<_ExecutionPolicy>(__exec), __reduce_fn, __transform_fn,
                      unseq_backend::__no_init_value{}, // no initial value
                      __buf.all_view())
@@ -650,6 +652,7 @@ __pattern_count(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Ite
 
     return oneapi::dpl::__par_backend_hetero::__parallel_transform_reduce<_ReduceValueType,
                                                                           ::std::true_type /*is_commutative*/>(
+               _BackendTag{},
                ::std::forward<_ExecutionPolicy>(__exec), __reduce_fn, __transform_fn,
                unseq_backend::__no_init_value{}, // no initial value
                __buf.all_view())
@@ -1116,6 +1119,7 @@ __pattern_is_partitioned(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __e
 
     auto __res = oneapi::dpl::__par_backend_hetero::__parallel_transform_reduce<_ReduceValueType,
                                                                                 ::std::false_type /*is_commutative*/>(
+                     _BackendTag{},
                      ::std::forward<_ExecutionPolicy>(__exec), __reduce_fn, __transform_fn,
                      unseq_backend::__no_init_value{}, // no initial value
                      __buf.all_view())
@@ -1417,6 +1421,7 @@ __pattern_lexicographical_compare(__hetero_tag<_BackendTag> __tag, _ExecutionPol
     auto __ret_idx =
         oneapi::dpl::__par_backend_hetero::__parallel_transform_reduce<_ReduceValueType,
                                                                        ::std::false_type /*is_commutative*/>(
+            _BackendTag{},
             ::std::forward<_ExecutionPolicy>(__exec), __reduce_fn, __transform_fn,
             unseq_backend::__no_init_value{}, // no initial value
             __buf1.all_view(), __buf2.all_view())
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
index 9aabb51117c..77fa3388f47 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
@@ -101,6 +101,7 @@ __parallel_transform_reduce(oneapi::dpl::__internal::__fpga_backend_tag, _Execut
 {
     // workaround until we implement more performant version for patterns
     return oneapi::dpl::__par_backend_hetero::__parallel_transform_reduce<_Tp, _Commutative>(
+        oneapi::dpl::__internal::__device_backend_tag{},
         __exec.__device_policy(), __reduce_op, __transform_op, __init, ::std::forward<_Ranges>(__rngs)...);
 }
 
diff --git a/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
index cabcc10eecd..84a49029182 100644
--- a/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
@@ -60,6 +60,7 @@ __pattern_transform_reduce(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& _
 
     return oneapi::dpl::__par_backend_hetero::__parallel_transform_reduce<_RepackedTp,
                                                                           ::std::true_type /*is_commutative*/>(
+               _BackendTag{},
                ::std::forward<_ExecutionPolicy>(__exec), __binary_op1, _Functor{__binary_op2},
                unseq_backend::__init_value<_RepackedTp>{__init}, // initial value
                __buf1.all_view(), __buf2.all_view())
@@ -88,6 +89,7 @@ __pattern_transform_reduce(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& _
 
     return oneapi::dpl::__par_backend_hetero::__parallel_transform_reduce<_RepackedTp,
                                                                           ::std::true_type /*is_commutative*/>(
+               _BackendTag{},
                ::std::forward<_ExecutionPolicy>(__exec), __binary_op, _Functor{__unary_op},
                unseq_backend::__init_value<_RepackedTp>{__init}, // initial value
                __buf.all_view())

From 8d70c650006b6ce4237bfe59d40b8131bfbacb4d Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 17:01:14 +0100
Subject: [PATCH 317/566] Fix compile  error: no matching function for call to
 '__parallel_transform_scan_base'

---
 include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h | 1 +
 include/oneapi/dpl/pstl/hetero/numeric_ranges_impl_hetero.h   | 1 +
 2 files changed, 2 insertions(+)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
index 74189556494..258dd76e2d6 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
@@ -363,6 +363,7 @@ __pattern_scan_copy(_ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&& __rng
 
     auto __res =
         __par_backend_hetero::__parallel_transform_scan_base(
+            oneapi::dpl::__internal::__device_backend_tag{}, // TODO required to fix backend tag eval
             ::std::forward<_ExecutionPolicy>(__exec),
             oneapi::dpl::__ranges::zip_view(
                 __rng1, oneapi::dpl::__ranges::all_view<int32_t, __par_backend_hetero::access_mode::read_write>(
diff --git a/include/oneapi/dpl/pstl/hetero/numeric_ranges_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/numeric_ranges_impl_hetero.h
index 0b104a5ff4b..3565ac22473 100644
--- a/include/oneapi/dpl/pstl/hetero/numeric_ranges_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/numeric_ranges_impl_hetero.h
@@ -106,6 +106,7 @@ __pattern_transform_scan_base(_ExecutionPolicy&& __exec, _Range1&& __rng1, _Rang
     _NoOpFunctor __get_data_op;
 
     oneapi::dpl::__par_backend_hetero::__parallel_transform_scan_base(
+        oneapi::dpl::__internal::__device_backend_tag{}, // TODO required to fix backend tag eval
         ::std::forward<_ExecutionPolicy>(__exec), ::std::forward<_Range1>(__rng1), ::std::forward<_Range2>(__rng2),
         __binary_op, __init,
         // local scan

From 40048ea32a4ca2dfe440c6958fd24f2cad0a21c3 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 17:15:49 +0100
Subject: [PATCH 318/566] 
 include/oneapi/dpl/internal/async_impl/glue_async_impl.h - fix compile error:
 no matching function for call to '__pattern_transform_scan_async'

---
 include/oneapi/dpl/internal/async_impl/glue_async_impl.h | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/include/oneapi/dpl/internal/async_impl/glue_async_impl.h b/include/oneapi/dpl/internal/async_impl/glue_async_impl.h
index d89da19493e..d9f06601b45 100644
--- a/include/oneapi/dpl/internal/async_impl/glue_async_impl.h
+++ b/include/oneapi/dpl/internal/async_impl/glue_async_impl.h
@@ -220,9 +220,13 @@ auto
 inclusive_scan_async(_ExecutionPolicy&& __exec, _ForwardIt1 __first1, _ForwardIt1 __last1, _ForwardIt2 __first2,
                      _Events&&... __dependencies)
 {
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIt1, _ForwardIt2>();
+
     using _ValueType = typename ::std::iterator_traits<_ForwardIt1>::value_type;
     wait_for_all(::std::forward<_Events>(__dependencies)...);
     return oneapi::dpl::__internal::__pattern_transform_scan_async(
+        __dispatch_tag,
         ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, oneapi::dpl::__internal::__no_op(),
         ::std::plus<_ValueType>(), /*inclusive=*/::std::true_type());
 }

From 82b7531b5e5ca8be8c8cefeded8e43dc4e877f55 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 17:21:00 +0100
Subject: [PATCH 319/566] 
 include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h - fix compile
 error: no matching function for call to '__parallel_for'

---
 .../dpl/pstl/hetero/algorithm_ranges_impl_hetero.h       | 9 ++++++---
 1 file changed, 6 insertions(+), 3 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
index 258dd76e2d6..e10f8c0a8f5 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
@@ -46,9 +46,10 @@ __pattern_walk_n(_ExecutionPolicy&& __exec, _Function __f, _Ranges&&... __rngs)
     auto __n = oneapi::dpl::__ranges::__get_first_range_size(__rngs...);
     if (__n > 0)
     {
-        oneapi::dpl::__par_backend_hetero::__parallel_for(::std::forward<_ExecutionPolicy>(__exec),
-                                                          unseq_backend::walk_n<_ExecutionPolicy, _Function>{__f}, __n,
-                                                          ::std::forward<_Ranges>(__rngs)...)
+        oneapi::dpl::__par_backend_hetero::__parallel_for(
+            oneapi::dpl::__internal::__device_backend_tag{}, // TODO required to fix backend tag eval
+            ::std::forward<_ExecutionPolicy>(__exec), unseq_backend::walk_n<_ExecutionPolicy, _Function>{__f}, __n,
+            ::std::forward<_Ranges>(__rngs)...)
             .wait();
     }
 }
@@ -753,6 +754,7 @@ __pattern_reduce_by_segment(_ExecutionPolicy&& __exec, _Range1&& __keys, _Range2
 
     //reduce by segment
     oneapi::dpl::__par_backend_hetero::__parallel_for(
+        oneapi::dpl::__internal::__device_backend_tag{}, // TODO required to fix backend tag eval
         oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__reduce1_wrapper>(__exec),
         unseq_backend::__brick_reduce_idx<_BinaryOperator, decltype(__n)>(__binary_op, __n), __intermediate_result_end,
         oneapi::dpl::__ranges::take_view_simple(experimental::ranges::views::all_read(__idx),
@@ -794,6 +796,7 @@ __pattern_reduce_by_segment(_ExecutionPolicy&& __exec, _Range1&& __keys, _Range2
 
     //reduce by segment
     oneapi::dpl::__par_backend_hetero::__parallel_for(
+        oneapi::dpl::__internal::__device_backend_tag{}, // TODO required to fix backend tag eval
         oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__reduce2_wrapper>(
             ::std::forward<_ExecutionPolicy>(__exec)),
         unseq_backend::__brick_reduce_idx<_BinaryOperator, decltype(__intermediate_result_end)>(

From 6fc106d0dc8491cb2ace052d43efe59a7a18c24e Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 16 Feb 2024 17:24:56 +0100
Subject: [PATCH 320/566] include/oneapi/dpl/pstl/algorithm_impl.h - fix
 compile error: no matching function for call to '__remove_elements'

---
 include/oneapi/dpl/pstl/algorithm_impl.h | 6 ++----
 1 file changed, 2 insertions(+), 4 deletions(-)

diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index acef0bd9cbc..3f6ae184678 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -1558,8 +1558,7 @@ __pattern_unique(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _Ra
             __internal::__brick_walk3(
                 __b, __e, __it - 1, __it,
                 [&__pred](bool& __x, _ReferenceType __y, _ReferenceType __z) { __x = !__pred(__y, __z); }, _IsVector{});
-        },
-        _IsVector{});
+        });
 }
 
 //------------------------------------------------------------------------
@@ -2991,8 +2990,7 @@ __pattern_remove_if(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec,
         [&__pred](bool* __b, bool* __e, _RandomAccessIterator __it) {
             __internal::__brick_walk2(
                 __b, __e, __it, [&__pred](bool& __x, _ReferenceType __y) { __x = !__pred(__y); }, _IsVector{});
-        },
-        _IsVector{});
+        });
 }
 
 //------------------------------------------------------------------------

From 2e94dfa3f7d4ef28ce3b64d52db21730a844c30a Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Mon, 19 Feb 2024 09:24:40 +0100
Subject: [PATCH 321/566] include/oneapi/dpl/pstl/algorithm_impl.h - fix
 compile error: no matching function for call to '__parallel_transform_reduce'

---
 include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h | 3 +++
 include/oneapi/dpl/pstl/hetero/numeric_ranges_impl_hetero.h   | 1 +
 2 files changed, 4 insertions(+)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
index e10f8c0a8f5..9c3882063f1 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
@@ -328,6 +328,7 @@ __pattern_count(_ExecutionPolicy&& __exec, _Range&& __rng, _Predicate __predicat
 
     return oneapi::dpl::__par_backend_hetero::__parallel_transform_reduce<_ReduceValueType,
                                                                           ::std::true_type /*is_commutative*/>(
+               oneapi::dpl::__internal::__device_backend_tag{}, // TODO required to fix backend tag eval
                ::std::forward<_ExecutionPolicy>(__exec), __reduce_fn, __transform_fn,
                unseq_backend::__no_init_value{}, // no initial value
                ::std::forward<_Range>(__rng))
@@ -571,6 +572,7 @@ __pattern_min_element(_ExecutionPolicy&& __exec, _Range&& __rng, _Compare __comp
     auto __ret_idx =
         oneapi::dpl::__par_backend_hetero::__parallel_transform_reduce<_ReduceValueType,
                                                                        ::std::false_type /*is_commutative*/>(
+            oneapi::dpl::__internal::__device_backend_tag{}, // TODO required to fix backend tag eval
             ::std::forward<_ExecutionPolicy>(__exec), __reduce_fn, __transform_fn,
             unseq_backend::__no_init_value{}, // no initial value
             ::std::forward<_Range>(__rng))
@@ -627,6 +629,7 @@ __pattern_minmax_element(_ExecutionPolicy&& __exec, _Range&& __rng, _Compare __c
     _ReduceValueType __ret =
         oneapi::dpl::__par_backend_hetero::__parallel_transform_reduce<_ReduceValueType,
                                                                        ::std::false_type /*is_commutative*/>(
+            oneapi::dpl::__internal::__device_backend_tag{}, // TODO required to fix backend tag eval
             ::std::forward<_ExecutionPolicy>(__exec), __reduce_fn, __transform_fn,
             unseq_backend::__no_init_value{}, // no initial value
             ::std::forward<_Range>(__rng))
diff --git a/include/oneapi/dpl/pstl/hetero/numeric_ranges_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/numeric_ranges_impl_hetero.h
index 3565ac22473..40f1f617eca 100644
--- a/include/oneapi/dpl/pstl/hetero/numeric_ranges_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/numeric_ranges_impl_hetero.h
@@ -74,6 +74,7 @@ __pattern_transform_reduce(_ExecutionPolicy&& __exec, _Range&& __rng, _Tp __init
 
     return oneapi::dpl::__par_backend_hetero::__parallel_transform_reduce<_RepackedTp,
                                                                           ::std::true_type /*is_commutative*/>(
+               oneapi::dpl::__internal::__device_backend_tag{}, // TODO required to fix backend tag eval
                ::std::forward<_ExecutionPolicy>(__exec), __binary_op, _Functor{__unary_op},
                unseq_backend::__init_value<_RepackedTp>{__init}, // initial value
                ::std::forward<_Range>(__rng))

From 214ba0f9ae19ae9f25048e4d84319d9dcdef75e2 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Mon, 19 Feb 2024 09:49:27 +0100
Subject: [PATCH 322/566] struct
 __parallel_transform_reduce_small_submitter::operator() + tag impls

---
 .../dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h     | 5 ++---
 1 file changed, 2 insertions(+), 3 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h
index 28b48333d52..0dd1a79a186 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h
@@ -106,11 +106,10 @@ struct __parallel_transform_reduce_small_submitter<_Tp, __work_group_size, __ite
                                                    __internal::__optional_kernel_name<_Name...>>
 {
     template <typename _ExecutionPolicy, typename _Size, typename _ReduceOp, typename _TransformOp, typename _InitType,
-              oneapi::dpl::__internal::__enable_if_device_execution_policy<_ExecutionPolicy, int> = 0,
               typename... _Ranges>
     auto
-    operator()(_ExecutionPolicy&& __exec, const _Size __n, _ReduceOp __reduce_op, _TransformOp __transform_op,
-               _InitType __init, _Ranges&&... __rngs) const
+    operator()(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&& __exec, const _Size __n,
+               _ReduceOp __reduce_op, _TransformOp __transform_op, _InitType __init, _Ranges&&... __rngs) const
     {
         auto __transform_pattern =
             unseq_backend::transform_reduce<_ExecutionPolicy, __iters_per_work_item, _ReduceOp, _TransformOp,

From a67e6ec5dbf1b7d24180634b171539a79f6c3c10 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Mon, 19 Feb 2024 09:50:06 +0100
Subject: [PATCH 323/566] struct
 __parallel_transform_reduce_small_submitter::operator() + tag calls

---
 .../dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h    | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h
index 0dd1a79a186..d1761d266d9 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h
@@ -148,9 +148,9 @@ __parallel_transform_reduce_small_impl(oneapi::dpl::__internal::__device_backend
         __reduce_small_kernel<::std::integral_constant<::std::uint8_t, __iters_per_work_item>, _CustomName>>;
 
     return __parallel_transform_reduce_small_submitter<_Tp, __work_group_size, __iters_per_work_item, _Commutative,
-                                                       _ReduceKernel>()(::std::forward<_ExecutionPolicy>(__exec), __n,
-                                                                        __reduce_op, __transform_op, __init,
-                                                                        ::std::forward<_Ranges>(__rngs)...);
+                                                       _ReduceKernel>()(
+        oneapi::dpl::__internal::__device_backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __n, __reduce_op,
+        __transform_op, __init, ::std::forward<_Ranges>(__rngs)...);
 }
 
 // Submits the first kernel of the parallel_transform_reduce for mid-sized arrays.

From 262f54ffd31dfdb98079f4c2f1536e7165bc508a Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Mon, 19 Feb 2024 09:50:30 +0100
Subject: [PATCH 324/566] struct
 __parallel_transform_reduce_device_kernel_submitter::operator() + tag impls

---
 .../dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h    | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h
index d1761d266d9..b30cb6ee6f3 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h
@@ -166,11 +166,11 @@ struct __parallel_transform_reduce_device_kernel_submitter<_Tp, __work_group_siz
                                                            __internal::__optional_kernel_name<_KernelName...>>
 {
     template <typename _ExecutionPolicy, typename _Size, typename _ReduceOp, typename _TransformOp, typename _InitType,
-              oneapi::dpl::__internal::__enable_if_device_execution_policy<_ExecutionPolicy, int> = 0,
               typename... _Ranges>
     auto
-    operator()(_ExecutionPolicy&& __exec, _Size __n, _ReduceOp __reduce_op, _TransformOp __transform_op,
-               _InitType __init, sycl::buffer<_Tp>& __temp, _Ranges&&... __rngs) const
+    operator()(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&& __exec, _Size __n,
+               _ReduceOp __reduce_op, _TransformOp __transform_op, _InitType __init, sycl::buffer<_Tp>& __temp,
+               _Ranges&&... __rngs) const
     {
         auto __transform_pattern =
             unseq_backend::transform_reduce<_ExecutionPolicy, __iters_per_work_item, _ReduceOp, _TransformOp,

From b5c682d39117668f27e3f66780e72fd56d9d0fd4 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Mon, 19 Feb 2024 09:53:12 +0100
Subject: [PATCH 325/566] struct
 __parallel_transform_reduce_device_kernel_submitter::operator() + tag calls

---
 .../dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h       | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h
index b30cb6ee6f3..c0785583975 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h
@@ -279,7 +279,8 @@ __parallel_transform_reduce_mid_impl(oneapi::dpl::__internal::__device_backend_t
     sycl::event __reduce_event =
         __parallel_transform_reduce_device_kernel_submitter<_Tp, __work_group_size, __iters_per_work_item_device_kernel,
                                                             _Commutative, _ReduceDeviceKernel>()(
-            __exec, __n, __reduce_op, __transform_op, __init, __temp, ::std::forward<_Ranges>(__rngs)...);
+            oneapi::dpl::__internal::__device_backend_tag{}, __exec, __n, __reduce_op, __transform_op, __init, __temp,
+            ::std::forward<_Ranges>(__rngs)...);
 
     __n = __n_groups; // Number of preliminary results from the device kernel.
     return __parallel_transform_reduce_work_group_kernel_submitter<

From fcf5105b192d0983f174b8c7fe910aae53e71be0 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Mon, 19 Feb 2024 09:53:37 +0100
Subject: [PATCH 326/566] struct
 __parallel_transform_reduce_work_group_kernel_submitter::operator() + tag
 impls

---
 .../dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h  | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h
index c0785583975..eac42441558 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h
@@ -207,11 +207,11 @@ template <typename _Tp, ::std::uint16_t __work_group_size, ::std::uint8_t __iter
 struct __parallel_transform_reduce_work_group_kernel_submitter<
     _Tp, __work_group_size, __iters_per_work_item, _Commutative, __internal::__optional_kernel_name<_KernelName...>>
 {
-    template <typename _ExecutionPolicy, typename _Size, typename _ReduceOp, typename _TransformOp, typename _InitType,
-              oneapi::dpl::__internal::__enable_if_device_execution_policy<_ExecutionPolicy, int> = 0>
+    template <typename _ExecutionPolicy, typename _Size, typename _ReduceOp, typename _TransformOp, typename _InitType>
     auto
-    operator()(_ExecutionPolicy&& __exec, sycl::event& __reduce_event, _Size __n, _ReduceOp __reduce_op,
-               _TransformOp __transform_op, _InitType __init, sycl::buffer<_Tp>& __temp) const
+    operator()(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&& __exec, sycl::event& __reduce_event,
+               _Size __n, _ReduceOp __reduce_op, _TransformOp __transform_op, _InitType __init,
+               sycl::buffer<_Tp>& __temp) const
     {
         using _NoOpFunctor = unseq_backend::walk_n<_ExecutionPolicy, oneapi::dpl::__internal::__no_op>;
         auto __transform_pattern =

From 0fce2a16216a3084fcd969c9b839f60be9fa98e2 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Mon, 19 Feb 2024 09:54:06 +0100
Subject: [PATCH 327/566] struct__parallel_transform_reduce_impl::submit + tag
 impls

---
 .../dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h | 9 +++++----
 1 file changed, 5 insertions(+), 4 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h
index eac42441558..04c201494f1 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h
@@ -285,7 +285,8 @@ __parallel_transform_reduce_mid_impl(oneapi::dpl::__internal::__device_backend_t
     __n = __n_groups; // Number of preliminary results from the device kernel.
     return __parallel_transform_reduce_work_group_kernel_submitter<
         _Tp, __work_group_size, __iters_per_work_item_work_group_kernel, _Commutative, _ReduceWorkGroupKernel>()(
-        ::std::forward<_ExecutionPolicy>(__exec), __reduce_event, __n, __reduce_op, __transform_op, __init, __temp);
+        oneapi::dpl::__internal::__device_backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __reduce_event, __n,
+        __reduce_op, __transform_op, __init, __temp);
 }
 
 // General implementation using a tree reduction
@@ -293,11 +294,11 @@ template <typename _Tp, ::std::uint8_t __iters_per_work_item, typename _Commutat
 struct __parallel_transform_reduce_impl
 {
     template <typename _ExecutionPolicy, typename _Size, typename _ReduceOp, typename _TransformOp, typename _InitType,
-              oneapi::dpl::__internal::__enable_if_device_execution_policy<_ExecutionPolicy, int> = 0,
               typename... _Ranges>
     static auto
-    submit(_ExecutionPolicy&& __exec, _Size __n, ::std::uint16_t __work_group_size, _ReduceOp __reduce_op,
-           _TransformOp __transform_op, _InitType __init, _Ranges&&... __rngs)
+    submit(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&& __exec, _Size __n,
+           ::std::uint16_t __work_group_size, _ReduceOp __reduce_op, _TransformOp __transform_op, _InitType __init,
+           _Ranges&&... __rngs)
     {
         using _CustomName = oneapi::dpl::__internal::__policy_kernel_name<_ExecutionPolicy>;
         using _NoOpFunctor = unseq_backend::walk_n<_ExecutionPolicy, oneapi::dpl::__internal::__no_op>;

From 1e2ed7d87e4a06fb124270b6d9aff2b677f8c2dd Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Mon, 19 Feb 2024 09:55:02 +0100
Subject: [PATCH 328/566] struct__parallel_transform_reduce_impl::submit + tag
 calls

---
 .../oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h  | 1 +
 1 file changed, 1 insertion(+)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h
index 04c201494f1..c6d164f9415 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h
@@ -520,6 +520,7 @@ __parallel_transform_reduce(oneapi::dpl::__internal::__device_backend_tag, _Exec
     }
     // Otherwise use a recursive tree reduction.
     return __parallel_transform_reduce_impl<_Tp, 32, _Commutative>::submit(
+        oneapi::dpl::__internal::__device_backend_tag{},
         ::std::forward<_ExecutionPolicy>(__exec), __n, __work_group_size, __reduce_op, __transform_op, __init,
         ::std::forward<_Ranges>(__rngs)...);
 }

From 88a3e14636aa87561fac77731735ae641172070e Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Mon, 19 Feb 2024 10:03:57 +0100
Subject: [PATCH 329/566] include/oneapi/dpl/pstl/algorithm_impl.h - fix
 compile error: no matching function for call to '__parallel_merge'

---
 .../oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h  | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
index 9c3882063f1..bf4087fcea3 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
@@ -517,9 +517,10 @@ __pattern_merge(_ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&& __rng2, _
     }
     else
     {
-        __par_backend_hetero::__parallel_merge(::std::forward<_ExecutionPolicy>(__exec),
-                                               ::std::forward<_Range1>(__rng1), ::std::forward<_Range2>(__rng2),
-                                               ::std::forward<_Range3>(__rng3), __comp)
+        __par_backend_hetero::__parallel_merge(
+            oneapi::dpl::__internal::__device_backend_tag{}, // TODO required to fix backend tag eval
+            ::std::forward<_ExecutionPolicy>(__exec), ::std::forward<_Range1>(__rng1), ::std::forward<_Range2>(__rng2),
+            ::std::forward<_Range3>(__rng3), __comp)
             .wait();
     }
 

From 666acb213c9a0f99aa60550403591890348d5332 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Mon, 19 Feb 2024 10:14:29 +0100
Subject: [PATCH 330/566] include/oneapi/dpl/pstl/algorithm_impl.h - fix
 compile error: no matching function for call to '__parallel_stable_sort'

---
 .../oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h    | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
index bf4087fcea3..24a5d428a77 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
@@ -536,8 +536,9 @@ oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy>
 __pattern_sort(_ExecutionPolicy&& __exec, _Range&& __rng, _Compare __comp, _Proj __proj)
 {
     if (__rng.size() >= 2)
-        __par_backend_hetero::__parallel_stable_sort(::std::forward<_ExecutionPolicy>(__exec),
-                                                     ::std::forward<_Range>(__rng), __comp, __proj)
+        __par_backend_hetero::__parallel_stable_sort(
+            oneapi::dpl::__internal::__device_backend_tag{}, // TODO required to fix backend tag eval
+            ::std::forward<_ExecutionPolicy>(__exec), ::std::forward<_Range>(__rng), __comp, __proj)
             .wait();
 }
 

From a8c0a10e2895698761227329ccc3005434c9c87f Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Mon, 19 Feb 2024 10:22:02 +0100
Subject: [PATCH 331/566] 
 include/oneapi/dpl/pstl/hetero/numeric_ranges_impl_hetero.h - fix compile
 error: no matching function for call to '__parallel_transform_reduce'

---
 include/oneapi/dpl/pstl/hetero/numeric_ranges_impl_hetero.h | 1 +
 1 file changed, 1 insertion(+)

diff --git a/include/oneapi/dpl/pstl/hetero/numeric_ranges_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/numeric_ranges_impl_hetero.h
index 40f1f617eca..42502a4b7b5 100644
--- a/include/oneapi/dpl/pstl/hetero/numeric_ranges_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/numeric_ranges_impl_hetero.h
@@ -51,6 +51,7 @@ __pattern_transform_reduce(_ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&
 
     return oneapi::dpl::__par_backend_hetero::__parallel_transform_reduce<_RepackedTp,
                                                                           ::std::true_type /*is_commutative*/>(
+               oneapi::dpl::__internal::__device_backend_tag{}, // TODO required to fix backend tag eval
                ::std::forward<_ExecutionPolicy>(__exec), __binary_op1, _Functor{__binary_op2},
                unseq_backend::__init_value<_RepackedTp>{__init}, // initial value
                ::std::forward<_Range1>(__rng1), ::std::forward<_Range2>(__rng2))

From d1b498de48a781e068cf149269d013cd91aa99ce Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Mon, 19 Feb 2024 10:35:05 +0100
Subject: [PATCH 332/566] include/oneapi/dpl/pstl/algorithm_fwd.h - fix compile
 error: no member named '__pattern_shift_right' in namespace
 'oneapi::dpl::__internal'

---
 include/oneapi/dpl/pstl/algorithm_fwd.h | 9 +++++++++
 1 file changed, 9 insertions(+)

diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index fd89678dd3b..7eba977aba1 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -1265,6 +1265,15 @@ _ForwardIterator
 __pattern_shift_left(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _ForwardIterator, _ForwardIterator,
                      typename ::std::iterator_traits<_ForwardIterator>::difference_type);
 
+//------------------------------------------------------------------------
+// shift_right
+//------------------------------------------------------------------------
+
+template <class _Tag, class _ExecutionPolicy, class _BidirectionalIterator>
+_BidirectionalIterator
+__pattern_shift_right(_Tag, _ExecutionPolicy&&, _BidirectionalIterator, _BidirectionalIterator,
+                      typename ::std::iterator_traits<_BidirectionalIterator>::difference_type);
+
 } // namespace __internal
 } // namespace dpl
 } // namespace oneapi

From 5f34eda9ec56354effcd6906fd0a02f50f3d783f Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Mon, 19 Feb 2024 10:49:46 +0100
Subject: [PATCH 333/566] async patterns tag impls and calls

---
 .../internal/async_impl/async_impl_hetero.h   | 84 ++++++++-----------
 .../dpl/internal/async_impl/glue_async_impl.h | 47 ++++++++---
 2 files changed, 67 insertions(+), 64 deletions(-)

diff --git a/include/oneapi/dpl/internal/async_impl/async_impl_hetero.h b/include/oneapi/dpl/internal/async_impl/async_impl_hetero.h
index dc08e7ba2d1..eaa92958c27 100644
--- a/include/oneapi/dpl/internal/async_impl/async_impl_hetero.h
+++ b/include/oneapi/dpl/internal/async_impl/async_impl_hetero.h
@@ -30,10 +30,10 @@ namespace dpl
 namespace __internal
 {
 
-template <typename _ExecutionPolicy, typename _ForwardIterator, typename _Function,
-          oneapi::dpl::__internal::__enable_if_device_execution_policy<_ExecutionPolicy, int> = 0>
+template <typename _BackendTag, typename _ExecutionPolicy, typename _ForwardIterator, typename _Function>
 auto
-__pattern_walk1_async(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last, _Function __f)
+__pattern_walk1_async(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _ForwardIterator __first,
+                      _ForwardIterator __last, _Function __f)
 {
     auto __n = __last - __first;
     assert(__n > 0);
@@ -42,23 +42,19 @@ __pattern_walk1_async(_ExecutionPolicy&& __exec, _ForwardIterator __first, _Forw
         oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read_write, _ForwardIterator>();
     auto __buf = __keep(__first, __last);
 
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
     auto __future_obj = oneapi::dpl::__par_backend_hetero::__parallel_for(
-        __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
+        _BackendTag{}, ::std::forward<_ExecutionPolicy>(__exec),
         unseq_backend::walk_n<_ExecutionPolicy, _Function>{__f}, __n, __buf.all_view());
     return __future_obj;
 }
 
-template <typename _IsSync = ::std::false_type,
+template <typename _BackendTag, typename _IsSync = ::std::false_type,
           __par_backend_hetero::access_mode __acc_mode1 = __par_backend_hetero::access_mode::read,
           __par_backend_hetero::access_mode __acc_mode2 = __par_backend_hetero::access_mode::write,
-          typename _ExecutionPolicy, typename _ForwardIterator1, typename _ForwardIterator2, typename _Function,
-          oneapi::dpl::__internal::__enable_if_device_execution_policy<_ExecutionPolicy, int> = 0>
+          typename _ExecutionPolicy, typename _ForwardIterator1, typename _ForwardIterator2, typename _Function>
 auto
-__pattern_walk2_async(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
-                      _ForwardIterator2 __first2, _Function __f)
+__pattern_walk2_async(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _ForwardIterator1 __first1,
+                      _ForwardIterator1 __last1, _ForwardIterator2 __first2, _Function __f)
 {
     auto __n = __last1 - __first1;
     assert(__n > 0);
@@ -69,12 +65,8 @@ __pattern_walk2_async(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _Fo
     auto __keep2 = oneapi::dpl::__ranges::__get_sycl_range<__acc_mode2, _ForwardIterator2>();
     auto __buf2 = __keep2(__first2, __first2 + __n);
 
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
     auto __future = oneapi::dpl::__par_backend_hetero::__parallel_for(
-        __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
+        _BackendTag{}, ::std::forward<_ExecutionPolicy>(__exec),
         unseq_backend::walk_n<_ExecutionPolicy, _Function>{__f}, __n, __buf1.all_view(), __buf2.all_view());
 
     if constexpr (_IsSync::value)
@@ -83,11 +75,11 @@ __pattern_walk2_async(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _Fo
     return __future.__make_future(__first2 + __n);
 }
 
-template <typename _ExecutionPolicy, typename _ForwardIterator1, typename _ForwardIterator2, typename _ForwardIterator3,
-          typename _Function, oneapi::dpl::__internal::__enable_if_device_execution_policy<_ExecutionPolicy, int> = 0>
+template <typename _BackendTag, typename _ExecutionPolicy, typename _ForwardIterator1, typename _ForwardIterator2,
+          typename _ForwardIterator3, typename _Function>
 auto
-__pattern_walk3_async(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
-                      _ForwardIterator2 __first2, _ForwardIterator3 __first3, _Function __f)
+__pattern_walk3_async(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _ForwardIterator1 __first1,
+                      _ForwardIterator1 __last1, _ForwardIterator2 __first2, _ForwardIterator3 __first3, _Function __f)
 {
     auto __n = __last1 - __first1;
     assert(__n > 0);
@@ -102,25 +94,22 @@ __pattern_walk3_async(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _Fo
         oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::write, _ForwardIterator3>();
     auto __buf3 = __keep3(__first3, __first3 + __n);
 
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1,
-                                                                              _ForwardIterator2, _ForwardIterator3>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
     auto __future =
-        oneapi::dpl::__par_backend_hetero::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
+        oneapi::dpl::__par_backend_hetero::__parallel_for(_BackendTag{}, ::std::forward<_ExecutionPolicy>(__exec),
                                                           unseq_backend::walk_n<_ExecutionPolicy, _Function>{__f}, __n,
                                                           __buf1.all_view(), __buf2.all_view(), __buf3.all_view());
 
     return __future.__make_future(__first3 + __n);
 }
 
-template <typename _ExecutionPolicy, typename _ForwardIterator1, typename _ForwardIterator2, typename _Brick,
-          oneapi::dpl::__internal::__enable_if_device_execution_policy<_ExecutionPolicy, int> = 0>
+template <typename _BackendTag, typename _ExecutionPolicy, typename _ForwardIterator1, typename _ForwardIterator2,
+          typename _Brick>
 auto
-__pattern_walk2_brick_async(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
-                            _ForwardIterator2 __first2, _Brick __brick)
+__pattern_walk2_brick_async(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _ForwardIterator1 __first1,
+                            _ForwardIterator1 __last1, _ForwardIterator2 __first2, _Brick __brick)
 {
     return __pattern_walk2_async(
+        __tag,
         __par_backend_hetero::make_wrapped_policy<__walk2_brick_wrapper>(::std::forward<_ExecutionPolicy>(__exec)),
         __first1, __last1, __first2, __brick);
 }
@@ -129,11 +118,10 @@ __pattern_walk2_brick_async(_ExecutionPolicy&& __exec, _ForwardIterator1 __first
 // transform_reduce (version with two binary functions)
 //------------------------------------------------------------------------
 
-template <typename _ExecutionPolicy, typename _RandomAccessIterator1, typename _RandomAccessIterator2, typename _Tp,
-          typename _BinaryOperation1, typename _BinaryOperation2,
-          oneapi::dpl::__internal::__enable_if_device_execution_policy<_ExecutionPolicy, int> = 0>
+template <typename _BackendTag, typename _ExecutionPolicy, typename _RandomAccessIterator1,
+          typename _RandomAccessIterator2, typename _Tp, typename _BinaryOperation1, typename _BinaryOperation2>
 auto
-__pattern_transform_reduce_async(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1,
+__pattern_transform_reduce_async(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1,
                                  _RandomAccessIterator1 __last1, _RandomAccessIterator2 __first2, _Tp __init,
                                  _BinaryOperation1 __binary_op1, _BinaryOperation2 __binary_op2)
 {
@@ -151,13 +139,9 @@ __pattern_transform_reduce_async(_ExecutionPolicy&& __exec, _RandomAccessIterato
         oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _RandomAccessIterator2>();
     auto __buf2 = __keep2(__first2, __first2 + __n);
 
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
     return oneapi::dpl::__par_backend_hetero::__parallel_transform_reduce<_RepackedTp,
                                                                           ::std::true_type /*is_commutative*/>(
-        __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __binary_op1, _Functor{__binary_op2},
+        _BackendTag{}, ::std::forward<_ExecutionPolicy>(__exec), __binary_op1, _Functor{__binary_op2},
         unseq_backend::__init_value<_RepackedTp>{__init}, // initial value
         __buf1.all_view(), __buf2.all_view());
 }
@@ -166,12 +150,12 @@ __pattern_transform_reduce_async(_ExecutionPolicy&& __exec, _RandomAccessIterato
 // transform_reduce (with unary and binary functions)
 //------------------------------------------------------------------------
 
-template <typename _ExecutionPolicy, typename _ForwardIterator, typename _Tp, typename _BinaryOperation,
-          typename _UnaryOperation,
-          oneapi::dpl::__internal::__enable_if_device_execution_policy<_ExecutionPolicy, int> = 0>
+template <typename _BackendTag, typename _ExecutionPolicy, typename _ForwardIterator, typename _Tp,
+          typename _BinaryOperation, typename _UnaryOperation>
 auto
-__pattern_transform_reduce_async(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last,
-                                 _Tp __init, _BinaryOperation __binary_op, _UnaryOperation __unary_op)
+__pattern_transform_reduce_async(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _ForwardIterator __first,
+                                 _ForwardIterator __last, _Tp __init, _BinaryOperation __binary_op,
+                                 _UnaryOperation __unary_op)
 {
     assert(__first < __last);
 
@@ -182,22 +166,20 @@ __pattern_transform_reduce_async(_ExecutionPolicy&& __exec, _ForwardIterator __f
     auto __keep = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _ForwardIterator>();
     auto __buf = __keep(__first, __last);
 
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
-    using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
     return oneapi::dpl::__par_backend_hetero::__parallel_transform_reduce<_RepackedTp,
                                                                           ::std::true_type /*is_commutative*/>(
-        __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __binary_op, _Functor{__unary_op},
+        _BackendTag{}, ::std::forward<_ExecutionPolicy>(__exec), __binary_op, _Functor{__unary_op},
         unseq_backend::__init_value<_RepackedTp>{__init}, // initial value
         __buf.all_view());
 }
 
-template <typename _ExecutionPolicy, typename _ForwardIterator, typename _T,
-          oneapi::dpl::__internal::__enable_if_device_execution_policy<_ExecutionPolicy, int> = 0>
+template <typename _BackendTag, typename _ExecutionPolicy, typename _ForwardIterator, typename _T>
 auto
-__pattern_fill_async(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last, const _T& __value)
+__pattern_fill_async(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _ForwardIterator __first,
+                     _ForwardIterator __last, const _T& __value)
 {
     return __pattern_walk1_async(
+        __tag,
         ::std::forward<_ExecutionPolicy>(__exec),
         __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::write>(__first),
         __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::write>(__last),
diff --git a/include/oneapi/dpl/internal/async_impl/glue_async_impl.h b/include/oneapi/dpl/internal/async_impl/glue_async_impl.h
index d9f06601b45..30a37fe229e 100644
--- a/include/oneapi/dpl/internal/async_impl/glue_async_impl.h
+++ b/include/oneapi/dpl/internal/async_impl/glue_async_impl.h
@@ -43,9 +43,12 @@ auto
 transform_async(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _ForwardIterator1 __last,
                 _ForwardIterator2 __result, _UnaryOperation __op, _Events&&... __dependencies)
 {
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
+
     wait_for_all(::std::forward<_Events>(__dependencies)...);
     auto ret_val = oneapi::dpl::__internal::__pattern_walk2_async(
-        ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result,
+        __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result,
         oneapi::dpl::__internal::__transform_functor<_UnaryOperation>{::std::move(__op)});
     return ret_val;
 }
@@ -59,9 +62,12 @@ transform_async(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardI
                 _ForwardIterator2 __first2, _ForwardIterator __result, _BinaryOperation __op,
                 _Events&&... __dependencies)
 {
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1,
+                                                                              _ForwardIterator2, _ForwardIterator>();
+
     wait_for_all(::std::forward<_Events>(__dependencies)...);
     auto ret_val = oneapi::dpl::__internal::__pattern_walk3_async(
-        ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __result,
+        __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __result,
         oneapi::dpl::__internal::__transform_functor<_BinaryOperation>(::std::move(__op)));
     return ret_val;
 }
@@ -73,9 +79,12 @@ auto
 copy_async(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _ForwardIterator1 __last, _ForwardIterator2 __result,
            _Events&&... __dependencies)
 {
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
+
     wait_for_all(::std::forward<_Events>(__dependencies)...);
     auto ret_val = oneapi::dpl::__internal::__pattern_walk2_brick_async(
-        ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result,
+        __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result,
         oneapi::dpl::__internal::__brick_copy<_ExecutionPolicy>{});
     return ret_val;
 }
@@ -118,9 +127,11 @@ auto
 for_each_async(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last, _Function __f,
                _Events&&... __dependencies)
 {
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+
     wait_for_all(::std::forward<_Events>(__dependencies)...);
-    auto ret_val =
-        oneapi::dpl::__internal::__pattern_walk1_async(::std::forward<_ExecutionPolicy>(__exec), __first, __last, __f);
+    auto ret_val = oneapi::dpl::__internal::__pattern_walk1_async(
+        __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __f);
     return ret_val;
 }
 
@@ -133,10 +144,12 @@ auto
 reduce_async(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last, _Tp __init,
              _BinaryOperation __binary_op, _Events&&... __dependencies)
 {
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+
     wait_for_all(::std::forward<_Events>(__dependencies)...);
-    auto ret_val = oneapi::dpl::__internal::__pattern_transform_reduce_async(::std::forward<_ExecutionPolicy>(__exec),
-                                                                             __first, __last, __init, __binary_op,
-                                                                             oneapi::dpl::__internal::__no_op());
+    auto ret_val = oneapi::dpl::__internal::__pattern_transform_reduce_async(
+        __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __init, __binary_op,
+        oneapi::dpl::__internal::__no_op());
     return ret_val;
 }
 
@@ -168,9 +181,11 @@ auto
 fill_async(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last, const _Tp& __value,
            _Events&&... __dependencies)
 {
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+
     wait_for_all(::std::forward<_Events>(__dependencies)...);
-    return oneapi::dpl::__internal::__pattern_fill_async(::std::forward<_ExecutionPolicy>(__exec), __first, __last,
-                                                         __value);
+    return oneapi::dpl::__internal::__pattern_fill_async(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
+                                                         __first, __last, __value);
 }
 
 // [async.transform_reduce]
@@ -183,9 +198,13 @@ auto
 transform_reduce_async(_ExecutionPolicy&& __exec, _ForwardIt1 __first1, _ForwardIt1 __last1, _ForwardIt2 __first2,
                        _T __init, _BinaryOp1 __binary_op1, _BinaryOp2 __binary_op2, _Events&&... __dependencies)
 {
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIt1, _ForwardIt2>();
+
     wait_for_all(::std::forward<_Events>(__dependencies)...);
     return oneapi::dpl::__internal::__pattern_transform_reduce_async(
-        ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __init, __binary_op1, __binary_op2);
+        __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __init, __binary_op1,
+        __binary_op2);
 }
 
 template <class _ExecutionPolicy, class _ForwardIt, class _T, class _BinaryOp, class _UnaryOp, class... _Events,
@@ -195,9 +214,11 @@ auto
 transform_reduce_async(_ExecutionPolicy&& __exec, _ForwardIt __first, _ForwardIt __last, _T __init,
                        _BinaryOp __binary_op, _UnaryOp __unary_op, _Events&&... __dependencies)
 {
+    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIt>();
+
     wait_for_all(::std::forward<_Events>(__dependencies)...);
-    return oneapi::dpl::__internal::__pattern_transform_reduce_async(::std::forward<_ExecutionPolicy>(__exec), __first,
-                                                                     __last, __init, __binary_op, __unary_op);
+    return oneapi::dpl::__internal::__pattern_transform_reduce_async(
+        __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __init, __binary_op, __unary_op);
 }
 
 template <class _ExecutionPolicy, class _ForwardIt1, class _ForwardIt2, class _T, class... _Events,

From 6dcbdff51a174ee9b71879db1f71e8a11b5e6faa Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Mon, 19 Feb 2024 11:05:44 +0100
Subject: [PATCH 334/566] 
 include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h - remove
 fpga patterns which simple calls
 oneapi::dpl::__internal::__device_backend_tag impls

---
 .../hetero/dpcpp/parallel_backend_sycl_fpga.h | 182 ------------------
 1 file changed, 182 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
index 77fa3388f47..6c88c276a66 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
@@ -89,188 +89,6 @@ __parallel_for(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionPolicy&& _
                                                                 __count, std::forward<_Ranges>(__rngs)...);
 }
 
-//------------------------------------------------------------------------
-// parallel_transform_reduce
-//------------------------------------------------------------------------
-
-template <typename _Tp, typename _Commutative, typename _ExecutionPolicy, typename _ReduceOp, typename _TransformOp,
-          typename _InitType, typename... _Ranges>
-auto
-__parallel_transform_reduce(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionPolicy&& __exec,
-                            _ReduceOp __reduce_op, _TransformOp __transform_op, _InitType __init, _Ranges&&... __rngs)
-{
-    // workaround until we implement more performant version for patterns
-    return oneapi::dpl::__par_backend_hetero::__parallel_transform_reduce<_Tp, _Commutative>(
-        oneapi::dpl::__internal::__device_backend_tag{},
-        __exec.__device_policy(), __reduce_op, __transform_op, __init, ::std::forward<_Ranges>(__rngs)...);
-}
-
-//------------------------------------------------------------------------
-// parallel_transform_scan
-//------------------------------------------------------------------------
-
-template <typename _ExecutionPolicy, typename _Range1, typename _Range2, typename _UnaryOperation, typename _InitType,
-          typename _BinaryOperation, typename _Inclusive>
-auto
-__parallel_transform_scan(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionPolicy&& __exec, _Range1&& __in_rng,
-                          _Range2&& __out_rng, ::std::size_t __n, _UnaryOperation __unary_op, _InitType __init,
-                          _BinaryOperation __binary_op, _Inclusive)
-{
-    // workaround until we implement more performant version for patterns
-    return oneapi::dpl::__par_backend_hetero::__parallel_transform_scan(
-        oneapi::dpl::__internal::__device_backend_tag{}, __exec.__device_policy(), ::std::forward<_Range1>(__in_rng),
-        ::std::forward<_Range2>(__out_rng), __n, __unary_op, __init, __binary_op, _Inclusive{});
-}
-
-template <typename _ExecutionPolicy, typename _Range1, typename _Range2, typename _BinaryOperation, typename _InitType,
-          typename _LocalScan, typename _GroupScan, typename _GlobalScan>
-auto
-__parallel_transform_scan_base(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionPolicy&& __exec, _Range1&& __rng1,
-                               _Range2&& __rng2, _BinaryOperation __binary_op, _InitType __init,
-                               _LocalScan __local_scan, _GroupScan __group_scan, _GlobalScan __global_scan)
-{
-    // workaround until we implement more performant version for patterns
-    return oneapi::dpl::__par_backend_hetero::__parallel_transform_scan_base(
-        oneapi::dpl::__internal::__device_backend_tag{}, __exec.__device_policy(), ::std::forward<_Range1>(__rng1),
-        ::std::forward<_Range2>(__rng2), __binary_op, __init, __local_scan, __group_scan, __global_scan);
-}
-
-template <typename _ExecutionPolicy, typename _InRng, typename _OutRng, typename _Size, typename _Pred>
-auto
-__parallel_copy_if(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionPolicy&& __exec, _InRng&& __in_rng,
-                   _OutRng&& __out_rng, _Size __n, _Pred __pred)
-{
-    // workaround until we implement more performant version for patterns
-    return oneapi::dpl::__par_backend_hetero::__parallel_copy_if(
-        oneapi::dpl::__internal::__device_backend_tag{}, __exec.__device_policy(), ::std::forward<_InRng>(__in_rng),
-        ::std::forward<_OutRng>(__out_rng), __n, __pred);
-}
-
-template <typename _ExecutionPolicy, typename _InRng, typename _OutRng, typename _Size, typename _CreateMaskOp,
-          typename _CopyByMaskOp>
-auto
-__parallel_scan_copy(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionPolicy&& __exec, _InRng&& __in_rng,
-                     _OutRng&& __out_rng, _Size __n, _CreateMaskOp __create_mask_op, _CopyByMaskOp __copy_by_mask_op)
-{
-    // workaround until we implement more performant version for patterns
-    return oneapi::dpl::__par_backend_hetero::__parallel_scan_copy(
-        oneapi::dpl::__internal::__device_backend_tag{}, __exec.__device_policy(), ::std::forward<_InRng>(__in_rng),
-        ::std::forward<_OutRng>(__out_rng), __n, __create_mask_op, __copy_by_mask_op);
-}
-
-//------------------------------------------------------------------------
-// __parallel_find_or
-//-----------------------------------------------------------------------
-
-template <typename _ExecutionPolicy, typename _Brick, typename _BrickTag, typename... _Ranges>
-::std::conditional_t<
-    ::std::is_same_v<_BrickTag, __parallel_or_tag>, bool,
-    oneapi::dpl::__internal::__difference_t<typename oneapi::dpl::__ranges::__get_first_range_type<_Ranges...>::type>>
-__parallel_find_or(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionPolicy&& __exec, _Brick __f,
-                   _BrickTag __brick_tag, _Ranges&&... __rngs)
-{
-    return oneapi::dpl::__par_backend_hetero::__parallel_find_or(oneapi::dpl::__internal::__device_backend_tag{},
-                                                                 __exec.__device_policy(), __f, __brick_tag,
-                                                                 ::std::forward<_Ranges>(__rngs)...);
-}
-
-//------------------------------------------------------------------------
-// parallel_or
-//-----------------------------------------------------------------------
-template <typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2, typename _Brick>
-bool
-__parallel_or(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionPolicy&& __exec, _Iterator1 __first,
-              _Iterator1 __last, _Iterator2 __s_first, _Iterator2 __s_last, _Brick __f)
-{
-    // workaround until we implement more performant version for patterns
-    return oneapi::dpl::__par_backend_hetero::__parallel_or(oneapi::dpl::__internal::__device_backend_tag{},
-                                                            __exec.__device_policy(), __first, __last, __s_first,
-                                                            __s_last, __f);
-}
-
-template <typename _ExecutionPolicy, typename _Iterator, typename _Brick>
-bool
-__parallel_or(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionPolicy&& __exec, _Iterator __first,
-              _Iterator __last, _Brick __f)
-{
-    // workaround until we implement more performant version for patterns
-    return oneapi::dpl::__par_backend_hetero::__parallel_or(oneapi::dpl::__internal::__device_backend_tag{},
-                                                            __exec.__device_policy(), __first, __last, __f);
-}
-
-//------------------------------------------------------------------------
-// parallel_find
-//-----------------------------------------------------------------------
-
-template <typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2, typename _Brick, typename _IsFirst>
-_Iterator1
-__parallel_find(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionPolicy&& __exec, _Iterator1 __first,
-                _Iterator1 __last, _Iterator2 __s_first, _Iterator2 __s_last, _Brick __f, _IsFirst __is_first)
-{
-    // workaround until we implement more performant version for patterns
-    return oneapi::dpl::__par_backend_hetero::__parallel_find(oneapi::dpl::__internal::__device_backend_tag{},
-                                                              __exec.__device_policy(), __first, __last, __s_first,
-                                                              __s_last, __f, __is_first);
-}
-
-template <typename _ExecutionPolicy, typename _Iterator, typename _Brick, typename _IsFirst>
-_Iterator
-__parallel_find(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionPolicy&& __exec, _Iterator __first,
-                _Iterator __last, _Brick __f, _IsFirst __is_first)
-{
-    // workaround until we implement more performant version for patterns
-    return oneapi::dpl::__par_backend_hetero::__parallel_find(
-        oneapi::dpl::__internal::__device_backend_tag{}, __exec.__device_policy(), __first, __last, __f, __is_first);
-}
-
-//------------------------------------------------------------------------
-// parallel_merge
-//-----------------------------------------------------------------------
-
-template <typename _ExecutionPolicy, typename _Range1, typename _Range2, typename _Range3, typename _Compare>
-auto
-__parallel_merge(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionPolicy&& __exec, _Range1&& __rng1,
-                 _Range2&& __rng2, _Range3&& __rng3, _Compare __comp)
-    -> decltype(oneapi::dpl::__par_backend_hetero::__parallel_merge(
-        oneapi::dpl::__internal::__device_backend_tag{}, __exec.__device_policy(), ::std::forward<_Range1>(__rng1),
-        ::std::forward<_Range2>(__rng2), ::std::forward<_Range3>(__rng3), __comp))
-{
-    // workaround until we implement more performant version for patterns
-    return oneapi::dpl::__par_backend_hetero::__parallel_merge(
-        oneapi::dpl::__internal::__device_backend_tag{}, __exec.__device_policy(), ::std::forward<_Range1>(__rng1),
-        ::std::forward<_Range2>(__rng2), ::std::forward<_Range3>(__rng3), __comp);
-}
-
-//------------------------------------------------------------------------
-// parallel_stable_sort
-//-----------------------------------------------------------------------
-
-template <typename _ExecutionPolicy, typename _Range, typename _Compare, typename _Proj>
-auto
-__parallel_stable_sort(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionPolicy&& __exec, _Range&& __rng,
-                       _Compare __comp, _Proj __proj)
-{
-    // workaround until we implement more performant version for patterns
-    return oneapi::dpl::__par_backend_hetero::__parallel_stable_sort(oneapi::dpl::__internal::__device_backend_tag{},
-                                                                     __exec.__device_policy(),
-                                                                     ::std::forward<_Range>(__rng), __comp, __proj);
-}
-
-//------------------------------------------------------------------------
-// parallel_partial_sort
-//-----------------------------------------------------------------------
-
-// TODO: check if it makes sense to move these wrappers out of backend to a common place
-template <typename _ExecutionPolicy, typename _Iterator, typename _Compare>
-auto
-__parallel_partial_sort(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionPolicy&& __exec, _Iterator __first,
-                        _Iterator __mid, _Iterator __last, _Compare __comp)
-{
-    // workaround until we implement more performant version for patterns
-    return oneapi::dpl::__par_backend_hetero::__parallel_partial_sort(
-        oneapi::dpl::__internal::__device_backend_tag{}, __exec.__device_policy(), __first, __mid, __last, __comp);
-}
-
 //------------------------------------------------------------------------
 // parallel_histogram
 //-----------------------------------------------------------------------

From a006b97a4ef3862593731f6f2cd2adae2a3438e4 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Mon, 19 Feb 2024 11:23:55 +0100
Subject: [PATCH 335/566] Apply GitHUB clang format

---
 .../internal/async_impl/async_impl_hetero.h   |  8 +--
 .../dpl/internal/async_impl/glue_async_impl.h |  5 +-
 .../internal/exclusive_scan_by_segment_impl.h | 14 ++--
 .../internal/inclusive_scan_by_segment_impl.h | 14 ++--
 .../dpl/internal/reduce_by_segment_impl.h     | 20 +++---
 include/oneapi/dpl/pstl/algorithm_impl.h      | 11 ++--
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 25 +++----
 .../hetero/algorithm_ranges_impl_hetero.h     |  4 +-
 .../pstl/hetero/dpcpp/parallel_backend_sycl.h | 12 ++--
 .../hetero/dpcpp/parallel_backend_sycl_fpga.h |  8 +--
 .../dpcpp/parallel_backend_sycl_histogram.h   | 16 ++---
 .../dpcpp/parallel_backend_sycl_reduce.h      | 65 ++++++++-----------
 .../dpl/pstl/hetero/numeric_impl_hetero.h     |  6 +-
 include/oneapi/dpl/pstl/numeric_impl.h        |  8 +--
 14 files changed, 95 insertions(+), 121 deletions(-)

diff --git a/include/oneapi/dpl/internal/async_impl/async_impl_hetero.h b/include/oneapi/dpl/internal/async_impl/async_impl_hetero.h
index eaa92958c27..02ee5f852f1 100644
--- a/include/oneapi/dpl/internal/async_impl/async_impl_hetero.h
+++ b/include/oneapi/dpl/internal/async_impl/async_impl_hetero.h
@@ -179,8 +179,7 @@ __pattern_fill_async(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec,
                      _ForwardIterator __last, const _T& __value)
 {
     return __pattern_walk1_async(
-        __tag,
-        ::std::forward<_ExecutionPolicy>(__exec),
+        __tag, ::std::forward<_ExecutionPolicy>(__exec),
         __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::write>(__first),
         __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::write>(__last),
         fill_functor<_T>{__value});
@@ -206,9 +205,8 @@ __pattern_transform_scan_base_async(__hetero_tag<_BackendTag>, _ExecutionPolicy&
     auto __buf2 = __keep2(__result, __result + __n);
 
     auto __res = oneapi::dpl::__par_backend_hetero::__parallel_transform_scan(
-        _BackendTag{},
-        ::std::forward<_ExecutionPolicy>(__exec), __buf1.all_view(), __buf2.all_view(), __n, __unary_op, __init,
-        __binary_op, _Inclusive{});
+        _BackendTag{}, ::std::forward<_ExecutionPolicy>(__exec), __buf1.all_view(), __buf2.all_view(), __n, __unary_op,
+        __init, __binary_op, _Inclusive{});
     return __res.__make_future(__result + __n);
 }
 
diff --git a/include/oneapi/dpl/internal/async_impl/glue_async_impl.h b/include/oneapi/dpl/internal/async_impl/glue_async_impl.h
index 30a37fe229e..586360f83af 100644
--- a/include/oneapi/dpl/internal/async_impl/glue_async_impl.h
+++ b/include/oneapi/dpl/internal/async_impl/glue_async_impl.h
@@ -247,9 +247,8 @@ inclusive_scan_async(_ExecutionPolicy&& __exec, _ForwardIt1 __first1, _ForwardIt
     using _ValueType = typename ::std::iterator_traits<_ForwardIt1>::value_type;
     wait_for_all(::std::forward<_Events>(__dependencies)...);
     return oneapi::dpl::__internal::__pattern_transform_scan_async(
-        __dispatch_tag,
-        ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, oneapi::dpl::__internal::__no_op(),
-        ::std::plus<_ValueType>(), /*inclusive=*/::std::true_type());
+        __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2,
+        oneapi::dpl::__internal::__no_op(), ::std::plus<_ValueType>(), /*inclusive=*/::std::true_type());
 }
 
 template <class _ExecutionPolicy, class _ForwardIt1, class _ForwardIt2, class _BinaryOperation, class... _Events,
diff --git a/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h b/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h
index 24d5db6b630..c447c0c199b 100644
--- a/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h
@@ -97,9 +97,10 @@ pattern_exclusive_scan_by_segment(_Tag, Policy&& policy, InputIterator1 first1,
 template <typename _BackendTag, typename Policy, typename InputIterator1, typename InputIterator2,
           typename OutputIterator, typename T, typename BinaryPredicate, typename Operator>
 OutputIterator
-exclusive_scan_by_segment_impl(__internal::__hetero_tag<_BackendTag> __tag, Policy&& policy, InputIterator1 first1, InputIterator1 last1,
-                               InputIterator2 first2, OutputIterator result, T init, BinaryPredicate binary_pred,
-                               Operator binary_op, ::std::true_type /* has_known_identity*/)
+exclusive_scan_by_segment_impl(__internal::__hetero_tag<_BackendTag> __tag, Policy&& policy, InputIterator1 first1,
+                               InputIterator1 last1, InputIterator2 first2, OutputIterator result, T init,
+                               BinaryPredicate binary_pred, Operator binary_op,
+                               ::std::true_type /* has_known_identity*/)
 {
     return internal::__scan_by_segment_impl_common(__tag, ::std::forward<Policy>(policy), first1, last1, first2, result,
                                                    init, binary_pred, binary_op, ::std::false_type{});
@@ -108,9 +109,10 @@ exclusive_scan_by_segment_impl(__internal::__hetero_tag<_BackendTag> __tag, Poli
 template <typename _BackendTag, typename Policy, typename InputIterator1, typename InputIterator2,
           typename OutputIterator, typename T, typename BinaryPredicate, typename Operator>
 OutputIterator
-exclusive_scan_by_segment_impl(__internal::__hetero_tag<_BackendTag>, Policy&& policy, InputIterator1 first1, InputIterator1 last1,
-                               InputIterator2 first2, OutputIterator result, T init, BinaryPredicate binary_pred,
-                               Operator binary_op, ::std::false_type /* has_known_identity*/)
+exclusive_scan_by_segment_impl(__internal::__hetero_tag<_BackendTag>, Policy&& policy, InputIterator1 first1,
+                               InputIterator1 last1, InputIterator2 first2, OutputIterator result, T init,
+                               BinaryPredicate binary_pred, Operator binary_op,
+                               ::std::false_type /* has_known_identity*/)
 {
 
     const auto n = ::std::distance(first1, last1);
diff --git a/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h b/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h
index f211f995b23..38b377bf601 100644
--- a/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h
@@ -78,9 +78,10 @@ pattern_inclusive_scan_by_segment(_Tag, Policy&& policy, InputIterator1 first1,
 template <typename _BackendTag, typename Policy, typename InputIterator1, typename InputIterator2,
           typename OutputIterator, typename BinaryPredicate, typename BinaryOperator>
 OutputIterator
-inclusive_scan_by_segment_impl(__internal::__hetero_tag<_BackendTag> __tag, Policy&& policy, InputIterator1 first1, InputIterator1 last1,
-                               InputIterator2 first2, OutputIterator result, BinaryPredicate binary_pred,
-                               BinaryOperator binary_op, ::std::true_type /* has_known_identity */)
+inclusive_scan_by_segment_impl(__internal::__hetero_tag<_BackendTag> __tag, Policy&& policy, InputIterator1 first1,
+                               InputIterator1 last1, InputIterator2 first2, OutputIterator result,
+                               BinaryPredicate binary_pred, BinaryOperator binary_op,
+                               ::std::true_type /* has_known_identity */)
 {
     using iter_value_t = typename ::std::iterator_traits<InputIterator2>::value_type;
     iter_value_t identity = unseq_backend::__known_identity<BinaryOperator, iter_value_t>;
@@ -91,9 +92,10 @@ inclusive_scan_by_segment_impl(__internal::__hetero_tag<_BackendTag> __tag, Poli
 template <typename _BackendTag, typename Policy, typename InputIterator1, typename InputIterator2,
           typename OutputIterator, typename BinaryPredicate, typename BinaryOperator>
 OutputIterator
-inclusive_scan_by_segment_impl(__internal::__hetero_tag<_BackendTag>, Policy&& policy, InputIterator1 first1, InputIterator1 last1,
-                               InputIterator2 first2, OutputIterator result, BinaryPredicate binary_pred,
-                               BinaryOperator binary_op, ::std::false_type /* has_known_identity */)
+inclusive_scan_by_segment_impl(__internal::__hetero_tag<_BackendTag>, Policy&& policy, InputIterator1 first1,
+                               InputIterator1 last1, InputIterator2 first2, OutputIterator result,
+                               BinaryPredicate binary_pred, BinaryOperator binary_op,
+                               ::std::false_type /* has_known_identity */)
 {
 
     typedef unsigned int FlagType;
diff --git a/include/oneapi/dpl/internal/reduce_by_segment_impl.h b/include/oneapi/dpl/internal/reduce_by_segment_impl.h
index e5ea2180cdb..546237b2ae8 100644
--- a/include/oneapi/dpl/internal/reduce_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/reduce_by_segment_impl.h
@@ -193,9 +193,10 @@ using _SegReducePrefixPhase = __seg_reduce_prefix_kernel<_Name...>;
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Range1, typename _Range2, typename _Range3,
           typename _Range4, typename _BinaryPredicate, typename _BinaryOperator>
 oneapi::dpl::__internal::__difference_t<_Range3>
-__sycl_reduce_by_segment(__internal::__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Range1&& __keys, _Range2&& __values,
-                         _Range3&& __out_keys, _Range4&& __out_values, _BinaryPredicate __binary_pred,
-                         _BinaryOperator __binary_op, ::std::false_type /* has_known_identity */)
+__sycl_reduce_by_segment(__internal::__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Range1&& __keys,
+                         _Range2&& __values, _Range3&& __out_keys, _Range4&& __out_values,
+                         _BinaryPredicate __binary_pred, _BinaryOperator __binary_op,
+                         ::std::false_type /* has_known_identity */)
 {
     return oneapi::dpl::experimental::ranges::reduce_by_segment(
         ::std::forward<_ExecutionPolicy>(__exec), ::std::forward<_Range1>(__keys), ::std::forward<_Range2>(__values),
@@ -205,9 +206,10 @@ __sycl_reduce_by_segment(__internal::__hetero_tag<_BackendTag>, _ExecutionPolicy
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Range1, typename _Range2, typename _Range3,
           typename _Range4, typename _BinaryPredicate, typename _BinaryOperator>
 oneapi::dpl::__internal::__difference_t<_Range3>
-__sycl_reduce_by_segment(__internal::__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Range1&& __keys, _Range2&& __values,
-                         _Range3&& __out_keys, _Range4&& __out_values, _BinaryPredicate __binary_pred,
-                         _BinaryOperator __binary_op, ::std::true_type /* has_known_identity */)
+__sycl_reduce_by_segment(__internal::__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Range1&& __keys,
+                         _Range2&& __values, _Range3&& __out_keys, _Range4&& __out_values,
+                         _BinaryPredicate __binary_pred, _BinaryOperator __binary_op,
+                         ::std::true_type /* has_known_identity */)
 {
     using _CustomName = oneapi::dpl::__internal::__policy_kernel_name<_ExecutionPolicy>;
 
@@ -573,9 +575,9 @@ __sycl_reduce_by_segment(__internal::__hetero_tag<_BackendTag>, _ExecutionPolicy
 template <typename _BackendTag, typename Policy, typename InputIterator1, typename InputIterator2,
           typename OutputIterator1, typename OutputIterator2, typename BinaryPred, typename BinaryOperator>
 ::std::pair<OutputIterator1, OutputIterator2>
-reduce_by_segment_impl(__internal::__hetero_tag<_BackendTag> __tag, Policy&& policy, InputIterator1 first1, InputIterator1 last1,
-                       InputIterator2 first2, OutputIterator1 result1, OutputIterator2 result2, BinaryPred binary_pred,
-                       BinaryOperator binary_op)
+reduce_by_segment_impl(__internal::__hetero_tag<_BackendTag> __tag, Policy&& policy, InputIterator1 first1,
+                       InputIterator1 last1, InputIterator2 first2, OutputIterator1 result1, OutputIterator2 result2,
+                       BinaryPred binary_pred, BinaryOperator binary_op)
 {
     // The algorithm reduces values in [first2, first2 + (last1-first1)) where the associated
     // keys for the values are equal to the adjacent key.
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 3f6ae184678..87415af4904 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -1470,7 +1470,7 @@ __remove_elements(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Forward
         _DifferenceType __min = __par_backend::__parallel_reduce(
             __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), _DifferenceType(0), __n, __n,
             [__first, __mask, &__calc_mask](_DifferenceType __i, _DifferenceType __j,
-                                                         _DifferenceType __local_min) -> _DifferenceType {
+                                            _DifferenceType __local_min) -> _DifferenceType {
                 // Create mask
                 __calc_mask(__mask + __i, __mask + __j, __first + __i);
 
@@ -1552,8 +1552,7 @@ __pattern_unique(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _Ra
         return __internal::__brick_unique(__first, __last, __pred, _IsVector{});
     }
     return __internal::__remove_elements(
-        __tag,
-        ::std::forward<_ExecutionPolicy>(__exec), ++__first, __last,
+        __tag, ::std::forward<_ExecutionPolicy>(__exec), ++__first, __last,
         [&__pred](bool* __b, bool* __e, _RandomAccessIterator __it) {
             __internal::__brick_walk3(
                 __b, __e, __it - 1, __it,
@@ -2985,8 +2984,7 @@ __pattern_remove_if(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec,
     }
 
     return __internal::__remove_elements(
-        __tag,
-        ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+        __tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
         [&__pred](bool* __b, bool* __e, _RandomAccessIterator __it) {
             __internal::__brick_walk2(
                 __b, __e, __it, [&__pred](bool& __x, _ReferenceType __y) { __x = !__pred(__y); }, _IsVector{});
@@ -4270,7 +4268,8 @@ __pattern_shift_left(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Forw
 
 template <class _Tag, class _ExecutionPolicy, class _BidirectionalIterator>
 _BidirectionalIterator
-__pattern_shift_right(_Tag __tag, _ExecutionPolicy&& __exec, _BidirectionalIterator __first, _BidirectionalIterator __last,
+__pattern_shift_right(_Tag __tag, _ExecutionPolicy&& __exec, _BidirectionalIterator __first,
+                      _BidirectionalIterator __last,
                       typename ::std::iterator_traits<_BidirectionalIterator>::difference_type __n)
 {
     static_assert(__is_backend_tag_v<_Tag>);
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 659fcc99155..ea13b6e53cf 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -488,8 +488,7 @@ __pattern_min_element(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec
     auto __buf = __keep(__first, __last);
 
     auto __ret_idx = oneapi::dpl::__par_backend_hetero::__parallel_transform_reduce<_ReduceValueType, _Commutative>(
-                         _BackendTag{},
-                         ::std::forward<_ExecutionPolicy>(__exec), __reduce_fn, __transform_fn,
+                         _BackendTag{}, ::std::forward<_ExecutionPolicy>(__exec), __reduce_fn, __transform_fn,
                          unseq_backend::__no_init_value{}, // no initial value
                          __buf.all_view())
                          .get();
@@ -557,8 +556,7 @@ __pattern_minmax_element(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __e
 
     auto __ret = oneapi::dpl::__par_backend_hetero::__parallel_transform_reduce<_ReduceValueType,
                                                                                 ::std::false_type /*is_commutative*/>(
-                     _BackendTag{},
-                     ::std::forward<_ExecutionPolicy>(__exec), __reduce_fn, __transform_fn,
+                     _BackendTag{}, ::std::forward<_ExecutionPolicy>(__exec), __reduce_fn, __transform_fn,
                      unseq_backend::__no_init_value{}, // no initial value
                      __buf.all_view())
                      .get();
@@ -652,8 +650,7 @@ __pattern_count(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Ite
 
     return oneapi::dpl::__par_backend_hetero::__parallel_transform_reduce<_ReduceValueType,
                                                                           ::std::true_type /*is_commutative*/>(
-               _BackendTag{},
-               ::std::forward<_ExecutionPolicy>(__exec), __reduce_fn, __transform_fn,
+               _BackendTag{}, ::std::forward<_ExecutionPolicy>(__exec), __reduce_fn, __transform_fn,
                unseq_backend::__no_init_value{}, // no initial value
                __buf.all_view())
         .get();
@@ -1119,8 +1116,7 @@ __pattern_is_partitioned(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __e
 
     auto __res = oneapi::dpl::__par_backend_hetero::__parallel_transform_reduce<_ReduceValueType,
                                                                                 ::std::false_type /*is_commutative*/>(
-                     _BackendTag{},
-                     ::std::forward<_ExecutionPolicy>(__exec), __reduce_fn, __transform_fn,
+                     _BackendTag{}, ::std::forward<_ExecutionPolicy>(__exec), __reduce_fn, __transform_fn,
                      unseq_backend::__no_init_value{}, // no initial value
                      __buf.all_view())
                      .get();
@@ -1234,8 +1230,8 @@ __pattern_merge(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Ite
 
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator, typename _Compare>
 void
-__pattern_inplace_merge(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Iterator __first,
-                        _Iterator __middle, _Iterator __last, _Compare __comp)
+__pattern_inplace_merge(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __middle,
+                        _Iterator __last, _Compare __comp)
 {
     using _ValueType = typename ::std::iterator_traits<_Iterator>::value_type;
 
@@ -1421,8 +1417,7 @@ __pattern_lexicographical_compare(__hetero_tag<_BackendTag> __tag, _ExecutionPol
     auto __ret_idx =
         oneapi::dpl::__par_backend_hetero::__parallel_transform_reduce<_ReduceValueType,
                                                                        ::std::false_type /*is_commutative*/>(
-            _BackendTag{},
-            ::std::forward<_ExecutionPolicy>(__exec), __reduce_fn, __transform_fn,
+            _BackendTag{}, ::std::forward<_ExecutionPolicy>(__exec), __reduce_fn, __transform_fn,
             unseq_backend::__no_init_value{}, // no initial value
             __buf1.all_view(), __buf2.all_view())
             .get();
@@ -1775,8 +1770,7 @@ __pattern_hetero_set_op(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _F
 
     auto __result_size =
         __par_backend_hetero::__parallel_transform_scan_base(
-            _BackendTag{},
-            ::std::forward<_ExecutionPolicy>(__exec),
+            _BackendTag{}, ::std::forward<_ExecutionPolicy>(__exec),
             oneapi::dpl::__ranges::make_zip_view(
                 __buf1.all_view(), __buf2.all_view(),
                 oneapi::dpl::__ranges::all_view<int32_t, __par_backend_hetero::access_mode::read_write>(
@@ -2087,7 +2081,8 @@ __pattern_shift_right(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec
 
     //A shift right is the shift left with a reverse logic.
     auto __rng = oneapi::dpl::__ranges::reverse_view_simple<decltype(__buf.all_view())>{__buf.all_view()};
-    auto __res = oneapi::dpl::__internal::__pattern_shift_left(__tag, ::std::forward<_ExecutionPolicy>(__exec), __rng, __n);
+    auto __res =
+        oneapi::dpl::__internal::__pattern_shift_left(__tag, ::std::forward<_ExecutionPolicy>(__exec), __rng, __n);
 
     return __last - __res;
 }
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
index 24a5d428a77..f227fd29383 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
@@ -239,8 +239,8 @@ __pattern_search(_ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&& __rng2,
 
     return oneapi::dpl::__par_backend_hetero::__parallel_find_or(
         oneapi::dpl::__internal::__device_backend_tag{}, // TODO required to fix backend tag eval
-        oneapi::dpl::__par_backend_hetero::make_wrapped_policy<oneapi::dpl::__par_backend_hetero::__find_policy_wrapper>
-            (::std::forward<_ExecutionPolicy>(__exec)),
+        oneapi::dpl::__par_backend_hetero::make_wrapped_policy<
+            oneapi::dpl::__par_backend_hetero::__find_policy_wrapper>(::std::forward<_ExecutionPolicy>(__exec)),
         _Predicate{__pred}, _TagType{}, ::std::forward<_Range1>(__rng1), ::std::forward<_Range2>(__rng2));
 }
 
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
index aba0a6ada17..ab31b05dc72 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
@@ -817,9 +817,8 @@ __parallel_transform_scan(oneapi::dpl::__internal::__device_backend_tag, _Execut
 
     return __future(
         __parallel_transform_scan_base(
-            oneapi::dpl::__internal::__device_backend_tag{},
-            ::std::forward<_ExecutionPolicy>(__exec), ::std::forward<_Range1>(__in_rng),
-            ::std::forward<_Range2>(__out_rng), __binary_op, __init,
+            oneapi::dpl::__internal::__device_backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
+            ::std::forward<_Range1>(__in_rng), ::std::forward<_Range2>(__out_rng), __binary_op, __init,
             // local scan
             unseq_backend::__scan<_Inclusive, _ExecutionPolicy, _BinaryOperation, _UnaryFunctor, _Assigner, _Assigner,
                                   _NoOpFunctor, _InitType>{__binary_op, _UnaryFunctor{__unary_op}, __assign_op,
@@ -895,8 +894,7 @@ __parallel_scan_copy(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPo
     oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, int32_t> __mask_buf(__exec, __n);
 
     return __parallel_transform_scan_base(
-        oneapi::dpl::__internal::__device_backend_tag{},
-        ::std::forward<_ExecutionPolicy>(__exec),
+        oneapi::dpl::__internal::__device_backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
         oneapi::dpl::__ranges::make_zip_view(
             ::std::forward<_InRng>(__in_rng),
             oneapi::dpl::__ranges::all_view<int32_t, __par_backend_hetero::access_mode::read_write>(
@@ -1814,8 +1812,8 @@ __parallel_stable_sort(oneapi::dpl::__internal::__device_backend_tag, _Execution
                        _Compare, _Proj __proj)
 {
     return __parallel_radix_sort<__internal::__is_comp_ascending<::std::decay_t<_Compare>>::value>(
-        oneapi::dpl::__internal::__device_backend_tag{},
-        ::std::forward<_ExecutionPolicy>(__exec), ::std::forward<_Range>(__rng), __proj);
+        oneapi::dpl::__internal::__device_backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
+        ::std::forward<_Range>(__rng), __proj);
 }
 #endif
 
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
index 6c88c276a66..08ccc58287c 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
@@ -97,17 +97,15 @@ __parallel_for(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionPolicy&& _
 template <typename _ExecutionPolicy, typename _Event, typename _Range1, typename _Range2, typename _BinHashMgr>
 auto
 __parallel_histogram(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionPolicy&& __exec, const _Event& __init_event,
-                     _Range1&& __input, _Range2&& __bins,
-                     const _BinHashMgr& __binhash_manager)
+                     _Range1&& __input, _Range2&& __bins, const _BinHashMgr& __binhash_manager)
 {
     static_assert(sizeof(oneapi::dpl::__internal::__value_t<_Range2>) <= sizeof(::std::uint32_t),
                   "histogram is not supported on FPGA devices with output types greater than 32 bits");
 
     // workaround until we implement more performant version for patterns
     return oneapi::dpl::__par_backend_hetero::__parallel_histogram(
-        oneapi::dpl::__internal::__device_backend_tag{},
-        __exec.__device_policy(), __init_event, ::std::forward<_Range1>(__input), ::std::forward<_Range2>(__bins),
-        __binhash_manager);
+        oneapi::dpl::__internal::__device_backend_tag{}, __exec.__device_policy(), __init_event,
+        ::std::forward<_Range1>(__input), ::std::forward<_Range2>(__bins), __binhash_manager);
 }
 
 } // namespace __par_backend_hetero
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h
index 8fbe0ec5af6..dd2de5ebc7d 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h
@@ -515,9 +515,9 @@ __parallel_histogram_select_kernel(oneapi::dpl::__internal::__device_backend_tag
     {
         return __future(
             __histogram_general_registers_local_reduction<__iters_per_work_item, __max_work_item_private_bins>(
-                oneapi::dpl::__internal::__device_backend_tag{},
-                ::std::forward<_ExecutionPolicy>(__exec), __init_event, __work_group_size,
-                ::std::forward<_Range1>(__input), ::std::forward<_Range2>(__bins), __binhash_manager));
+                oneapi::dpl::__internal::__device_backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __init_event,
+                __work_group_size, ::std::forward<_Range1>(__input), ::std::forward<_Range2>(__bins),
+                __binhash_manager));
     }
     // if bins fit into SLM, use local atomics
     else if (__num_bins * sizeof(_local_histogram_type) +
@@ -551,16 +551,14 @@ __parallel_histogram(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPo
     if (__input.size() < 1048576) // 2^20
     {
         return __parallel_histogram_select_kernel</*iters_per_workitem = */ 4>(
-            oneapi::dpl::__internal::__device_backend_tag{},
-            ::std::forward<_ExecutionPolicy>(__exec), __init_event, ::std::forward<_Range1>(__input),
-            ::std::forward<_Range2>(__bins), __binhash_manager);
+            oneapi::dpl::__internal::__device_backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __init_event,
+            ::std::forward<_Range1>(__input), ::std::forward<_Range2>(__bins), __binhash_manager);
     }
     else
     {
         return __parallel_histogram_select_kernel</*iters_per_workitem = */ 32>(
-            oneapi::dpl::__internal::__device_backend_tag{},
-            ::std::forward<_ExecutionPolicy>(__exec), __init_event, ::std::forward<_Range1>(__input),
-            ::std::forward<_Range2>(__bins), __binhash_manager);
+            oneapi::dpl::__internal::__device_backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __init_event,
+            ::std::forward<_Range1>(__input), ::std::forward<_Range2>(__bins), __binhash_manager);
     }
 }
 
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h
index c6d164f9415..726c1743301 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h
@@ -432,44 +432,38 @@ __parallel_transform_reduce(oneapi::dpl::__internal::__device_backend_tag, _Exec
         if (__n <= 256)
         {
             return __parallel_transform_reduce_small_impl<_Tp, 256, 1, _Commutative>(
-                oneapi::dpl::__internal::__device_backend_tag{},
-                ::std::forward<_ExecutionPolicy>(__exec), __n, __reduce_op, __transform_op, __init,
-                ::std::forward<_Ranges>(__rngs)...);
+                oneapi::dpl::__internal::__device_backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __n,
+                __reduce_op, __transform_op, __init, ::std::forward<_Ranges>(__rngs)...);
         }
         else if (__n <= 512)
         {
             return __parallel_transform_reduce_small_impl<_Tp, 256, 2, _Commutative>(
-                oneapi::dpl::__internal::__device_backend_tag{},
-                ::std::forward<_ExecutionPolicy>(__exec), __n, __reduce_op, __transform_op, __init,
-                ::std::forward<_Ranges>(__rngs)...);
+                oneapi::dpl::__internal::__device_backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __n,
+                __reduce_op, __transform_op, __init, ::std::forward<_Ranges>(__rngs)...);
         }
         else if (__n <= 1024)
         {
             return __parallel_transform_reduce_small_impl<_Tp, 256, 4, _Commutative>(
-                oneapi::dpl::__internal::__device_backend_tag{},
-                ::std::forward<_ExecutionPolicy>(__exec), __n, __reduce_op, __transform_op, __init,
-                ::std::forward<_Ranges>(__rngs)...);
+                oneapi::dpl::__internal::__device_backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __n,
+                __reduce_op, __transform_op, __init, ::std::forward<_Ranges>(__rngs)...);
         }
         else if (__n <= 2048)
         {
             return __parallel_transform_reduce_small_impl<_Tp, 256, 8, _Commutative>(
-                oneapi::dpl::__internal::__device_backend_tag{},
-                ::std::forward<_ExecutionPolicy>(__exec), __n, __reduce_op, __transform_op, __init,
-                ::std::forward<_Ranges>(__rngs)...);
+                oneapi::dpl::__internal::__device_backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __n,
+                __reduce_op, __transform_op, __init, ::std::forward<_Ranges>(__rngs)...);
         }
         else if (__n <= 4096)
         {
             return __parallel_transform_reduce_small_impl<_Tp, 256, 16, _Commutative>(
-                oneapi::dpl::__internal::__device_backend_tag{},
-                ::std::forward<_ExecutionPolicy>(__exec), __n, __reduce_op, __transform_op, __init,
-                ::std::forward<_Ranges>(__rngs)...);
+                oneapi::dpl::__internal::__device_backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __n,
+                __reduce_op, __transform_op, __init, ::std::forward<_Ranges>(__rngs)...);
         }
         else if (__n <= 8192)
         {
             return __parallel_transform_reduce_small_impl<_Tp, 256, 32, _Commutative>(
-                oneapi::dpl::__internal::__device_backend_tag{},
-                ::std::forward<_ExecutionPolicy>(__exec), __n, __reduce_op, __transform_op, __init,
-                ::std::forward<_Ranges>(__rngs)...);
+                oneapi::dpl::__internal::__device_backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __n,
+                __reduce_op, __transform_op, __init, ::std::forward<_Ranges>(__rngs)...);
         }
 
         // Use two-step tree reduction.
@@ -478,51 +472,44 @@ __parallel_transform_reduce(oneapi::dpl::__internal::__device_backend_tag, _Exec
         else if (__n <= 2097152)
         {
             return __parallel_transform_reduce_mid_impl<_Tp, 256, 32, 1, _Commutative>(
-                oneapi::dpl::__internal::__device_backend_tag{},
-                ::std::forward<_ExecutionPolicy>(__exec), __n, __reduce_op, __transform_op, __init,
-                ::std::forward<_Ranges>(__rngs)...);
+                oneapi::dpl::__internal::__device_backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __n,
+                __reduce_op, __transform_op, __init, ::std::forward<_Ranges>(__rngs)...);
         }
         else if (__n <= 4194304)
         {
             return __parallel_transform_reduce_mid_impl<_Tp, 256, 32, 2, _Commutative>(
-                oneapi::dpl::__internal::__device_backend_tag{},
-                ::std::forward<_ExecutionPolicy>(__exec), __n, __reduce_op, __transform_op, __init,
-                ::std::forward<_Ranges>(__rngs)...);
+                oneapi::dpl::__internal::__device_backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __n,
+                __reduce_op, __transform_op, __init, ::std::forward<_Ranges>(__rngs)...);
         }
         else if (__n <= 8388608)
         {
             return __parallel_transform_reduce_mid_impl<_Tp, 256, 32, 4, _Commutative>(
-                oneapi::dpl::__internal::__device_backend_tag{},
-                ::std::forward<_ExecutionPolicy>(__exec), __n, __reduce_op, __transform_op, __init,
-                ::std::forward<_Ranges>(__rngs)...);
+                oneapi::dpl::__internal::__device_backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __n,
+                __reduce_op, __transform_op, __init, ::std::forward<_Ranges>(__rngs)...);
         }
         else if (__n <= 16777216)
         {
             return __parallel_transform_reduce_mid_impl<_Tp, 256, 32, 8, _Commutative>(
-                oneapi::dpl::__internal::__device_backend_tag{},
-                ::std::forward<_ExecutionPolicy>(__exec), __n, __reduce_op, __transform_op, __init,
-                ::std::forward<_Ranges>(__rngs)...);
+                oneapi::dpl::__internal::__device_backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __n,
+                __reduce_op, __transform_op, __init, ::std::forward<_Ranges>(__rngs)...);
         }
         else if (__n <= 33554432)
         {
             return __parallel_transform_reduce_mid_impl<_Tp, 256, 32, 16, _Commutative>(
-                oneapi::dpl::__internal::__device_backend_tag{},
-                ::std::forward<_ExecutionPolicy>(__exec), __n, __reduce_op, __transform_op, __init,
-                ::std::forward<_Ranges>(__rngs)...);
+                oneapi::dpl::__internal::__device_backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __n,
+                __reduce_op, __transform_op, __init, ::std::forward<_Ranges>(__rngs)...);
         }
         else if (__n <= 67108864)
         {
             return __parallel_transform_reduce_mid_impl<_Tp, 256, 32, 32, _Commutative>(
-                oneapi::dpl::__internal::__device_backend_tag{},
-                ::std::forward<_ExecutionPolicy>(__exec), __n, __reduce_op, __transform_op, __init,
-                ::std::forward<_Ranges>(__rngs)...);
+                oneapi::dpl::__internal::__device_backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __n,
+                __reduce_op, __transform_op, __init, ::std::forward<_Ranges>(__rngs)...);
         }
     }
     // Otherwise use a recursive tree reduction.
     return __parallel_transform_reduce_impl<_Tp, 32, _Commutative>::submit(
-        oneapi::dpl::__internal::__device_backend_tag{},
-        ::std::forward<_ExecutionPolicy>(__exec), __n, __work_group_size, __reduce_op, __transform_op, __init,
-        ::std::forward<_Ranges>(__rngs)...);
+        oneapi::dpl::__internal::__device_backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __n,
+        __work_group_size, __reduce_op, __transform_op, __init, ::std::forward<_Ranges>(__rngs)...);
 }
 
 } // namespace __par_backend_hetero
diff --git a/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
index 84a49029182..2955815d29d 100644
--- a/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
@@ -60,8 +60,7 @@ __pattern_transform_reduce(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& _
 
     return oneapi::dpl::__par_backend_hetero::__parallel_transform_reduce<_RepackedTp,
                                                                           ::std::true_type /*is_commutative*/>(
-               _BackendTag{},
-               ::std::forward<_ExecutionPolicy>(__exec), __binary_op1, _Functor{__binary_op2},
+               _BackendTag{}, ::std::forward<_ExecutionPolicy>(__exec), __binary_op1, _Functor{__binary_op2},
                unseq_backend::__init_value<_RepackedTp>{__init}, // initial value
                __buf1.all_view(), __buf2.all_view())
         .get();
@@ -89,8 +88,7 @@ __pattern_transform_reduce(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& _
 
     return oneapi::dpl::__par_backend_hetero::__parallel_transform_reduce<_RepackedTp,
                                                                           ::std::true_type /*is_commutative*/>(
-               _BackendTag{},
-               ::std::forward<_ExecutionPolicy>(__exec), __binary_op, _Functor{__unary_op},
+               _BackendTag{}, ::std::forward<_ExecutionPolicy>(__exec), __binary_op, _Functor{__unary_op},
                unseq_backend::__init_value<_RepackedTp>{__init}, // initial value
                __buf.all_view())
         .get();
diff --git a/include/oneapi/dpl/pstl/numeric_impl.h b/include/oneapi/dpl/pstl/numeric_impl.h
index d5dfe6a09bb..5e7f4dcbc0d 100644
--- a/include/oneapi/dpl/pstl/numeric_impl.h
+++ b/include/oneapi/dpl/pstl/numeric_impl.h
@@ -263,8 +263,7 @@ __pattern_transform_scan(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _
 
     return __internal::__except_handler([&]() {
         __par_backend::__parallel_transform_scan(
-            __backend_tag{},
-            ::std::forward<_ExecutionPolicy>(__exec), __last - __first,
+            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __last - __first,
             [__first, __unary_op](_DifferenceType __i) mutable { return __unary_op(__first[__i]); }, __init,
             __binary_op,
             [__first, __unary_op, __binary_op](_DifferenceType __i, _DifferenceType __j, _Tp __init) {
@@ -388,9 +387,8 @@ __pattern_adjacent_difference(_Tag, _ExecutionPolicy&&, _ForwardIterator __first
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
           class _BinaryOperation>
 _RandomAccessIterator2
-__pattern_adjacent_difference(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec,
-                              _RandomAccessIterator1 __first, _RandomAccessIterator1 __last,
-                              _RandomAccessIterator2 __d_first, _BinaryOperation __op)
+__pattern_adjacent_difference(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator1 __first,
+                              _RandomAccessIterator1 __last, _RandomAccessIterator2 __d_first, _BinaryOperation __op)
 {
     using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
 

From fd191e25e3d0642154cdc157af9e2f49d230f75d Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Mon, 19 Feb 2024 11:47:54 +0100
Subject: [PATCH 336/566] Fix self review comment: using
 oneapi::dpl::__internal::__device_backend_tag instance from function param
 instead of creating the new one

---
 .../pstl/hetero/dpcpp/parallel_backend_sycl.h | 60 +++++++++++--------
 .../hetero/dpcpp/parallel_backend_sycl_fpga.h |  3 +-
 .../dpcpp/parallel_backend_sycl_histogram.h   | 15 ++---
 .../dpcpp/parallel_backend_sycl_reduce.h      | 40 +++++++------
 4 files changed, 65 insertions(+), 53 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
index ab31b05dc72..0459126fbed 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
@@ -265,15 +265,14 @@ struct __parallel_for_submitter<__internal::__optional_kernel_name<_Name...>>
 //for some algorithms happens that size of processing range is n, but amount of iterations is n/2.
 template <typename _ExecutionPolicy, typename _Fp, typename _Index, typename... _Ranges>
 auto
-__parallel_for(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&& __exec, _Fp __brick, _Index __count,
-               _Ranges&&... __rngs)
+__parallel_for(oneapi::dpl::__internal::__device_backend_tag __backend_tag, _ExecutionPolicy&& __exec, _Fp __brick,
+               _Index __count, _Ranges&&... __rngs)
 {
     using _CustomName = oneapi::dpl::__internal::__policy_kernel_name<_ExecutionPolicy>;
     using _ForKernel = oneapi::dpl::__par_backend_hetero::__internal::__kernel_name_provider<_CustomName>;
 
-    return __parallel_for_submitter<_ForKernel>()(oneapi::dpl::__internal::__device_backend_tag{},
-                                                  ::std::forward<_ExecutionPolicy>(__exec), __brick, __count,
-                                                  ::std::forward<_Ranges>(__rngs)...);
+    return __parallel_for_submitter<_ForKernel>()(__backend_tag, ::std::forward<_ExecutionPolicy>(__exec), __brick,
+                                                  __count, ::std::forward<_Ranges>(__rngs)...);
 }
 
 //------------------------------------------------------------------------
@@ -775,7 +774,8 @@ __parallel_transform_scan_base(oneapi::dpl::__internal::__device_backend_tag, _E
 template <typename _ExecutionPolicy, typename _Range1, typename _Range2, typename _UnaryOperation, typename _InitType,
           typename _BinaryOperation, typename _Inclusive>
 auto
-__parallel_transform_scan(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&& __exec, _Range1&& __in_rng,
+__parallel_transform_scan(oneapi::dpl::__internal::__device_backend_tag __backend_tag, _ExecutionPolicy&& __exec,
+                          _Range1&& __in_rng,
                           _Range2&& __out_rng, ::std::size_t __n, _UnaryOperation __unary_op, _InitType __init,
                           _BinaryOperation __binary_op, _Inclusive)
 {
@@ -799,7 +799,7 @@ __parallel_transform_scan(oneapi::dpl::__internal::__device_backend_tag, _Execut
         if (__n <= __single_group_upper_limit && __max_slm_size >= __req_slm_size)
         {
             return __parallel_transform_scan_single_group(
-                oneapi::dpl::__internal::__device_backend_tag{}, std::forward<_ExecutionPolicy>(__exec),
+                __backend_tag, std::forward<_ExecutionPolicy>(__exec),
                 ::std::forward<_Range1>(__in_rng), ::std::forward<_Range2>(__out_rng), __n, __unary_op, __init,
                 __binary_op, _Inclusive{});
         }
@@ -817,7 +817,7 @@ __parallel_transform_scan(oneapi::dpl::__internal::__device_backend_tag, _Execut
 
     return __future(
         __parallel_transform_scan_base(
-            oneapi::dpl::__internal::__device_backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
+            __backend_tag, ::std::forward<_ExecutionPolicy>(__exec),
             ::std::forward<_Range1>(__in_rng), ::std::forward<_Range2>(__out_rng), __binary_op, __init,
             // local scan
             unseq_backend::__scan<_Inclusive, _ExecutionPolicy, _BinaryOperation, _UnaryFunctor, _Assigner, _Assigner,
@@ -875,7 +875,8 @@ struct __invoke_single_group_copy_if
 template <typename _ExecutionPolicy, typename _InRng, typename _OutRng, typename _Size, typename _CreateMaskOp,
           typename _CopyByMaskOp>
 auto
-__parallel_scan_copy(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&& __exec, _InRng&& __in_rng,
+__parallel_scan_copy(oneapi::dpl::__internal::__device_backend_tag __backend_tag, _ExecutionPolicy&& __exec,
+                     _InRng&& __in_rng,
                      _OutRng&& __out_rng, _Size __n, _CreateMaskOp __create_mask_op, _CopyByMaskOp __copy_by_mask_op)
 {
     using _ReduceOp = ::std::plus<_Size>;
@@ -894,7 +895,7 @@ __parallel_scan_copy(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPo
     oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, int32_t> __mask_buf(__exec, __n);
 
     return __parallel_transform_scan_base(
-        oneapi::dpl::__internal::__device_backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
+        __backend_tag, ::std::forward<_ExecutionPolicy>(__exec),
         oneapi::dpl::__ranges::make_zip_view(
             ::std::forward<_InRng>(__in_rng),
             oneapi::dpl::__ranges::all_view<int32_t, __par_backend_hetero::access_mode::read_write>(
@@ -914,7 +915,8 @@ __parallel_scan_copy(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPo
 
 template <typename _ExecutionPolicy, typename _InRng, typename _OutRng, typename _Size, typename _Pred>
 auto
-__parallel_copy_if(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&& __exec, _InRng&& __in_rng,
+__parallel_copy_if(oneapi::dpl::__internal::__device_backend_tag __backend_tag, _ExecutionPolicy&& __exec,
+                   _InRng&& __in_rng,
                    _OutRng&& __out_rng, _Size __n, _Pred __pred)
 {
     using _SingleGroupInvoker = __invoke_single_group_copy_if<_Size>;
@@ -950,7 +952,7 @@ __parallel_copy_if(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPoli
         using CopyOp = unseq_backend::__copy_by_mask<_ReduceOp, oneapi::dpl::__internal::__pstl_assign,
                                                      /*inclusive*/ ::std::true_type, 1>;
 
-        return __parallel_scan_copy(oneapi::dpl::__internal::__device_backend_tag{},
+        return __parallel_scan_copy(__backend_tag,
                                     ::std::forward<_ExecutionPolicy>(__exec), ::std::forward<_InRng>(__in_rng),
                                     ::std::forward<_OutRng>(__out_rng), __n, CreateOp{__pred}, CopyOp{});
     }
@@ -1203,7 +1205,8 @@ class __or_policy_wrapper
 
 template <typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2, typename _Brick>
 bool
-__parallel_or(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&& __exec, _Iterator1 __first,
+__parallel_or(oneapi::dpl::__internal::__device_backend_tag __backend_tag, _ExecutionPolicy&& __exec,
+              _Iterator1 __first,
               _Iterator1 __last, _Iterator2 __s_first, _Iterator2 __s_last, _Brick __f)
 {
     auto __keep = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _Iterator1>();
@@ -1212,7 +1215,7 @@ __parallel_or(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&&
     auto __s_buf = __s_keep(__s_first, __s_last);
 
     return oneapi::dpl::__par_backend_hetero::__parallel_find_or(
-        oneapi::dpl::__internal::__device_backend_tag{},
+        __backend_tag,
         __par_backend_hetero::make_wrapped_policy<__or_policy_wrapper>(::std::forward<_ExecutionPolicy>(__exec)), __f,
         __parallel_or_tag{}, __buf.all_view(), __s_buf.all_view());
 }
@@ -1222,14 +1225,14 @@ __parallel_or(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&&
 // backend code.
 template <typename _ExecutionPolicy, typename _Iterator, typename _Brick>
 bool
-__parallel_or(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&& __exec, _Iterator __first,
+__parallel_or(oneapi::dpl::__internal::__device_backend_tag __backend_tag, _ExecutionPolicy&& __exec, _Iterator __first,
               _Iterator __last, _Brick __f)
 {
     auto __keep = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _Iterator>();
     auto __buf = __keep(__first, __last);
 
     return oneapi::dpl::__par_backend_hetero::__parallel_find_or(
-        oneapi::dpl::__internal::__device_backend_tag{},
+        __backend_tag,
         __par_backend_hetero::make_wrapped_policy<__or_policy_wrapper>(::std::forward<_ExecutionPolicy>(__exec)), __f,
         __parallel_or_tag{}, __buf.all_view());
 }
@@ -1245,7 +1248,8 @@ class __find_policy_wrapper
 
 template <typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2, typename _Brick, typename _IsFirst>
 _Iterator1
-__parallel_find(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&& __exec, _Iterator1 __first,
+__parallel_find(oneapi::dpl::__internal::__device_backend_tag __backend_tag, _ExecutionPolicy&& __exec,
+                _Iterator1 __first,
                 _Iterator1 __last, _Iterator2 __s_first, _Iterator2 __s_last, _Brick __f, _IsFirst)
 {
     auto __keep = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _Iterator1>();
@@ -1256,7 +1260,7 @@ __parallel_find(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&
     using _TagType = ::std::conditional_t<_IsFirst::value, __parallel_find_forward_tag<decltype(__buf.all_view())>,
                                           __parallel_find_backward_tag<decltype(__buf.all_view())>>;
     return __first + oneapi::dpl::__par_backend_hetero::__parallel_find_or(
-                         oneapi::dpl::__internal::__device_backend_tag{},
+                         __backend_tag,
                          __par_backend_hetero::make_wrapped_policy<__find_policy_wrapper>(
                              ::std::forward<_ExecutionPolicy>(__exec)),
                          __f, _TagType{}, __buf.all_view(), __s_buf.all_view());
@@ -1267,7 +1271,8 @@ __parallel_find(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&
 // backend code.
 template <typename _ExecutionPolicy, typename _Iterator, typename _Brick, typename _IsFirst>
 _Iterator
-__parallel_find(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&& __exec, _Iterator __first,
+__parallel_find(oneapi::dpl::__internal::__device_backend_tag __backend_tag, _ExecutionPolicy&& __exec,
+                _Iterator __first,
                 _Iterator __last, _Brick __f, _IsFirst)
 {
     auto __keep = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _Iterator>();
@@ -1277,7 +1282,7 @@ __parallel_find(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&
         typename ::std::conditional<_IsFirst::value, __parallel_find_forward_tag<decltype(__buf.all_view())>,
                                     __parallel_find_backward_tag<decltype(__buf.all_view())>>::type;
     return __first + oneapi::dpl::__par_backend_hetero::__parallel_find_or(
-                         oneapi::dpl::__internal::__device_backend_tag{},
+                         __backend_tag,
                          __par_backend_hetero::make_wrapped_policy<__find_policy_wrapper>(
                              ::std::forward<_ExecutionPolicy>(__exec)),
                          __f, _TagType{}, __buf.all_view());
@@ -1808,11 +1813,12 @@ template <
     ::std::enable_if_t<
         __is_radix_sort_usable_for_type<oneapi::dpl::__internal::__key_t<_Proj, _Range>, _Compare>::value, int> = 0>
 auto
-__parallel_stable_sort(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&& __exec, _Range&& __rng,
+__parallel_stable_sort(oneapi::dpl::__internal::__device_backend_tag __backend_tag, _ExecutionPolicy&& __exec,
+                       _Range&& __rng,
                        _Compare, _Proj __proj)
 {
     return __parallel_radix_sort<__internal::__is_comp_ascending<::std::decay_t<_Compare>>::value>(
-        oneapi::dpl::__internal::__device_backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
+        __backend_tag, ::std::forward<_ExecutionPolicy>(__exec),
         ::std::forward<_Range>(__rng), __proj);
 }
 #endif
@@ -1822,13 +1828,14 @@ template <
     ::std::enable_if_t<
         !__is_radix_sort_usable_for_type<oneapi::dpl::__internal::__key_t<_Proj, _Range>, _Compare>::value, int> = 0>
 auto
-__parallel_stable_sort(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&& __exec, _Range&& __rng,
+__parallel_stable_sort(oneapi::dpl::__internal::__device_backend_tag __backend_tag, _ExecutionPolicy&& __exec,
+                       _Range&& __rng,
                        _Compare __comp, _Proj __proj)
 {
     auto __cmp_f = [__comp, __proj](const auto& __a, const auto& __b) mutable {
         return __comp(__proj(__a), __proj(__b));
     };
-    return __parallel_sort_impl(oneapi::dpl::__internal::__device_backend_tag{},
+    return __parallel_sort_impl(__backend_tag,
                                 ::std::forward<_ExecutionPolicy>(__exec), ::std::forward<_Range>(__rng), __cmp_f);
 }
 
@@ -1841,7 +1848,8 @@ __parallel_stable_sort(oneapi::dpl::__internal::__device_backend_tag, _Execution
 //       __full_merge_kernel in order to use __parallel_sort_impl routine
 template <typename _ExecutionPolicy, typename _Iterator, typename _Compare>
 auto
-__parallel_partial_sort(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&& __exec, _Iterator __first,
+__parallel_partial_sort(oneapi::dpl::__internal::__device_backend_tag __backend_tag, _ExecutionPolicy&& __exec,
+                        _Iterator __first,
                         _Iterator __mid, _Iterator __last, _Compare __comp)
 {
     const auto __mid_idx = __mid - __first;
@@ -1849,7 +1857,7 @@ __parallel_partial_sort(oneapi::dpl::__internal::__device_backend_tag, _Executio
     auto __keep = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read_write, _Iterator>();
     auto __buf = __keep(__first, __last);
 
-    return __parallel_partial_sort_impl(oneapi::dpl::__internal::__device_backend_tag{},
+    return __parallel_partial_sort_impl(__backend_tag,
                                         ::std::forward<_ExecutionPolicy>(__exec), __buf.all_view(),
                                         __partial_merge_kernel<decltype(__mid_idx)>{__mid_idx}, __comp);
 }
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
index 08ccc58287c..d3970e5f5a5 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
@@ -96,7 +96,8 @@ __parallel_for(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionPolicy&& _
 // TODO: check if it makes sense to move these wrappers out of backend to a common place
 template <typename _ExecutionPolicy, typename _Event, typename _Range1, typename _Range2, typename _BinHashMgr>
 auto
-__parallel_histogram(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionPolicy&& __exec, const _Event& __init_event,
+__parallel_histogram(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionPolicy&& __exec,
+                     const _Event& __init_event,
                      _Range1&& __input, _Range2&& __bins, const _BinHashMgr& __binhash_manager)
 {
     static_assert(sizeof(oneapi::dpl::__internal::__value_t<_Range2>) <= sizeof(::std::uint32_t),
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h
index dd2de5ebc7d..54d6e54bb8e 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h
@@ -495,7 +495,8 @@ __histogram_general_private_global_atomics(oneapi::dpl::__internal::__device_bac
 template <::std::uint16_t __iters_per_work_item, typename _ExecutionPolicy, typename _Range1, typename _Range2,
           typename _BinHashMgr>
 auto
-__parallel_histogram_select_kernel(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&& __exec,
+__parallel_histogram_select_kernel(oneapi::dpl::__internal::__device_backend_tag __backend_tag,
+                                   _ExecutionPolicy&& __exec,
                                    const sycl::event& __init_event, _Range1&& __input, _Range2&& __bins,
                                    const _BinHashMgr& __binhash_manager)
 {
@@ -515,7 +516,7 @@ __parallel_histogram_select_kernel(oneapi::dpl::__internal::__device_backend_tag
     {
         return __future(
             __histogram_general_registers_local_reduction<__iters_per_work_item, __max_work_item_private_bins>(
-                oneapi::dpl::__internal::__device_backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __init_event,
+                __backend_tag, ::std::forward<_ExecutionPolicy>(__exec), __init_event,
                 __work_group_size, ::std::forward<_Range1>(__input), ::std::forward<_Range2>(__bins),
                 __binhash_manager));
     }
@@ -525,7 +526,7 @@ __parallel_histogram_select_kernel(oneapi::dpl::__internal::__device_backend_tag
              __local_mem_size)
     {
         return __future(__histogram_general_local_atomics<__iters_per_work_item>(
-            oneapi::dpl::__internal::__device_backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __init_event,
+            __backend_tag, ::std::forward<_ExecutionPolicy>(__exec), __init_event,
             __work_group_size, ::std::forward<_Range1>(__input), ::std::forward<_Range2>(__bins), __binhash_manager));
     }
     else // otherwise, use global atomics (private copies per workgroup)
@@ -536,7 +537,7 @@ __parallel_histogram_select_kernel(oneapi::dpl::__internal::__device_backend_tag
         // private copies of the histogram bins in global memory.  No unrolling is taken advantage of here because it
         // is a runtime argument.
         return __future(__histogram_general_private_global_atomics(
-            oneapi::dpl::__internal::__device_backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __init_event,
+            __backend_tag, ::std::forward<_ExecutionPolicy>(__exec), __init_event,
             __iters_per_work_item, __work_group_size, ::std::forward<_Range1>(__input), ::std::forward<_Range2>(__bins),
             __binhash_manager));
     }
@@ -544,20 +545,20 @@ __parallel_histogram_select_kernel(oneapi::dpl::__internal::__device_backend_tag
 
 template <typename _ExecutionPolicy, typename _Range1, typename _Range2, typename _BinHashMgr>
 auto
-__parallel_histogram(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&& __exec,
+__parallel_histogram(oneapi::dpl::__internal::__device_backend_tag __backend_tag, _ExecutionPolicy&& __exec,
                      const sycl::event& __init_event, _Range1&& __input, _Range2&& __bins,
                      const _BinHashMgr& __binhash_manager)
 {
     if (__input.size() < 1048576) // 2^20
     {
         return __parallel_histogram_select_kernel</*iters_per_workitem = */ 4>(
-            oneapi::dpl::__internal::__device_backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __init_event,
+            __backend_tag, ::std::forward<_ExecutionPolicy>(__exec), __init_event,
             ::std::forward<_Range1>(__input), ::std::forward<_Range2>(__bins), __binhash_manager);
     }
     else
     {
         return __parallel_histogram_select_kernel</*iters_per_workitem = */ 32>(
-            oneapi::dpl::__internal::__device_backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __init_event,
+            __backend_tag, ::std::forward<_ExecutionPolicy>(__exec), __init_event,
             ::std::forward<_Range1>(__input), ::std::forward<_Range2>(__bins), __binhash_manager);
     }
 }
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h
index 726c1743301..100369680bf 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h
@@ -139,7 +139,8 @@ template <typename _Tp, ::std::uint16_t __work_group_size, ::std::uint8_t __iter
           typename _ExecutionPolicy, typename _Size, typename _ReduceOp, typename _TransformOp, typename _InitType,
           typename... _Ranges>
 auto
-__parallel_transform_reduce_small_impl(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&& __exec,
+__parallel_transform_reduce_small_impl(oneapi::dpl::__internal::__device_backend_tag __backend_tag,
+                                       _ExecutionPolicy&& __exec,
                                        const _Size __n, _ReduceOp __reduce_op, _TransformOp __transform_op,
                                        _InitType __init, _Ranges&&... __rngs)
 {
@@ -149,7 +150,7 @@ __parallel_transform_reduce_small_impl(oneapi::dpl::__internal::__device_backend
 
     return __parallel_transform_reduce_small_submitter<_Tp, __work_group_size, __iters_per_work_item, _Commutative,
                                                        _ReduceKernel>()(
-        oneapi::dpl::__internal::__device_backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __n, __reduce_op,
+        __backend_tag, ::std::forward<_ExecutionPolicy>(__exec), __n, __reduce_op,
         __transform_op, __init, ::std::forward<_Ranges>(__rngs)...);
 }
 
@@ -257,7 +258,8 @@ template <typename _Tp, ::std::uint16_t __work_group_size, ::std::uint8_t __iter
           ::std::uint8_t __iters_per_work_item_work_group_kernel, typename _Commutative, typename _ExecutionPolicy,
           typename _Size, typename _ReduceOp, typename _TransformOp, typename _InitType, typename... _Ranges>
 auto
-__parallel_transform_reduce_mid_impl(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&& __exec,
+__parallel_transform_reduce_mid_impl(oneapi::dpl::__internal::__device_backend_tag __backend_tag,
+                                     _ExecutionPolicy&& __exec,
                                      _Size __n, _ReduceOp __reduce_op, _TransformOp __transform_op, _InitType __init,
                                      _Ranges&&... __rngs)
 {
@@ -279,13 +281,13 @@ __parallel_transform_reduce_mid_impl(oneapi::dpl::__internal::__device_backend_t
     sycl::event __reduce_event =
         __parallel_transform_reduce_device_kernel_submitter<_Tp, __work_group_size, __iters_per_work_item_device_kernel,
                                                             _Commutative, _ReduceDeviceKernel>()(
-            oneapi::dpl::__internal::__device_backend_tag{}, __exec, __n, __reduce_op, __transform_op, __init, __temp,
+            __backend_tag, __exec, __n, __reduce_op, __transform_op, __init, __temp,
             ::std::forward<_Ranges>(__rngs)...);
 
     __n = __n_groups; // Number of preliminary results from the device kernel.
     return __parallel_transform_reduce_work_group_kernel_submitter<
         _Tp, __work_group_size, __iters_per_work_item_work_group_kernel, _Commutative, _ReduceWorkGroupKernel>()(
-        oneapi::dpl::__internal::__device_backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __reduce_event, __n,
+        __backend_tag, ::std::forward<_ExecutionPolicy>(__exec), __reduce_event, __n,
         __reduce_op, __transform_op, __init, __temp);
 }
 
@@ -415,7 +417,7 @@ struct __parallel_transform_reduce_impl
 template <typename _Tp, typename _Commutative, typename _ExecutionPolicy, typename _ReduceOp, typename _TransformOp,
           typename _InitType, typename... _Ranges>
 auto
-__parallel_transform_reduce(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&& __exec,
+__parallel_transform_reduce(oneapi::dpl::__internal::__device_backend_tag __backend_tag, _ExecutionPolicy&& __exec,
                             _ReduceOp __reduce_op, _TransformOp __transform_op, _InitType __init, _Ranges&&... __rngs)
 {
     auto __n = oneapi::dpl::__ranges::__get_first_range_size(__rngs...);
@@ -432,37 +434,37 @@ __parallel_transform_reduce(oneapi::dpl::__internal::__device_backend_tag, _Exec
         if (__n <= 256)
         {
             return __parallel_transform_reduce_small_impl<_Tp, 256, 1, _Commutative>(
-                oneapi::dpl::__internal::__device_backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __n,
+                __backend_tag, ::std::forward<_ExecutionPolicy>(__exec), __n,
                 __reduce_op, __transform_op, __init, ::std::forward<_Ranges>(__rngs)...);
         }
         else if (__n <= 512)
         {
             return __parallel_transform_reduce_small_impl<_Tp, 256, 2, _Commutative>(
-                oneapi::dpl::__internal::__device_backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __n,
+                __backend_tag, ::std::forward<_ExecutionPolicy>(__exec), __n,
                 __reduce_op, __transform_op, __init, ::std::forward<_Ranges>(__rngs)...);
         }
         else if (__n <= 1024)
         {
             return __parallel_transform_reduce_small_impl<_Tp, 256, 4, _Commutative>(
-                oneapi::dpl::__internal::__device_backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __n,
+                __backend_tag, ::std::forward<_ExecutionPolicy>(__exec), __n,
                 __reduce_op, __transform_op, __init, ::std::forward<_Ranges>(__rngs)...);
         }
         else if (__n <= 2048)
         {
             return __parallel_transform_reduce_small_impl<_Tp, 256, 8, _Commutative>(
-                oneapi::dpl::__internal::__device_backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __n,
+                __backend_tag, ::std::forward<_ExecutionPolicy>(__exec), __n,
                 __reduce_op, __transform_op, __init, ::std::forward<_Ranges>(__rngs)...);
         }
         else if (__n <= 4096)
         {
             return __parallel_transform_reduce_small_impl<_Tp, 256, 16, _Commutative>(
-                oneapi::dpl::__internal::__device_backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __n,
+                __backend_tag, ::std::forward<_ExecutionPolicy>(__exec), __n,
                 __reduce_op, __transform_op, __init, ::std::forward<_Ranges>(__rngs)...);
         }
         else if (__n <= 8192)
         {
             return __parallel_transform_reduce_small_impl<_Tp, 256, 32, _Commutative>(
-                oneapi::dpl::__internal::__device_backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __n,
+                __backend_tag, ::std::forward<_ExecutionPolicy>(__exec), __n,
                 __reduce_op, __transform_op, __init, ::std::forward<_Ranges>(__rngs)...);
         }
 
@@ -472,43 +474,43 @@ __parallel_transform_reduce(oneapi::dpl::__internal::__device_backend_tag, _Exec
         else if (__n <= 2097152)
         {
             return __parallel_transform_reduce_mid_impl<_Tp, 256, 32, 1, _Commutative>(
-                oneapi::dpl::__internal::__device_backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __n,
+                __backend_tag, ::std::forward<_ExecutionPolicy>(__exec), __n,
                 __reduce_op, __transform_op, __init, ::std::forward<_Ranges>(__rngs)...);
         }
         else if (__n <= 4194304)
         {
             return __parallel_transform_reduce_mid_impl<_Tp, 256, 32, 2, _Commutative>(
-                oneapi::dpl::__internal::__device_backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __n,
+                __backend_tag, ::std::forward<_ExecutionPolicy>(__exec), __n,
                 __reduce_op, __transform_op, __init, ::std::forward<_Ranges>(__rngs)...);
         }
         else if (__n <= 8388608)
         {
             return __parallel_transform_reduce_mid_impl<_Tp, 256, 32, 4, _Commutative>(
-                oneapi::dpl::__internal::__device_backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __n,
+                __backend_tag, ::std::forward<_ExecutionPolicy>(__exec), __n,
                 __reduce_op, __transform_op, __init, ::std::forward<_Ranges>(__rngs)...);
         }
         else if (__n <= 16777216)
         {
             return __parallel_transform_reduce_mid_impl<_Tp, 256, 32, 8, _Commutative>(
-                oneapi::dpl::__internal::__device_backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __n,
+                __backend_tag, ::std::forward<_ExecutionPolicy>(__exec), __n,
                 __reduce_op, __transform_op, __init, ::std::forward<_Ranges>(__rngs)...);
         }
         else if (__n <= 33554432)
         {
             return __parallel_transform_reduce_mid_impl<_Tp, 256, 32, 16, _Commutative>(
-                oneapi::dpl::__internal::__device_backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __n,
+                __backend_tag, ::std::forward<_ExecutionPolicy>(__exec), __n,
                 __reduce_op, __transform_op, __init, ::std::forward<_Ranges>(__rngs)...);
         }
         else if (__n <= 67108864)
         {
             return __parallel_transform_reduce_mid_impl<_Tp, 256, 32, 32, _Commutative>(
-                oneapi::dpl::__internal::__device_backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __n,
+                __backend_tag, ::std::forward<_ExecutionPolicy>(__exec), __n,
                 __reduce_op, __transform_op, __init, ::std::forward<_Ranges>(__rngs)...);
         }
     }
     // Otherwise use a recursive tree reduction.
     return __parallel_transform_reduce_impl<_Tp, 32, _Commutative>::submit(
-        oneapi::dpl::__internal::__device_backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __n,
+        __backend_tag, ::std::forward<_ExecutionPolicy>(__exec), __n,
         __work_group_size, __reduce_op, __transform_op, __init, ::std::forward<_Ranges>(__rngs)...);
 }
 

From 90827c27ec1fe6e645313ea51c9608a5d9f96465 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Mon, 19 Feb 2024 11:54:53 +0100
Subject: [PATCH 337/566] Apply GitHUB clang format

---
 .../pstl/hetero/dpcpp/parallel_backend_sycl.h | 55 ++++++---------
 .../hetero/dpcpp/parallel_backend_sycl_fpga.h |  3 +-
 .../dpcpp/parallel_backend_sycl_histogram.h   | 27 ++++---
 .../dpcpp/parallel_backend_sycl_reduce.h      | 70 +++++++++----------
 4 files changed, 69 insertions(+), 86 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
index 0459126fbed..e0f7e3a4a5c 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
@@ -775,9 +775,8 @@ template <typename _ExecutionPolicy, typename _Range1, typename _Range2, typenam
           typename _BinaryOperation, typename _Inclusive>
 auto
 __parallel_transform_scan(oneapi::dpl::__internal::__device_backend_tag __backend_tag, _ExecutionPolicy&& __exec,
-                          _Range1&& __in_rng,
-                          _Range2&& __out_rng, ::std::size_t __n, _UnaryOperation __unary_op, _InitType __init,
-                          _BinaryOperation __binary_op, _Inclusive)
+                          _Range1&& __in_rng, _Range2&& __out_rng, ::std::size_t __n, _UnaryOperation __unary_op,
+                          _InitType __init, _BinaryOperation __binary_op, _Inclusive)
 {
     using _Type = typename _InitType::__value_type;
 
@@ -799,9 +798,8 @@ __parallel_transform_scan(oneapi::dpl::__internal::__device_backend_tag __backen
         if (__n <= __single_group_upper_limit && __max_slm_size >= __req_slm_size)
         {
             return __parallel_transform_scan_single_group(
-                __backend_tag, std::forward<_ExecutionPolicy>(__exec),
-                ::std::forward<_Range1>(__in_rng), ::std::forward<_Range2>(__out_rng), __n, __unary_op, __init,
-                __binary_op, _Inclusive{});
+                __backend_tag, std::forward<_ExecutionPolicy>(__exec), ::std::forward<_Range1>(__in_rng),
+                ::std::forward<_Range2>(__out_rng), __n, __unary_op, __init, __binary_op, _Inclusive{});
         }
     }
 
@@ -817,8 +815,8 @@ __parallel_transform_scan(oneapi::dpl::__internal::__device_backend_tag __backen
 
     return __future(
         __parallel_transform_scan_base(
-            __backend_tag, ::std::forward<_ExecutionPolicy>(__exec),
-            ::std::forward<_Range1>(__in_rng), ::std::forward<_Range2>(__out_rng), __binary_op, __init,
+            __backend_tag, ::std::forward<_ExecutionPolicy>(__exec), ::std::forward<_Range1>(__in_rng),
+            ::std::forward<_Range2>(__out_rng), __binary_op, __init,
             // local scan
             unseq_backend::__scan<_Inclusive, _ExecutionPolicy, _BinaryOperation, _UnaryFunctor, _Assigner, _Assigner,
                                   _NoOpFunctor, _InitType>{__binary_op, _UnaryFunctor{__unary_op}, __assign_op,
@@ -876,8 +874,8 @@ template <typename _ExecutionPolicy, typename _InRng, typename _OutRng, typename
           typename _CopyByMaskOp>
 auto
 __parallel_scan_copy(oneapi::dpl::__internal::__device_backend_tag __backend_tag, _ExecutionPolicy&& __exec,
-                     _InRng&& __in_rng,
-                     _OutRng&& __out_rng, _Size __n, _CreateMaskOp __create_mask_op, _CopyByMaskOp __copy_by_mask_op)
+                     _InRng&& __in_rng, _OutRng&& __out_rng, _Size __n, _CreateMaskOp __create_mask_op,
+                     _CopyByMaskOp __copy_by_mask_op)
 {
     using _ReduceOp = ::std::plus<_Size>;
     using _Assigner = unseq_backend::__scan_assigner;
@@ -916,8 +914,7 @@ __parallel_scan_copy(oneapi::dpl::__internal::__device_backend_tag __backend_tag
 template <typename _ExecutionPolicy, typename _InRng, typename _OutRng, typename _Size, typename _Pred>
 auto
 __parallel_copy_if(oneapi::dpl::__internal::__device_backend_tag __backend_tag, _ExecutionPolicy&& __exec,
-                   _InRng&& __in_rng,
-                   _OutRng&& __out_rng, _Size __n, _Pred __pred)
+                   _InRng&& __in_rng, _OutRng&& __out_rng, _Size __n, _Pred __pred)
 {
     using _SingleGroupInvoker = __invoke_single_group_copy_if<_Size>;
 
@@ -952,9 +949,9 @@ __parallel_copy_if(oneapi::dpl::__internal::__device_backend_tag __backend_tag,
         using CopyOp = unseq_backend::__copy_by_mask<_ReduceOp, oneapi::dpl::__internal::__pstl_assign,
                                                      /*inclusive*/ ::std::true_type, 1>;
 
-        return __parallel_scan_copy(__backend_tag,
-                                    ::std::forward<_ExecutionPolicy>(__exec), ::std::forward<_InRng>(__in_rng),
-                                    ::std::forward<_OutRng>(__out_rng), __n, CreateOp{__pred}, CopyOp{});
+        return __parallel_scan_copy(__backend_tag, ::std::forward<_ExecutionPolicy>(__exec),
+                                    ::std::forward<_InRng>(__in_rng), ::std::forward<_OutRng>(__out_rng), __n,
+                                    CreateOp{__pred}, CopyOp{});
     }
 }
 
@@ -1206,8 +1203,7 @@ class __or_policy_wrapper
 template <typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2, typename _Brick>
 bool
 __parallel_or(oneapi::dpl::__internal::__device_backend_tag __backend_tag, _ExecutionPolicy&& __exec,
-              _Iterator1 __first,
-              _Iterator1 __last, _Iterator2 __s_first, _Iterator2 __s_last, _Brick __f)
+              _Iterator1 __first, _Iterator1 __last, _Iterator2 __s_first, _Iterator2 __s_last, _Brick __f)
 {
     auto __keep = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _Iterator1>();
     auto __buf = __keep(__first, __last);
@@ -1249,8 +1245,7 @@ class __find_policy_wrapper
 template <typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2, typename _Brick, typename _IsFirst>
 _Iterator1
 __parallel_find(oneapi::dpl::__internal::__device_backend_tag __backend_tag, _ExecutionPolicy&& __exec,
-                _Iterator1 __first,
-                _Iterator1 __last, _Iterator2 __s_first, _Iterator2 __s_last, _Brick __f, _IsFirst)
+                _Iterator1 __first, _Iterator1 __last, _Iterator2 __s_first, _Iterator2 __s_last, _Brick __f, _IsFirst)
 {
     auto __keep = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _Iterator1>();
     auto __buf = __keep(__first, __last);
@@ -1272,8 +1267,7 @@ __parallel_find(oneapi::dpl::__internal::__device_backend_tag __backend_tag, _Ex
 template <typename _ExecutionPolicy, typename _Iterator, typename _Brick, typename _IsFirst>
 _Iterator
 __parallel_find(oneapi::dpl::__internal::__device_backend_tag __backend_tag, _ExecutionPolicy&& __exec,
-                _Iterator __first,
-                _Iterator __last, _Brick __f, _IsFirst)
+                _Iterator __first, _Iterator __last, _Brick __f, _IsFirst)
 {
     auto __keep = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _Iterator>();
     auto __buf = __keep(__first, __last);
@@ -1814,12 +1808,10 @@ template <
         __is_radix_sort_usable_for_type<oneapi::dpl::__internal::__key_t<_Proj, _Range>, _Compare>::value, int> = 0>
 auto
 __parallel_stable_sort(oneapi::dpl::__internal::__device_backend_tag __backend_tag, _ExecutionPolicy&& __exec,
-                       _Range&& __rng,
-                       _Compare, _Proj __proj)
+                       _Range&& __rng, _Compare, _Proj __proj)
 {
     return __parallel_radix_sort<__internal::__is_comp_ascending<::std::decay_t<_Compare>>::value>(
-        __backend_tag, ::std::forward<_ExecutionPolicy>(__exec),
-        ::std::forward<_Range>(__rng), __proj);
+        __backend_tag, ::std::forward<_ExecutionPolicy>(__exec), ::std::forward<_Range>(__rng), __proj);
 }
 #endif
 
@@ -1829,14 +1821,13 @@ template <
         !__is_radix_sort_usable_for_type<oneapi::dpl::__internal::__key_t<_Proj, _Range>, _Compare>::value, int> = 0>
 auto
 __parallel_stable_sort(oneapi::dpl::__internal::__device_backend_tag __backend_tag, _ExecutionPolicy&& __exec,
-                       _Range&& __rng,
-                       _Compare __comp, _Proj __proj)
+                       _Range&& __rng, _Compare __comp, _Proj __proj)
 {
     auto __cmp_f = [__comp, __proj](const auto& __a, const auto& __b) mutable {
         return __comp(__proj(__a), __proj(__b));
     };
-    return __parallel_sort_impl(__backend_tag,
-                                ::std::forward<_ExecutionPolicy>(__exec), ::std::forward<_Range>(__rng), __cmp_f);
+    return __parallel_sort_impl(__backend_tag, ::std::forward<_ExecutionPolicy>(__exec), ::std::forward<_Range>(__rng),
+                                __cmp_f);
 }
 
 //------------------------------------------------------------------------
@@ -1849,16 +1840,14 @@ __parallel_stable_sort(oneapi::dpl::__internal::__device_backend_tag __backend_t
 template <typename _ExecutionPolicy, typename _Iterator, typename _Compare>
 auto
 __parallel_partial_sort(oneapi::dpl::__internal::__device_backend_tag __backend_tag, _ExecutionPolicy&& __exec,
-                        _Iterator __first,
-                        _Iterator __mid, _Iterator __last, _Compare __comp)
+                        _Iterator __first, _Iterator __mid, _Iterator __last, _Compare __comp)
 {
     const auto __mid_idx = __mid - __first;
 
     auto __keep = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read_write, _Iterator>();
     auto __buf = __keep(__first, __last);
 
-    return __parallel_partial_sort_impl(__backend_tag,
-                                        ::std::forward<_ExecutionPolicy>(__exec), __buf.all_view(),
+    return __parallel_partial_sort_impl(__backend_tag, ::std::forward<_ExecutionPolicy>(__exec), __buf.all_view(),
                                         __partial_merge_kernel<decltype(__mid_idx)>{__mid_idx}, __comp);
 }
 } // namespace __par_backend_hetero
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
index d3970e5f5a5..08ccc58287c 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
@@ -96,8 +96,7 @@ __parallel_for(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionPolicy&& _
 // TODO: check if it makes sense to move these wrappers out of backend to a common place
 template <typename _ExecutionPolicy, typename _Event, typename _Range1, typename _Range2, typename _BinHashMgr>
 auto
-__parallel_histogram(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionPolicy&& __exec,
-                     const _Event& __init_event,
+__parallel_histogram(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionPolicy&& __exec, const _Event& __init_event,
                      _Range1&& __input, _Range2&& __bins, const _BinHashMgr& __binhash_manager)
 {
     static_assert(sizeof(oneapi::dpl::__internal::__value_t<_Range2>) <= sizeof(::std::uint32_t),
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h
index 54d6e54bb8e..2e7223ec062 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h
@@ -496,9 +496,8 @@ template <::std::uint16_t __iters_per_work_item, typename _ExecutionPolicy, type
           typename _BinHashMgr>
 auto
 __parallel_histogram_select_kernel(oneapi::dpl::__internal::__device_backend_tag __backend_tag,
-                                   _ExecutionPolicy&& __exec,
-                                   const sycl::event& __init_event, _Range1&& __input, _Range2&& __bins,
-                                   const _BinHashMgr& __binhash_manager)
+                                   _ExecutionPolicy&& __exec, const sycl::event& __init_event, _Range1&& __input,
+                                   _Range2&& __bins, const _BinHashMgr& __binhash_manager)
 {
     using _private_histogram_type = ::std::uint16_t;
     using _local_histogram_type = ::std::uint32_t;
@@ -516,9 +515,8 @@ __parallel_histogram_select_kernel(oneapi::dpl::__internal::__device_backend_tag
     {
         return __future(
             __histogram_general_registers_local_reduction<__iters_per_work_item, __max_work_item_private_bins>(
-                __backend_tag, ::std::forward<_ExecutionPolicy>(__exec), __init_event,
-                __work_group_size, ::std::forward<_Range1>(__input), ::std::forward<_Range2>(__bins),
-                __binhash_manager));
+                __backend_tag, ::std::forward<_ExecutionPolicy>(__exec), __init_event, __work_group_size,
+                ::std::forward<_Range1>(__input), ::std::forward<_Range2>(__bins), __binhash_manager));
     }
     // if bins fit into SLM, use local atomics
     else if (__num_bins * sizeof(_local_histogram_type) +
@@ -526,8 +524,8 @@ __parallel_histogram_select_kernel(oneapi::dpl::__internal::__device_backend_tag
              __local_mem_size)
     {
         return __future(__histogram_general_local_atomics<__iters_per_work_item>(
-            __backend_tag, ::std::forward<_ExecutionPolicy>(__exec), __init_event,
-            __work_group_size, ::std::forward<_Range1>(__input), ::std::forward<_Range2>(__bins), __binhash_manager));
+            __backend_tag, ::std::forward<_ExecutionPolicy>(__exec), __init_event, __work_group_size,
+            ::std::forward<_Range1>(__input), ::std::forward<_Range2>(__bins), __binhash_manager));
     }
     else // otherwise, use global atomics (private copies per workgroup)
     {
@@ -537,9 +535,8 @@ __parallel_histogram_select_kernel(oneapi::dpl::__internal::__device_backend_tag
         // private copies of the histogram bins in global memory.  No unrolling is taken advantage of here because it
         // is a runtime argument.
         return __future(__histogram_general_private_global_atomics(
-            __backend_tag, ::std::forward<_ExecutionPolicy>(__exec), __init_event,
-            __iters_per_work_item, __work_group_size, ::std::forward<_Range1>(__input), ::std::forward<_Range2>(__bins),
-            __binhash_manager));
+            __backend_tag, ::std::forward<_ExecutionPolicy>(__exec), __init_event, __iters_per_work_item,
+            __work_group_size, ::std::forward<_Range1>(__input), ::std::forward<_Range2>(__bins), __binhash_manager));
     }
 }
 
@@ -552,14 +549,14 @@ __parallel_histogram(oneapi::dpl::__internal::__device_backend_tag __backend_tag
     if (__input.size() < 1048576) // 2^20
     {
         return __parallel_histogram_select_kernel</*iters_per_workitem = */ 4>(
-            __backend_tag, ::std::forward<_ExecutionPolicy>(__exec), __init_event,
-            ::std::forward<_Range1>(__input), ::std::forward<_Range2>(__bins), __binhash_manager);
+            __backend_tag, ::std::forward<_ExecutionPolicy>(__exec), __init_event, ::std::forward<_Range1>(__input),
+            ::std::forward<_Range2>(__bins), __binhash_manager);
     }
     else
     {
         return __parallel_histogram_select_kernel</*iters_per_workitem = */ 32>(
-            __backend_tag, ::std::forward<_ExecutionPolicy>(__exec), __init_event,
-            ::std::forward<_Range1>(__input), ::std::forward<_Range2>(__bins), __binhash_manager);
+            __backend_tag, ::std::forward<_ExecutionPolicy>(__exec), __init_event, ::std::forward<_Range1>(__input),
+            ::std::forward<_Range2>(__bins), __binhash_manager);
     }
 }
 
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h
index 100369680bf..da1d37e22f1 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce.h
@@ -140,9 +140,8 @@ template <typename _Tp, ::std::uint16_t __work_group_size, ::std::uint8_t __iter
           typename... _Ranges>
 auto
 __parallel_transform_reduce_small_impl(oneapi::dpl::__internal::__device_backend_tag __backend_tag,
-                                       _ExecutionPolicy&& __exec,
-                                       const _Size __n, _ReduceOp __reduce_op, _TransformOp __transform_op,
-                                       _InitType __init, _Ranges&&... __rngs)
+                                       _ExecutionPolicy&& __exec, const _Size __n, _ReduceOp __reduce_op,
+                                       _TransformOp __transform_op, _InitType __init, _Ranges&&... __rngs)
 {
     using _CustomName = oneapi::dpl::__internal::__policy_kernel_name<_ExecutionPolicy>;
     using _ReduceKernel = oneapi::dpl::__par_backend_hetero::__internal::__kernel_name_provider<
@@ -150,8 +149,8 @@ __parallel_transform_reduce_small_impl(oneapi::dpl::__internal::__device_backend
 
     return __parallel_transform_reduce_small_submitter<_Tp, __work_group_size, __iters_per_work_item, _Commutative,
                                                        _ReduceKernel>()(
-        __backend_tag, ::std::forward<_ExecutionPolicy>(__exec), __n, __reduce_op,
-        __transform_op, __init, ::std::forward<_Ranges>(__rngs)...);
+        __backend_tag, ::std::forward<_ExecutionPolicy>(__exec), __n, __reduce_op, __transform_op, __init,
+        ::std::forward<_Ranges>(__rngs)...);
 }
 
 // Submits the first kernel of the parallel_transform_reduce for mid-sized arrays.
@@ -259,9 +258,8 @@ template <typename _Tp, ::std::uint16_t __work_group_size, ::std::uint8_t __iter
           typename _Size, typename _ReduceOp, typename _TransformOp, typename _InitType, typename... _Ranges>
 auto
 __parallel_transform_reduce_mid_impl(oneapi::dpl::__internal::__device_backend_tag __backend_tag,
-                                     _ExecutionPolicy&& __exec,
-                                     _Size __n, _ReduceOp __reduce_op, _TransformOp __transform_op, _InitType __init,
-                                     _Ranges&&... __rngs)
+                                     _ExecutionPolicy&& __exec, _Size __n, _ReduceOp __reduce_op,
+                                     _TransformOp __transform_op, _InitType __init, _Ranges&&... __rngs)
 {
     using _CustomName = oneapi::dpl::__internal::__policy_kernel_name<_ExecutionPolicy>;
 
@@ -287,8 +285,8 @@ __parallel_transform_reduce_mid_impl(oneapi::dpl::__internal::__device_backend_t
     __n = __n_groups; // Number of preliminary results from the device kernel.
     return __parallel_transform_reduce_work_group_kernel_submitter<
         _Tp, __work_group_size, __iters_per_work_item_work_group_kernel, _Commutative, _ReduceWorkGroupKernel>()(
-        __backend_tag, ::std::forward<_ExecutionPolicy>(__exec), __reduce_event, __n,
-        __reduce_op, __transform_op, __init, __temp);
+        __backend_tag, ::std::forward<_ExecutionPolicy>(__exec), __reduce_event, __n, __reduce_op, __transform_op,
+        __init, __temp);
 }
 
 // General implementation using a tree reduction
@@ -434,38 +432,38 @@ __parallel_transform_reduce(oneapi::dpl::__internal::__device_backend_tag __back
         if (__n <= 256)
         {
             return __parallel_transform_reduce_small_impl<_Tp, 256, 1, _Commutative>(
-                __backend_tag, ::std::forward<_ExecutionPolicy>(__exec), __n,
-                __reduce_op, __transform_op, __init, ::std::forward<_Ranges>(__rngs)...);
+                __backend_tag, ::std::forward<_ExecutionPolicy>(__exec), __n, __reduce_op, __transform_op, __init,
+                ::std::forward<_Ranges>(__rngs)...);
         }
         else if (__n <= 512)
         {
             return __parallel_transform_reduce_small_impl<_Tp, 256, 2, _Commutative>(
-                __backend_tag, ::std::forward<_ExecutionPolicy>(__exec), __n,
-                __reduce_op, __transform_op, __init, ::std::forward<_Ranges>(__rngs)...);
+                __backend_tag, ::std::forward<_ExecutionPolicy>(__exec), __n, __reduce_op, __transform_op, __init,
+                ::std::forward<_Ranges>(__rngs)...);
         }
         else if (__n <= 1024)
         {
             return __parallel_transform_reduce_small_impl<_Tp, 256, 4, _Commutative>(
-                __backend_tag, ::std::forward<_ExecutionPolicy>(__exec), __n,
-                __reduce_op, __transform_op, __init, ::std::forward<_Ranges>(__rngs)...);
+                __backend_tag, ::std::forward<_ExecutionPolicy>(__exec), __n, __reduce_op, __transform_op, __init,
+                ::std::forward<_Ranges>(__rngs)...);
         }
         else if (__n <= 2048)
         {
             return __parallel_transform_reduce_small_impl<_Tp, 256, 8, _Commutative>(
-                __backend_tag, ::std::forward<_ExecutionPolicy>(__exec), __n,
-                __reduce_op, __transform_op, __init, ::std::forward<_Ranges>(__rngs)...);
+                __backend_tag, ::std::forward<_ExecutionPolicy>(__exec), __n, __reduce_op, __transform_op, __init,
+                ::std::forward<_Ranges>(__rngs)...);
         }
         else if (__n <= 4096)
         {
             return __parallel_transform_reduce_small_impl<_Tp, 256, 16, _Commutative>(
-                __backend_tag, ::std::forward<_ExecutionPolicy>(__exec), __n,
-                __reduce_op, __transform_op, __init, ::std::forward<_Ranges>(__rngs)...);
+                __backend_tag, ::std::forward<_ExecutionPolicy>(__exec), __n, __reduce_op, __transform_op, __init,
+                ::std::forward<_Ranges>(__rngs)...);
         }
         else if (__n <= 8192)
         {
             return __parallel_transform_reduce_small_impl<_Tp, 256, 32, _Commutative>(
-                __backend_tag, ::std::forward<_ExecutionPolicy>(__exec), __n,
-                __reduce_op, __transform_op, __init, ::std::forward<_Ranges>(__rngs)...);
+                __backend_tag, ::std::forward<_ExecutionPolicy>(__exec), __n, __reduce_op, __transform_op, __init,
+                ::std::forward<_Ranges>(__rngs)...);
         }
 
         // Use two-step tree reduction.
@@ -474,44 +472,44 @@ __parallel_transform_reduce(oneapi::dpl::__internal::__device_backend_tag __back
         else if (__n <= 2097152)
         {
             return __parallel_transform_reduce_mid_impl<_Tp, 256, 32, 1, _Commutative>(
-                __backend_tag, ::std::forward<_ExecutionPolicy>(__exec), __n,
-                __reduce_op, __transform_op, __init, ::std::forward<_Ranges>(__rngs)...);
+                __backend_tag, ::std::forward<_ExecutionPolicy>(__exec), __n, __reduce_op, __transform_op, __init,
+                ::std::forward<_Ranges>(__rngs)...);
         }
         else if (__n <= 4194304)
         {
             return __parallel_transform_reduce_mid_impl<_Tp, 256, 32, 2, _Commutative>(
-                __backend_tag, ::std::forward<_ExecutionPolicy>(__exec), __n,
-                __reduce_op, __transform_op, __init, ::std::forward<_Ranges>(__rngs)...);
+                __backend_tag, ::std::forward<_ExecutionPolicy>(__exec), __n, __reduce_op, __transform_op, __init,
+                ::std::forward<_Ranges>(__rngs)...);
         }
         else if (__n <= 8388608)
         {
             return __parallel_transform_reduce_mid_impl<_Tp, 256, 32, 4, _Commutative>(
-                __backend_tag, ::std::forward<_ExecutionPolicy>(__exec), __n,
-                __reduce_op, __transform_op, __init, ::std::forward<_Ranges>(__rngs)...);
+                __backend_tag, ::std::forward<_ExecutionPolicy>(__exec), __n, __reduce_op, __transform_op, __init,
+                ::std::forward<_Ranges>(__rngs)...);
         }
         else if (__n <= 16777216)
         {
             return __parallel_transform_reduce_mid_impl<_Tp, 256, 32, 8, _Commutative>(
-                __backend_tag, ::std::forward<_ExecutionPolicy>(__exec), __n,
-                __reduce_op, __transform_op, __init, ::std::forward<_Ranges>(__rngs)...);
+                __backend_tag, ::std::forward<_ExecutionPolicy>(__exec), __n, __reduce_op, __transform_op, __init,
+                ::std::forward<_Ranges>(__rngs)...);
         }
         else if (__n <= 33554432)
         {
             return __parallel_transform_reduce_mid_impl<_Tp, 256, 32, 16, _Commutative>(
-                __backend_tag, ::std::forward<_ExecutionPolicy>(__exec), __n,
-                __reduce_op, __transform_op, __init, ::std::forward<_Ranges>(__rngs)...);
+                __backend_tag, ::std::forward<_ExecutionPolicy>(__exec), __n, __reduce_op, __transform_op, __init,
+                ::std::forward<_Ranges>(__rngs)...);
         }
         else if (__n <= 67108864)
         {
             return __parallel_transform_reduce_mid_impl<_Tp, 256, 32, 32, _Commutative>(
-                __backend_tag, ::std::forward<_ExecutionPolicy>(__exec), __n,
-                __reduce_op, __transform_op, __init, ::std::forward<_Ranges>(__rngs)...);
+                __backend_tag, ::std::forward<_ExecutionPolicy>(__exec), __n, __reduce_op, __transform_op, __init,
+                ::std::forward<_Ranges>(__rngs)...);
         }
     }
     // Otherwise use a recursive tree reduction.
     return __parallel_transform_reduce_impl<_Tp, 32, _Commutative>::submit(
-        __backend_tag, ::std::forward<_ExecutionPolicy>(__exec), __n,
-        __work_group_size, __reduce_op, __transform_op, __init, ::std::forward<_Ranges>(__rngs)...);
+        __backend_tag, ::std::forward<_ExecutionPolicy>(__exec), __n, __work_group_size, __reduce_op, __transform_op,
+        __init, ::std::forward<_Ranges>(__rngs)...);
 }
 
 } // namespace __par_backend_hetero

From 508b95d667d5c2b566f16354dd862c937d023162 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Mon, 19 Feb 2024 17:15:35 +0100
Subject: [PATCH 338/566] 
 include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h - fix self review
 comment: remove extra __select_backend calls

---
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 112 +++++-------------
 1 file changed, 29 insertions(+), 83 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index ea13b6e53cf..105f805c313 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -1039,14 +1039,9 @@ __pattern_remove_if(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec,
 
     auto __copy_last = __pattern_copy_if(__tag, __exec, __first, __last, __copy_first, __not_pred<_Predicate>{__pred});
 
-    constexpr auto __dispatch_tag1 =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, decltype(__copy_first), decltype(__copy_last),
-                                                  decltype(__first)>();
-
     //TODO: optimize copy back depending on Iterator, i.e. set_final_data for host iterator/pointer
     return __pattern_walk2(
-        __dispatch_tag1,
-        __par_backend_hetero::make_wrapped_policy<copy_back_wrapper>(::std::forward<_ExecutionPolicy>(__exec)),
+        __tag, __par_backend_hetero::make_wrapped_policy<copy_back_wrapper>(::std::forward<_ExecutionPolicy>(__exec)),
         __copy_first, __copy_last, __first, __brick_copy<_ExecutionPolicy>{});
 }
 
@@ -1062,21 +1057,12 @@ __pattern_unique(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _It
 
     oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __buf(__exec, __last - __first);
     auto __copy_first = __buf.get();
-    auto __copy_last =
-        __pattern_unique_copy(oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, decltype(__first),
-                                                                        decltype(__last), decltype(__copy_first)>(),
-                              __exec, __first, __last, __copy_first, __pred);
-
-    constexpr auto __dispatch_tag1 =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, decltype(__copy_first), decltype(__copy_last),
-                                                  decltype(__first)>();
-    using __backend_tag1 = typename decltype(__dispatch_tag1)::__backend_tag;
+    auto __copy_last = __pattern_unique_copy(__tag, __exec, __first, __last, __copy_first, __pred);
 
     //TODO: optimize copy back depending on Iterator, i.e. set_final_data for host iterator/pointer
-    return __pattern_walk2<__backend_tag1, /*_IsSync=*/::std::true_type, __par_backend_hetero::access_mode::read_write,
+    return __pattern_walk2<_BackendTag, /*_IsSync=*/::std::true_type, __par_backend_hetero::access_mode::read_write,
                            __par_backend_hetero::access_mode::read_write>(
-        __dispatch_tag1,
-        __par_backend_hetero::make_wrapped_policy<copy_back_wrapper>(::std::forward<_ExecutionPolicy>(__exec)),
+        __tag, __par_backend_hetero::make_wrapped_policy<copy_back_wrapper>(::std::forward<_ExecutionPolicy>(__exec)),
         __copy_first, __copy_last, __first, __brick_copy<_ExecutionPolicy>{});
 }
 
@@ -1230,7 +1216,7 @@ __pattern_merge(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Ite
 
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator, typename _Compare>
 void
-__pattern_inplace_merge(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __middle,
+__pattern_inplace_merge(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __middle,
                         _Iterator __last, _Compare __comp)
 {
     using _ValueType = typename ::std::iterator_traits<_Iterator>::value_type;
@@ -1245,20 +1231,16 @@ __pattern_inplace_merge(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _I
     auto __copy_first = __buf.get();
     auto __copy_last = __copy_first + __n;
 
-    __pattern_merge(oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, decltype(__first), decltype(__middle),
-                                                              decltype(__last), decltype(__copy_first)>(),
-                    __exec, __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__first),
-                    __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__middle),
-                    __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__middle),
-                    __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__last),
-                    __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::write>(__copy_first),
-                    __comp);
+    __pattern_merge(
+        __tag, __exec, __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__first),
+        __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__middle),
+        __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__middle),
+        __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read>(__last),
+        __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::write>(__copy_first), __comp);
 
     //TODO: optimize copy back depending on Iterator, i.e. set_final_data for host iterator/pointer
     __pattern_walk2(
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, decltype(__copy_first), decltype(__copy_last),
-                                                  decltype(__first)>(),
-        __par_backend_hetero::make_wrapped_policy<copy_back_wrapper>(::std::forward<_ExecutionPolicy>(__exec)),
+        __tag, __par_backend_hetero::make_wrapped_policy<copy_back_wrapper>(::std::forward<_ExecutionPolicy>(__exec)),
         __copy_first, __copy_last, __first, __brick_move<_ExecutionPolicy>{});
 }
 
@@ -1338,27 +1320,16 @@ __pattern_stable_partition(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& _
     auto __true_result = __true_buf.get();
     auto __false_result = __false_buf.get();
 
-    constexpr auto __dispatch_tag = __select_backend<_ExecutionPolicy, decltype(__first), decltype(__last),
-                                                     decltype(__true_result), decltype(__false_result)>();
-    auto copy_result =
-        __pattern_partition_copy(__dispatch_tag, __exec, __first, __last, __true_result, __false_result, __pred);
+    auto copy_result = __pattern_partition_copy(__tag, __exec, __first, __last, __true_result, __false_result, __pred);
     auto true_count = copy_result.first - __true_result;
 
     //TODO: optimize copy back if possible (inplace, decrease number of submits)
-    constexpr auto __dispatch_tag1 =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, decltype(__true_result),
-                                                  decltype(copy_result.first), decltype(__first)>();
-    using __backend_tag1 = typename decltype(__dispatch_tag1)::__backend_tag;
-    __pattern_walk2<__backend_tag1, /*_IsSync=*/::std::false_type>(
-        __dispatch_tag1, __par_backend_hetero::make_wrapped_policy<copy_back_wrapper>(__exec), __true_result,
-        copy_result.first, __first, __brick_move<_ExecutionPolicy>{});
-
-    constexpr auto __dispatch_tag2 =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, decltype(__false_result),
-                                                  decltype(copy_result.second), decltype(__first + true_count)>();
+    __pattern_walk2<_BackendTag, /*_IsSync=*/::std::false_type>(
+        __tag, __par_backend_hetero::make_wrapped_policy<copy_back_wrapper>(__exec), __true_result, copy_result.first,
+        __first, __brick_move<_ExecutionPolicy>{});
+
     __pattern_walk2(
-        __dispatch_tag2,
-        __par_backend_hetero::make_wrapped_policy<copy_back_wrapper2>(::std::forward<_ExecutionPolicy>(__exec)),
+        __tag, __par_backend_hetero::make_wrapped_policy<copy_back_wrapper2>(::std::forward<_ExecutionPolicy>(__exec)),
         __false_result, copy_result.second, __first + true_count, __brick_move<_ExecutionPolicy>{});
 
     return __first + true_count;
@@ -1545,35 +1516,20 @@ __pattern_partial_sort_copy(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&&
 
         auto __buf_first = __buf.get();
 
-        constexpr auto __dispatch_tag1 =
-            oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, decltype(__first), decltype(__last),
-                                                      decltype(__buf_first)>();
-        using __backend_tag1 = typename decltype(__dispatch_tag1)::__backend_tag;
-
-        auto __buf_last = __pattern_walk2<__backend_tag1, /*_IsSync=*/::std::false_type>(
-            __dispatch_tag1, __par_backend_hetero::make_wrapped_policy<__initial_copy_2>(__exec), __first, __last,
-            __buf_first, __brick_copy<_ExecutionPolicy>{});
+        auto __buf_last = __pattern_walk2<_BackendTag, /*_IsSync=*/::std::false_type>(
+            __tag, __par_backend_hetero::make_wrapped_policy<__initial_copy_2>(__exec), __first, __last, __buf_first,
+            __brick_copy<_ExecutionPolicy>{});
 
         auto __buf_mid = __buf_first + __out_size;
 
-        constexpr auto __dispatch_tag11 =
-            oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, decltype(__buf_first), decltype(__buf_mid),
-                                                      decltype(__buf_last)>();
-        using __backend_tag11 = typename decltype(__dispatch_tag11)::__backend_tag;
-
         __par_backend_hetero::__parallel_partial_sort(
-            __backend_tag11{}, __par_backend_hetero::make_wrapped_policy<__partial_sort_2>(__exec),
+            _BackendTag{}, __par_backend_hetero::make_wrapped_policy<__partial_sort_2>(__exec),
             __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read_write>(__buf_first),
             __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read_write>(__buf_mid),
             __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::read_write>(__buf_last), __comp);
 
-        constexpr auto __dispatch_tag2 =
-            oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, decltype(__buf_first), decltype(__buf_mid),
-                                                      decltype(__out_first)>();
-
         return __pattern_walk2(
-            __dispatch_tag2,
-            __par_backend_hetero::make_wrapped_policy<__copy_back>(::std::forward<_ExecutionPolicy>(__exec)),
+            __tag, __par_backend_hetero::make_wrapped_policy<__copy_back>(::std::forward<_ExecutionPolicy>(__exec)),
             __buf_first, __buf_mid, __out_first, __brick_copy<_ExecutionPolicy>{});
     }
 }
@@ -1887,17 +1843,13 @@ __pattern_set_union(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec,
 
     //1. Calc difference {2} \ {1}
     const auto __n_diff =
-        oneapi::dpl::__internal::__pattern_hetero_set_op(
-            oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, decltype(__first2), decltype(__last2),
-                                                      decltype(__first1), decltype(__last1), decltype(__buf)>(),
-            __exec, __first2, __last2, __first1, __last1, __buf, __comp, unseq_backend::_DifferenceTag()) -
+        oneapi::dpl::__internal::__pattern_hetero_set_op(__tag, __exec, __first2, __last2, __first1, __last1, __buf,
+                                                         __comp, unseq_backend::_DifferenceTag()) -
         __buf;
 
     //2. Merge {1} and the difference
-    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, decltype(__first1),
-                                                                          decltype(__buf), decltype(__result)>();
     return oneapi::dpl::__internal::__pattern_merge(
-        __dispatch_tag,
+        __tag,
         oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__set_union_copy_case_2>(
             ::std::forward<_ExecutionPolicy>(__exec)),
         __first1, __last1, __buf, __buf + __n_diff, __result, __comp);
@@ -1974,25 +1926,19 @@ __pattern_set_symmetric_difference(__hetero_tag<_BackendTag> __tag, _ExecutionPo
     //1. Calc difference {1} \ {2}
     const auto __n_diff_1 =
         oneapi::dpl::__internal::__pattern_hetero_set_op(
-            oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, decltype(__first1), decltype(__last1),
-                                                      decltype(__first2), decltype(__last2), decltype(__buf_1)>(),
-            oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__set_symmetric_difference_phase_1>(__exec),
+            __tag, oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__set_symmetric_difference_phase_1>(__exec),
             __first1, __last1, __first2, __last2, __buf_1, __comp, unseq_backend::_DifferenceTag()) -
         __buf_1;
 
     //2. Calc difference {2} \ {1}
     const auto __n_diff_2 =
         oneapi::dpl::__internal::__pattern_hetero_set_op(
-            oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, decltype(__first2), decltype(__last2),
-                                                      decltype(__first1), decltype(__last1), decltype(__buf_2)>(),
-            oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__set_symmetric_difference_phase_2>(__exec),
+            __tag, oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__set_symmetric_difference_phase_2>(__exec),
             __first2, __last2, __first1, __last1, __buf_2, __comp, unseq_backend::_DifferenceTag()) -
         __buf_2;
 
     //3. Merge the differences
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, decltype(__buf_1),
-                                                                              decltype(__buf_2), decltype(__result)>();
-    return oneapi::dpl::__internal::__pattern_merge(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __buf_1,
+    return oneapi::dpl::__internal::__pattern_merge(__tag, ::std::forward<_ExecutionPolicy>(__exec), __buf_1,
                                                     __buf_1 + __n_diff_1, __buf_2, __buf_2 + __n_diff_2, __result,
                                                     __comp);
 }

From 6d04ccbc605ff4f9a79c29cd01e61a7720822376 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Mon, 19 Feb 2024 17:21:28 +0100
Subject: [PATCH 339/566] Fix an error: remove extra __pattern_replace_if and
 it's usage

---
 include/oneapi/dpl/pstl/algorithm_impl.h      | 14 --------------
 include/oneapi/dpl/pstl/glue_algorithm_impl.h |  7 +++++--
 2 files changed, 5 insertions(+), 16 deletions(-)

diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 87415af4904..008c261042b 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -174,20 +174,6 @@ __pattern_walk1(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _ForwardIt
     });
 }
 
-template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _UnaryPredicate, class _Tp>
-void
-__pattern_replace_if(_Tag __tag, _ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last,
-                     _UnaryPredicate __pred, const _Tp& __new_value)
-{
-    static_assert(__is_backend_tag_v<_Tag>);
-
-    oneapi::dpl::__internal::__pattern_walk1(
-        __tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
-        oneapi::dpl::__internal::__replace_functor<
-            oneapi::dpl::__internal::__ref_or_copy<_ExecutionPolicy, const _Tp>,
-            oneapi::dpl::__internal::__ref_or_copy<_ExecutionPolicy, _UnaryPredicate>>(__new_value, __pred));
-}
-
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Brick>
 void
 __pattern_walk_brick(_Tag, _ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last,
diff --git a/include/oneapi/dpl/pstl/glue_algorithm_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
index 495ff06be8c..abb6d6f6c7f 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
@@ -395,8 +395,11 @@ replace_if(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator
 {
     constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
 
-    __pattern_replace_if(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __pred,
-                         __new_value);
+    oneapi::dpl::__internal::__pattern_walk1(
+        __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
+        oneapi::dpl::__internal::__replace_functor<
+            oneapi::dpl::__internal::__ref_or_copy<_ExecutionPolicy, const _Tp>,
+            oneapi::dpl::__internal::__ref_or_copy<_ExecutionPolicy, _UnaryPredicate>>(__new_value, __pred));
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator, class _Tp>

From 70cdfeba361f64bb30965aa7372c224af174ed9b Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 09:35:46 +0100
Subject: [PATCH 340/566] 
 include/oneapi/dpl/pstl/hetero/dpcpp/utils_ranges_sycl.h - implementation of
 __ranges::__select_backend()

---
 include/oneapi/dpl/pstl/hetero/dpcpp/utils_ranges_sycl.h | 9 +++++++++
 1 file changed, 9 insertions(+)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/utils_ranges_sycl.h b/include/oneapi/dpl/pstl/hetero/dpcpp/utils_ranges_sycl.h
index b102ab58123..9e44033a330 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/utils_ranges_sycl.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/utils_ranges_sycl.h
@@ -716,6 +716,15 @@ struct __get_sycl_range
     }
 };
 
+template <class _ExecutionPolicy, typename... _Ranges>
+constexpr oneapi::dpl::__internal::__hetero_tag<
+    typename oneapi::dpl::__internal::__select_backend_for_hetero_policy_trait<
+        ::std::decay_t<_ExecutionPolicy>>::__backend_tag>
+__select_backend()
+{
+    return {}; // return __hetero_tag<__device_backend_tag> or __hetero_tag<__fpga_backend_tag>
+}
+
 } // namespace __ranges
 } // namespace dpl
 } // namespace oneapi

From c299b06aba91854a94207910d3bf2fdd19396d09 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 09:37:05 +0100
Subject: [PATCH 341/566] __ranges::__pattern_transform_reduce + tag impls

---
 .../pstl/hetero/numeric_ranges_impl_hetero.h  | 19 ++++++++++---------
 1 file changed, 10 insertions(+), 9 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/numeric_ranges_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/numeric_ranges_impl_hetero.h
index 42502a4b7b5..4abed65341b 100644
--- a/include/oneapi/dpl/pstl/hetero/numeric_ranges_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/numeric_ranges_impl_hetero.h
@@ -37,11 +37,11 @@ namespace __ranges
 // transform_reduce (version with two binary functions)
 //------------------------------------------------------------------------
 
-template <typename _ExecutionPolicy, typename _Range1, typename _Range2, typename _Tp, typename _BinaryOperation1,
-          typename _BinaryOperation2>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, _Tp>
-__pattern_transform_reduce(_ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&& __rng2, _Tp __init,
-                           _BinaryOperation1 __binary_op1, _BinaryOperation2 __binary_op2)
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Range1, typename _Range2, typename _Tp,
+          typename _BinaryOperation1, typename _BinaryOperation2>
+_Tp
+__pattern_transform_reduce(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&& __rng2,
+                           _Tp __init, _BinaryOperation1 __binary_op1, _BinaryOperation2 __binary_op2)
 {
     if (__rng1.empty())
         return __init;
@@ -62,10 +62,11 @@ __pattern_transform_reduce(_ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&
 // transform_reduce (with unary and binary functions)
 //------------------------------------------------------------------------
 
-template <typename _ExecutionPolicy, typename _Range, typename _Tp, typename _BinaryOperation, typename _UnaryOperation>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, _Tp>
-__pattern_transform_reduce(_ExecutionPolicy&& __exec, _Range&& __rng, _Tp __init, _BinaryOperation __binary_op,
-                           _UnaryOperation __unary_op)
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Range, typename _Tp, typename _BinaryOperation,
+          typename _UnaryOperation>
+_Tp
+__pattern_transform_reduce(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Range&& __rng, _Tp __init,
+                           _BinaryOperation __binary_op, _UnaryOperation __unary_op)
 {
     if (__rng.empty())
         return __init;

From b4f7eab4e84d595987d45e80d3aead0d267f84ae Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 09:37:35 +0100
Subject: [PATCH 342/566] __ranges::__pattern_transform_scan_base + tag impls

---
 .../dpl/pstl/hetero/numeric_ranges_impl_hetero.h      | 11 +++++------
 1 file changed, 5 insertions(+), 6 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/numeric_ranges_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/numeric_ranges_impl_hetero.h
index 4abed65341b..9cfe02afe6e 100644
--- a/include/oneapi/dpl/pstl/hetero/numeric_ranges_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/numeric_ranges_impl_hetero.h
@@ -87,12 +87,11 @@ __pattern_transform_reduce(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec,
 // transform_scan
 //------------------------------------------------------------------------
 
-template <typename _ExecutionPolicy, typename _Range1, typename _Range2, typename _UnaryOperation, typename _InitType,
-          typename _BinaryOperation, typename _Inclusive>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy,
-                                                             oneapi::dpl::__internal::__difference_t<_Range2>>
-__pattern_transform_scan_base(_ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&& __rng2, _UnaryOperation __unary_op,
-                              _InitType __init, _BinaryOperation __binary_op, _Inclusive)
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Range1, typename _Range2, typename _UnaryOperation,
+          typename _InitType, typename _BinaryOperation, typename _Inclusive>
+oneapi::dpl::__internal::__difference_t<_Range2>
+__pattern_transform_scan_base(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&& __rng2,
+                              _UnaryOperation __unary_op, _InitType __init, _BinaryOperation __binary_op, _Inclusive)
 {
     if (__rng1.empty())
         return 0;

From 646c63b2d82eb63601bcd0671acf9e7243f4fef6 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 09:37:52 +0100
Subject: [PATCH 343/566] __ranges::__pattern_transform_scan + tag impls

---
 .../pstl/hetero/numeric_ranges_impl_hetero.h  | 20 +++++++++----------
 1 file changed, 9 insertions(+), 11 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/numeric_ranges_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/numeric_ranges_impl_hetero.h
index 9cfe02afe6e..4223facecfb 100644
--- a/include/oneapi/dpl/pstl/hetero/numeric_ranges_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/numeric_ranges_impl_hetero.h
@@ -125,12 +125,11 @@ __pattern_transform_scan_base(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __ex
     return __rng1_size;
 }
 
-template <typename _ExecutionPolicy, typename _Range1, typename _Range2, typename _UnaryOperation, typename _Type,
-          typename _BinaryOperation, typename _Inclusive>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy,
-                                                             oneapi::dpl::__internal::__difference_t<_Range2>>
-__pattern_transform_scan(_ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&& __rng2, _UnaryOperation __unary_op,
-                         _Type __init, _BinaryOperation __binary_op, _Inclusive)
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Range1, typename _Range2, typename _UnaryOperation,
+          typename _Type, typename _BinaryOperation, typename _Inclusive>
+oneapi::dpl::__internal::__difference_t<_Range2>
+__pattern_transform_scan(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&& __rng2,
+                         _UnaryOperation __unary_op, _Type __init, _BinaryOperation __binary_op, _Inclusive)
 {
     using _RepackedType = __par_backend_hetero::__repacked_tuple_t<_Type>;
     using _InitType = unseq_backend::__init_value<_RepackedType>;
@@ -141,12 +140,11 @@ __pattern_transform_scan(_ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&&
 }
 
 // scan without initial element
-template <typename _ExecutionPolicy, typename _Range1, typename _Range2, typename _UnaryOperation,
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Range1, typename _Range2, typename _UnaryOperation,
           typename _BinaryOperation, typename _Inclusive>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy,
-                                                             oneapi::dpl::__internal::__difference_t<_Range2>>
-__pattern_transform_scan(_ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&& __rng2, _UnaryOperation __unary_op,
-                         _BinaryOperation __binary_op, _Inclusive)
+oneapi::dpl::__internal::__difference_t<_Range2>
+__pattern_transform_scan(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&& __rng2,
+                         _UnaryOperation __unary_op, _BinaryOperation __binary_op, _Inclusive)
 {
     using _Type = oneapi::dpl::__internal::__value_t<_Range1>;
     using _RepackedType = __par_backend_hetero::__repacked_tuple_t<_Type>;

From a41760a2f1baa5eec241550ea123cca2728a0930 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 09:39:09 +0100
Subject: [PATCH 344/566] __ranges::__pattern_walk_n + tag impls

---
 .../oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h   | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
index f227fd29383..e7f743c7172 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
@@ -39,9 +39,9 @@ namespace __ranges
 // walk_n
 //------------------------------------------------------------------------
 
-template <typename _ExecutionPolicy, typename _Function, typename... _Ranges>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy>
-__pattern_walk_n(_ExecutionPolicy&& __exec, _Function __f, _Ranges&&... __rngs)
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Function, typename... _Ranges>
+void
+__pattern_walk_n(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Function __f, _Ranges&&... __rngs)
 {
     auto __n = oneapi::dpl::__ranges::__get_first_range_size(__rngs...);
     if (__n > 0)

From 97ff7b2e5c3d07e2bf1cc53d9c2e16ca2e0bcff4 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 09:39:21 +0100
Subject: [PATCH 345/566] __ranges::__pattern_swap + tag impls

---
 .../oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h   | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
index e7f743c7172..6ae7d5b87f4 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
@@ -68,9 +68,9 @@ class __swap2_wrapper
 {
 };
 
-template <typename _ExecutionPolicy, typename _Range1, typename _Range2, typename _Function>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, bool>
-__pattern_swap(_ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&& __rng2, _Function __f)
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Range1, typename _Range2, typename _Function>
+bool
+__pattern_swap(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&& __rng2, _Function __f)
 {
     if (__rng1.size() <= __rng2.size())
     {

From bfec7cc04b6318ac6b6ebb78eaa73d125186e649 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 09:39:32 +0100
Subject: [PATCH 346/566] __ranges::__pattern_equal + tag impls

---
 .../oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h   | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
index 6ae7d5b87f4..4f08f8f9fc4 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
@@ -92,9 +92,9 @@ __pattern_swap(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Rang
 // equal
 //------------------------------------------------------------------------
 
-template <typename _ExecutionPolicy, typename _Range1, typename _Range2, typename _Pred>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, bool>
-__pattern_equal(_ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&& __rng2, _Pred __pred)
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Range1, typename _Range2, typename _Pred>
+bool
+__pattern_equal(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&& __rng2, _Pred __pred)
 {
     if (__rng1.empty() || __rng2.empty() || __rng1.size() != __rng2.size())
         return false;

From 2313a2bb9729559d2ea0e43826cd50fe927f3e5b Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 09:39:46 +0100
Subject: [PATCH 347/566] __ranges::__pattern_find_if + tag impls

---
 .../oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h  | 7 +++----
 1 file changed, 3 insertions(+), 4 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
index 4f08f8f9fc4..5cf8d4be6aa 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
@@ -114,10 +114,9 @@ __pattern_equal(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Range1&&
 // find_if
 //------------------------------------------------------------------------
 
-template <typename _ExecutionPolicy, typename _Range, typename _Pred>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy,
-                                                             oneapi::dpl::__internal::__difference_t<_Range>>
-__pattern_find_if(_ExecutionPolicy&& __exec, _Range&& __rng, _Pred __pred)
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Range, typename _Pred>
+oneapi::dpl::__internal::__difference_t<_Range>
+__pattern_find_if(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Range&& __rng, _Pred __pred)
 {
     //trivial pre-checks
     if (__rng.empty())

From fbfe04f2206c4f003ba54b57c41a194d236c2fd9 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 09:40:00 +0100
Subject: [PATCH 348/566] __ranges::__pattern_find_end + tag impls

---
 .../oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
index 5cf8d4be6aa..9f31e62c354 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
@@ -136,10 +136,10 @@ __pattern_find_if(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Range&&
 // find_end
 //------------------------------------------------------------------------
 
-template <typename _ExecutionPolicy, typename _Range1, typename _Range2, typename _Pred>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy,
-                                                             oneapi::dpl::__internal::__difference_t<_Range1>>
-__pattern_find_end(_ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&& __rng2, _Pred __pred)
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Range1, typename _Range2, typename _Pred>
+oneapi::dpl::__internal::__difference_t<_Range1>
+__pattern_find_end(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&& __rng2,
+                   _Pred __pred)
 {
     //trivial pre-checks
     if (__rng1.empty() || __rng2.empty() || __rng1.size() < __rng2.size())

From 1844c14c615b7694246bb4610371aea34c808cda Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 09:40:13 +0100
Subject: [PATCH 349/566] __ranges::__pattern_find_first_of + tag impls

---
 .../oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
index 9f31e62c354..9087f68aa4f 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
@@ -166,10 +166,10 @@ __pattern_find_end(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _
 // find_first_of
 //------------------------------------------------------------------------
 
-template <typename _ExecutionPolicy, typename _Range1, typename _Range2, typename _Pred>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy,
-                                                             oneapi::dpl::__internal::__difference_t<_Range1>>
-__pattern_find_first_of(_ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&& __rng2, _Pred __pred)
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Range1, typename _Range2, typename _Pred>
+oneapi::dpl::__internal::__difference_t<_Range1>
+__pattern_find_first_of(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&& __rng2,
+                        _Pred __pred)
 {
     //trivial pre-checks
     if (__rng1.empty() || __rng2.empty())

From aef87590e27bb80e51e77f0c860e925678ddacce Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 09:40:28 +0100
Subject: [PATCH 350/566] __ranges::__pattern_any_of + tag impls

---
 .../oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h   | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
index 9087f68aa4f..16d4e9410af 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
@@ -190,9 +190,9 @@ __pattern_find_first_of(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _R
 // any_of
 //------------------------------------------------------------------------
 
-template <typename _ExecutionPolicy, typename _Range, typename _Pred>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy, bool>
-__pattern_any_of(_ExecutionPolicy&& __exec, _Range&& __rng, _Pred __pred)
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Range, typename _Pred>
+bool
+__pattern_any_of(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Range&& __rng, _Pred __pred)
 {
     if (__rng.empty())
         return false;

From b3fb59ff5a32e3a1e656926839430a233ea70db5 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 09:40:41 +0100
Subject: [PATCH 351/566] __ranges::__pattern_search + tag impls

---
 .../oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
index 16d4e9410af..f502d78a775 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
@@ -214,10 +214,10 @@ class equal_wrapper
 {
 };
 
-template <typename _ExecutionPolicy, typename _Range1, typename _Range2, typename _Pred>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy,
-                                                             oneapi::dpl::__internal::__difference_t<_Range1>>
-__pattern_search(_ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&& __rng2, _Pred __pred)
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Range1, typename _Range2, typename _Pred>
+oneapi::dpl::__internal::__difference_t<_Range1>
+__pattern_search(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&& __rng2,
+                 _Pred __pred)
 {
     //trivial pre-checks
     if (__rng2.empty())

From b49ec8a19b2511581c1a18201dc098f3d28bddbb Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 09:40:55 +0100
Subject: [PATCH 352/566] __ranges::__pattern_search_n + tag impls

---
 .../oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
index f502d78a775..0919ecd644a 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
@@ -247,10 +247,10 @@ __pattern_search(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Ra
 // search_n
 //------------------------------------------------------------------------
 
-template <typename _ExecutionPolicy, typename _Range, typename _Size, typename _Tp, typename _BinaryPredicate>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy,
-                                                             oneapi::dpl::__internal::__difference_t<_Range>>
-__pattern_search_n(_ExecutionPolicy&& __exec, _Range&& __rng, _Size __count, const _Tp& __value,
+template <template <typename _BackendTag, typename _ExecutionPolicy, typename _Range, typename _Size, typename _Tp,
+                    typename _BinaryPredicate>
+oneapi::dpl::__internal::__difference_t<_Range>
+__pattern_search_n(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Range&& __rng, _Size __count, const _Tp& __value,
                    _BinaryPredicate __pred)
 {
     //TODO: To consider definition a kind of special factory "multiple_view" (addition to standard "single_view").

From 1fb9702052b7cc39fccbcc9d5511edfe86e40695 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 09:41:07 +0100
Subject: [PATCH 353/566] __ranges::__pattern_adjacent_find + tag impls

---
 .../dpl/pstl/hetero/algorithm_ranges_impl_hetero.h     | 10 +++++-----
 1 file changed, 5 insertions(+), 5 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
index 0919ecd644a..d3f4138efaa 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
@@ -279,11 +279,11 @@ return_value(_Size __res, _Size __size, ::std::false_type)
 // adjacent_find
 //------------------------------------------------------------------------
 
-template <typename _ExecutionPolicy, typename _Range, typename _BinaryPredicate, typename _OrFirstTag>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy,
-                                                             oneapi::dpl::__internal::__difference_t<_Range>>
-__pattern_adjacent_find(_ExecutionPolicy&& __exec, _Range&& __rng, _BinaryPredicate __predicate,
-                        _OrFirstTag __is__or_semantic)
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Range, typename _BinaryPredicate,
+          typename _OrFirstTag>
+oneapi::dpl::__internal::__difference_t<_Range>
+__pattern_adjacent_find(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Range&& __rng,
+                        _BinaryPredicate __predicate, _OrFirstTag __is__or_semantic)
 {
     if (__rng.size() < 2)
         return __rng.size();

From 2e72a4bc8102cbe817656eab541ff5d38e3c5601 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 09:41:23 +0100
Subject: [PATCH 354/566] __ranges::__pattern_count + tag impls

---
 .../oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h  | 7 +++----
 1 file changed, 3 insertions(+), 4 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
index d3f4138efaa..33163ad400b 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
@@ -308,10 +308,9 @@ __pattern_adjacent_find(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _R
     return return_value(result, __rng.size(), __is__or_semantic);
 }
 
-template <typename _ExecutionPolicy, typename _Range, typename _Predicate>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy,
-                                                             oneapi::dpl::__internal::__difference_t<_Range>>
-__pattern_count(_ExecutionPolicy&& __exec, _Range&& __rng, _Predicate __predicate)
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Range, typename _Predicate>
+oneapi::dpl::__internal::__difference_t<_Range>
+__pattern_count(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Range&& __rng, _Predicate __predicate)
 {
     if (__rng.size() == 0)
         return 0;

From 812baf8bad91780d131942b4fb7ed0d7810828e1 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 09:41:38 +0100
Subject: [PATCH 355/566] __ranges::__pattern_scan_copy + tag impls

---
 .../dpl/pstl/hetero/algorithm_ranges_impl_hetero.h     | 10 +++++-----
 1 file changed, 5 insertions(+), 5 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
index 33163ad400b..f67a9bfbfe9 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
@@ -337,11 +337,11 @@ __pattern_count(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Range&& _
 // copy_if
 //------------------------------------------------------------------------
 
-template <typename _ExecutionPolicy, typename _Range1, typename _Range2, typename _CreateMaskOp, typename _CopyByMaskOp>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy,
-                                                             oneapi::dpl::__internal::__difference_t<_Range1>>
-__pattern_scan_copy(_ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&& __rng2, _CreateMaskOp __create_mask_op,
-                    _CopyByMaskOp __copy_by_mask_op)
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Range1, typename _Range2, typename _CreateMaskOp,
+          typename _CopyByMaskOp>
+oneapi::dpl::__internal::__difference_t<_Range1>
+__pattern_scan_copy(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&& __rng2,
+                    _CreateMaskOp __create_mask_op, _CopyByMaskOp __copy_by_mask_op)
 {
     if (__rng1.size() == 0)
         return __rng1.size();

From d2bfb4c21caf432ea47a5c61b50891ebe72c9e5e Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 09:41:57 +0100
Subject: [PATCH 356/566] __ranges::__pattern_copy_if + tag impls

---
 .../oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
index f67a9bfbfe9..51407f3e2e6 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
@@ -384,11 +384,11 @@ __pattern_scan_copy(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Range
     return __res;
 }
 
-template <typename _ExecutionPolicy, typename _Range1, typename _Range2, typename _Predicate,
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Range1, typename _Range2, typename _Predicate,
           typename _Assign = oneapi::dpl::__internal::__pstl_assign>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy,
-                                                             oneapi::dpl::__internal::__difference_t<_Range2>>
-__pattern_copy_if(_ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&& __rng2, _Predicate __pred, _Assign)
+oneapi::dpl::__internal::__difference_t<_Range2>
+__pattern_copy_if(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&& __rng2,
+                  _Predicate __pred, _Assign)
 {
     using _SizeType = decltype(__rng1.size());
     using _ReduceOp = ::std::plus<_SizeType>;

From 02ea7a2bb2f6dbecb1f271856142361a3235c5b8 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 09:42:14 +0100
Subject: [PATCH 357/566] __ranges::__pattern_remove_if + tag impls

---
 .../oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h  | 7 +++----
 1 file changed, 3 insertions(+), 4 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
index 51407f3e2e6..8baf84e7506 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
@@ -404,10 +404,9 @@ __pattern_copy_if(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _R
 // remove_if
 //------------------------------------------------------------------------
 
-template <typename _ExecutionPolicy, typename _Range, typename _Predicate>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy,
-                                                             oneapi::dpl::__internal::__difference_t<_Range>>
-__pattern_remove_if(_ExecutionPolicy&& __exec, _Range&& __rng, _Predicate __pred)
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Range, typename _Predicate>
+oneapi::dpl::__internal::__difference_t<_Range>
+__pattern_remove_if(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Range&& __rng, _Predicate __pred)
 {
     if (__rng.size() == 0)
         return __rng.size();

From 4a118ba5f38f9462ab6d6d73a9ede9d89003cc7f Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 09:42:32 +0100
Subject: [PATCH 358/566] __ranges::__pattern_unique_copy + tag impls

---
 .../dpl/pstl/hetero/algorithm_ranges_impl_hetero.h     | 10 +++++-----
 1 file changed, 5 insertions(+), 5 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
index 8baf84e7506..5384b761aa7 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
@@ -431,11 +431,11 @@ __pattern_remove_if(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec,
 // unique_copy
 //------------------------------------------------------------------------
 
-template <typename _ExecutionPolicy, typename _Range1, typename _Range2, typename _BinaryPredicate,
-          typename _Assign = oneapi::dpl::__internal::__pstl_assign>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy,
-                                                             oneapi::dpl::__internal::__difference_t<_Range2>>
-__pattern_unique_copy(_ExecutionPolicy&& __exec, _Range1&& __rng, _Range2&& __result, _BinaryPredicate __pred, _Assign)
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Range1, typename _Range2,
+          typename _BinaryPredicate, typename _Assign = oneapi::dpl::__internal::__pstl_assign>
+oneapi::dpl::__internal::__difference_t<_Range2>
+__pattern_unique_copy(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Range1&& __rng, _Range2&& __result,
+                      _BinaryPredicate __pred, _Assign)
 {
     using _It1DifferenceType = oneapi::dpl::__internal::__difference_t<_Range1>;
     unseq_backend::__copy_by_mask<::std::plus<_It1DifferenceType>, _Assign, /*inclusive*/ ::std::true_type, 1>

From 2094d0b2ae905a67ede674c2e6241155f9ddfda9 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 09:42:46 +0100
Subject: [PATCH 359/566] __ranges::__pattern_unique + tag impls

---
 .../oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h  | 7 +++----
 1 file changed, 3 insertions(+), 4 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
index 5384b761aa7..85a244a935f 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
@@ -451,10 +451,9 @@ __pattern_unique_copy(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec
 // unique
 //------------------------------------------------------------------------
 
-template <typename _ExecutionPolicy, typename _Range, typename _BinaryPredicate>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy,
-                                                             oneapi::dpl::__internal::__difference_t<_Range>>
-__pattern_unique(_ExecutionPolicy&& __exec, _Range&& __rng, _BinaryPredicate __pred)
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Range, typename _BinaryPredicate>
+oneapi::dpl::__internal::__difference_t<_Range>
+__pattern_unique(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Range&& __rng, _BinaryPredicate __pred)
 {
     if (__rng.size() == 0)
         return __rng.size();

From c667ecd9e4fd9c2c0c35d8ff5e93846d45680b4e Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 09:42:59 +0100
Subject: [PATCH 360/566] __ranges::__pattern_merge + tag impls

---
 .../dpl/pstl/hetero/algorithm_ranges_impl_hetero.h       | 9 +++++----
 1 file changed, 5 insertions(+), 4 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
index 85a244a935f..78ce483fb94 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
@@ -483,10 +483,11 @@ class __copy2_wrapper
 {
 };
 
-template <typename _ExecutionPolicy, typename _Range1, typename _Range2, typename _Range3, typename _Compare>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy,
-                                                             oneapi::dpl::__internal::__difference_t<_Range3>>
-__pattern_merge(_ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&& __rng2, _Range3&& __rng3, _Compare __comp)
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Range1, typename _Range2, typename _Range3,
+          typename _Compare>
+oneapi::dpl::__internal::__difference_t<_Range3>
+__pattern_merge(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&& __rng2,
+                _Range3&& __rng3, _Compare __comp)
 {
     auto __n1 = __rng1.size();
     auto __n2 = __rng2.size();

From 6980f0e6d4e3735f0dcfbb96dc26ab808c2fd92e Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 09:43:29 +0100
Subject: [PATCH 361/566] __ranges::__pattern_sort + tag impls

---
 .../oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h   | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
index 78ce483fb94..6a28faac3dc 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
@@ -528,9 +528,9 @@ __pattern_merge(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Ran
 // sort
 //------------------------------------------------------------------------
 
-template <typename _ExecutionPolicy, typename _Range, typename _Compare, typename _Proj>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy>
-__pattern_sort(_ExecutionPolicy&& __exec, _Range&& __rng, _Compare __comp, _Proj __proj)
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Range, typename _Compare, typename _Proj>
+void
+__pattern_sort(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Range&& __rng, _Compare __comp, _Proj __proj)
 {
     if (__rng.size() >= 2)
         __par_backend_hetero::__parallel_stable_sort(

From b7a51ff8424c952a515b38604e2e2eacbaa21023 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 09:43:58 +0100
Subject: [PATCH 362/566] __ranges::__pattern_minmax_element + tag impls

---
 .../oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h | 8 +++-----
 1 file changed, 3 insertions(+), 5 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
index 6a28faac3dc..f5ba605fb25 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
@@ -585,11 +585,9 @@ __pattern_min_element(_ExecutionPolicy&& __exec, _Range&& __rng, _Compare __comp
 // minmax_element
 //------------------------------------------------------------------------
 
-template <typename _ExecutionPolicy, typename _Range, typename _Compare>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<
-    _ExecutionPolicy,
-    ::std::pair<oneapi::dpl::__internal::__difference_t<_Range>, oneapi::dpl::__internal::__difference_t<_Range>>>
-__pattern_minmax_element(_ExecutionPolicy&& __exec, _Range&& __rng, _Compare __comp)
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Range, typename _Compare>
+::std::pair<oneapi::dpl::__internal::__difference_t<_Range>, oneapi::dpl::__internal::__difference_t<_Range>>
+__pattern_minmax_element(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Range&& __rng, _Compare __comp)
 {
     //If size == 1, result is the zero-indexed element. If size == 0, result is 0.
     if (__rng.size() < 2)

From f3c553c5252c216ab8fc1dab7216ce2951d47b4a Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 09:44:18 +0100
Subject: [PATCH 363/566] __ranges::__pattern_min_element + tag impls

---
 .../oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h  | 7 +++----
 1 file changed, 3 insertions(+), 4 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
index f5ba605fb25..da6752d6063 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
@@ -543,10 +543,9 @@ __pattern_sort(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Range&& __
 // min_element
 //------------------------------------------------------------------------
 
-template <typename _ExecutionPolicy, typename _Range, typename _Compare>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy,
-                                                             oneapi::dpl::__internal::__difference_t<_Range>>
-__pattern_min_element(_ExecutionPolicy&& __exec, _Range&& __rng, _Compare __comp)
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Range, typename _Compare>
+oneapi::dpl::__internal::__difference_t<_Range>
+__pattern_min_element(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Range&& __rng, _Compare __comp)
 {
     //If size == 1, result is the zero-indexed element. If size == 0, result is 0.
     if (__rng.size() < 2)

From 9de0038633077033e042c78c91903c57f76b4ea9 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 09:44:53 +0100
Subject: [PATCH 364/566] __ranges::__pattern_reduce_by_segment + tag impls

---
 .../dpl/pstl/hetero/algorithm_ranges_impl_hetero.h   | 12 ++++++------
 1 file changed, 6 insertions(+), 6 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
index da6752d6063..ba20e54b5ab 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
@@ -669,12 +669,12 @@ class __assign_key2_wrapper
 {
 };
 
-template <typename _ExecutionPolicy, typename _Range1, typename _Range2, typename _Range3, typename _Range4,
-          typename _BinaryPredicate, typename _BinaryOperator>
-oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy,
-                                                             oneapi::dpl::__internal::__difference_t<_Range3>>
-__pattern_reduce_by_segment(_ExecutionPolicy&& __exec, _Range1&& __keys, _Range2&& __values, _Range3&& __out_keys,
-                            _Range4&& __out_values, _BinaryPredicate __binary_pred, _BinaryOperator __binary_op)
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Range1, typename _Range2, typename _Range3,
+          typename _Range4, typename _BinaryPredicate, typename _BinaryOperator>
+oneapi::dpl::__internal::__difference_t<_Range3>
+__pattern_reduce_by_segment(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Range1&& __keys, _Range2&& __values,
+                            _Range3&& __out_keys, _Range4&& __out_values, _BinaryPredicate __binary_pred,
+                            _BinaryOperator __binary_op)
 {
     // The algorithm reduces values in __values where the
     // associated keys for the values are equal to the adjacent key.

From 1b889f6526a55c7d35bb1f86f743758e974569f1 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 09:45:41 +0100
Subject: [PATCH 365/566] __ranges::__pattern_any_of + tag calls

---
 include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
index bbe5cc72ac8..8391a0795cc 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
@@ -39,7 +39,9 @@ template <typename _ExecutionPolicy, typename _Range, typename _Predicate>
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, bool>
 any_of(_ExecutionPolicy&& __exec, _Range&& __rng, _Predicate __pred)
 {
-    return oneapi::dpl::__internal::__ranges::__pattern_any_of(::std::forward<_ExecutionPolicy>(__exec),
+    constexpr auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range>();
+
+    return oneapi::dpl::__internal::__ranges::__pattern_any_of(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
                                                                views::all_read(::std::forward<_Range>(__rng)), __pred);
 }
 

From 70efe25e3032d62ca6d7f95e1b6cc77572951ad3 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 09:45:50 +0100
Subject: [PATCH 366/566] __ranges::__pattern_walk_n + tag calls

---
 .../dpl/pstl/glue_algorithm_ranges_impl.h     | 34 ++++++++++++++-----
 .../hetero/algorithm_ranges_impl_hetero.h     | 13 +++++--
 2 files changed, 37 insertions(+), 10 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
index 8391a0795cc..cede9b70b8d 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
@@ -72,7 +72,9 @@ template <typename _ExecutionPolicy, typename _Range, typename _Function>
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy>
 for_each(_ExecutionPolicy&& __exec, _Range&& __rng, _Function __f)
 {
-    oneapi::dpl::__internal::__ranges::__pattern_walk_n(::std::forward<_ExecutionPolicy>(__exec), __f,
+    constexpr auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range>();
+
+    oneapi::dpl::__internal::__ranges::__pattern_walk_n(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __f,
                                                         views::all(::std::forward<_Range>(__rng)));
 }
 
@@ -232,9 +234,12 @@ template <typename _ExecutionPolicy, typename _Range1, typename _Range2>
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy>
 copy(_ExecutionPolicy&& __exec, _Range1&& __rng, _Range2&& __result)
 {
-    oneapi::dpl::__internal::__ranges::__pattern_walk_n(
-        ::std::forward<_ExecutionPolicy>(__exec), oneapi::dpl::__internal::__brick_copy<_ExecutionPolicy>{},
-        views::all_read(::std::forward<_Range1>(__rng)), views::all_write(::std::forward<_Range2>(__result)));
+    constexpr auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range1, _Range2>();
+
+    oneapi::dpl::__internal::__ranges::__pattern_walk_n(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
+                                                        oneapi::dpl::__internal::__brick_copy<_ExecutionPolicy>{},
+                                                        views::all_read(::std::forward<_Range1>(__rng)),
+                                                        views::all_write(::std::forward<_Range2>(__result)));
 }
 
 template <typename _ExecutionPolicy, typename _Range1, typename _Range2, typename _Predicate>
@@ -271,7 +276,10 @@ template <typename _ExecutionPolicy, typename _Range1, typename _Range2, typenam
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy>
 transform(_ExecutionPolicy&& __exec, _Range1&& __rng, _Range2&& __result, _UnaryOperation __op)
 {
+    constexpr auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range1, _Range2>();
+
     oneapi::dpl::__internal::__ranges::__pattern_walk_n(
+        __dispatch_tag,
         ::std::forward<_ExecutionPolicy>(__exec), [__op](auto x, auto& z) { z = __op(x); },
         views::all_read(::std::forward<_Range1>(__rng)), views::all_write(::std::forward<_Range2>(__result)));
 }
@@ -280,8 +288,11 @@ template <typename _ExecutionPolicy, typename _Range1, typename _Range2, typenam
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy>
 transform(_ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&& __rng2, _Range3&& __result, _BinaryOperation __op)
 {
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range1, _Range2, _Range3>();
+
     oneapi::dpl::__internal::__ranges::__pattern_walk_n(
-        ::std::forward<_ExecutionPolicy>(__exec), [__op](auto x, auto y, auto& z) { z = __op(x, y); },
+        __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), [__op](auto x, auto y, auto& z) { z = __op(x, y); },
         views::all_read(::std::forward<_Range1>(__rng1)), views::all_read(::std::forward<_Range2>(__rng2)),
         views::all_write(::std::forward<_Range3>(__result)));
 }
@@ -408,8 +419,10 @@ template <typename _ExecutionPolicy, typename _Range, typename _UnaryPredicate,
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy>
 replace_if(_ExecutionPolicy&& __exec, _Range&& __rng, _UnaryPredicate __pred, const _Tp& __new_value)
 {
+    constexpr auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range>();
+
     oneapi::dpl::__internal::__ranges::__pattern_walk_n(
-        ::std::forward<_ExecutionPolicy>(__exec),
+        __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
         oneapi::dpl::__internal::__replace_functor<
             oneapi::dpl::__internal::__ref_or_copy<_ExecutionPolicy, const _Tp>,
             oneapi::dpl::__internal::__ref_or_copy<_ExecutionPolicy, _UnaryPredicate>>(__new_value, __pred),
@@ -433,9 +446,11 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy,
 replace_copy_if(_ExecutionPolicy&& __exec, _Range1&& __rng, _Range2&& __result, _UnaryPredicate __pred,
                 const _Tp& __new_value)
 {
+    constexpr auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range1, _Range2>();
+
     auto __src = views::all_read(::std::forward<_Range1>(__rng));
     oneapi::dpl::__internal::__ranges::__pattern_walk_n(
-        ::std::forward<_ExecutionPolicy>(__exec),
+        __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
         oneapi::dpl::__internal::__replace_copy_functor<
             oneapi::dpl::__internal::__ref_or_copy<_ExecutionPolicy, const _Tp>,
             ::std::conditional_t<oneapi::dpl::__internal::__is_const_callable_object_v<_UnaryPredicate>,
@@ -560,10 +575,13 @@ template <typename _ExecutionPolicy, typename _Range1, typename _Range2>
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy>
 move(_ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&& __rng2)
 {
+    constexpr auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range1, _Range2>();
+
     using _DecayedExecutionPolicy = ::std::decay_t<_ExecutionPolicy>;
 
     oneapi::dpl::__internal::__ranges::__pattern_walk_n(
-        ::std::forward<_ExecutionPolicy>(__exec), oneapi::dpl::__internal::__brick_move<_DecayedExecutionPolicy>{},
+        __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
+        oneapi::dpl::__internal::__brick_move<_DecayedExecutionPolicy>{},
         views::all_read(::std::forward<_Range1>(__rng1)), views::all_write(::std::forward<_Range2>(__rng2)));
 }
 
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
index ba20e54b5ab..12af66fb457 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
@@ -75,6 +75,7 @@ __pattern_swap(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Rang
     if (__rng1.size() <= __rng2.size())
     {
         oneapi::dpl::__internal::__ranges::__pattern_walk_n(
+            __tag,
             oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__swap1_wrapper>(
                 ::std::forward<_ExecutionPolicy>(__exec)),
             __f, __rng1, __rng2);
@@ -82,6 +83,7 @@ __pattern_swap(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Rang
     }
 
     oneapi::dpl::__internal::__ranges::__pattern_walk_n(
+        __tag,
         oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__swap2_wrapper>(
             ::std::forward<_ExecutionPolicy>(__exec)),
         __f, __rng2, __rng1);
@@ -420,7 +422,7 @@ __pattern_remove_if(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec,
                                             oneapi::dpl::__internal::__pstl_assign());
     auto __copy_rng_truncated = __copy_rng | oneapi::dpl::experimental::ranges::views::take(__copy_last_id);
 
-    oneapi::dpl::__internal::__ranges::__pattern_walk_n(::std::forward<_ExecutionPolicy>(__exec),
+    oneapi::dpl::__internal::__ranges::__pattern_walk_n(__tag, ::std::forward<_ExecutionPolicy>(__exec),
                                                         oneapi::dpl::__internal::__brick_copy<_ExecutionPolicy>{},
                                                         __copy_rng_truncated, ::std::forward<_Range>(__rng));
 
@@ -464,7 +466,7 @@ __pattern_unique(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Ra
     auto res_rng = oneapi::dpl::__ranges::views::all(__buf.get_buffer());
     auto res = __pattern_unique_copy(__exec, __rng, res_rng, __pred, oneapi::dpl::__internal::__pstl_assign());
 
-    __pattern_walk_n(::std::forward<_ExecutionPolicy>(__exec), __brick_copy<_ExecutionPolicy>{}, res_rng,
+    __pattern_walk_n(__tag, ::std::forward<_ExecutionPolicy>(__exec), __brick_copy<_ExecutionPolicy>{}, res_rng,
                      ::std::forward<_Range>(__rng));
     return res;
 }
@@ -499,6 +501,7 @@ __pattern_merge(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Ran
     if (__n1 == 0)
     {
         oneapi::dpl::__internal::__ranges::__pattern_walk_n(
+            __tag,
             oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__copy1_wrapper>(
                 ::std::forward<_ExecutionPolicy>(__exec)),
             oneapi::dpl::__internal::__brick_copy<_ExecutionPolicy>{}, ::std::forward<_Range2>(__rng2),
@@ -507,6 +510,7 @@ __pattern_merge(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Ran
     else if (__n2 == 0)
     {
         oneapi::dpl::__internal::__ranges::__pattern_walk_n(
+            __tag,
             oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__copy2_wrapper>(
                 ::std::forward<_ExecutionPolicy>(__exec)),
             oneapi::dpl::__internal::__brick_copy<_ExecutionPolicy>{}, ::std::forward<_Range1>(__rng1),
@@ -685,6 +689,9 @@ __pattern_reduce_by_segment(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&&
     //          __out_keys   = { 1, 2, 3, 4, 1, 3, 1, 3, 0 }
     //          __out_values = { 1, 2, 3, 4, 2, 6, 2, 6, 0 }
 
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range1, _Range2, _Range3>();
+
     const auto __n = __keys.size();
 
     if (__n == 0)
@@ -695,10 +702,12 @@ __pattern_reduce_by_segment(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&&
         __brick_copy<_ExecutionPolicy> __copy_range{};
 
         oneapi::dpl::__internal::__ranges::__pattern_walk_n(
+            __tag,
             oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__copy_keys_wrapper>(__exec),
             __copy_range, ::std::forward<_Range1>(__keys), ::std::forward<_Range3>(__out_keys));
 
         oneapi::dpl::__internal::__ranges::__pattern_walk_n(
+            __tag,
             oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__copy_values_wrapper>
                 (::std::forward<_ExecutionPolicy>(__exec)),
             __copy_range, ::std::forward<_Range2>(__values), ::std::forward<_Range4>(__out_values));

From 0e3738ec17f5438db6b778b184cfc5986cebebca Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 09:46:00 +0100
Subject: [PATCH 367/566] __ranges::__pattern_find_if + tag calls

---
 include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
index cede9b70b8d..ffa591b7aaf 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
@@ -84,7 +84,9 @@ template <typename _ExecutionPolicy, typename _Range, typename _Predicate>
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, oneapi::dpl::__internal::__difference_t<_Range>>
 find_if(_ExecutionPolicy&& __exec, _Range&& __rng, _Predicate __pred)
 {
-    return oneapi::dpl::__internal::__ranges::__pattern_find_if(::std::forward<_ExecutionPolicy>(__exec),
+    constexpr auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range>();
+
+    return oneapi::dpl::__internal::__ranges::__pattern_find_if(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
                                                                 views::all_read(::std::forward<_Range>(__rng)), __pred);
 }
 

From 543cff4d682845f8f8e30c52b0626dbdb7818f33 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 09:46:12 +0100
Subject: [PATCH 368/566] __ranges::__pattern_find_end + tag calls

---
 include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
index ffa591b7aaf..fdbab090fee 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
@@ -117,8 +117,10 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy,
                                                       oneapi::dpl::__internal::__difference_t<_Range1>>
 find_end(_ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&& __rng2, _BinaryPredicate __pred)
 {
+    constexpr auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range1, _Range2>();
+
     return oneapi::dpl::__internal::__ranges::__pattern_find_end(
-        ::std::forward<_ExecutionPolicy>(__exec), views::all_read(::std::forward<_Range1>(__rng1)),
+        __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), views::all_read(::std::forward<_Range1>(__rng1)),
         views::all_read(::std::forward<_Range2>(__rng2)), __pred);
 }
 

From 9f081e5aa4b9dac19110c9612bc9619e47e59aa2 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 09:46:24 +0100
Subject: [PATCH 369/566] __ranges::__pattern_find_first_of + tag calls

---
 include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
index fdbab090fee..d2b746a2b7c 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
@@ -140,8 +140,10 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy,
                                                       oneapi::dpl::__internal::__difference_t<_Range1>>
 find_first_of(_ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&& __rng2, _BinaryPredicate __pred)
 {
+    constexpr auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range1, _Range2>();
+
     return oneapi::dpl::__internal::__ranges::__pattern_find_first_of(
-        ::std::forward<_ExecutionPolicy>(__exec), views::all_read(::std::forward<_Range1>(__rng1)),
+        __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), views::all_read(::std::forward<_Range1>(__rng1)),
         views::all_read(::std::forward<_Range2>(__rng2)), __pred);
 }
 

From ae9984eb522d9264d5a13aa5f2d1d4d7b81e75a4 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 09:46:33 +0100
Subject: [PATCH 370/566] __ranges::__pattern_adjacent_find + tag calls

---
 .../oneapi/dpl/pstl/glue_algorithm_ranges_impl.h | 16 +++++++++++-----
 1 file changed, 11 insertions(+), 5 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
index d2b746a2b7c..4891a40a49e 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
@@ -162,9 +162,11 @@ template <typename _ExecutionPolicy, typename _Range, typename _BinaryPredicate>
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, oneapi::dpl::__internal::__difference_t<_Range>>
 adjacent_find(_ExecutionPolicy&& __exec, _Range&& __rng, _BinaryPredicate __pred)
 {
+    constexpr auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range>();
+
     return oneapi::dpl::__internal::__ranges::__pattern_adjacent_find(
-        ::std::forward<_ExecutionPolicy>(__exec), views::all_read(::std::forward<_Range>(__rng)), __pred,
-        oneapi::dpl::__internal::__first_semantic());
+        __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), views::all_read(::std::forward<_Range>(__rng)),
+        __pred, oneapi::dpl::__internal::__first_semantic());
 }
 
 template <typename _ExecutionPolicy, typename _Range>
@@ -521,10 +523,12 @@ template <typename _ExecutionPolicy, typename _Range, typename _Compare>
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, oneapi::dpl::__internal::__difference_t<_Range>>
 is_sorted_until(_ExecutionPolicy&& __exec, _Range&& __rng, _Compare __comp)
 {
+    constexpr auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range>();
+
     auto __view = views::all_read(::std::forward<_Range>(__rng));
     const auto __res = oneapi::dpl::__internal::__ranges::__pattern_adjacent_find(
-        ::std::forward<_ExecutionPolicy>(__exec), __view, oneapi::dpl::__internal::__reorder_pred<_Compare>(__comp),
-        oneapi::dpl::__internal::__first_semantic());
+        __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __view,
+        oneapi::dpl::__internal::__reorder_pred<_Compare>(__comp), oneapi::dpl::__internal::__first_semantic());
 
     return __res == __view.size() ? __res : __res + 1;
 }
@@ -541,9 +545,11 @@ template <typename _ExecutionPolicy, typename _Range, typename _Compare>
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, bool>
 is_sorted(_ExecutionPolicy&& __exec, _Range&& __rng, _Compare __comp)
 {
+    constexpr auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range>();
+
     auto __view = views::all_read(::std::forward<_Range>(__rng));
     return oneapi::dpl::__internal::__ranges::__pattern_adjacent_find(
-               ::std::forward<_ExecutionPolicy>(__exec), __view,
+               __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __view,
                oneapi::dpl::__internal::__reorder_pred<_Compare>(__comp),
                oneapi::dpl::__internal::__or_semantic()) == __view.size();
 }

From 2937bbb3e4d10bd1b346c27244e3149a54ae47da Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 09:46:43 +0100
Subject: [PATCH 371/566] __ranges::__pattern_count + tag calls

---
 include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
index 4891a40a49e..afa2ca6c574 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
@@ -184,7 +184,9 @@ template <typename _ExecutionPolicy, typename _Range, typename _Predicate>
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, oneapi::dpl::__internal::__difference_t<_Range>>
 count_if(_ExecutionPolicy&& __exec, _Range&& __rng, _Predicate __pred)
 {
-    return oneapi::dpl::__internal::__ranges::__pattern_count(::std::forward<_ExecutionPolicy>(__exec),
+    constexpr auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range>();
+
+    return oneapi::dpl::__internal::__ranges::__pattern_count(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
                                                               views::all_read(::std::forward<_Range>(__rng)), __pred);
 }
 

From 02a26162e6a15894625fe59d62cc2b0bd3a82591 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 09:46:52 +0100
Subject: [PATCH 372/566] __ranges::__pattern_search + tag calls

---
 include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h          | 4 +++-
 include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h | 3 ++-
 2 files changed, 5 insertions(+), 2 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
index afa2ca6c574..7116a7c84ad 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
@@ -207,8 +207,10 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy,
                                                       oneapi::dpl::__internal::__difference_t<_Range1>>
 search(_ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&& __rng2, _BinaryPredicate __pred)
 {
+    constexpr auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range1, _Range2>();
+
     return oneapi::dpl::__internal::__ranges::__pattern_search(
-        ::std::forward<_ExecutionPolicy>(__exec), views::all_read(::std::forward<_Range1>(__rng1)),
+        __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), views::all_read(::std::forward<_Range1>(__rng1)),
         views::all_read(::std::forward<_Range2>(__rng2)), __pred);
 }
 
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
index 12af66fb457..73719dee464 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
@@ -260,7 +260,8 @@ __pattern_search_n(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _
     auto __s_rng = oneapi::dpl::experimental::ranges::views::iota(0, __count) |
                    oneapi::dpl::experimental::ranges::views::transform([__value](auto) { return __value; });
 
-    return __pattern_search(::std::forward<_ExecutionPolicy>(__exec), ::std::forward<_Range>(__rng), __s_rng, __pred);
+    return __pattern_search(__tag, ::std::forward<_ExecutionPolicy>(__exec), ::std::forward<_Range>(__rng), __s_rng,
+                            __pred);
 }
 
 template <typename _Size>

From 38f1047908fdf00a7dda724a7c7970bc667054cd Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 09:46:59 +0100
Subject: [PATCH 373/566] __ranges::__pattern_search_n + tag calls

---
 include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h | 8 +++++---
 1 file changed, 5 insertions(+), 3 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
index 7116a7c84ad..77350bdfc7c 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
@@ -227,9 +227,11 @@ template <typename _ExecutionPolicy, typename _Range, typename _Size, typename _
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, oneapi::dpl::__internal::__difference_t<_Range>>
 search_n(_ExecutionPolicy&& __exec, _Range&& __rng, _Size __count, const _Tp& __value, _BinaryPredicate __pred)
 {
-    return oneapi::dpl::__internal::__ranges::__pattern_search_n(::std::forward<_ExecutionPolicy>(__exec),
-                                                                 views::all_read(::std::forward<_Range>(__rng)),
-                                                                 __count, __value, __pred);
+    constexpr auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range>();
+
+    return oneapi::dpl::__internal::__ranges::__pattern_search_n(
+        __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), views::all_read(::std::forward<_Range>(__rng)),
+        __count, __value, __pred);
 }
 
 template <typename _ExecutionPolicy, typename _Range, typename _Size, typename _Tp>

From 0c3343b43f7997fd68f67d0e5a186f92936d4f50 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 09:47:30 +0100
Subject: [PATCH 374/566] __ranges::__pattern_copy_if + tag calls

---
 include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h      | 4 +++-
 .../oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h | 8 +++++---
 2 files changed, 8 insertions(+), 4 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
index 77350bdfc7c..dd24f574b81 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
@@ -261,8 +261,10 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy,
                                                       oneapi::dpl::__internal::__difference_t<_Range2>>
 copy_if(_ExecutionPolicy&& __exec, _Range1&& __rng, _Range2&& __result, _Predicate __pred)
 {
+    constexpr auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range1, _Range2>();
+
     return oneapi::dpl::__internal::__ranges::__pattern_copy_if(
-        ::std::forward<_ExecutionPolicy>(__exec), views::all_read(::std::forward<_Range1>(__rng)),
+        __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), views::all_read(::std::forward<_Range1>(__rng)),
         views::all_write(::std::forward<_Range2>(__result)), __pred, oneapi::dpl::__internal::__pstl_assign());
 }
 
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
index 73719dee464..4baa2a46a43 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
@@ -419,7 +419,7 @@ __pattern_remove_if(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec,
     oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __buf(__exec, __rng.size());
     auto __copy_rng = oneapi::dpl::__ranges::views::all(__buf.get_buffer());
 
-    auto __copy_last_id = __pattern_copy_if(__exec, __rng, __copy_rng, __not_pred<_Predicate>{__pred},
+    auto __copy_last_id = __pattern_copy_if(__tag, __exec, __rng, __copy_rng, __not_pred<_Predicate>{__pred},
                                             oneapi::dpl::__internal::__pstl_assign());
     auto __copy_rng_truncated = __copy_rng | oneapi::dpl::experimental::ranges::views::take(__copy_last_id);
 
@@ -749,7 +749,8 @@ __pattern_reduce_by_segment(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&&
     // adjacent element (marks end of real segments)
     // TODO: replace wgroup size with segment size based on platform specifics.
     auto __intermediate_result_end = __pattern_copy_if(
-        oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__assign_key1_wrapper>(__exec), __view1, __view2,
+        __tag, oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__assign_key1_wrapper>(__exec), __view1,
+        __view2,
         [__n, __binary_pred, __wgroup_size](const auto& __a) {
             // The size of key range for the (i-1) view is one less, so for the 0th index we do not check the keys
             // for (i-1), but we still need to get its key value as it is the start of a segment
@@ -793,7 +794,8 @@ __pattern_reduce_by_segment(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&&
     // element is copied if it is the 0th element (marks beginning of first segment), or has a key not equal to
     // the adjacent element (end of a segment). Artificial segments based on wg size are not created.
     auto __result_end = __pattern_copy_if(
-        oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__assign_key2_wrapper>(__exec), __view3, __view4,
+        __tag, oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__assign_key2_wrapper>(__exec), __view3,
+        __view4,
         [__binary_pred](const auto& __a) {
             // The size of key range for the (i-1) view is one less, so for the 0th index we do not check the keys
             // for (i-1), but we still need to get its key value as it is the start of a segment

From 0499f0cac8aeafee9815adf4498e7a7f2ad305ad Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 09:47:42 +0100
Subject: [PATCH 375/566] __ranges::__pattern_swap + tag calls

---
 include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
index dd24f574b81..de8bff59fa9 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
@@ -275,11 +275,13 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy,
                                                       oneapi::dpl::__internal::__difference_t<_Range1>>
 swap_ranges(_ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&& __rng2)
 {
+    constexpr auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range1, _Range2>();
+
     using _ReferenceType1 = oneapi::dpl::__internal::__value_t<_Range1>&;
     using _ReferenceType2 = oneapi::dpl::__internal::__value_t<_Range2>&;
 
     return oneapi::dpl::__internal::__ranges::__pattern_swap(
-        ::std::forward<_ExecutionPolicy>(__exec), views::all(::std::forward<_Range1>(__rng1)),
+        __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), views::all(::std::forward<_Range1>(__rng1)),
         views::all(::std::forward<_Range2>(__rng2)), [](_ReferenceType1 __x, _ReferenceType2 __y) {
             using ::std::swap;
             swap(__x, __y);

From 4135e7e7907d46025d1bcd2f6b1d761502375c2b Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 09:48:13 +0100
Subject: [PATCH 376/566] __ranges::__pattern_remove_if + tag calls

---
 include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
index de8bff59fa9..de4fcafd21f 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
@@ -321,8 +321,10 @@ template <typename _ExecutionPolicy, typename _Range, typename _UnaryPredicate>
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, oneapi::dpl::__internal::__difference_t<_Range>>
 remove_if(_ExecutionPolicy&& __exec, _Range&& __rng, _UnaryPredicate __pred)
 {
-    return oneapi::dpl::__internal::__ranges::__pattern_remove_if(::std::forward<_ExecutionPolicy>(__exec),
-                                                                  views::all(::std::forward<_Range>(__rng)), __pred);
+    constexpr auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range>();
+
+    return oneapi::dpl::__internal::__ranges::__pattern_remove_if(
+        __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), views::all(::std::forward<_Range>(__rng)), __pred);
 }
 
 template <typename _ExecutionPolicy, typename _Range, typename _Tp>

From 94f5e8f7829a235bd3133ae7f56c4a11432defda Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 09:48:22 +0100
Subject: [PATCH 377/566] __ranges::__pattern_unique + tag calls

---
 include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
index de4fcafd21f..2749189fc52 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
@@ -365,7 +365,9 @@ template <typename _ExecutionPolicy, typename _Range, typename _BinaryPredicate>
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, oneapi::dpl::__internal::__difference_t<_Range>>
 unique(_ExecutionPolicy&& __exec, _Range&& __rng, _BinaryPredicate __pred)
 {
-    return oneapi::dpl::__internal::__ranges::__pattern_unique(::std::forward<_ExecutionPolicy>(__exec),
+    constexpr auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range>();
+
+    return oneapi::dpl::__internal::__ranges::__pattern_unique(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
                                                                views::all(::std::forward<_Range>(__rng)), __pred);
 }
 

From 8af241a627e701a59ab9831added9d079cb7d714 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 09:48:32 +0100
Subject: [PATCH 378/566] __ranges::__pattern_unique_copy + tag calls

---
 include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h          | 4 +++-
 include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h | 2 +-
 2 files changed, 4 insertions(+), 2 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
index 2749189fc52..f0ea9f5c4a1 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
@@ -384,8 +384,10 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy,
                                                       oneapi::dpl::__internal::__difference_t<_Range2>>
 unique_copy(_ExecutionPolicy&& __exec, _Range1&& __rng, _Range2&& __result, _BinaryPredicate __pred)
 {
+    constexpr auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range1, _Range2>();
+
     return oneapi::dpl::__internal::__ranges::__pattern_unique_copy(
-        ::std::forward<_ExecutionPolicy>(__exec), views::all_read(::std::forward<_Range1>(__rng)),
+        __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), views::all_read(::std::forward<_Range1>(__rng)),
         views::all_write(::std::forward<_Range2>(__result)), __pred, oneapi::dpl::__internal::__pstl_assign());
 }
 
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
index 4baa2a46a43..0d319256706 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
@@ -465,7 +465,7 @@ __pattern_unique(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Ra
 
     oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __buf(__exec, __rng.size());
     auto res_rng = oneapi::dpl::__ranges::views::all(__buf.get_buffer());
-    auto res = __pattern_unique_copy(__exec, __rng, res_rng, __pred, oneapi::dpl::__internal::__pstl_assign());
+    auto res = __pattern_unique_copy(__tag, __exec, __rng, res_rng, __pred, oneapi::dpl::__internal::__pstl_assign());
 
     __pattern_walk_n(__tag, ::std::forward<_ExecutionPolicy>(__exec), __brick_copy<_ExecutionPolicy>{}, res_rng,
                      ::std::forward<_Range>(__rng));

From 77ea0e6a1e110da0457d0d3440b21167cd236109 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 09:49:02 +0100
Subject: [PATCH 379/566] __ranges::__pattern_sort + tag calls

---
 include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
index f0ea9f5c4a1..80e4c7c0396 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
@@ -504,7 +504,9 @@ template <typename _ExecutionPolicy, typename _Range, typename _Compare, typenam
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy>
 sort(_ExecutionPolicy&& __exec, _Range&& __rng, _Compare __comp, _Proj __proj)
 {
-    oneapi::dpl::__internal::__ranges::__pattern_sort(::std::forward<_ExecutionPolicy>(__exec),
+    constexpr auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range>();
+
+    oneapi::dpl::__internal::__ranges::__pattern_sort(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
                                                       views::all(::std::forward<_Range>(__rng)), __comp, __proj);
 }
 

From 8b9e0f434bb50ee2a958f23aa5eb1bb1c16d107e Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 09:49:37 +0100
Subject: [PATCH 380/566] __ranges::__pattern_equal + tag calls

---
 include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h          | 4 +++-
 include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h | 4 ++--
 2 files changed, 5 insertions(+), 3 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
index 80e4c7c0396..a01e5bf9590 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
@@ -586,7 +586,9 @@ template <typename _ExecutionPolicy, typename _Range1, typename _Range2, typenam
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, bool>
 equal(_ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&& __rng2, _BinaryPredicate __p)
 {
-    return oneapi::dpl::__internal::__ranges::__pattern_equal(::std::forward<_ExecutionPolicy>(__exec),
+    constexpr auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range1, _Range2>();
+
+    return oneapi::dpl::__internal::__ranges::__pattern_equal(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
                                                               views::all_read(::std::forward<_Range1>(__rng1)),
                                                               views::all_read(::std::forward<_Range2>(__rng2)), __p);
 }
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
index 0d319256706..1d2ec4462c5 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
@@ -149,7 +149,7 @@ __pattern_find_end(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _
 
     if (__rng1.size() == __rng2.size())
     {
-        const bool __res = __pattern_equal(::std::forward<_ExecutionPolicy>(__exec), __rng1,
+        const bool __res = __pattern_equal(__tag, ::std::forward<_ExecutionPolicy>(__exec), __rng1,
                                            ::std::forward<_Range2>(__rng2), __pred);
         return __res ? 0 : __rng1.size();
     }
@@ -230,7 +230,7 @@ __pattern_search(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Ra
     if (__rng1.size() == __rng2.size())
     {
         const bool __res = __pattern_equal(
-            __par_backend_hetero::make_wrapped_policy<equal_wrapper>(::std::forward<_ExecutionPolicy>(__exec)),
+            __tag, __par_backend_hetero::make_wrapped_policy<equal_wrapper>(::std::forward<_ExecutionPolicy>(__exec)),
             __rng1, ::std::forward<_Range2>(__rng2), __pred);
         return __res ? 0 : __rng1.size();
     }

From 2e7d206e9749b0ef6247ff1149eca0d83ca34360 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 09:49:57 +0100
Subject: [PATCH 381/566] __ranges::__pattern_merge + tag calls

---
 include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)

diff --git a/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
index a01e5bf9590..e3f32e90241 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
@@ -624,8 +624,11 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy,
                                                       oneapi::dpl::__internal::__difference_t<_Range3>>
 merge(_ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&& __rng2, _Range3&& __rng3, _Compare __comp)
 {
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range1, _Range2, _Range3>();
+
     return oneapi::dpl::__internal::__ranges::__pattern_merge(
-        ::std::forward<_ExecutionPolicy>(__exec), views::all_read(::std::forward<_Range1>(__rng1)),
+        __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), views::all_read(::std::forward<_Range1>(__rng1)),
         views::all_read(::std::forward<_Range2>(__rng2)), views::all_write(::std::forward<_Range3>(__rng3)), __comp);
 }
 

From 8cee65a9cd279cc73f86050627f1d6c25a94b7af Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 09:50:06 +0100
Subject: [PATCH 382/566] __ranges::__pattern_min_element + tag calls

---
 include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)

diff --git a/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
index e3f32e90241..f61e0f986cc 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
@@ -648,8 +648,11 @@ template <typename _ExecutionPolicy, typename _Range, typename _Compare>
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, oneapi::dpl::__internal::__difference_t<_Range>>
 min_element(_ExecutionPolicy&& __exec, _Range&& __rng, _Compare __comp)
 {
+    constexpr auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range>();
+
     return oneapi::dpl::__internal::__ranges::__pattern_min_element(
-        ::std::forward<_ExecutionPolicy>(__exec), views::all_read(::std::forward<_Range>(__rng)), __comp);
+        __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), views::all_read(::std::forward<_Range>(__rng)),
+        __comp);
 }
 
 template <typename _ExecutionPolicy, typename _Range>

From 4d7e0626a7cab71a14a812704fbd793211dbea8b Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 09:50:16 +0100
Subject: [PATCH 383/566] __ranges::__pattern_minmax_element + tag calls

---
 include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)

diff --git a/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
index f61e0f986cc..06226b9e70f 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
@@ -686,8 +686,11 @@ oneapi::dpl::__internal::__enable_if_execution_policy<
     ::std::pair<oneapi::dpl::__internal::__difference_t<_Range>, oneapi::dpl::__internal::__difference_t<_Range>>>
 minmax_element(_ExecutionPolicy&& __exec, _Range&& __rng, _Compare __comp)
 {
+    constexpr auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range>();
+
     return oneapi::dpl::__internal::__ranges::__pattern_minmax_element(
-        ::std::forward<_ExecutionPolicy>(__exec), views::all_read(::std::forward<_Range>(__rng)), __comp);
+        __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), views::all_read(::std::forward<_Range>(__rng)),
+        __comp);
 }
 
 template <typename _ExecutionPolicy, typename _Range>

From 2a55f39cd5a1df827d2a6f3ddf72620a0098edcb Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 09:50:26 +0100
Subject: [PATCH 384/566] __ranges::__pattern_reduce_by_segment + tag calls

---
 include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)

diff --git a/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
index 06226b9e70f..6f93fc7971c 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
@@ -710,8 +710,11 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy,
 reduce_by_segment(_ExecutionPolicy&& __exec, _Range1&& __keys, _Range2&& __values, _Range3&& __out_keys,
                   _Range4&& __out_values, _BinaryPredicate __binary_pred, _BinaryOperator __binary_op)
 {
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range1, _Range2, _Range3>();
+
     return oneapi::dpl::__internal::__ranges::__pattern_reduce_by_segment(
-        ::std::forward<_ExecutionPolicy>(__exec), views::all_read(::std::forward<_Range1>(__keys)),
+        __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), views::all_read(::std::forward<_Range1>(__keys)),
         views::all_read(::std::forward<_Range2>(__values)), views::all_write(::std::forward<_Range3>(__out_keys)),
         views::all_write(::std::forward<_Range4>(__out_values)), __binary_pred, __binary_op);
 }

From a1e1a9557172ca9821a66bcaa22edeb09ee4a35c Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 09:51:13 +0100
Subject: [PATCH 385/566] __ranges::__pattern_transform_scan_base + tag calls

---
 .../dpl/pstl/hetero/numeric_ranges_impl_hetero.h   | 14 ++++++++------
 1 file changed, 8 insertions(+), 6 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/numeric_ranges_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/numeric_ranges_impl_hetero.h
index 4223facecfb..5acdddab085 100644
--- a/include/oneapi/dpl/pstl/hetero/numeric_ranges_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/numeric_ranges_impl_hetero.h
@@ -131,12 +131,14 @@ oneapi::dpl::__internal::__difference_t<_Range2>
 __pattern_transform_scan(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&& __rng2,
                          _UnaryOperation __unary_op, _Type __init, _BinaryOperation __binary_op, _Inclusive)
 {
+    constexpr auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range1, _Range2>();
+
     using _RepackedType = __par_backend_hetero::__repacked_tuple_t<_Type>;
     using _InitType = unseq_backend::__init_value<_RepackedType>;
 
-    return __pattern_transform_scan_base(::std::forward<_ExecutionPolicy>(__exec), ::std::forward<_Range1>(__rng1),
-                                         ::std::forward<_Range2>(__rng2), __unary_op, _InitType{__init}, __binary_op,
-                                         _Inclusive{});
+    return __pattern_transform_scan_base(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
+                                         ::std::forward<_Range1>(__rng1), ::std::forward<_Range2>(__rng2), __unary_op,
+                                         _InitType{__init}, __binary_op, _Inclusive{});
 }
 
 // scan without initial element
@@ -150,9 +152,9 @@ __pattern_transform_scan(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __e
     using _RepackedType = __par_backend_hetero::__repacked_tuple_t<_Type>;
     using _InitType = unseq_backend::__no_init_value<_RepackedType>;
 
-    return __pattern_transform_scan_base(::std::forward<_ExecutionPolicy>(__exec), ::std::forward<_Range1>(__rng1),
-                                         ::std::forward<_Range2>(__rng2), __unary_op, _InitType{}, __binary_op,
-                                         _Inclusive{});
+    return __pattern_transform_scan_base(__tag, ::std::forward<_ExecutionPolicy>(__exec),
+                                         ::std::forward<_Range1>(__rng1), ::std::forward<_Range2>(__rng2), __unary_op,
+                                         _InitType{}, __binary_op, _Inclusive{});
 }
 
 } // namespace __ranges

From e22bbcc84adc1e7999fd37e4ccc1d0557e64a11f Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 09:51:42 +0100
Subject: [PATCH 386/566] __ranges::__pattern_transform_reduce + tag calls

---
 .../oneapi/dpl/pstl/glue_numeric_ranges_impl.h  | 17 ++++++++++++-----
 1 file changed, 12 insertions(+), 5 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_numeric_ranges_impl.h b/include/oneapi/dpl/pstl/glue_numeric_ranges_impl.h
index 42d7c6e15a3..04f3f4cd1c5 100644
--- a/include/oneapi/dpl/pstl/glue_numeric_ranges_impl.h
+++ b/include/oneapi/dpl/pstl/glue_numeric_ranges_impl.h
@@ -63,9 +63,12 @@ template <typename _ExecutionPolicy, typename _Range1, typename _Range2, typenam
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _Tp>
 transform_reduce(_ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&& __rng2, _Tp __init)
 {
+    constexpr auto __dispatch_tag =
+        oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range1, _Range2>();
+
     using _ValueType = oneapi::dpl::__internal::__value_t<_Range1>;
     return oneapi::dpl::__internal::__ranges::__pattern_transform_reduce(
-        ::std::forward<_ExecutionPolicy>(__exec), views::all_read(::std::forward<_Range1>(__rng1)),
+        __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), views::all_read(::std::forward<_Range1>(__rng1)),
         views::all_read(::std::forward<_Range2>(__rng2)), __init, ::std::plus<_ValueType>(),
         ::std::multiplies<_ValueType>());
 }
@@ -76,8 +79,10 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _Tp>
 transform_reduce(_ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&& __rng2, _Tp __init,
                  _BinaryOperation1 __binary_op1, _BinaryOperation2 __binary_op2)
 {
+    constexpr auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range1, _Range2>();
+
     return oneapi::dpl::__internal::__ranges::__pattern_transform_reduce(
-        ::std::forward<_ExecutionPolicy>(__exec), views::all_read(::std::forward<_Range1>(__rng1)),
+        __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), views::all_read(::std::forward<_Range1>(__rng1)),
         views::all_read(::std::forward<_Range2>(__rng2)), __init, __binary_op1, __binary_op2);
 }
 
@@ -86,9 +91,11 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _Tp>
 transform_reduce(_ExecutionPolicy&& __exec, _Range&& __rng, _Tp __init, _BinaryOperation __binary_op,
                  _UnaryOperation __unary_op)
 {
-    return oneapi::dpl::__internal::__ranges::__pattern_transform_reduce(::std::forward<_ExecutionPolicy>(__exec),
-                                                                         views::all_read(::std::forward<_Range>(__rng)),
-                                                                         __init, __binary_op, __unary_op);
+    constexpr auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range>();
+
+    return oneapi::dpl::__internal::__ranges::__pattern_transform_reduce(
+        __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), views::all_read(::std::forward<_Range>(__rng)),
+        __init, __binary_op, __unary_op);
 }
 
 // [exclusive.scan]

From 3c01d9c754b3ca8caa84fd542fc341debbb50fa5 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 09:52:15 +0100
Subject: [PATCH 387/566] __ranges::__pattern_transform_scan + tag calls

---
 include/oneapi/dpl/pstl/glue_numeric_ranges_impl.h | 12 +++++++++---
 1 file changed, 9 insertions(+), 3 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_numeric_ranges_impl.h b/include/oneapi/dpl/pstl/glue_numeric_ranges_impl.h
index 04f3f4cd1c5..c8b4f19b560 100644
--- a/include/oneapi/dpl/pstl/glue_numeric_ranges_impl.h
+++ b/include/oneapi/dpl/pstl/glue_numeric_ranges_impl.h
@@ -161,8 +161,10 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy,
 transform_exclusive_scan(_ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&& __rng2, _Tp __init,
                          _BinaryOperation __binary_op, _UnaryOperation __unary_op)
 {
+    constexpr auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range1, _Range2>();
+
     return oneapi::dpl::__internal::__ranges::__pattern_transform_scan(
-        ::std::forward<_ExecutionPolicy>(__exec), views::all_read(::std::forward<_Range1>(__rng1)),
+        __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), views::all_read(::std::forward<_Range1>(__rng1)),
         views::all_write(::std::forward<_Range2>(__rng2)), __unary_op, __init, __binary_op,
         /*inclusive=*/::std::false_type());
 }
@@ -176,8 +178,10 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy,
 transform_inclusive_scan(_ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&& __rng2, _BinaryOperation __binary_op,
                          _UnaryOperation __unary_op, _Tp __init)
 {
+    constexpr auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range1, _Range2>();
+
     return oneapi::dpl::__internal::__ranges::__pattern_transform_scan(
-        ::std::forward<_ExecutionPolicy>(__exec), views::all_read(::std::forward<_Range1>(__rng1)),
+        __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), views::all_read(::std::forward<_Range1>(__rng1)),
         views::all_write(::std::forward<_Range2>(__rng2)), __unary_op, __init, __binary_op,
         /*inclusive=*/::std::true_type());
 }
@@ -189,8 +193,10 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy,
 transform_inclusive_scan(_ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&& __rng2, _BinaryOperation __binary_op,
                          _UnaryOperation __unary_op)
 {
+    constexpr auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range1, _Range2>();
+
     return oneapi::dpl::__internal::__ranges::__pattern_transform_scan(
-        ::std::forward<_ExecutionPolicy>(__exec), views::all_read(::std::forward<_Range1>(__rng1)),
+        __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), views::all_read(::std::forward<_Range1>(__rng1)),
         views::all_write(::std::forward<_Range2>(__rng2)), __unary_op, __binary_op, /*inclusive=*/::std::true_type());
 }
 

From d7290b19ac0ea37194657e0a091fede88c96c9a5 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 09:54:16 +0100
Subject: [PATCH 388/566] __ranges::__pattern_scan_copy + tag calls

---
 include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
index 1d2ec4462c5..32e016390a5 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
@@ -399,7 +399,7 @@ __pattern_copy_if(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _R
     unseq_backend::__create_mask<_Predicate, _SizeType> __create_mask_op{__pred};
     unseq_backend::__copy_by_mask<_ReduceOp, _Assign, /*inclusive*/ ::std::true_type, 1> __copy_by_mask_op;
 
-    return __pattern_scan_copy(::std::forward<_ExecutionPolicy>(__exec), ::std::forward<_Range1>(__rng1),
+    return __pattern_scan_copy(__tag, ::std::forward<_ExecutionPolicy>(__exec), ::std::forward<_Range1>(__rng1),
                                ::std::forward<_Range2>(__rng2), __create_mask_op, __copy_by_mask_op);
 }
 
@@ -446,7 +446,7 @@ __pattern_unique_copy(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec
     __create_mask_unique_copy<__not_pred<_BinaryPredicate>, _It1DifferenceType> __create_mask_op{
         __not_pred<_BinaryPredicate>{__pred}};
 
-    return __pattern_scan_copy(::std::forward<_ExecutionPolicy>(__exec), ::std::forward<_Range1>(__rng),
+    return __pattern_scan_copy(__tag, ::std::forward<_ExecutionPolicy>(__exec), ::std::forward<_Range1>(__rng),
                                ::std::forward<_Range2>(__result), __create_mask_op, __copy_by_mask_op);
 }
 

From cf9dbd0332de9e6a1000a6e834fcb9aa5da43dbd Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 09:57:24 +0100
Subject: [PATCH 389/566] __par_backend_hetero::__parallel_transform_reduce +
 tag calls

---
 .../oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h   | 6 ++----
 include/oneapi/dpl/pstl/hetero/numeric_ranges_impl_hetero.h | 6 ++----
 2 files changed, 4 insertions(+), 8 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
index 32e016390a5..0a1019a3154 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
@@ -575,8 +575,7 @@ __pattern_min_element(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Ran
     auto __ret_idx =
         oneapi::dpl::__par_backend_hetero::__parallel_transform_reduce<_ReduceValueType,
                                                                        ::std::false_type /*is_commutative*/>(
-            oneapi::dpl::__internal::__device_backend_tag{}, // TODO required to fix backend tag eval
-            ::std::forward<_ExecutionPolicy>(__exec), __reduce_fn, __transform_fn,
+            _BackendTag{}, ::std::forward<_ExecutionPolicy>(__exec), __reduce_fn, __transform_fn,
             unseq_backend::__no_init_value{}, // no initial value
             ::std::forward<_Range>(__rng))
             .get();
@@ -630,8 +629,7 @@ __pattern_minmax_element(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _
     _ReduceValueType __ret =
         oneapi::dpl::__par_backend_hetero::__parallel_transform_reduce<_ReduceValueType,
                                                                        ::std::false_type /*is_commutative*/>(
-            oneapi::dpl::__internal::__device_backend_tag{}, // TODO required to fix backend tag eval
-            ::std::forward<_ExecutionPolicy>(__exec), __reduce_fn, __transform_fn,
+            _BackendTag{}, ::std::forward<_ExecutionPolicy>(__exec), __reduce_fn, __transform_fn,
             unseq_backend::__no_init_value{}, // no initial value
             ::std::forward<_Range>(__rng))
             .get();
diff --git a/include/oneapi/dpl/pstl/hetero/numeric_ranges_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/numeric_ranges_impl_hetero.h
index 5acdddab085..5e2d2efb075 100644
--- a/include/oneapi/dpl/pstl/hetero/numeric_ranges_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/numeric_ranges_impl_hetero.h
@@ -51,8 +51,7 @@ __pattern_transform_reduce(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec,
 
     return oneapi::dpl::__par_backend_hetero::__parallel_transform_reduce<_RepackedTp,
                                                                           ::std::true_type /*is_commutative*/>(
-               oneapi::dpl::__internal::__device_backend_tag{}, // TODO required to fix backend tag eval
-               ::std::forward<_ExecutionPolicy>(__exec), __binary_op1, _Functor{__binary_op2},
+               _BackendTag{}, ::std::forward<_ExecutionPolicy>(__exec), __binary_op1, _Functor{__binary_op2},
                unseq_backend::__init_value<_RepackedTp>{__init}, // initial value
                ::std::forward<_Range1>(__rng1), ::std::forward<_Range2>(__rng2))
         .get();
@@ -76,8 +75,7 @@ __pattern_transform_reduce(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec,
 
     return oneapi::dpl::__par_backend_hetero::__parallel_transform_reduce<_RepackedTp,
                                                                           ::std::true_type /*is_commutative*/>(
-               oneapi::dpl::__internal::__device_backend_tag{}, // TODO required to fix backend tag eval
-               ::std::forward<_ExecutionPolicy>(__exec), __binary_op, _Functor{__unary_op},
+               _BackendTag{}, ::std::forward<_ExecutionPolicy>(__exec), __binary_op, _Functor{__unary_op},
                unseq_backend::__init_value<_RepackedTp>{__init}, // initial value
                ::std::forward<_Range>(__rng))
         .get();

From bb4bbe44b0c44b440adb467123204a5e549122df Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 09:57:53 +0100
Subject: [PATCH 390/566] __par_backend_hetero::__parallel_transform_scan_base
 + tag calls

---
 .../oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h    | 3 +--
 include/oneapi/dpl/pstl/hetero/numeric_ranges_impl_hetero.h  | 5 ++---
 2 files changed, 3 insertions(+), 5 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
index 0a1019a3154..025c4500e48 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
@@ -366,8 +366,7 @@ __pattern_scan_copy(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Range
 
     auto __res =
         __par_backend_hetero::__parallel_transform_scan_base(
-            oneapi::dpl::__internal::__device_backend_tag{}, // TODO required to fix backend tag eval
-            ::std::forward<_ExecutionPolicy>(__exec),
+            _BackendTag{}, ::std::forward<_ExecutionPolicy>(__exec),
             oneapi::dpl::__ranges::zip_view(
                 __rng1, oneapi::dpl::__ranges::all_view<int32_t, __par_backend_hetero::access_mode::read_write>(
                             __mask_buf.get_buffer())),
diff --git a/include/oneapi/dpl/pstl/hetero/numeric_ranges_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/numeric_ranges_impl_hetero.h
index 5e2d2efb075..a06a3eb18fe 100644
--- a/include/oneapi/dpl/pstl/hetero/numeric_ranges_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/numeric_ranges_impl_hetero.h
@@ -106,9 +106,8 @@ __pattern_transform_scan_base(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __ex
     _NoOpFunctor __get_data_op;
 
     oneapi::dpl::__par_backend_hetero::__parallel_transform_scan_base(
-        oneapi::dpl::__internal::__device_backend_tag{}, // TODO required to fix backend tag eval
-        ::std::forward<_ExecutionPolicy>(__exec), ::std::forward<_Range1>(__rng1), ::std::forward<_Range2>(__rng2),
-        __binary_op, __init,
+        _BackendTag{}, ::std::forward<_ExecutionPolicy>(__exec), ::std::forward<_Range1>(__rng1),
+        ::std::forward<_Range2>(__rng2), __binary_op, __init,
         // local scan
         unseq_backend::__scan<_Inclusive, _ExecutionPolicy, _BinaryOperation, _UnaryFunctor, _Assigner, _Assigner,
                               _NoOpFunctor, _InitType>{__binary_op, _UnaryFunctor{__unary_op}, __assign_op, __assign_op,

From 1e02cad411132d43c59b820f63ee5f867c75c87d Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 09:58:10 +0100
Subject: [PATCH 391/566] __par_backend_hetero::__parallel_for + tag calls

---
 .../dpl/pstl/hetero/algorithm_ranges_impl_hetero.h   | 12 +++++-------
 1 file changed, 5 insertions(+), 7 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
index 025c4500e48..14cc82d5741 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
@@ -46,10 +46,9 @@ __pattern_walk_n(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Function
     auto __n = oneapi::dpl::__ranges::__get_first_range_size(__rngs...);
     if (__n > 0)
     {
-        oneapi::dpl::__par_backend_hetero::__parallel_for(
-            oneapi::dpl::__internal::__device_backend_tag{}, // TODO required to fix backend tag eval
-            ::std::forward<_ExecutionPolicy>(__exec), unseq_backend::walk_n<_ExecutionPolicy, _Function>{__f}, __n,
-            ::std::forward<_Ranges>(__rngs)...)
+        oneapi::dpl::__par_backend_hetero::__parallel_for(_BackendTag{}, ::std::forward<_ExecutionPolicy>(__exec),
+                                                          unseq_backend::walk_n<_ExecutionPolicy, _Function>{__f}, __n,
+                                                          ::std::forward<_Ranges>(__rngs)...)
             .wait();
     }
 }
@@ -761,8 +760,7 @@ __pattern_reduce_by_segment(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&&
 
     //reduce by segment
     oneapi::dpl::__par_backend_hetero::__parallel_for(
-        oneapi::dpl::__internal::__device_backend_tag{}, // TODO required to fix backend tag eval
-        oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__reduce1_wrapper>(__exec),
+        _BackendTag{}, oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__reduce1_wrapper>(__exec),
         unseq_backend::__brick_reduce_idx<_BinaryOperator, decltype(__n)>(__binary_op, __n), __intermediate_result_end,
         oneapi::dpl::__ranges::take_view_simple(experimental::ranges::views::all_read(__idx),
                                                 __intermediate_result_end),
@@ -804,7 +802,7 @@ __pattern_reduce_by_segment(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&&
 
     //reduce by segment
     oneapi::dpl::__par_backend_hetero::__parallel_for(
-        oneapi::dpl::__internal::__device_backend_tag{}, // TODO required to fix backend tag eval
+        _BackendTag{},
         oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__reduce2_wrapper>(
             ::std::forward<_ExecutionPolicy>(__exec)),
         unseq_backend::__brick_reduce_idx<_BinaryOperator, decltype(__intermediate_result_end)>(

From 926084b73340ad9ec37ea9f4231a28fa21368cb3 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 09:58:22 +0100
Subject: [PATCH 392/566] __par_backend_hetero::__parallel_find_or + tag calls

---
 .../hetero/algorithm_ranges_impl_hetero.h     | 19 +++++++++----------
 1 file changed, 9 insertions(+), 10 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
index 14cc82d5741..04ce591bbac 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
@@ -105,8 +105,7 @@ __pattern_equal(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Range1&&
     // TODO: in case of confilicting names
     // __par_backend_hetero::make_wrapped_policy<__par_backend_hetero::__or_policy_wrapper>()
     return !oneapi::dpl::__par_backend_hetero::__parallel_find_or(
-        oneapi::dpl::__internal::__device_backend_tag{}, // TODO required to fix backend tag eval
-        ::std::forward<_ExecutionPolicy>(__exec), _Predicate{equal_predicate<_Pred>{__pred}},
+        _BackendTag{}, ::std::forward<_ExecutionPolicy>(__exec), _Predicate{equal_predicate<_Pred>{__pred}},
         oneapi::dpl::__par_backend_hetero::__parallel_or_tag{},
         oneapi::dpl::__ranges::zip_view(::std::forward<_Range1>(__rng1), ::std::forward<_Range2>(__rng2)));
 }
@@ -127,7 +126,7 @@ __pattern_find_if(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Range&&
     using _TagType = oneapi::dpl::__par_backend_hetero::__parallel_find_forward_tag<_Range>;
 
     return oneapi::dpl::__par_backend_hetero::__parallel_find_or(
-        oneapi::dpl::__internal::__device_backend_tag{}, // TODO required to fix backend tag eval
+        _BackendTag{},
         __par_backend_hetero::make_wrapped_policy<__par_backend_hetero::__find_policy_wrapper>(
             ::std::forward<_ExecutionPolicy>(__exec)),
         _Predicate{__pred}, _TagType{}, ::std::forward<_Range>(__rng));
@@ -157,7 +156,7 @@ __pattern_find_end(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _
     using _TagType = __par_backend_hetero::__parallel_find_backward_tag<_Range1>;
 
     return oneapi::dpl::__par_backend_hetero::__parallel_find_or(
-        oneapi::dpl::__internal::__device_backend_tag{}, // TODO required to fix backend tag eval
+        _BackendTag{},
         __par_backend_hetero::make_wrapped_policy<__par_backend_hetero::__find_policy_wrapper>(
             ::std::forward<_ExecutionPolicy>(__exec)),
         _Predicate{__pred}, _TagType{}, ::std::forward<_Range1>(__rng1), ::std::forward<_Range2>(__rng2));
@@ -181,7 +180,7 @@ __pattern_find_first_of(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _R
 
     //TODO: To check whether it makes sense to iterate over the second sequence in case of __rng1.size() < __rng2.size()
     return oneapi::dpl::__par_backend_hetero::__parallel_find_or(
-        oneapi::dpl::__internal::__device_backend_tag{}, // TODO required to fix backend tag eval
+        _BackendTag{},
         __par_backend_hetero::make_wrapped_policy<__par_backend_hetero::__find_policy_wrapper>(
             ::std::forward<_ExecutionPolicy>(__exec)),
         _Predicate{__pred}, _TagType{}, ::std::forward<_Range1>(__rng1), ::std::forward<_Range2>(__rng2));
@@ -200,7 +199,7 @@ __pattern_any_of(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Range&&
 
     using _Predicate = oneapi::dpl::unseq_backend::single_match_pred<_ExecutionPolicy, _Pred>;
     return oneapi::dpl::__par_backend_hetero::__parallel_find_or(
-        oneapi::dpl::__internal::__device_backend_tag{}, // TODO required to fix backend tag eval
+        _BackendTag{},
         __par_backend_hetero::make_wrapped_policy<oneapi::dpl::__par_backend_hetero::__or_policy_wrapper>(
             ::std::forward<_ExecutionPolicy>(__exec)),
         _Predicate{__pred}, oneapi::dpl::__par_backend_hetero::__parallel_or_tag{}, ::std::forward<_Range>(__rng));
@@ -238,7 +237,7 @@ __pattern_search(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Ra
     using _TagType = oneapi::dpl::__par_backend_hetero::__parallel_find_forward_tag<_Range1>;
 
     return oneapi::dpl::__par_backend_hetero::__parallel_find_or(
-        oneapi::dpl::__internal::__device_backend_tag{}, // TODO required to fix backend tag eval
+        _BackendTag{},
         oneapi::dpl::__par_backend_hetero::make_wrapped_policy<
             oneapi::dpl::__par_backend_hetero::__find_policy_wrapper>(::std::forward<_ExecutionPolicy>(__exec)),
         _Predicate{__pred}, _TagType{}, ::std::forward<_Range1>(__rng1), ::std::forward<_Range2>(__rng2));
@@ -301,9 +300,9 @@ __pattern_adjacent_find(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _R
     // TODO: in case of confilicting names
     // __par_backend_hetero::make_wrapped_policy<__par_backend_hetero::__or_policy_wrapper>()
     auto result = oneapi::dpl::__par_backend_hetero::__parallel_find_or(
-        oneapi::dpl::__internal::__device_backend_tag{}, // TODO required to fix backend tag eval
-        ::std::forward<_ExecutionPolicy>(__exec), _Predicate{adjacent_find_fn<_BinaryPredicate>{__predicate}},
-        _TagType{}, oneapi::dpl::__ranges::zip_view(__rng1, __rng2));
+        _BackendTag{}, ::std::forward<_ExecutionPolicy>(__exec),
+        _Predicate{adjacent_find_fn<_BinaryPredicate>{__predicate}}, _TagType{},
+        oneapi::dpl::__ranges::zip_view(__rng1, __rng2));
 
     // inverted conditional because of
     // reorder_predicate in glue_algorithm_impl.h

From acf4090ed3c80a4395c616e78a3df37dc9e48c64 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 09:59:09 +0100
Subject: [PATCH 393/566] _par_backend_hetero::__parallel_transform_reduce +
 tag calls

---
 include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
index 04ce591bbac..50dfa3b60e4 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
@@ -327,8 +327,7 @@ __pattern_count(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Range&& _
 
     return oneapi::dpl::__par_backend_hetero::__parallel_transform_reduce<_ReduceValueType,
                                                                           ::std::true_type /*is_commutative*/>(
-               oneapi::dpl::__internal::__device_backend_tag{}, // TODO required to fix backend tag eval
-               ::std::forward<_ExecutionPolicy>(__exec), __reduce_fn, __transform_fn,
+               _BackendTag{}, ::std::forward<_ExecutionPolicy>(__exec), __reduce_fn, __transform_fn,
                unseq_backend::__no_init_value{}, // no initial value
                ::std::forward<_Range>(__rng))
         .get();

From ecb002af02f6df1002e3f2ad6cd6d87f026a9b43 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 09:59:37 +0100
Subject: [PATCH 394/566] __par_backend_hetero::__parallel_merge + tag calls

---
 .../oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h  | 7 +++----
 1 file changed, 3 insertions(+), 4 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
index 50dfa3b60e4..6875fd7861d 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
@@ -515,10 +515,9 @@ __pattern_merge(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Ran
     }
     else
     {
-        __par_backend_hetero::__parallel_merge(
-            oneapi::dpl::__internal::__device_backend_tag{}, // TODO required to fix backend tag eval
-            ::std::forward<_ExecutionPolicy>(__exec), ::std::forward<_Range1>(__rng1), ::std::forward<_Range2>(__rng2),
-            ::std::forward<_Range3>(__rng3), __comp)
+        __par_backend_hetero::__parallel_merge(_BackendTag{}, ::std::forward<_ExecutionPolicy>(__exec),
+                                               ::std::forward<_Range1>(__rng1), ::std::forward<_Range2>(__rng2),
+                                               ::std::forward<_Range3>(__rng3), __comp)
             .wait();
     }
 

From 6af11b74e533fc134a37129b3a04bfff6d71fa09 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 09:59:50 +0100
Subject: [PATCH 395/566] __par_backend_hetero::__parallel_stable_sort + tag
 calls

---
 .../oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h    | 5 ++---
 1 file changed, 2 insertions(+), 3 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
index 6875fd7861d..14ca7e82b8a 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
@@ -533,9 +533,8 @@ void
 __pattern_sort(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Range&& __rng, _Compare __comp, _Proj __proj)
 {
     if (__rng.size() >= 2)
-        __par_backend_hetero::__parallel_stable_sort(
-            oneapi::dpl::__internal::__device_backend_tag{}, // TODO required to fix backend tag eval
-            ::std::forward<_ExecutionPolicy>(__exec), ::std::forward<_Range>(__rng), __comp, __proj)
+        __par_backend_hetero::__parallel_stable_sort(_BackendTag{}, ::std::forward<_ExecutionPolicy>(__exec),
+                                                     ::std::forward<_Range>(__rng), __comp, __proj)
             .wait();
 }
 

From c195fb88ed5250eb7b2c9e70860480406b728c08 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 10:16:31 +0100
Subject: [PATCH 396/566] 
 include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h - fix spelling
 check error

---
 include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
index 14ca7e82b8a..9f6ccd5a2dd 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
@@ -297,7 +297,7 @@ __pattern_adjacent_find(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _R
     auto __rng1 = __rng | oneapi::dpl::experimental::ranges::views::take(__rng.size() - 1);
     auto __rng2 = __rng | oneapi::dpl::experimental::ranges::views::drop(1);
 
-    // TODO: in case of confilicting names
+    // TODO: in case of conflicting names
     // __par_backend_hetero::make_wrapped_policy<__par_backend_hetero::__or_policy_wrapper>()
     auto result = oneapi::dpl::__par_backend_hetero::__parallel_find_or(
         _BackendTag{}, ::std::forward<_ExecutionPolicy>(__exec),

From b9209400a8c38137ad9b4aedd416813dcfce9c11 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 10:23:34 +0100
Subject: [PATCH 397/566] 
 include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h - remove extra
 __ranges::__select_backend

---
 .../oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h   | 3 ---
 include/oneapi/dpl/pstl/hetero/numeric_ranges_impl_hetero.h | 6 ++----
 2 files changed, 2 insertions(+), 7 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
index 9f6ccd5a2dd..b6357fb5159 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
@@ -682,9 +682,6 @@ __pattern_reduce_by_segment(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&&
     //          __out_keys   = { 1, 2, 3, 4, 1, 3, 1, 3, 0 }
     //          __out_values = { 1, 2, 3, 4, 2, 6, 2, 6, 0 }
 
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range1, _Range2, _Range3>();
-
     const auto __n = __keys.size();
 
     if (__n == 0)
diff --git a/include/oneapi/dpl/pstl/hetero/numeric_ranges_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/numeric_ranges_impl_hetero.h
index a06a3eb18fe..969b05ab914 100644
--- a/include/oneapi/dpl/pstl/hetero/numeric_ranges_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/numeric_ranges_impl_hetero.h
@@ -125,15 +125,13 @@ __pattern_transform_scan_base(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __ex
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Range1, typename _Range2, typename _UnaryOperation,
           typename _Type, typename _BinaryOperation, typename _Inclusive>
 oneapi::dpl::__internal::__difference_t<_Range2>
-__pattern_transform_scan(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&& __rng2,
+__pattern_transform_scan(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&& __rng2,
                          _UnaryOperation __unary_op, _Type __init, _BinaryOperation __binary_op, _Inclusive)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range1, _Range2>();
-
     using _RepackedType = __par_backend_hetero::__repacked_tuple_t<_Type>;
     using _InitType = unseq_backend::__init_value<_RepackedType>;
 
-    return __pattern_transform_scan_base(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
+    return __pattern_transform_scan_base(__tag, ::std::forward<_ExecutionPolicy>(__exec),
                                          ::std::forward<_Range1>(__rng1), ::std::forward<_Range2>(__rng2), __unary_op,
                                          _InitType{__init}, __binary_op, _Inclusive{});
 }

From a95f9eed264ef6f8b6af07be56ffa4b4d6479a14 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 10:31:12 +0100
Subject: [PATCH 398/566] __ranges::__pattern_search + tag impls

---
 include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
index b6357fb5159..ed6f816782e 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
@@ -247,8 +247,8 @@ __pattern_search(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Ra
 // search_n
 //------------------------------------------------------------------------
 
-template <template <typename _BackendTag, typename _ExecutionPolicy, typename _Range, typename _Size, typename _Tp,
-                    typename _BinaryPredicate>
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Range, typename _Size, typename _Tp,
+          typename _BinaryPredicate>
 oneapi::dpl::__internal::__difference_t<_Range>
 __pattern_search_n(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Range&& __rng, _Size __count, const _Tp& __value,
                    _BinaryPredicate __pred)

From 19523343c13d27b72b7b6749a4b31164d7f84b16 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 11:03:58 +0100
Subject: [PATCH 399/566] 
 include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h - fix compile
 error: invalid operands to binary expression

/oneDPL/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h:420:44: error: invalid operands to binary expression ('oneapi::dpl::__ranges::all_view<int, (mode)1026, (target)2014, (placeholder)1>' and 'detail::rao_proxy<(lambda at oneDPL/include/oneapi/dpl/pstl/ranges/nanorange.hpp:18745:34)>' (aka '__nanorange::nano::detail::rao_proxy<(lambda at /home/sdp/skopienko/oneDPL/include/oneapi/dpl/pstl/ranges/nanorange.hpp:18745:34)>'))
  420 |     auto __copy_rng_truncated = __copy_rng | oneapi::dpl::experimental::ranges::views::take(__copy_last_id);
      |                                 ~~~~~~~~~~ ^ ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
---
 include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
index ed6f816782e..a54ecf46a54 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
@@ -415,7 +415,7 @@ __pattern_remove_if(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec,
     oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __buf(__exec, __rng.size());
     auto __copy_rng = oneapi::dpl::__ranges::views::all(__buf.get_buffer());
 
-    auto __copy_last_id = __pattern_copy_if(__tag, __exec, __rng, __copy_rng, __not_pred<_Predicate>{__pred},
+    auto __copy_last_id = __ranges::__pattern_copy_if(__tag, __exec, __rng, __copy_rng, __not_pred<_Predicate>{__pred},
                                             oneapi::dpl::__internal::__pstl_assign());
     auto __copy_rng_truncated = __copy_rng | oneapi::dpl::experimental::ranges::views::take(__copy_last_id);
 

From 402062c24ba42f55d45366e5dda2ef16a962238d Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 11:06:43 +0100
Subject: [PATCH 400/566] 
 include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h - fix compile
 error: error: no viable conversion

oneDPL/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h:468:12: error: no viable conversion from returned value of type 'oneapi::dpl::__internal::__pstl_equal' to function return type 'oneapi::dpl::__internal::__difference_t<all_view<int, sycl::access::mode::read_write, sycl::access::target::global_buffer, sycl::access::placeholder::true_t>>' (aka 'long')
    return res;
           ^~~
---
 include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
index a54ecf46a54..2c45c2cf541 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
@@ -461,7 +461,7 @@ __pattern_unique(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Ra
 
     oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __buf(__exec, __rng.size());
     auto res_rng = oneapi::dpl::__ranges::views::all(__buf.get_buffer());
-    auto res = __pattern_unique_copy(__tag, __exec, __rng, res_rng, __pred, oneapi::dpl::__internal::__pstl_assign());
+    auto res = __ranges::__pattern_unique_copy(__tag, __exec, __rng, res_rng, __pred, oneapi::dpl::__internal::__pstl_assign());
 
     __pattern_walk_n(__tag, ::std::forward<_ExecutionPolicy>(__exec), __brick_copy<_ExecutionPolicy>{}, res_rng,
                      ::std::forward<_Range>(__rng));

From c11ed9a339563ed4b62ae5e5aec2db03a2920f8e Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 11:09:42 +0100
Subject: [PATCH 401/566] 
 include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h - add
 `__ranges` namespace name for paterrn calls (to oalign with fixed compile
 errors)

---
 .../pstl/hetero/algorithm_ranges_impl_hetero.h   | 16 +++++++---------
 1 file changed, 7 insertions(+), 9 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
index 2c45c2cf541..daefb904b51 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
@@ -147,8 +147,8 @@ __pattern_find_end(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _
 
     if (__rng1.size() == __rng2.size())
     {
-        const bool __res = __pattern_equal(__tag, ::std::forward<_ExecutionPolicy>(__exec), __rng1,
-                                           ::std::forward<_Range2>(__rng2), __pred);
+        const bool __res = __ranges::__pattern_equal(__tag, ::std::forward<_ExecutionPolicy>(__exec), __rng1,
+                                                     ::std::forward<_Range2>(__rng2), __pred);
         return __res ? 0 : __rng1.size();
     }
 
@@ -227,7 +227,7 @@ __pattern_search(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Ra
 
     if (__rng1.size() == __rng2.size())
     {
-        const bool __res = __pattern_equal(
+        const bool __res = __ranges::__pattern_equal(
             __tag, __par_backend_hetero::make_wrapped_policy<equal_wrapper>(::std::forward<_ExecutionPolicy>(__exec)),
             __rng1, ::std::forward<_Range2>(__rng2), __pred);
         return __res ? 0 : __rng1.size();
@@ -737,9 +737,8 @@ __pattern_reduce_by_segment(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&&
     // evenly divisible by wg size (ensures segments are not long), or has a key not equal to the
     // adjacent element (marks end of real segments)
     // TODO: replace wgroup size with segment size based on platform specifics.
-    auto __intermediate_result_end = __pattern_copy_if(
-        __tag, oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__assign_key1_wrapper>(__exec), __view1,
-        __view2,
+    auto __intermediate_result_end = __ranges::__pattern_copy_if(
+        __tag, oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__assign_key1_wrapper>(__exec), __view1, __view2,
         [__n, __binary_pred, __wgroup_size](const auto& __a) {
             // The size of key range for the (i-1) view is one less, so for the 0th index we do not check the keys
             // for (i-1), but we still need to get its key value as it is the start of a segment
@@ -781,9 +780,8 @@ __pattern_reduce_by_segment(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&&
 
     // element is copied if it is the 0th element (marks beginning of first segment), or has a key not equal to
     // the adjacent element (end of a segment). Artificial segments based on wg size are not created.
-    auto __result_end = __pattern_copy_if(
-        __tag, oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__assign_key2_wrapper>(__exec), __view3,
-        __view4,
+    auto __result_end = __ranges::__pattern_copy_if(
+        __tag, oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__assign_key2_wrapper>(__exec), __view3, __view4,
         [__binary_pred](const auto& __a) {
             // The size of key range for the (i-1) view is one less, so for the 0th index we do not check the keys
             // for (i-1), but we still need to get its key value as it is the start of a segment

From e393899d49e7bc1ba4462926b035c23983074fb2 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 12:28:01 +0100
Subject: [PATCH 402/566] __cancel_execution + tag impls

---
 include/oneapi/dpl/pstl/omp/util.h                | 2 +-
 include/oneapi/dpl/pstl/parallel_backend_serial.h | 2 +-
 include/oneapi/dpl/pstl/parallel_backend_tbb.h    | 2 +-
 3 files changed, 3 insertions(+), 3 deletions(-)

diff --git a/include/oneapi/dpl/pstl/omp/util.h b/include/oneapi/dpl/pstl/omp/util.h
index bcbfecc23e4..292ee6a5c0b 100644
--- a/include/oneapi/dpl/pstl/omp/util.h
+++ b/include/oneapi/dpl/pstl/omp/util.h
@@ -48,7 +48,7 @@ namespace __omp_backend
 // use to cancel execution
 //------------------------------------------------------------------------
 inline void
-__cancel_execution()
+__cancel_execution(oneapi::dpl::__internal::__omp_backend_tag)
 {
     // TODO: Figure out how to make cancellation work.
 }
diff --git a/include/oneapi/dpl/pstl/parallel_backend_serial.h b/include/oneapi/dpl/pstl/parallel_backend_serial.h
index 5876e3383d7..fb33353c002 100644
--- a/include/oneapi/dpl/pstl/parallel_backend_serial.h
+++ b/include/oneapi/dpl/pstl/parallel_backend_serial.h
@@ -60,7 +60,7 @@ template <typename _ExecutionPolicy, typename _Tp>
 using __buffer = __buffer_impl<::std::decay_t<_ExecutionPolicy>, _Tp>;
 
 inline void
-__cancel_execution()
+__cancel_execution(oneapi::dpl::__internal::__serial_backend_tag)
 {
 }
 
diff --git a/include/oneapi/dpl/pstl/parallel_backend_tbb.h b/include/oneapi/dpl/pstl/parallel_backend_tbb.h
index b67e7c7931d..2bdd6b466e5 100644
--- a/include/oneapi/dpl/pstl/parallel_backend_tbb.h
+++ b/include/oneapi/dpl/pstl/parallel_backend_tbb.h
@@ -86,7 +86,7 @@ using __buffer = __buffer_impl<::std::decay_t<_ExecutionPolicy>, _Tp>;
 
 // Wrapper for tbb::task
 inline void
-__cancel_execution()
+__cancel_execution(oneapi::dpl::__internal::__tbb_backend_tag)
 {
 #if TBB_INTERFACE_VERSION <= 12000
     tbb::task::self().group()->cancel_group_execution();

From 069653ec62130ae6574d7694354bb9c1f9146ad3 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 12:28:19 +0100
Subject: [PATCH 403/566] __cancel_execution + tag calls

---
 include/oneapi/dpl/pstl/parallel_impl.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/include/oneapi/dpl/pstl/parallel_impl.h b/include/oneapi/dpl/pstl/parallel_impl.h
index fc5507f50d1..66d9d8d1741 100644
--- a/include/oneapi/dpl/pstl/parallel_impl.h
+++ b/include/oneapi/dpl/pstl/parallel_impl.h
@@ -85,7 +85,7 @@ __parallel_or(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _Index
                                       if (!__found.load(::std::memory_order_relaxed) && __f(__i, __j))
                                       {
                                           __found.store(true, ::std::memory_order_relaxed);
-                                          __par_backend::__cancel_execution();
+                                          __par_backend::__cancel_execution(__backend_tag{});
                                       }
                                   });
     return __found;

From c14933473728c349dfbf0f352d437aec30ac0406 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 14:56:45 +0100
Subject: [PATCH 404/566] include/oneapi/dpl/pstl/execution_impl.h - add using
 __backend_tag = __par_backend_tag; into struct __serial_tag

---
 include/oneapi/dpl/pstl/execution_impl.h | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/include/oneapi/dpl/pstl/execution_impl.h b/include/oneapi/dpl/pstl/execution_impl.h
index dae807121cf..319dfef617f 100644
--- a/include/oneapi/dpl/pstl/execution_impl.h
+++ b/include/oneapi/dpl/pstl/execution_impl.h
@@ -111,6 +111,9 @@ template <class _IsVector>
 struct __serial_tag
 {
     using __is_vector = _IsVector;
+    // backend tag can be change depending on
+    // TBB availability in the environment
+    using __backend_tag = __par_backend_tag;
 };
 
 template <class _IsVector>

From b6531933f1a517f46edfbd713eafc026f7e17ecc Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 13:38:37 +0100
Subject: [PATCH 405/566] using __buffer + tag impls

---
 .../dpl/pstl/hetero/dpcpp/parallel_backend_sycl_utils.h      | 5 +++--
 include/oneapi/dpl/pstl/omp/util.h                           | 5 +++--
 include/oneapi/dpl/pstl/parallel_backend_serial.h            | 5 +++--
 include/oneapi/dpl/pstl/parallel_backend_tbb.h               | 5 +++--
 4 files changed, 12 insertions(+), 8 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_utils.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_utils.h
index 62025411b24..dc3a53305a8 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_utils.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_utils.h
@@ -448,8 +448,9 @@ struct __memobj_traits<_T*>
 
 } // namespace __internal
 
-template <typename _ExecutionPolicy, typename _T>
-using __buffer = __internal::__buffer_impl<::std::decay_t<_ExecutionPolicy>, _T>;
+template <class _Tag, typename _ExecutionPolicy, typename _T>
+using __buffer = ::std::enable_if_t<::std::is_base_of<oneapi::dpl::__internal::__device_backend_tag, _Tag>::value,
+                                    __internal::__buffer_impl<::std::decay_t<_ExecutionPolicy>, _T>>;
 
 template <typename T>
 struct __repacked_tuple
diff --git a/include/oneapi/dpl/pstl/omp/util.h b/include/oneapi/dpl/pstl/omp/util.h
index 292ee6a5c0b..923d4504ea7 100644
--- a/include/oneapi/dpl/pstl/omp/util.h
+++ b/include/oneapi/dpl/pstl/omp/util.h
@@ -82,8 +82,9 @@ class __buffer_impl
     ~__buffer_impl() { __allocator_.deallocate(__ptr_, __buf_size_); }
 };
 
-template <typename _ExecutionPolicy, typename _Tp>
-using __buffer = __buffer_impl<::std::decay_t<_ExecutionPolicy>, _Tp>;
+template <class _Tag, typename _ExecutionPolicy, typename _Tp>
+using __buffer = ::std::enable_if_t<::std::is_same_v<_Tag, oneapi::dpl::__internal::__omp_backend_tag>,
+                                    __buffer_impl<::std::decay_t<_ExecutionPolicy>, _Tp>>;
 
 // Preliminary size of each chunk: requires further discussion
 constexpr std::size_t __default_chunk_size = 2048;
diff --git a/include/oneapi/dpl/pstl/parallel_backend_serial.h b/include/oneapi/dpl/pstl/parallel_backend_serial.h
index fb33353c002..a3f0ce2e38f 100644
--- a/include/oneapi/dpl/pstl/parallel_backend_serial.h
+++ b/include/oneapi/dpl/pstl/parallel_backend_serial.h
@@ -56,8 +56,9 @@ class __buffer_impl
     ~__buffer_impl() { __allocator_.deallocate(__ptr_, __buf_size_); }
 };
 
-template <typename _ExecutionPolicy, typename _Tp>
-using __buffer = __buffer_impl<::std::decay_t<_ExecutionPolicy>, _Tp>;
+template <class _Tag, typename _ExecutionPolicy, typename _Tp>
+using __buffer = ::std::enable_if_t<::std::is_same_v<_Tag, oneapi::dpl::__internal::__serial_backend_tag>,
+                                    __buffer_impl<::std::decay_t<_ExecutionPolicy>, _Tp>>;
 
 inline void
 __cancel_execution(oneapi::dpl::__internal::__serial_backend_tag)
diff --git a/include/oneapi/dpl/pstl/parallel_backend_tbb.h b/include/oneapi/dpl/pstl/parallel_backend_tbb.h
index 2bdd6b466e5..c2f94aab98d 100644
--- a/include/oneapi/dpl/pstl/parallel_backend_tbb.h
+++ b/include/oneapi/dpl/pstl/parallel_backend_tbb.h
@@ -81,8 +81,9 @@ class __buffer_impl
     ~__buffer_impl() { _M_allocator.deallocate(_M_ptr, _M_buf_size); }
 };
 
-template <typename _ExecutionPolicy, typename _Tp>
-using __buffer = __buffer_impl<::std::decay_t<_ExecutionPolicy>, _Tp>;
+template <class _Tag, typename _ExecutionPolicy, typename _Tp>
+using __buffer = ::std::enable_if_t<::std::is_same_v<_Tag, oneapi::dpl::__internal::__tbb_backend_tag>,
+                                    __buffer_impl<::std::decay_t<_ExecutionPolicy>, _Tp>>;
 
 // Wrapper for tbb::task
 inline void

From 6426c628604d227890d6e70b94de4683979f0fd0 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 13:27:55 +0100
Subject: [PATCH 406/566] using __buffer + tag calls

---
 .../internal/exclusive_scan_by_segment_impl.h | 10 +++++----
 .../internal/inclusive_scan_by_segment_impl.h |  6 +++--
 .../dpl/internal/reduce_by_segment_impl.h     | 20 +++++++++++------
 .../dpl/internal/scan_by_segment_impl.h       |  7 ++++--
 include/oneapi/dpl/pstl/algorithm_impl.h      | 20 ++++++++---------
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 22 +++++++++----------
 .../hetero/algorithm_ranges_impl_hetero.h     | 12 +++++-----
 .../pstl/hetero/dpcpp/parallel_backend_sycl.h | 13 ++++++++---
 .../dpcpp/parallel_backend_sycl_histogram.h   |  5 ++++-
 .../dpcpp/parallel_backend_sycl_radix_sort.h  |  2 +-
 .../dpl/pstl/hetero/numeric_impl_hetero.h     |  2 +-
 include/oneapi/dpl/pstl/omp/parallel_scan.h   |  2 +-
 .../oneapi/dpl/pstl/parallel_backend_tbb.h    |  4 ++--
 13 files changed, 74 insertions(+), 51 deletions(-)

diff --git a/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h b/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h
index c447c0c199b..b7e15b1123e 100644
--- a/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h
@@ -46,6 +46,8 @@ pattern_exclusive_scan_by_segment(_Tag, Policy&& policy, InputIterator1 first1,
 {
     static_assert(__internal::__is_backend_tag_v<_Tag>);
 
+    using __backend_tag = typename _Tag::__backend_tag;
+
     const auto n = ::std::distance(first1, last1);
 
     // Check for empty and single element ranges
@@ -64,7 +66,7 @@ pattern_exclusive_scan_by_segment(_Tag, Policy&& policy, InputIterator1 first1,
     InputIterator2 last2 = first2 + n;
 
     // compute head flags
-    oneapi::dpl::__par_backend::__buffer<Policy, FlagType> _flags(n);
+    oneapi::dpl::__par_backend::__buffer<__backend_tag, Policy, FlagType> _flags(n);
     auto flags = _flags.get();
     flags[0] = 1;
 
@@ -72,7 +74,7 @@ pattern_exclusive_scan_by_segment(_Tag, Policy&& policy, InputIterator1 first1,
               oneapi::dpl::__internal::__not_pred<BinaryPredicate>(binary_pred));
 
     // shift input one to the right and initialize segments with init
-    oneapi::dpl::__par_backend::__buffer<Policy, OutputType> _temp(n);
+    oneapi::dpl::__par_backend::__buffer<__backend_tag, Policy, OutputType> _temp(n);
     auto temp = _temp.get();
 
     temp[0] = init;
@@ -123,7 +125,7 @@ exclusive_scan_by_segment_impl(__internal::__hetero_tag<_BackendTag>, Policy&& p
     InputIterator2 last2 = first2 + n;
 
     // compute head flags
-    oneapi::dpl::__par_backend_hetero::__buffer<Policy, FlagType> _flags(policy, n);
+    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, Policy, FlagType> _flags(policy, n);
     {
         auto flag_buf = _flags.get_buffer();
         auto flags = flag_buf.get_host_access(sycl::read_write);
@@ -134,7 +136,7 @@ exclusive_scan_by_segment_impl(__internal::__hetero_tag<_BackendTag>, Policy&& p
               oneapi::dpl::__internal::__not_pred<BinaryPredicate>(binary_pred));
 
     // shift input one to the right and initialize segments with init
-    oneapi::dpl::__par_backend_hetero::__buffer<Policy, OutputType> _temp(policy, n);
+    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, Policy, OutputType> _temp(policy, n);
     {
         auto temp_buf = _temp.get_buffer();
         auto temp = temp_buf.get_host_access(sycl::read_write);
diff --git a/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h b/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h
index 38b377bf601..c15715a0823 100644
--- a/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h
@@ -45,6 +45,8 @@ pattern_inclusive_scan_by_segment(_Tag, Policy&& policy, InputIterator1 first1,
 {
     static_assert(__internal::__is_backend_tag_v<_Tag>);
 
+    using __backend_tag = typename _Tag::__backend_tag;
+
     const auto n = ::std::distance(first1, last1);
 
     // Check for empty and single element ranges
@@ -59,7 +61,7 @@ pattern_inclusive_scan_by_segment(_Tag, Policy&& policy, InputIterator1 first1,
     typedef unsigned int FlagType;
     typedef typename ::std::iterator_traits<InputIterator2>::value_type ValueType;
 
-    oneapi::dpl::__par_backend::__buffer<Policy, FlagType> _mask(n);
+    oneapi::dpl::__par_backend::__buffer<__backend_tag, Policy, FlagType> _mask(n);
     auto mask = _mask.get();
 
     mask[0] = 1;
@@ -109,7 +111,7 @@ inclusive_scan_by_segment_impl(__internal::__hetero_tag<_BackendTag>, Policy&& p
 
     FlagType initial_mask = 1;
 
-    oneapi::dpl::__par_backend_hetero::__buffer<Policy, FlagType> _mask(policy, n);
+    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, Policy, FlagType> _mask(policy, n);
     {
         auto mask_buf = _mask.get_buffer();
         auto mask = mask_buf.get_host_access(sycl::read_write);
diff --git a/include/oneapi/dpl/internal/reduce_by_segment_impl.h b/include/oneapi/dpl/internal/reduce_by_segment_impl.h
index 546237b2ae8..907db9c6b9f 100644
--- a/include/oneapi/dpl/internal/reduce_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/reduce_by_segment_impl.h
@@ -87,6 +87,8 @@ reduce_by_segment_impl(_Tag, Policy&& policy, InputIterator1 first1, InputIterat
 {
     static_assert(__internal::__is_backend_tag_v<_Tag>);
 
+    using __backend_tag = typename _Tag::__backend_tag;
+
     // The algorithm reduces values in [first2, first2 + (last1-first1)) where the associated
     // keys for the values are equal to the adjacent key. This function's implementation is a derivative work
     // and responsible for the second copyright notice in this header.
@@ -114,7 +116,7 @@ reduce_by_segment_impl(_Tag, Policy&& policy, InputIterator1 first1, InputIterat
 
     // buffer that is used to store a flag indicating if the associated key is not equal to
     // the next key, and thus its associated sum should be part of the final result
-    oneapi::dpl::__par_backend::__buffer<Policy, FlagType> _mask(n + 1);
+    oneapi::dpl::__par_backend::__buffer<__backend_tag, Policy, FlagType> _mask(n + 1);
     auto mask = _mask.get();
     mask[0] = 1;
 
@@ -130,11 +132,11 @@ reduce_by_segment_impl(_Tag, Policy&& policy, InputIterator1 first1, InputIterat
     // buffer stores the sums of values associated with a given key. Sums are copied with
     // a shift into result2, and the shift is computed at the same time as the sums, so the
     // sums can't be written to result2 directly.
-    oneapi::dpl::__par_backend::__buffer<Policy, ValueType> _scanned_values(n);
+    oneapi::dpl::__par_backend::__buffer<__backend_tag, Policy, ValueType> _scanned_values(n);
 
     // Buffer is used to store results of the scan of the mask. Values indicate which position
     // in result2 needs to be written with the scanned_values element.
-    oneapi::dpl::__par_backend::__buffer<Policy, FlagType> _scanned_tail_flags(n);
+    oneapi::dpl::__par_backend::__buffer<__backend_tag, Policy, FlagType> _scanned_tail_flags(n);
 
     // Compute the sum of the segments. scanned_tail_flags values are not used.
     inclusive_scan(policy, make_zip_iterator(first2, _mask.get()), make_zip_iterator(first2, _mask.get()) + n,
@@ -263,17 +265,21 @@ __sycl_reduce_by_segment(__internal::__hetero_tag<_BackendTag>, _ExecutionPolicy
 
     // intermediate reductions within a workgroup
     auto __partials =
-        oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, __val_type>(__exec, __n_groups).get_buffer();
+        oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, __val_type>(__exec, __n_groups)
+            .get_buffer();
 
-    auto __end_idx = oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, __diff_type>(__exec, 1).get_buffer();
+    auto __end_idx =
+        oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, __diff_type>(__exec, 1).get_buffer();
 
     // the number of segment ends found in each work group
     auto __seg_ends =
-        oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, __diff_type>(__exec, __n_groups).get_buffer();
+        oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, __diff_type>(__exec, __n_groups)
+            .get_buffer();
 
     // buffer that stores an exclusive scan of the results
     auto __seg_ends_scanned =
-        oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, __diff_type>(__exec, __n_groups).get_buffer();
+        oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, __diff_type>(__exec, __n_groups)
+            .get_buffer();
 
     // 1. Count the segment ends in each workgroup
     auto __seg_end_identification = __exec.queue().submit([&](sycl::handler& __cgh) {
diff --git a/include/oneapi/dpl/internal/scan_by_segment_impl.h b/include/oneapi/dpl/internal/scan_by_segment_impl.h
index 7007e875861..856493f2d2a 100644
--- a/include/oneapi/dpl/internal/scan_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/scan_by_segment_impl.h
@@ -109,6 +109,9 @@ struct __sycl_scan_by_segment_impl
     operator()(_ExecutionPolicy&& __exec, _Range1&& __keys, _Range2&& __values, _Range3&& __out_values,
                _BinaryPredicate __binary_pred, _BinaryOperator __binary_op, _T __init, _T __identity)
     {
+        constexpr auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range1, _Range2, _Range3>();
+        using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
         using _CustomName = oneapi::dpl::__internal::__policy_kernel_name<_ExecutionPolicy>;
 
         using _SegScanWgKernel = oneapi::dpl::__par_backend_hetero::__internal::__kernel_name_generator<
@@ -145,11 +148,11 @@ struct __sycl_scan_by_segment_impl
         ::std::size_t __n_groups = __internal::__dpl_ceiling_div(__n, __wgroup_size * __vals_per_item);
 
         auto __partials =
-            oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, __val_type>(__exec, __n_groups).get_buffer();
+            oneapi::dpl::__par_backend_hetero::__buffer<__backend_tag, _ExecutionPolicy, __val_type>(__exec, __n_groups).get_buffer();
 
         // the number of segment ends found in each work group
         auto __seg_ends =
-            oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, bool>(__exec, __n_groups).get_buffer();
+            oneapi::dpl::__par_backend_hetero::__buffer<__backend_tag, _ExecutionPolicy, bool>(__exec, __n_groups).get_buffer();
 
         // 1. Work group reduction
         auto __wg_scan = __exec.queue().submit([&](sycl::handler& __cgh) {
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 008c261042b..6a742b59b48 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -1330,7 +1330,7 @@ __pattern_copy_if(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomA
     const _DifferenceType __n = __last - __first;
     if (_DifferenceType(1) < __n)
     {
-        __par_backend::__buffer<_ExecutionPolicy, bool> __mask_buf(__n);
+        __par_backend::__buffer<__backend_tag, _ExecutionPolicy, bool> __mask_buf(__n);
         return __internal::__except_handler([&__exec, __n, __first, __result, __pred, &__mask_buf]() {
             bool* __mask = __mask_buf.get();
             _DifferenceType __m{};
@@ -1449,7 +1449,7 @@ __remove_elements(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Forward
     typedef typename ::std::iterator_traits<_ForwardIterator>::difference_type _DifferenceType;
     typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _Tp;
     _DifferenceType __n = __last - __first;
-    __par_backend::__buffer<_ExecutionPolicy, bool> __mask_buf(__n);
+    __par_backend::__buffer<__backend_tag, _ExecutionPolicy, bool> __mask_buf(__n);
     // 1. find a first iterator that should be removed
     return __internal::__except_handler([&]() {
         bool* __mask = __mask_buf.get();
@@ -1486,7 +1486,7 @@ __remove_elements(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Forward
         __n -= __min;
         __first += __min;
 
-        __par_backend::__buffer<_ExecutionPolicy, _Tp> __buf(__n);
+        __par_backend::__buffer<__backend_tag, _ExecutionPolicy, _Tp> __buf(__n);
         _Tp* __result = __buf.get();
         __mask += __min;
         _DifferenceType __m{};
@@ -1614,7 +1614,7 @@ __pattern_unique_copy(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Ran
     const _DifferenceType __n = __last - __first;
     if (_DifferenceType(2) < __n)
     {
-        __par_backend::__buffer<_ExecutionPolicy, bool> __mask_buf(__n);
+        __par_backend::__buffer<__backend_tag, _ExecutionPolicy, bool> __mask_buf(__n);
         if (_DifferenceType(2) < __n)
         {
             return __internal::__except_handler([&__exec, __n, __first, __result, __pred, &__mask_buf]() {
@@ -1855,7 +1855,7 @@ __pattern_rotate(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAc
     auto __m = __middle - __first;
     if (__m <= __n / 2)
     {
-        __par_backend::__buffer<_ExecutionPolicy, _Tp> __buf(__n - __m);
+        __par_backend::__buffer<__backend_tag, _ExecutionPolicy, _Tp> __buf(__n - __m);
         return __internal::__except_handler([&__exec, __n, __m, __first, __middle, __last, &__buf]() {
             _Tp* __result = __buf.get();
             __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __middle, __last,
@@ -1881,7 +1881,7 @@ __pattern_rotate(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAc
     }
     else
     {
-        __par_backend::__buffer<_ExecutionPolicy, _Tp> __buf(__m);
+        __par_backend::__buffer<__backend_tag, _ExecutionPolicy, _Tp> __buf(__m);
         return __internal::__except_handler([&__exec, __n, __m, __first, __middle, __last, &__buf]() {
             _Tp* __result = __buf.get();
             __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __middle,
@@ -2356,7 +2356,7 @@ __pattern_partition_copy(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _
     const _DifferenceType __n = __last - __first;
     if (_DifferenceType(1) < __n)
     {
-        __par_backend::__buffer<_ExecutionPolicy, bool> __mask_buf(__n);
+        __par_backend::__buffer<__backend_tag, _ExecutionPolicy, bool> __mask_buf(__n);
         return __internal::__except_handler([&__exec, __n, __first, __out_true, __out_false, __pred, &__mask_buf]() {
             bool* __mask = __mask_buf.get();
             _ReturnType __m{};
@@ -2582,7 +2582,7 @@ __pattern_partial_sort_copy(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec
         {
             typedef typename ::std::iterator_traits<_RandomAccessIterator1>::value_type _T1;
             typedef typename ::std::iterator_traits<_RandomAccessIterator2>::value_type _T2;
-            __par_backend::__buffer<_ExecutionPolicy, _T1> __buf(__n1);
+            __par_backend::__buffer<__backend_tag, _ExecutionPolicy, _T1> __buf(__n1);
             _T1* __r = __buf.get();
 
             __par_backend::__parallel_stable_sort(
@@ -3075,7 +3075,7 @@ __pattern_inplace_merge(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _R
 
     typedef typename ::std::iterator_traits<_RandomAccessIterator>::value_type _Tp;
     auto __n = __last - __first;
-    __par_backend::__buffer<_ExecutionPolicy, _Tp> __buf(__n);
+    __par_backend::__buffer<__backend_tag, _ExecutionPolicy, _Tp> __buf(__n);
     _Tp* __r = __buf.get();
     __internal::__except_handler([&]() {
         auto __move_values = [](_RandomAccessIterator __x, _Tp* __z) {
@@ -3209,7 +3209,7 @@ __parallel_set_op(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Forward
     const _DifferenceType __n1 = __last1 - __first1;
     const _DifferenceType __n2 = __last2 - __first2;
 
-    __par_backend::__buffer<_ExecutionPolicy, _T> __buf(__size_func(__n1, __n2));
+    __par_backend::__buffer<__backend_tag, _ExecutionPolicy, _T> __buf(__size_func(__n1, __n2));
 
     return __internal::__except_handler([&__exec, __n1, __first1, __last1, __first2, __last2, __result, __comp,
                                          __size_func, __set_op, &__buf]() {
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 105f805c313..6611378952a 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -1034,7 +1034,7 @@ __pattern_remove_if(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec,
 
     using _ValueType = typename ::std::iterator_traits<_Iterator>::value_type;
 
-    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __buf(__exec, __last - __first);
+    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _ValueType> __buf(__exec, __last - __first);
     auto __copy_first = __buf.get();
 
     auto __copy_last = __pattern_copy_if(__tag, __exec, __first, __last, __copy_first, __not_pred<_Predicate>{__pred});
@@ -1055,7 +1055,7 @@ __pattern_unique(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _It
 
     using _ValueType = typename ::std::iterator_traits<_Iterator>::value_type;
 
-    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __buf(__exec, __last - __first);
+    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _ValueType> __buf(__exec, __last - __first);
     auto __copy_first = __buf.get();
     auto __copy_last = __pattern_unique_copy(__tag, __exec, __first, __last, __copy_first, __pred);
 
@@ -1227,7 +1227,7 @@ __pattern_inplace_merge(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __ex
     assert(__first < __middle && __middle < __last);
 
     auto __n = __last - __first;
-    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __buf(__exec, __n);
+    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _ValueType> __buf(__exec, __n);
     auto __copy_first = __buf.get();
     auto __copy_last = __copy_first + __n;
 
@@ -1315,8 +1315,8 @@ __pattern_stable_partition(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& _
 
     auto __n = __last - __first;
 
-    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __true_buf(__exec, __n);
-    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __false_buf(__exec, __n);
+    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _ValueType> __true_buf(__exec, __n);
+    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _ValueType> __false_buf(__exec, __n);
     auto __true_result = __true_buf.get();
     auto __false_result = __false_buf.get();
 
@@ -1512,7 +1512,7 @@ __pattern_partial_sort_copy(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&&
         // - create a temporary buffer and copy all the elements from the input buffer there
         // - run partial sort on the temporary buffer
         // - copy k elements from the temporary buffer to the output buffer.
-        oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __buf(__exec, __in_size);
+        oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _ValueType> __buf(__exec, __in_size);
 
         auto __buf_first = __buf.get();
 
@@ -1629,7 +1629,7 @@ __pattern_rotate(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Iterator
 
     auto __keep = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read_write, _Iterator>();
     auto __buf = __keep(__first, __last);
-    auto __temp_buf = oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _Tp>(__exec, __n);
+    auto __temp_buf = oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _Tp>(__exec, __n);
 
     auto __temp_rng =
         oneapi::dpl::__ranges::all_view<_Tp, __par_backend_hetero::access_mode::write>(__temp_buf.get_buffer());
@@ -1712,7 +1712,7 @@ __pattern_hetero_set_op(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _F
         __comp, __n1, __n2};
 
     // temporary buffer to store boolean mask
-    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, int32_t> __mask_buf(__exec, __n1);
+    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, int32_t> __mask_buf(__exec, __n1);
 
     auto __keep1 =
         oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _ForwardIterator1>();
@@ -1838,7 +1838,7 @@ __pattern_set_union(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec,
 
     // temporary buffer to store intermediate result
     const auto __n2 = __last2 - __first2;
-    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __diff(__exec, __n2);
+    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _ValueType> __diff(__exec, __n2);
     auto __buf = __diff.get();
 
     //1. Calc difference {2} \ {1}
@@ -1917,10 +1917,10 @@ __pattern_set_symmetric_difference(__hetero_tag<_BackendTag> __tag, _ExecutionPo
 
     // temporary buffers to store intermediate result
     const auto __n1 = __last1 - __first1;
-    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __diff_1(__exec, __n1);
+    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _ValueType> __diff_1(__exec, __n1);
     auto __buf_1 = __diff_1.get();
     const auto __n2 = __last2 - __first2;
-    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __diff_2(__exec, __n2);
+    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _ValueType> __diff_2(__exec, __n2);
     auto __buf_2 = __diff_2.get();
 
     //1. Calc difference {1} \ {2}
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
index daefb904b51..73d93a15086 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
@@ -359,7 +359,7 @@ __pattern_scan_copy(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Range
     _DataAcc __get_data_op;
     _MaskAssigner __add_mask_op;
 
-    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, int32_t> __mask_buf(__exec, __rng1.size());
+    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, int32_t> __mask_buf(__exec, __rng1.size());
 
     auto __res =
         __par_backend_hetero::__parallel_transform_scan_base(
@@ -412,7 +412,7 @@ __pattern_remove_if(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec,
 
     using _ValueType = oneapi::dpl::__internal::__value_t<_Range>;
 
-    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __buf(__exec, __rng.size());
+    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _ValueType> __buf(__exec, __rng.size());
     auto __copy_rng = oneapi::dpl::__ranges::views::all(__buf.get_buffer());
 
     auto __copy_last_id = __ranges::__pattern_copy_if(__tag, __exec, __rng, __copy_rng, __not_pred<_Predicate>{__pred},
@@ -459,7 +459,7 @@ __pattern_unique(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Ra
 
     using _ValueType = oneapi::dpl::__internal::__value_t<_Range>;
 
-    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __buf(__exec, __rng.size());
+    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _ValueType> __buf(__exec, __rng.size());
     auto res_rng = oneapi::dpl::__ranges::views::all(__buf.get_buffer());
     auto res = __ranges::__pattern_unique_copy(__tag, __exec, __rng, res_rng, __pred, oneapi::dpl::__internal::__pstl_assign());
 
@@ -712,11 +712,11 @@ __pattern_reduce_by_segment(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&&
     // Round 1: reduce with extra indices added to avoid long segments
     // TODO: At threshold points check if the key is equal to the key at the previous threshold point, indicating a long sequence.
     // Skip a round of copy_if and reduces if there are none.
-    auto __idx = oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, __diff_type>(__exec, __n).get_buffer();
+    auto __idx = oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, __diff_type>(__exec, __n).get_buffer();
     auto __tmp_out_keys =
-        oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, __key_type>(__exec, __n).get_buffer();
+        oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, __key_type>(__exec, __n).get_buffer();
     auto __tmp_out_values =
-        oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, __val_type>(__exec, __n).get_buffer();
+        oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, __val_type>(__exec, __n).get_buffer();
 
     // Replicating first element of keys view to be able to compare (i-1)-th and (i)-th key with aligned sequences,
     //  dropping the last key for the i-1 sequence.
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
index e0f7e3a4a5c..dd2505c57ec 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
@@ -890,7 +890,8 @@ __parallel_scan_copy(oneapi::dpl::__internal::__device_backend_tag __backend_tag
     _MaskAssigner __add_mask_op;
 
     // temporary buffer to store boolean mask
-    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, int32_t> __mask_buf(__exec, __n);
+    oneapi::dpl::__par_backend_hetero::__buffer<oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy, int32_t>
+        __mask_buf(__exec, __n);
 
     return __parallel_transform_scan_base(
         __backend_tag, ::std::forward<_ExecutionPolicy>(__exec),
@@ -1575,6 +1576,9 @@ struct __parallel_sort_submitter<_IdType, __internal::__optional_kernel_name<_Le
     auto
     operator()(_ExecutionPolicy&& __exec, _Range&& __rng, _Compare __comp) const
     {
+        constexpr auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range>();
+        using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
         using _Tp = oneapi::dpl::__internal::__value_t<_Range>;
         using _Size = oneapi::dpl::__internal::__difference_t<_Range>;
 
@@ -1596,7 +1600,7 @@ struct __parallel_sort_submitter<_IdType, __internal::__optional_kernel_name<_Le
         });
 
         // 2. Merge sorting
-        oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _Tp> __temp_buf(__exec, __n);
+        oneapi::dpl::__par_backend_hetero::__buffer<__backend_tag, _ExecutionPolicy, _Tp> __temp_buf(__exec, __n);
         auto __temp = __temp_buf.get_buffer();
         bool __data_in_temp = false;
         _IdType __n_sorted = __leaf;
@@ -1709,13 +1713,16 @@ struct __parallel_partial_sort_submitter<__internal::__optional_kernel_name<_Glo
     auto
     operator()(_ExecutionPolicy&& __exec, _Range&& __rng, _Merge __merge, _Compare __comp) const
     {
+        constexpr auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range>();
+        using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
         using _Tp = oneapi::dpl::__internal::__value_t<_Range>;
         using _Size = oneapi::dpl::__internal::__difference_t<_Range>;
 
         _Size __n = __rng.size();
         assert(__n > 1);
 
-        oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _Tp> __temp_buf(__exec, __n);
+        oneapi::dpl::__par_backend_hetero::__buffer<__backend_tag, _ExecutionPolicy, _Tp> __temp_buf(__exec, __n);
         auto __temp = __temp_buf.get_buffer();
         _PRINT_INFO_IN_DEBUG_MODE(__exec);
 
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h
index 2e7223ec062..0c3cb721ed9 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h
@@ -411,6 +411,9 @@ struct __histogram_general_private_global_atomics_submitter<__internal::__option
                ::std::uint16_t __work_group_size, _Range1&& __input, _Range2&& __bins,
                const _BinHashMgr& __binhash_manager)
     {
+        constexpr auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range1, _Range2>();
+        using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
+
         const ::std::size_t __n = __input.size();
         const ::std::size_t __num_bins = __bins.size();
         using _bin_type = oneapi::dpl::__internal::__value_t<_Range2>;
@@ -426,7 +429,7 @@ struct __histogram_general_private_global_atomics_submitter<__internal::__option
             oneapi::dpl::__internal::__dpl_ceiling_div(__n, __work_group_size * __iters_per_work_item);
 
         auto __private_histograms =
-            oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _bin_type>(__exec, __segments * __num_bins)
+            oneapi::dpl::__par_backend_hetero::__buffer<__backend_tag, _ExecutionPolicy, _bin_type>(__exec, __segments * __num_bins)
                 .get_buffer();
 
         return __exec.queue().submit([&](auto& __h) {
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_radix_sort.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_radix_sort.h
index df262d54dfc..9ae0b9f788c 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_radix_sort.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_radix_sort.h
@@ -813,7 +813,7 @@ __parallel_radix_sort(oneapi::dpl::__internal::__device_backend_tag, _ExecutionP
         sycl::buffer<::std::uint32_t, 1> __tmp_buf{sycl::range<1>(__tmp_buf_size)};
 
         // memory for storing values sorted for an iteration
-        oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueT> __out_buffer_holder{__exec, __n};
+        oneapi::dpl::__par_backend_hetero::__buffer<oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy, _ValueT> __out_buffer_holder{__exec, __n};
         auto __out_rng = oneapi::dpl::__ranges::all_view<_ValueT, __par_backend_hetero::access_mode::read_write>(
             __out_buffer_holder.get_buffer());
 
diff --git a/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
index 2955815d29d..ff089c083d3 100644
--- a/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
@@ -164,7 +164,7 @@ __pattern_transform_scan_base(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&
         using _NewExecutionPolicy = decltype(__policy);
 
         // Create temporary buffer
-        oneapi::dpl::__par_backend_hetero::__buffer<_NewExecutionPolicy, _Type> __tmp_buf(__policy, __n);
+        oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _NewExecutionPolicy, _Type> __tmp_buf(__policy, __n);
         auto __first_tmp = __tmp_buf.get();
         auto __last_tmp = __first_tmp + __n;
         auto __keep2 = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::write, _Iterator2>();
diff --git a/include/oneapi/dpl/pstl/omp/parallel_scan.h b/include/oneapi/dpl/pstl/omp/parallel_scan.h
index d3609152319..90e0a150152 100644
--- a/include/oneapi/dpl/pstl/omp/parallel_scan.h
+++ b/include/oneapi/dpl/pstl/omp/parallel_scan.h
@@ -88,7 +88,7 @@ __parallel_strict_scan_body(_Index __n, _Tp __initial, _Rp __reduce, _Cp __combi
     const _Index __slack = 4;
     _Index __tilesize = (__n - 1) / (__slack * __p) + 1;
     _Index __m = (__n - 1) / __tilesize;
-    __buffer<_ExecutionPolicy, _Tp> __buf(__m + 1);
+    __buffer<oneapi::dpl::__internal::__omp_backend_tag, _ExecutionPolicy, _Tp> __buf(__m + 1);
     _Tp* __r = __buf.get();
 
     oneapi::dpl::__omp_backend::__upsweep(_Index(0), _Index(__m + 1), __tilesize, __r, __n - __m * __tilesize, __reduce,
diff --git a/include/oneapi/dpl/pstl/parallel_backend_tbb.h b/include/oneapi/dpl/pstl/parallel_backend_tbb.h
index c2f94aab98d..659596d932c 100644
--- a/include/oneapi/dpl/pstl/parallel_backend_tbb.h
+++ b/include/oneapi/dpl/pstl/parallel_backend_tbb.h
@@ -391,7 +391,7 @@ __parallel_strict_scan(oneapi::dpl::__internal::__tbb_backend_tag, _ExecutionPol
             const _Index __slack = 4;
             _Index __tilesize = (__n - 1) / (__slack * __p) + 1;
             _Index __m = (__n - 1) / __tilesize;
-            __tbb_backend::__buffer<_ExecutionPolicy, _Tp> __buf(__m + 1);
+            __tbb_backend::__buffer<oneapi::dpl::__internal::__tbb_backend_tag, _ExecutionPolicy, _Tp> __buf(__m + 1);
             _Tp* __r = __buf.get();
             __tbb_backend::__upsweep(_Index(0), _Index(__m + 1), __tilesize, __r, __n - __m * __tilesize, __reduce,
                                      __combine);
@@ -1196,7 +1196,7 @@ __parallel_stable_sort(oneapi::dpl::__internal::__tbb_backend_tag, _ExecutionPol
         const _DifferenceType __sort_cut_off = _ONEDPL_STABLE_SORT_CUT_OFF;
         if (__n > __sort_cut_off)
         {
-            __tbb_backend::__buffer<_ExecutionPolicy, _ValueType> __buf(__n);
+            __tbb_backend::__buffer<oneapi::dpl::__internal::__tbb_backend_tag, _ExecutionPolicy, _ValueType> __buf(__n);
             __root_task<__stable_sort_func<_RandomAccessIterator, _ValueType*, _Compare, _LeafSort>> __root{
                 __xs, __xe, __buf.get(), true, __comp, __leaf_sort, __nsort, __xs, __buf.get()};
             __task::spawn_root_and_wait(__root);

From 32b7a3a6da9af5354864581308c9576bc8078642 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 15:14:54 +0100
Subject: [PATCH 407/566] Remove fpga_policy::__device_policy() and it's usage
 as not required anymore

---
 .../oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h    | 8 --------
 .../dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h    | 4 ++--
 2 files changed, 2 insertions(+), 10 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h b/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h
index 9f72220225a..ce935ff1ddb 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h
@@ -107,14 +107,6 @@ class fpga_policy : public device_policy<KernelName>
     fpga_policy(const fpga_policy<other_factor, OtherName>& other) : base(other.queue()){};
     explicit fpga_policy(sycl::queue q) : base(q) {}
     explicit fpga_policy(sycl::device d) : base(d) {}
-
-    // For internal use only
-
-    const base&
-    __device_policy() const
-    {
-        return static_cast<const base&>(*this);
-    };
 };
 
 #endif // _ONEDPL_FPGA_DEVICE
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
index 08ccc58287c..0aea9efd28a 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
@@ -104,8 +104,8 @@ __parallel_histogram(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionPoli
 
     // workaround until we implement more performant version for patterns
     return oneapi::dpl::__par_backend_hetero::__parallel_histogram(
-        oneapi::dpl::__internal::__device_backend_tag{}, __exec.__device_policy(), __init_event,
-        ::std::forward<_Range1>(__input), ::std::forward<_Range2>(__bins), __binhash_manager);
+        oneapi::dpl::__internal::__device_backend_tag{}, __exec, __init_event, ::std::forward<_Range1>(__input),
+        ::std::forward<_Range2>(__bins), __binhash_manager);
 }
 
 } // namespace __par_backend_hetero

From a11b4173bc5a7c61fbf92d91a8b6107dfb724c6d Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 15:59:27 +0100
Subject: [PATCH 408/566] include/oneapi/dpl/pstl/execution_impl.h - fix self
 review comments

---
 include/oneapi/dpl/pstl/execution_impl.h | 45 ++++++++++++++++++------
 1 file changed, 35 insertions(+), 10 deletions(-)

diff --git a/include/oneapi/dpl/pstl/execution_impl.h b/include/oneapi/dpl/pstl/execution_impl.h
index 319dfef617f..32d63dd20f9 100644
--- a/include/oneapi/dpl/pstl/execution_impl.h
+++ b/include/oneapi/dpl/pstl/execution_impl.h
@@ -174,41 +174,66 @@ __select_backend()
     return {};
 }
 
+//----------------------------------------------------------
+// __is_backend_tag, __is_backend_tag_v
+//----------------------------------------------------------
+
 template <class _Tag>
 struct __is_backend_tag : ::std::false_type
 {
 };
 
-template <class _Tag>
-struct __is_backend_tag_serial : ::std::false_type
+template <class _IsVector>
+struct __is_backend_tag<__serial_tag<_IsVector>> : ::std::true_type
 {
 };
 
 template <class _IsVector>
-struct __is_backend_tag<__serial_tag<_IsVector>> : ::std::true_type
+struct __is_backend_tag<__parallel_tag<_IsVector>> : ::std::true_type
 {
 };
 
-template <class _IsVector>
-struct __is_backend_tag_serial<__serial_tag<_IsVector>> : ::std::true_type
+template <>
+struct __is_backend_tag<__parallel_forward_tag> : ::std::true_type
 {
 };
 
-template <class _IsVector>
-struct __is_backend_tag<__parallel_tag<_IsVector>> : ::std::true_type
+template <class _Tag>
+inline constexpr bool __is_backend_tag_v = __is_backend_tag<::std::decay_t<_Tag>>::value;
+
+//----------------------------------------------------------
+// __is_backend_tag_serial, __is_backend_tag_serial_v
+//----------------------------------------------------------
+
+template <class _Tag>
+struct __is_backend_tag_serial : ::std::false_type
 {
 };
 
-template <>
-struct __is_backend_tag<__parallel_forward_tag> : ::std::true_type
+template <class _IsVector>
+struct __is_backend_tag_serial<__serial_tag<_IsVector>> : ::std::true_type
 {
 };
 
 template <class _Tag>
 inline constexpr bool __is_backend_tag_serial_v = __is_backend_tag_serial<::std::decay_t<_Tag>>::value;
 
+//----------------------------------------------------------
+// __is_backend_tag_parallel_forward, __is_backend_tag_parallel_forward_v
+//----------------------------------------------------------
+
 template <class _Tag>
-inline constexpr bool __is_backend_tag_v = __is_backend_tag<::std::decay_t<_Tag>>::value;
+struct __is_backend_tag_parallel_forward : ::std::false_type
+{
+};
+
+template <>
+struct __is_backend_tag_parallel_forward<__parallel_forward_tag> : ::std::true_type
+{
+};
+
+template <class _Tag>
+inline constexpr bool __is_backend_tag_parallel_forward_v = __is_backend_tag_parallel_forward<::std::decay_t<_Tag>>::value;
 
 } // namespace __internal
 } // namespace dpl

From 42e970268b03c2a1126850ca0d0b9b01c68db70d Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 17:33:37 +0100
Subject: [PATCH 409/566] Fix review comment: required to check
 __is_backend_tag_serial_v or __is_backend_tag_parallel_forward_v instead of
 __is_backend_tag_v

---
 .../oneapi/dpl/internal/binary_search_impl.h  |  12 +++
 .../internal/exclusive_scan_by_segment_impl.h |   4 +
 .../dpl/internal/reduce_by_segment_impl.h     |   2 +-
 include/oneapi/dpl/pstl/algorithm_impl.h      | 102 +++++++++---------
 .../experimental/internal/for_loop_impl.h     |   8 +-
 include/oneapi/dpl/pstl/histogram_impl.h      |   2 +-
 include/oneapi/dpl/pstl/numeric_impl.h        |  12 ++-
 7 files changed, 81 insertions(+), 61 deletions(-)

diff --git a/include/oneapi/dpl/internal/binary_search_impl.h b/include/oneapi/dpl/internal/binary_search_impl.h
index 11e47d01417..821e5e869b4 100644
--- a/include/oneapi/dpl/internal/binary_search_impl.h
+++ b/include/oneapi/dpl/internal/binary_search_impl.h
@@ -74,7 +74,11 @@ OutputIterator
 lower_bound_impl(_Tag, Policy&& policy, InputIterator1 start, InputIterator1 end, InputIterator2 value_start,
                  InputIterator2 value_end, OutputIterator result, StrictWeakOrdering comp)
 {
+//#if _ONEDPL_BACKEND_SYCL
+//    static_assert(__internal::__is_backend_tag_serial_v<_Tag> || __internal::__is_backend_tag_parallel_forward_v<_Tag>);
+//#else
     static_assert(__internal::__is_backend_tag_v<_Tag>);
+//#endif
 
     return oneapi::dpl::transform(policy, value_start, value_end, result,
                                   [=](typename ::std::iterator_traits<InputIterator2>::reference val) {
@@ -88,7 +92,11 @@ OutputIterator
 upper_bound_impl(_Tag, Policy&& policy, InputIterator1 start, InputIterator1 end, InputIterator2 value_start,
                  InputIterator2 value_end, OutputIterator result, StrictWeakOrdering comp)
 {
+//#if _ONEDPL_BACKEND_SYCL
+//    static_assert(__internal::__is_backend_tag_serial_v<_Tag> || __internal::__is_backend_tag_parallel_forward_v<_Tag>);
+//#else
     static_assert(__internal::__is_backend_tag_v<_Tag>);
+//#endif
 
     return oneapi::dpl::transform(policy, value_start, value_end, result,
                                   [=](typename ::std::iterator_traits<InputIterator2>::reference val) {
@@ -102,7 +110,11 @@ OutputIterator
 binary_search_impl(_Tag, Policy&& policy, InputIterator1 start, InputIterator1 end, InputIterator2 value_start,
                    InputIterator2 value_end, OutputIterator result, StrictWeakOrdering comp)
 {
+//#if _ONEDPL_BACKEND_SYCL
+//    static_assert(__internal::__is_backend_tag_serial_v<_Tag> || __internal::__is_backend_tag_parallel_forward_v<_Tag>);
+//#else
     static_assert(__internal::__is_backend_tag_v<_Tag>);
+//#endif
 
     return oneapi::dpl::transform(policy, value_start, value_end, result,
                                   [=](typename ::std::iterator_traits<InputIterator2>::reference val) {
diff --git a/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h b/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h
index b7e15b1123e..2ff180251a0 100644
--- a/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h
@@ -44,7 +44,11 @@ pattern_exclusive_scan_by_segment(_Tag, Policy&& policy, InputIterator1 first1,
                                   InputIterator2 first2, OutputIterator result, T init, BinaryPredicate binary_pred,
                                   Operator binary_op)
 {
+//#if _ONEDPL_BACKEND_SYCL
+//    static_assert(__internal::__is_backend_tag_serial_v<_Tag> || __internal::__is_backend_tag_parallel_forward_v<_Tag>);
+//#else
     static_assert(__internal::__is_backend_tag_v<_Tag>);
+//#endif
 
     using __backend_tag = typename _Tag::__backend_tag;
 
diff --git a/include/oneapi/dpl/internal/reduce_by_segment_impl.h b/include/oneapi/dpl/internal/reduce_by_segment_impl.h
index 907db9c6b9f..1a613089e21 100644
--- a/include/oneapi/dpl/internal/reduce_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/reduce_by_segment_impl.h
@@ -85,7 +85,7 @@ reduce_by_segment_impl(_Tag, Policy&& policy, InputIterator1 first1, InputIterat
                        OutputIterator1 result1, OutputIterator2 result2, BinaryPred binary_pred,
                        BinaryOperator binary_op)
 {
-    static_assert(__internal::__is_backend_tag_v<_Tag>);
+    static_assert(__internal::__is_backend_tag_serial_v<_Tag> || __internal::__is_backend_tag_parallel_forward_v<_Tag>);
 
     using __backend_tag = typename _Tag::__backend_tag;
 
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 6a742b59b48..15fd9d6fed3 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -67,7 +67,7 @@ template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Pre
 bool
 __pattern_any_of(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _Pred __pred) noexcept
 {
-    static_assert(__is_backend_tag_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_any_of(__first, __last, __pred, typename _Tag::__is_vector{});
 }
@@ -179,7 +179,7 @@ void
 __pattern_walk_brick(_Tag, _ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last,
                      _Brick __brick) noexcept
 {
-    static_assert(__is_backend_tag_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     __brick(__first, __last, typename _Tag::__is_vector{});
 }
@@ -221,7 +221,7 @@ template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Siz
 _ForwardIterator
 __pattern_walk1_n(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _Size __n, _Function __f) noexcept
 {
-    static_assert(__is_backend_tag_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_walk1_n(__first, __n, __f, typename _Tag::__is_vector{});
 }
@@ -240,7 +240,7 @@ template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Siz
 _ForwardIterator
 __pattern_walk_brick_n(_Tag, _ExecutionPolicy&& __exec, _ForwardIterator __first, _Size __n, _Brick __brick) noexcept
 {
-    static_assert(__is_backend_tag_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __brick(__first, __n, typename _Tag::__is_vector{});
 }
@@ -364,7 +364,7 @@ _ForwardIterator2
 __pattern_walk2_n(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _Size __n, _ForwardIterator2 __first2,
                   _Function __f) noexcept
 {
-    static_assert(__is_backend_tag_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_walk2_n(__first1, __n, __first2, __f, typename _Tag::__is_vector{});
 }
@@ -461,7 +461,7 @@ _ForwardIterator2
 __pattern_walk2_brick_n(_Tag, _ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _Size __n,
                         _ForwardIterator2 __first2, _Brick __brick) noexcept
 {
-    static_assert(__is_backend_tag_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __brick(__first1, __n, __first2, typename _Tag::__is_vector{});
 }
@@ -608,7 +608,7 @@ bool
 __pattern_equal(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
                 _ForwardIterator2 __first2, _ForwardIterator2 __last2, _BinaryPredicate __p) noexcept
 {
-    static_assert(__is_backend_tag_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_equal(__first1, __last1, __first2, __last2, __p, typename _Tag::__is_vector{});
 }
@@ -659,7 +659,7 @@ bool
 __pattern_equal(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
                 _ForwardIterator2 __first2, _BinaryPredicate __p) noexcept
 {
-    static_assert(__is_backend_tag_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_equal(__first1, __last1, __first2, __p, typename _Tag::__is_vector{});
 }
@@ -706,7 +706,7 @@ _ForwardIterator
 __pattern_find_if(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                   _Predicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_find_if(__first, __last, __pred, typename _Tag::__is_vector{});
 }
@@ -849,7 +849,7 @@ _ForwardIterator1
 __pattern_find_end(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first, _ForwardIterator1 __last,
                    _ForwardIterator2 __s_first, _ForwardIterator2 __s_last, _BinaryPredicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_find_end(__first, __last, __s_first, __s_last, __pred, typename _Tag::__is_vector{});
 }
@@ -905,7 +905,7 @@ _ForwardIterator1
 __pattern_find_first_of(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first, _ForwardIterator1 __last,
                         _ForwardIterator2 __s_first, _ForwardIterator2 __s_last, _BinaryPredicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_find_first_of(__first, __last, __s_first, __s_last, __pred,
                                              typename _Tag::__is_vector{});
@@ -954,7 +954,7 @@ _ForwardIterator1
 __pattern_search(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first, _ForwardIterator1 __last,
                  _ForwardIterator2 __s_first, _ForwardIterator2 __s_last, _BinaryPredicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_search(__first, __last, __s_first, __s_last, __pred, typename _Tag::__is_vector{});
 }
@@ -1010,7 +1010,7 @@ _ForwardIterator
 __pattern_search_n(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _Size __count,
                    const _Tp& __value, _BinaryPredicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_search_n(__first, __last, __count, __value, __pred, typename _Tag::__is_vector{});
 }
@@ -1313,7 +1313,7 @@ _OutputIterator
 __pattern_copy_if(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _OutputIterator __result,
                   _UnaryPredicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_copy_if(__first, __last, __result, __pred, typename _Tag::__is_vector{});
 }
@@ -1378,7 +1378,7 @@ template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Pre
 typename ::std::iterator_traits<_ForwardIterator>::difference_type
 __pattern_count(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _Predicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_count(__first, __last, __pred, typename _Tag::__is_vector{});
 }
@@ -1432,7 +1432,7 @@ _ForwardIterator
 __pattern_unique(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                  _BinaryPredicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_unique(__first, __last, __pred, typename _Tag::__is_vector{});
 }
@@ -1575,7 +1575,7 @@ _OutputIterator
 __pattern_unique_copy(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                       _OutputIterator __result, _BinaryPredicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_unique_copy(__first, __last, __result, __pred, typename _Tag::__is_vector{});
 }
@@ -1711,7 +1711,7 @@ template <class _Tag, class _ExecutionPolicy, class _BidirectionalIterator>
 void
 __pattern_reverse(_Tag, _ExecutionPolicy&&, _BidirectionalIterator __first, _BidirectionalIterator __last) noexcept
 {
-    static_assert(__is_backend_tag_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     __internal::__brick_reverse(__first, __last, typename _Tag::__is_vector{});
 }
@@ -1759,7 +1759,7 @@ _OutputIterator
 __pattern_reverse_copy(_Tag, _ExecutionPolicy&&, _BidirectionalIterator __first, _BidirectionalIterator __last,
                        _OutputIterator __d_first) noexcept
 {
-    static_assert(__is_backend_tag_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_reverse_copy(__first, __last, __d_first, typename _Tag::__is_vector{});
 }
@@ -1838,7 +1838,7 @@ _ForwardIterator
 __pattern_rotate(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __middle,
                  _ForwardIterator __last) noexcept
 {
-    static_assert(__is_backend_tag_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_rotate(__first, __middle, __last, typename _Tag::__is_vector{});
 }
@@ -1934,7 +1934,7 @@ _OutputIterator
 __pattern_rotate_copy(_Tag, _ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __middle,
                       _ForwardIterator __last, _OutputIterator __result) noexcept
 {
-    static_assert(__is_backend_tag_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_rotate_copy(::std::forward<_ExecutionPolicy>(__exec), __first, __middle, __last,
                                            __result, typename _Tag::__is_vector{});
@@ -2016,7 +2016,7 @@ bool
 __pattern_is_partitioned(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                          _UnaryPredicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_is_partitioned(__first, __last, __pred, typename _Tag::__is_vector{});
 }
@@ -2146,7 +2146,7 @@ _ForwardIterator
 __pattern_partition(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                     _UnaryPredicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_partition(__first, __last, __pred, typename _Tag::__is_vector{});
 }
@@ -2246,7 +2246,7 @@ _BidirectionalIterator
 __pattern_stable_partition(_Tag, _ExecutionPolicy&&, _BidirectionalIterator __first, _BidirectionalIterator __last,
                            _UnaryPredicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_stable_partition(__first, __last, __pred, typename _Tag::__is_vector{});
 }
@@ -2336,7 +2336,7 @@ ::std::pair<_OutputIterator1, _OutputIterator2>
 __pattern_partition_copy(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                          _OutputIterator1 __out_true, _OutputIterator2 __out_false, _UnaryPredicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_partition_copy(__first, __last, __out_true, __out_false, __pred,
                                               typename _Tag::__is_vector{});
@@ -2392,7 +2392,7 @@ void
 __pattern_sort(_Tag, _ExecutionPolicy&&, _RandomAccessIterator __first, _RandomAccessIterator __last, _Compare __comp,
                _IsMoveConstructible) noexcept
 {
-    static_assert(__is_backend_tag_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     ::std::sort(__first, __last, __comp);
 }
@@ -2424,7 +2424,7 @@ void
 __pattern_stable_sort(_Tag, _ExecutionPolicy&&, _RandomAccessIterator __first, _RandomAccessIterator __last,
                       _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     ::std::stable_sort(__first, __last, __comp);
 }
@@ -2457,7 +2457,7 @@ __pattern_sort_by_key(_Tag, _ExecutionPolicy&& __exec, _RandomAccessIterator1 __
                       _RandomAccessIterator1 __keys_last, _RandomAccessIterator2 __values_first,
                       _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     auto __beg = oneapi::dpl::make_zip_iterator(__keys_first, __values_first);
     auto __end = __beg + (__keys_last - __keys_first);
@@ -2503,7 +2503,7 @@ void
 __pattern_partial_sort(_Tag, _ExecutionPolicy&&, _RandomAccessIterator __first, _RandomAccessIterator __middle,
                        _RandomAccessIterator __last, _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     ::std::partial_sort(__first, __middle, __last, __comp);
 }
@@ -2541,7 +2541,7 @@ _RandomAccessIterator
 __pattern_partial_sort_copy(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                             _RandomAccessIterator __d_first, _RandomAccessIterator __d_last, _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return ::std::partial_sort_copy(__first, __last, __d_first, __d_last, __comp);
 }
@@ -2645,7 +2645,7 @@ _ForwardIterator
 __pattern_adjacent_find(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                         _BinaryPredicate __pred, _Semantic) noexcept
 {
-    static_assert(__is_backend_tag_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_adjacent_find(__first, __last, __pred, typename _Tag::__is_vector{}, _Semantic::value);
 }
@@ -2705,7 +2705,7 @@ void
 __pattern_nth_element(_Tag, _ExecutionPolicy&&, _RandomAccessIterator __first, _RandomAccessIterator __nth,
                       _RandomAccessIterator __last, _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     ::std::nth_element(__first, __nth, __last, __comp);
 }
@@ -2780,7 +2780,7 @@ template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Tp>
 void
 __pattern_fill(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, const _Tp& __value) noexcept
 {
-    static_assert(__is_backend_tag_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     __internal::__brick_fill<_Tp, _ExecutionPolicy>{__value}(__first, __last, typename _Tag::__is_vector{});
 }
@@ -2829,7 +2829,7 @@ template <class _Tag, class _ExecutionPolicy, class _OutputIterator, class _Size
 _OutputIterator
 __pattern_fill_n(_Tag, _ExecutionPolicy&&, _OutputIterator __first, _Size __count, const _Tp& __value) noexcept
 {
-    static_assert(__is_backend_tag_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_fill_n<_Tp, _ExecutionPolicy>{__value}(__first, __count, typename _Tag::__is_vector{});
 }
@@ -2866,7 +2866,7 @@ template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Gen
 void
 __pattern_generate(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _Generator __g) noexcept
 {
-    static_assert(__is_backend_tag_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     __internal::__brick_generate(__first, __last, __g, typename _Tag::__is_vector{});
 }
@@ -2906,7 +2906,7 @@ template <class _Tag, class _ExecutionPolicy, class _OutputIterator, class _Size
 _OutputIterator
 __pattern_generate_n(_Tag, _ExecutionPolicy&&, _OutputIterator __first, _Size __count, _Generator __g) noexcept
 {
-    static_assert(__is_backend_tag_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_generate_n(__first, __count, __g, typename _Tag::__is_vector{});
 }
@@ -2951,7 +2951,7 @@ _ForwardIterator
 __pattern_remove_if(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                     _UnaryPredicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_remove_if(__first, __last, __pred, typename _Tag::__is_vector{});
 }
@@ -3007,7 +3007,7 @@ __pattern_merge(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIt
                 _ForwardIterator2 __first2, _ForwardIterator2 __last2, _OutputIterator __d_first,
                 _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_merge(__first1, __last1, __first2, __last2, __d_first, __comp,
                                      typename _Tag::__is_vector{});
@@ -3056,7 +3056,7 @@ void
 __pattern_inplace_merge(_Tag, _ExecutionPolicy&&, _BidirectionalIterator __first, _BidirectionalIterator __middle,
                         _BidirectionalIterator __last, _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     __internal::__brick_inplace_merge(__first, __middle, __last, __comp, typename _Tag::__is_vector{});
 }
@@ -3115,7 +3115,7 @@ bool
 __pattern_includes(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
                    _ForwardIterator2 __first2, _ForwardIterator2 __last2, _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return ::std::includes(__first1, __last1, __first2, __last2, __comp);
 }
@@ -3430,7 +3430,7 @@ __pattern_set_union(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _Forwa
                     _ForwardIterator2 __first2, _ForwardIterator2 __last2, _OutputIterator __result,
                     _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_set_union(__first1, __last1, __first2, __last2, __result, __comp,
                                          typename _Tag::__is_vector{});
@@ -3491,7 +3491,7 @@ __pattern_set_intersection(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1,
                            _ForwardIterator2 __first2, _ForwardIterator2 __last2, _OutputIterator __result,
                            _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_set_intersection(__first1, __last1, __first2, __last2, __result, __comp,
                                                 typename _Tag::__is_vector{});
@@ -3589,7 +3589,7 @@ __pattern_set_difference(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _
                          _ForwardIterator2 __first2, _ForwardIterator2 __last2, _OutputIterator __result,
                          _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_set_difference(__first1, __last1, __first2, __last2, __result, __comp,
                                               typename _Tag::__is_vector{});
@@ -3676,7 +3676,7 @@ __pattern_set_symmetric_difference(_Tag, _ExecutionPolicy&&, _ForwardIterator1 _
                                    _ForwardIterator2 __first2, _ForwardIterator2 __last2, _OutputIterator __result,
                                    _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_set_symmetric_difference(__first1, __last1, __first2, __last2, __result, __comp,
                                                         typename _Tag::__is_vector{});
@@ -3735,7 +3735,7 @@ _RandomAccessIterator
 __pattern_is_heap_until(_Tag, _ExecutionPolicy&&, _RandomAccessIterator __first, _RandomAccessIterator __last,
                         _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_is_heap_until(__first, __last, __comp, typename _Tag::__is_vector{});
 }
@@ -3823,7 +3823,7 @@ bool
 __pattern_is_heap(_Tag, _ExecutionPolicy&&, _RandomAccessIterator __first, _RandomAccessIterator __last,
                   _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_is_heap(__first, __last, __comp, typename _Tag::__is_vector{});
 }
@@ -3871,7 +3871,7 @@ _ForwardIterator
 __pattern_min_element(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                       _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_min_element(__first, __last, __comp, typename _Tag::__is_vector{});
 }
@@ -3938,7 +3938,7 @@ ::std::pair<_ForwardIterator, _ForwardIterator>
 __pattern_minmax_element(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                          _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_minmax_element(__first, __last, __comp, typename _Tag::__is_vector{});
 }
@@ -4019,7 +4019,7 @@ ::std::pair<_ForwardIterator1, _ForwardIterator2>
 __pattern_mismatch(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
                    _ForwardIterator2 __first2, _ForwardIterator2 __last2, _Predicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_mismatch(__first1, __last1, __first2, __last2, __pred, typename _Tag::__is_vector{});
 }
@@ -4100,7 +4100,7 @@ bool
 __pattern_lexicographical_compare(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
                                   _ForwardIterator2 __first2, _ForwardIterator2 __last2, _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_lexicographical_compare(__first1, __last1, __first2, __last2, __comp,
                                                        typename _Tag::__is_vector{});
@@ -4202,7 +4202,7 @@ _ForwardIterator
 __pattern_shift_left(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                      typename ::std::iterator_traits<_ForwardIterator>::difference_type __n) noexcept
 {
-    static_assert(__is_backend_tag_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __brick_shift_left(__first, __last, __n, typename _Tag::__is_vector{});
 }
diff --git a/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h b/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
index 13d72c4fab4..d3b6094c288 100644
--- a/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
+++ b/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
@@ -167,7 +167,7 @@ void
 __pattern_for_loop_n(_Tag, _ExecutionPolicy&&, _Ip __first, _Size __n, _Function __f, __single_stride_type,
                      _Rest&&... __rest) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     __reduction_pack<_Rest...> __pack{__reduction_pack_tag(), ::std::forward<_Rest>(__rest)...};
 
@@ -192,7 +192,7 @@ void
 __pattern_for_loop_n(_Tag, _ExecutionPolicy&&, _Ip __first, _Size __n, _Function __f, _Sp __stride,
                      _Rest&&... __rest) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     __reduction_pack<_Rest...> __pack{__reduction_pack_tag(), ::std::forward<_Rest>(__rest)...};
 
@@ -265,7 +265,7 @@ void
 __pattern_for_loop(_Tag __tag, _ExecutionPolicy&& __exec, _Ip __first, _Ip __last, _Function __f, _Sp __stride,
                    _Rest&&... __rest) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     if constexpr (__is_random_access_or_integral_v<_Ip>)
     {
@@ -366,7 +366,7 @@ ::std::enable_if_t<!__is_random_access_or_integral_v<_Ip>>
 __pattern_for_loop(_Tag, _ExecutionPolicy&&, _Ip __first, _Ip __last, _Function __f, __single_stride_type,
                    _Rest&&... __rest) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     __reduction_pack<_Rest...> __pack{__reduction_pack_tag(), ::std::forward<_Rest>(__rest)...};
 
diff --git a/include/oneapi/dpl/pstl/histogram_impl.h b/include/oneapi/dpl/pstl/histogram_impl.h
index b9fe40422f9..8e75800e2af 100644
--- a/include/oneapi/dpl/pstl/histogram_impl.h
+++ b/include/oneapi/dpl/pstl/histogram_impl.h
@@ -38,7 +38,7 @@ void
 __pattern_histogram(_Tag, _ExecutionPolicy&& exec, _RandomAccessIterator1 __first, _RandomAccessIterator1 __last,
                     _Size __num_bins, _IdxHashFunc __func, _RandomAccessIterator2 __histogram_first)
 {
-    static_assert(__is_backend_tag_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     static_assert(sizeof(_Size) == 0 /*false*/,
                   "Histogram API is currently unsupported for policies other than device execution policies");
diff --git a/include/oneapi/dpl/pstl/numeric_impl.h b/include/oneapi/dpl/pstl/numeric_impl.h
index 5e7f4dcbc0d..7aeac1afc4b 100644
--- a/include/oneapi/dpl/pstl/numeric_impl.h
+++ b/include/oneapi/dpl/pstl/numeric_impl.h
@@ -68,7 +68,7 @@ __pattern_transform_reduce(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1,
                            _ForwardIterator2 __first2, _Tp __init, _BinaryOperation1 __binary_op1,
                            _BinaryOperation2 __binary_op2) noexcept
 {
-    static_assert(__is_backend_tag_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __brick_transform_reduce(__first1, __last1, __first2, __init, __binary_op1, __binary_op2,
                                     typename _Tag::__is_vector{});
@@ -133,7 +133,7 @@ _Tp
 __pattern_transform_reduce(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _Tp __init,
                            _BinaryOperation __binary_op, _UnaryOperation __unary_op) noexcept
 {
-    static_assert(__is_backend_tag_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_transform_reduce(__first, __last, __init, __binary_op, __unary_op,
                                                 typename _Tag::__is_vector{});
@@ -243,7 +243,7 @@ __pattern_transform_scan(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _Fo
                          _OutputIterator __result, _UnaryOperation __unary_op, _Tp __init, _BinaryOperation __binary_op,
                          _Inclusive) noexcept
 {
-    static_assert(__is_backend_tag_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_transform_scan(__first, __last, __result, __unary_op, __init, __binary_op, _Inclusive(),
                                               typename _Tag::__is_vector{})
@@ -327,7 +327,11 @@ _OutputIterator
 __pattern_transform_scan(_Tag __tag, _ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last,
                          _OutputIterator __result, _UnaryOperation __unary_op, _BinaryOperation __binary_op, _Inclusive)
 {
+//#if _ONEDPL_BACKEND_SYCL
+//    static_assert(is_backend_tag_serial_v<_Tag> || is_backend_tag_parallel_forward_v<_Tag>);
+//#else
     static_assert(__is_backend_tag_v<_Tag>);
+//#endif
 
     typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _ValueType;
     if (__first != __last)
@@ -379,7 +383,7 @@ _OutputIterator
 __pattern_adjacent_difference(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                               _OutputIterator __d_first, _BinaryOperation __op) noexcept
 {
-    static_assert(__is_backend_tag_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_adjacent_difference(__first, __last, __d_first, __op, typename _Tag::__is_vector{});
 }

From b70d807dd743725eb08baee6305e2672a60aac80 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 17:30:55 +0100
Subject: [PATCH 410/566] include/oneapi/dpl/pstl/algorithm_impl.h - remove
 extra static_assert(__is_backend_tag_v<_Tag>); checks

---
 include/oneapi/dpl/pstl/algorithm_impl.h | 8 --------
 1 file changed, 8 deletions(-)

diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 15fd9d6fed3..d6430ead636 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -561,8 +561,6 @@ _ForwardIterator2
 __pattern_walk2_transform_if(_Tag __tag, _ExecutionPolicy&& __exec, _ForwardIterator1 __first1,
                              _ForwardIterator1 __last1, _ForwardIterator2 __first2, _Function __func) noexcept
 {
-    static_assert(__is_backend_tag_v<_Tag>);
-
     return __pattern_walk2(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __func);
 }
 
@@ -573,8 +571,6 @@ __pattern_walk3_transform_if(_Tag __tag, _ExecutionPolicy&& __exec, _ForwardIter
                              _ForwardIterator1 __last1, _ForwardIterator2 __first2, _ForwardIterator3 __first3,
                              _Function __func) noexcept
 {
-    static_assert(__is_backend_tag_v<_Tag>);
-
     return __pattern_walk3(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __first3,
                            __func);
 }
@@ -4124,8 +4120,6 @@ _ForwardIterator2
 __pattern_swap(_Tag __tag, _ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
                _ForwardIterator2 __first2, _Function __f)
 {
-    static_assert(__is_backend_tag_v<_Tag>);
-
     return __pattern_walk2(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __f);
 }
 
@@ -4258,8 +4252,6 @@ __pattern_shift_right(_Tag __tag, _ExecutionPolicy&& __exec, _BidirectionalItera
                       _BidirectionalIterator __last,
                       typename ::std::iterator_traits<_BidirectionalIterator>::difference_type __n)
 {
-    static_assert(__is_backend_tag_v<_Tag>);
-
     using _ReverseIterator = typename ::std::reverse_iterator<_BidirectionalIterator>;
 
     auto __res = oneapi::dpl::__internal::__pattern_shift_left(

From ef69023449eb5f3f0f4e06163b0919624ce028bb Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 15:16:37 +0100
Subject: [PATCH 411/566] test/general/test_tag_dispatching.pass.cpp - new test
 for tag_dispatching checks

---
 test/general/test_tag_dispatching.pass.cpp | 100 +++++++++++++++++++++
 1 file changed, 100 insertions(+)
 create mode 100644 test/general/test_tag_dispatching.pass.cpp

diff --git a/test/general/test_tag_dispatching.pass.cpp b/test/general/test_tag_dispatching.pass.cpp
new file mode 100644
index 00000000000..c5eeca41e5f
--- /dev/null
+++ b/test/general/test_tag_dispatching.pass.cpp
@@ -0,0 +1,100 @@
+// -*- C++ -*-
+//===----------------------------------------------------------------------===//
+//
+// Copyright (C) Intel Corporation
+//
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+// This file incorporates work covered by the following copyright and permission
+// notice:
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+//
+//===----------------------------------------------------------------------===//
+
+#include "support/test_config.h"
+
+#include _PSTL_TEST_HEADER(execution)
+#include _PSTL_TEST_HEADER(algorithm)
+
+#include "support/utils.h"
+
+void
+test_backend_tags()
+{
+    using namespace oneapi::dpl::__internal;
+
+    // template <class _IsVector>
+    // struct __serial_tag;
+    static_assert(__is_backend_tag_v<__serial_tag<std::true_type>>);
+    static_assert(__is_backend_tag_v<__serial_tag<std::false_type>>);
+
+    // template <class _IsVector>
+    // struct __parallel_tag
+    static_assert(__is_backend_tag_v<__parallel_tag<std::true_type>>);
+    static_assert(__is_backend_tag_v<__parallel_tag<std::false_type>>);
+
+    // struct __parallel_forward_tag
+    static_assert(__is_backend_tag_v<__parallel_forward_tag>);
+
+
+#if TEST_DPCPP_BACKEND_PRESENT
+
+    // template <typename _BackendTag>
+    // struct __hetero_tag
+    static_assert(__is_backend_tag_v<__hetero_tag<__device_backend_tag>>);
+
+#if ONEDPL_FPGA_DEVICE
+
+    // template <typename _BackendTag>
+    // struct __hetero_tag
+    static_assert(__is_backend_tag_v<__hetero_tag<__fpga_backend_tag>>);
+
+#endif // ONEDPL_FPGA_DEVICE
+#endif // TEST_DPCPP_BACKEND_PRESENT
+}
+
+void
+test_backend_tags_serial()
+{
+    using namespace oneapi::dpl::__internal;
+
+    // template <class _IsVector>
+    // struct __serial_tag;
+    static_assert(__is_backend_tag_serial_v<__serial_tag<std::true_type>>);
+    static_assert(__is_backend_tag_serial_v<__serial_tag<std::false_type>>);
+
+    // struct __parallel_forward_tag
+    static_assert(!__is_backend_tag_serial_v<__parallel_forward_tag>);
+
+    // template <class _IsVector>
+    // struct __parallel_tag
+    static_assert(!__is_backend_tag_serial_v<__parallel_tag<std::true_type>>);
+    static_assert(!__is_backend_tag_serial_v<__parallel_tag<std::false_type>>);
+
+#if TEST_DPCPP_BACKEND_PRESENT
+
+    // template <typename _BackendTag>
+    // struct __hetero_tag
+    static_assert(!__is_backend_tag_serial_v<__hetero_tag<__device_backend_tag>>);
+
+#if ONEDPL_FPGA_DEVICE
+
+    // template <typename _BackendTag>
+    // struct __hetero_tag
+    static_assert(!__is_backend_tag_serial_v<__hetero_tag<__fpga_backend_tag>>);
+
+#endif // ONEDPL_FPGA_DEVICE
+#endif // TEST_DPCPP_BACKEND_PRESENT
+}
+
+std::int32_t
+main()
+{
+    test_backend_tags();
+    test_backend_tags_serial();
+
+    return TestUtils::done(TEST_DPCPP_BACKEND_PRESENT);
+}
+

From f92eb6151c732af6dc5a1be7e841a446faa19846 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 19:40:56 +0100
Subject: [PATCH 412/566] Revert "test/general/test_tag_dispatching.pass.cpp -
 new test for tag_dispatching checks"

This reverts commit ef69023449eb5f3f0f4e06163b0919624ce028bb.
---
 test/general/test_tag_dispatching.pass.cpp | 100 ---------------------
 1 file changed, 100 deletions(-)
 delete mode 100644 test/general/test_tag_dispatching.pass.cpp

diff --git a/test/general/test_tag_dispatching.pass.cpp b/test/general/test_tag_dispatching.pass.cpp
deleted file mode 100644
index c5eeca41e5f..00000000000
--- a/test/general/test_tag_dispatching.pass.cpp
+++ /dev/null
@@ -1,100 +0,0 @@
-// -*- C++ -*-
-//===----------------------------------------------------------------------===//
-//
-// Copyright (C) Intel Corporation
-//
-// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
-//
-// This file incorporates work covered by the following copyright and permission
-// notice:
-//
-// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
-// See https://llvm.org/LICENSE.txt for license information.
-//
-//===----------------------------------------------------------------------===//
-
-#include "support/test_config.h"
-
-#include _PSTL_TEST_HEADER(execution)
-#include _PSTL_TEST_HEADER(algorithm)
-
-#include "support/utils.h"
-
-void
-test_backend_tags()
-{
-    using namespace oneapi::dpl::__internal;
-
-    // template <class _IsVector>
-    // struct __serial_tag;
-    static_assert(__is_backend_tag_v<__serial_tag<std::true_type>>);
-    static_assert(__is_backend_tag_v<__serial_tag<std::false_type>>);
-
-    // template <class _IsVector>
-    // struct __parallel_tag
-    static_assert(__is_backend_tag_v<__parallel_tag<std::true_type>>);
-    static_assert(__is_backend_tag_v<__parallel_tag<std::false_type>>);
-
-    // struct __parallel_forward_tag
-    static_assert(__is_backend_tag_v<__parallel_forward_tag>);
-
-
-#if TEST_DPCPP_BACKEND_PRESENT
-
-    // template <typename _BackendTag>
-    // struct __hetero_tag
-    static_assert(__is_backend_tag_v<__hetero_tag<__device_backend_tag>>);
-
-#if ONEDPL_FPGA_DEVICE
-
-    // template <typename _BackendTag>
-    // struct __hetero_tag
-    static_assert(__is_backend_tag_v<__hetero_tag<__fpga_backend_tag>>);
-
-#endif // ONEDPL_FPGA_DEVICE
-#endif // TEST_DPCPP_BACKEND_PRESENT
-}
-
-void
-test_backend_tags_serial()
-{
-    using namespace oneapi::dpl::__internal;
-
-    // template <class _IsVector>
-    // struct __serial_tag;
-    static_assert(__is_backend_tag_serial_v<__serial_tag<std::true_type>>);
-    static_assert(__is_backend_tag_serial_v<__serial_tag<std::false_type>>);
-
-    // struct __parallel_forward_tag
-    static_assert(!__is_backend_tag_serial_v<__parallel_forward_tag>);
-
-    // template <class _IsVector>
-    // struct __parallel_tag
-    static_assert(!__is_backend_tag_serial_v<__parallel_tag<std::true_type>>);
-    static_assert(!__is_backend_tag_serial_v<__parallel_tag<std::false_type>>);
-
-#if TEST_DPCPP_BACKEND_PRESENT
-
-    // template <typename _BackendTag>
-    // struct __hetero_tag
-    static_assert(!__is_backend_tag_serial_v<__hetero_tag<__device_backend_tag>>);
-
-#if ONEDPL_FPGA_DEVICE
-
-    // template <typename _BackendTag>
-    // struct __hetero_tag
-    static_assert(!__is_backend_tag_serial_v<__hetero_tag<__fpga_backend_tag>>);
-
-#endif // ONEDPL_FPGA_DEVICE
-#endif // TEST_DPCPP_BACKEND_PRESENT
-}
-
-std::int32_t
-main()
-{
-    test_backend_tags();
-    test_backend_tags_serial();
-
-    return TestUtils::done(TEST_DPCPP_BACKEND_PRESENT);
-}
-

From 8986026315841b04c6d693d88b2ac5e9b9ff91be Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 18:32:21 +0100
Subject: [PATCH 413/566] Implementation and usage of
 __select_backend_for_loop()

---
 .../experimental/internal/for_loop_impl.h     | 44 +++++++++----------
 1 file changed, 20 insertions(+), 24 deletions(-)

diff --git a/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h b/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
index d3b6094c288..263b9effebc 100644
--- a/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
+++ b/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
@@ -529,48 +529,44 @@ __use_parallelization() -> decltype(__use_par_vec_helper<_Ip>::template __use_pa
     return __use_par_vec_helper<_Ip>::template __use_parallel<_ExecutionPolicy>();
 }
 
-// Helper functions to extract to separate a Callable object from the pack of reductions and inductions
-template <typename _ExecutionPolicy, typename _Ip, typename _Fp, typename _Sp, typename... _Rest, ::std::size_t... _Is>
-void
-__for_loop_impl(_ExecutionPolicy&& __exec, _Ip __start, _Ip __finish, _Fp&& __f, _Sp __stride,
-                ::std::tuple<_Rest...>&& __t, ::std::index_sequence<_Is...>)
+template <typename _ExecutionPolicy, typename _Ip>
+constexpr auto
+__select_backend_for_loop()
 {
     using _IsVector = decltype(oneapi::dpl::__internal::__use_vectorization<_ExecutionPolicy, _Ip>());
 
     if constexpr (oneapi::dpl::__internal::__use_parallelization<_ExecutionPolicy, _Ip>())
     {
-        oneapi::dpl::__internal::__pattern_for_loop(__parallel_tag<_IsVector>{},
-                                                    ::std::forward<_ExecutionPolicy>(__exec), __start, __finish, __f,
-                                                    __stride, ::std::get<_Is>(::std::move(__t))...);
+        return __parallel_tag<_IsVector>{};
     }
     else
     {
-        oneapi::dpl::__internal::__pattern_for_loop(__serial_tag<_IsVector>{}, ::std::forward<_ExecutionPolicy>(__exec),
-                                                    __start, __finish, __f, __stride,
-                                                    ::std::get<_Is>(::std::move(__t))...);
+        return __serial_tag<_IsVector>{};
     }
 }
 
+// Helper functions to extract to separate a Callable object from the pack of reductions and inductions
+template <typename _ExecutionPolicy, typename _Ip, typename _Fp, typename _Sp, typename... _Rest, ::std::size_t... _Is>
+void
+__for_loop_impl(_ExecutionPolicy&& __exec, _Ip __start, _Ip __finish, _Fp&& __f, _Sp __stride,
+                ::std::tuple<_Rest...>&& __t, ::std::index_sequence<_Is...>)
+{
+    constexpr auto __dispatch_tag = __select_backend_for_loop<_ExecutionPolicy, _Ip>();
+
+    oneapi::dpl::__internal::__pattern_for_loop(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __start,
+                                                __finish, __f, __stride, ::std::get<_Is>(::std::move(__t))...);
+}
+
 template <typename _ExecutionPolicy, typename _Ip, typename _Size, typename _Fp, typename _Sp, typename... _Rest,
           ::std::size_t... _Is>
 void
 __for_loop_n_impl(_ExecutionPolicy&& __exec, _Ip __start, _Size __n, _Fp&& __f, _Sp __stride,
                   ::std::tuple<_Rest...>&& __t, ::std::index_sequence<_Is...>)
 {
-    using _IsVector = decltype(oneapi::dpl::__internal::__use_vectorization<_ExecutionPolicy, _Ip>());
+    constexpr auto __dispatch_tag = __select_backend_for_loop<_ExecutionPolicy, _Ip>();
 
-    if constexpr (oneapi::dpl::__internal::__use_parallelization<_ExecutionPolicy, _Ip>())
-    {
-        oneapi::dpl::__internal::__pattern_for_loop_n(__parallel_tag<_IsVector>{},
-                                                      ::std::forward<_ExecutionPolicy>(__exec), __start, __n, __f,
-                                                      __stride, ::std::get<_Is>(::std::move(__t))...);
-    }
-    else
-    {
-        oneapi::dpl::__internal::__pattern_for_loop_n(__serial_tag<_IsVector>{},
-                                                      ::std::forward<_ExecutionPolicy>(__exec), __start, __n, __f,
-                                                      __stride, ::std::get<_Is>(::std::move(__t))...);
-    }
+    oneapi::dpl::__internal::__pattern_for_loop_n(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __start,
+                                                  __n, __f, __stride, ::std::get<_Is>(::std::move(__t))...);
 }
 
 template <typename _ExecutionPolicy, typename _Ip, typename _Sp, typename... _Rest>

From 094e2b08b51a7926b09c177bd6474cb19e87cb09 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 18:38:03 +0100
Subject: [PATCH 414/566] Remove __use_vectorization() and
 __use_parallelization()

---
 .../pstl/experimental/internal/for_loop_impl.h | 18 ++----------------
 1 file changed, 2 insertions(+), 16 deletions(-)

diff --git a/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h b/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
index 263b9effebc..ef444fb6cac 100644
--- a/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
+++ b/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
@@ -515,27 +515,13 @@ struct __use_par_vec_helper<_Ip, ::std::enable_if_t<!::std::is_integral_v<_Ip>>>
 };
 
 // Special versions for for_loop: handles both iterators and integral types(treated as random access iterators)
-template <typename _ExecutionPolicy, typename _Ip>
-constexpr auto
-__use_vectorization() -> decltype(__use_par_vec_helper<_Ip>::template __use_vector<_ExecutionPolicy>())
-{
-    return __use_par_vec_helper<_Ip>::template __use_vector<_ExecutionPolicy>();
-}
-
-template <typename _ExecutionPolicy, typename _Ip>
-constexpr auto
-__use_parallelization() -> decltype(__use_par_vec_helper<_Ip>::template __use_parallel<_ExecutionPolicy>())
-{
-    return __use_par_vec_helper<_Ip>::template __use_parallel<_ExecutionPolicy>();
-}
-
 template <typename _ExecutionPolicy, typename _Ip>
 constexpr auto
 __select_backend_for_loop()
 {
-    using _IsVector = decltype(oneapi::dpl::__internal::__use_vectorization<_ExecutionPolicy, _Ip>());
+    using _IsVector = decltype(__use_par_vec_helper<_Ip>::template __use_vector<_ExecutionPolicy>());
 
-    if constexpr (oneapi::dpl::__internal::__use_parallelization<_ExecutionPolicy, _Ip>())
+    if constexpr (__use_par_vec_helper<_Ip>::template __use_parallel<_ExecutionPolicy>())
     {
         return __parallel_tag<_IsVector>{};
     }

From e85cde0d51e9a4149a1277b298b333a080cd9d5e Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 18:38:25 +0100
Subject: [PATCH 415/566] Remove __is_vectorization_preferred()

---
 include/oneapi/dpl/pstl/execution_impl.h | 8 --------
 1 file changed, 8 deletions(-)

diff --git a/include/oneapi/dpl/pstl/execution_impl.h b/include/oneapi/dpl/pstl/execution_impl.h
index 32d63dd20f9..d9f966474b0 100644
--- a/include/oneapi/dpl/pstl/execution_impl.h
+++ b/include/oneapi/dpl/pstl/execution_impl.h
@@ -77,14 +77,6 @@ using __allow_unsequenced = typename __internal::__policy_traits<::std::decay_t<
 template <typename _ExecutionPolicy>
 using __allow_parallel = typename __internal::__policy_traits<::std::decay_t<_ExecutionPolicy>>::__allow_parallel;
 
-template <typename _ExecutionPolicy, typename... _IteratorTypes>
-constexpr auto
-__is_vectorization_preferred()
-{
-    return ::std::conjunction<decltype(::std::decay_t<_ExecutionPolicy>::__allow_vector()),
-                              __internal::__is_random_access_iterator_t<_IteratorTypes...>>();
-}
-
 template <typename _ExecutionPolicy, typename... _IteratorTypes>
 constexpr auto
 __is_parallelization_preferred()

From e69d9587b7bf5cc489abc1b1ae531cd8651a8e40 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 18:38:47 +0100
Subject: [PATCH 416/566] Remove __is_parallelization_preferred()

---
 include/oneapi/dpl/pstl/execution_impl.h | 8 --------
 1 file changed, 8 deletions(-)

diff --git a/include/oneapi/dpl/pstl/execution_impl.h b/include/oneapi/dpl/pstl/execution_impl.h
index d9f966474b0..865959bba38 100644
--- a/include/oneapi/dpl/pstl/execution_impl.h
+++ b/include/oneapi/dpl/pstl/execution_impl.h
@@ -77,14 +77,6 @@ using __allow_unsequenced = typename __internal::__policy_traits<::std::decay_t<
 template <typename _ExecutionPolicy>
 using __allow_parallel = typename __internal::__policy_traits<::std::decay_t<_ExecutionPolicy>>::__allow_parallel;
 
-template <typename _ExecutionPolicy, typename... _IteratorTypes>
-constexpr auto
-__is_parallelization_preferred()
-{
-    return ::std::conjunction<decltype(::std::decay_t<_ExecutionPolicy>::__allow_parallel()),
-                              __internal::__is_random_access_iterator_t<_IteratorTypes...>>();
-}
-
 //------------------------------------------------------------------------
 // backend selector with tags
 //------------------------------------------------------------------------

From fc6b7a7ac722b246c40ca12ad4a74c09132bc2ee Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 18:40:42 +0100
Subject: [PATCH 417/566] Remove struct __use_par_vec_helper

---
 .../experimental/internal/for_loop_impl.h     | 49 ++-----------------
 1 file changed, 3 insertions(+), 46 deletions(-)

diff --git a/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h b/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
index ef444fb6cac..58511ed2375 100644
--- a/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
+++ b/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
@@ -473,61 +473,18 @@ __pattern_for_loop(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _
         ::std::forward<_Rest>(__rest)...);
 }
 
-// Helper structure to split code functions for integral and iterator types so the return
-// value can be successfully deduced.
-template <typename _Ip, typename = void>
-struct __use_par_vec_helper;
-
-template <typename _Ip>
-struct __use_par_vec_helper<_Ip, ::std::enable_if_t<::std::is_integral_v<_Ip>>>
-{
-    template <typename _ExecutionPolicy>
-    static constexpr auto
-    __use_vector()
-    {
-        return ::std::decay_t<_ExecutionPolicy>::__allow_vector();
-    }
-
-    template <typename _ExecutionPolicy>
-    static constexpr auto
-    __use_parallel()
-    {
-        return ::std::decay_t<_ExecutionPolicy>::__allow_parallel();
-    }
-};
-
-template <typename _Ip>
-struct __use_par_vec_helper<_Ip, ::std::enable_if_t<!::std::is_integral_v<_Ip>>>
-{
-    template <typename _ExecutionPolicy>
-    static constexpr auto
-    __use_vector()
-    {
-        return oneapi::dpl::__internal::__is_vectorization_preferred<_ExecutionPolicy, _Ip>();
-    }
-
-    template <typename _ExecutionPolicy>
-    static constexpr auto
-    __use_parallel()
-    {
-        return oneapi::dpl::__internal::__is_parallelization_preferred<_ExecutionPolicy, _Ip>();
-    }
-};
-
 // Special versions for for_loop: handles both iterators and integral types(treated as random access iterators)
 template <typename _ExecutionPolicy, typename _Ip>
 constexpr auto
 __select_backend_for_loop()
 {
-    using _IsVector = decltype(__use_par_vec_helper<_Ip>::template __use_vector<_ExecutionPolicy>());
-
-    if constexpr (__use_par_vec_helper<_Ip>::template __use_parallel<_ExecutionPolicy>())
+    if constexpr (::std::is_integral_v<_Ip>)
     {
-        return __parallel_tag<_IsVector>{};
+        return __select_backend<_ExecutionPolicy>();
     }
     else
     {
-        return __serial_tag<_IsVector>{};
+        return __select_backend<_ExecutionPolicy, _Ip>();
     }
 }
 

From 73724c35e5e0fdd73b61e226182f49394598e365 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 18:41:33 +0100
Subject: [PATCH 418/566] Remove __allow_unsequenced(), __allow_vector() and
 __allow_parallel() from oneDPL execution policies

---
 include/oneapi/dpl/pstl/execution_defs.h      | 68 -------------------
 .../pstl/hetero/dpcpp/execution_sycl_defs.h   | 18 -----
 2 files changed, 86 deletions(-)

diff --git a/include/oneapi/dpl/pstl/execution_defs.h b/include/oneapi/dpl/pstl/execution_defs.h
index 6d1d0efd569..98b99f6cce5 100644
--- a/include/oneapi/dpl/pstl/execution_defs.h
+++ b/include/oneapi/dpl/pstl/execution_defs.h
@@ -31,88 +31,20 @@ inline namespace v1
 // 2.4, Sequential execution policy
 class sequenced_policy
 {
-  public:
-    // For internal use only
-    static constexpr ::std::false_type
-    __allow_unsequenced()
-    {
-        return ::std::false_type{};
-    }
-    static constexpr ::std::false_type
-    __allow_vector()
-    {
-        return ::std::false_type{};
-    }
-    static constexpr ::std::false_type
-    __allow_parallel()
-    {
-        return ::std::false_type{};
-    }
 };
 
 // 2.5, Parallel execution policy
 class parallel_policy
 {
-  public:
-    // For internal use only
-    static constexpr ::std::false_type
-    __allow_unsequenced()
-    {
-        return ::std::false_type{};
-    }
-    static constexpr ::std::false_type
-    __allow_vector()
-    {
-        return ::std::false_type{};
-    }
-    static constexpr ::std::true_type
-    __allow_parallel()
-    {
-        return ::std::true_type{};
-    }
 };
 
 // 2.6, Parallel+Vector execution policy
 class parallel_unsequenced_policy
 {
-  public:
-    // For internal use only
-    static constexpr ::std::true_type
-    __allow_unsequenced()
-    {
-        return ::std::true_type{};
-    }
-    static constexpr ::std::true_type
-    __allow_vector()
-    {
-        return ::std::true_type{};
-    }
-    static constexpr ::std::true_type
-    __allow_parallel()
-    {
-        return ::std::true_type{};
-    }
 };
 
 class unsequenced_policy
 {
-  public:
-    // For internal use only
-    static constexpr ::std::true_type
-    __allow_unsequenced()
-    {
-        return ::std::true_type{};
-    }
-    static constexpr ::std::true_type
-    __allow_vector()
-    {
-        return ::std::true_type{};
-    }
-    static constexpr ::std::false_type
-    __allow_parallel()
-    {
-        return ::std::false_type{};
-    }
 };
 
 // 2.8, Execution policy objects
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h b/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h
index ce935ff1ddb..3cdddc16efc 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h
@@ -60,24 +60,6 @@ class device_policy
         return q;
     }
 
-    // For internal use only
-    static constexpr ::std::true_type
-    __allow_unsequenced()
-    {
-        return ::std::true_type{};
-    }
-    // __allow_vector is needed for __is_vectorization_preferred
-    static constexpr ::std::true_type
-    __allow_vector()
-    {
-        return ::std::true_type{};
-    }
-    static constexpr ::std::true_type
-    __allow_parallel()
-    {
-        return ::std::true_type{};
-    }
-
   private:
     sycl::queue q;
 };

From eecf1e6e14a6ca6ae3b166e7b01105770354da28 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 18:43:18 +0100
Subject: [PATCH 419/566] Remove template aliases __allow_vector,
 __allow_unsequenced and __allow_parallel

---
 include/oneapi/dpl/pstl/execution_impl.h | 9 ---------
 1 file changed, 9 deletions(-)

diff --git a/include/oneapi/dpl/pstl/execution_impl.h b/include/oneapi/dpl/pstl/execution_impl.h
index 865959bba38..39f096d9de4 100644
--- a/include/oneapi/dpl/pstl/execution_impl.h
+++ b/include/oneapi/dpl/pstl/execution_impl.h
@@ -68,15 +68,6 @@ struct __policy_traits<oneapi::dpl::execution::parallel_unsequenced_policy>
     typedef ::std::true_type __allow_vector;
 };
 
-template <typename _ExecutionPolicy>
-using __allow_vector = typename __internal::__policy_traits<::std::decay_t<_ExecutionPolicy>>::__allow_vector;
-
-template <typename _ExecutionPolicy>
-using __allow_unsequenced = typename __internal::__policy_traits<::std::decay_t<_ExecutionPolicy>>::__allow_unsequenced;
-
-template <typename _ExecutionPolicy>
-using __allow_parallel = typename __internal::__policy_traits<::std::decay_t<_ExecutionPolicy>>::__allow_parallel;
-
 //------------------------------------------------------------------------
 // backend selector with tags
 //------------------------------------------------------------------------

From e55c415906fc3c2d15cf0388acdda335f7a6b9bc Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 20 Feb 2024 18:44:02 +0100
Subject: [PATCH 420/566] Remove struct __policy_traits

---
 include/oneapi/dpl/pstl/execution_impl.h      | 38 -------------------
 .../dpl/pstl/experimental/internal/for_loop.h |  9 -----
 2 files changed, 47 deletions(-)

diff --git a/include/oneapi/dpl/pstl/execution_impl.h b/include/oneapi/dpl/pstl/execution_impl.h
index 39f096d9de4..51ee4ed9fd1 100644
--- a/include/oneapi/dpl/pstl/execution_impl.h
+++ b/include/oneapi/dpl/pstl/execution_impl.h
@@ -30,44 +30,6 @@ namespace dpl
 namespace __internal
 {
 
-/* policy */
-template <typename Policy>
-struct __policy_traits
-{
-};
-
-template <>
-struct __policy_traits<oneapi::dpl::execution::sequenced_policy>
-{
-    typedef ::std::false_type __allow_parallel;
-    typedef ::std::false_type __allow_unsequenced;
-    typedef ::std::false_type __allow_vector;
-};
-
-template <>
-struct __policy_traits<oneapi::dpl::execution::unsequenced_policy>
-{
-    typedef ::std::false_type __allow_parallel;
-    typedef ::std::true_type __allow_unsequenced;
-    typedef ::std::true_type __allow_vector;
-};
-
-template <>
-struct __policy_traits<oneapi::dpl::execution::parallel_policy>
-{
-    typedef ::std::true_type __allow_parallel;
-    typedef ::std::false_type __allow_unsequenced;
-    typedef ::std::false_type __allow_vector;
-};
-
-template <>
-struct __policy_traits<oneapi::dpl::execution::parallel_unsequenced_policy>
-{
-    typedef ::std::true_type __allow_parallel;
-    typedef ::std::true_type __allow_unsequenced;
-    typedef ::std::true_type __allow_vector;
-};
-
 //------------------------------------------------------------------------
 // backend selector with tags
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/experimental/internal/for_loop.h b/include/oneapi/dpl/pstl/experimental/internal/for_loop.h
index 4a61dd7c09b..b0975209674 100644
--- a/include/oneapi/dpl/pstl/experimental/internal/for_loop.h
+++ b/include/oneapi/dpl/pstl/experimental/internal/for_loop.h
@@ -42,15 +42,6 @@ struct type_identity
 template <typename _Tp>
 using type_identity_t = typename type_identity<_Tp>::type;
 
-//hide vector policy till the algorithms implementation
-// template <>
-// struct __policy_traits<oneapi::dpl::execution::vector_policy>
-// {
-//     typedef ::std::false_type __allow_parallel;
-//     typedef ::std::false_type __allow_unsequenced;
-//     typedef ::std::true_type __allow_vector;
-// };
-
 // TODO: add static asserts for parameters according to the requirements
 template <typename _ExecutionPolicy, typename _Ip, typename... _Rest>
 void

From efcd7e18dbc6a662f9abeded66600ae81e0c98bf Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 22 Feb 2024 10:13:06 +0100
Subject: [PATCH 421/566] __pattern_for_loop_n + tag impls

---
 .../experimental/internal/for_loop_impl.h     | 34 +++++++++----------
 1 file changed, 17 insertions(+), 17 deletions(-)

diff --git a/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h b/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
index 58511ed2375..fec71d17642 100644
--- a/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
+++ b/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
@@ -161,7 +161,7 @@ class __reduction_pack
     }
 };
 
-// Sequenced version of for_loop_n
+// Sequenced + vectorized version of for_loop_n
 template <class _Tag, typename _ExecutionPolicy, typename _Ip, typename _Size, typename _Function, typename... _Rest>
 void
 __pattern_for_loop_n(_Tag, _ExecutionPolicy&&, _Ip __first, _Size __n, _Function __f, __single_stride_type,
@@ -171,17 +171,17 @@ __pattern_for_loop_n(_Tag, _ExecutionPolicy&&, _Ip __first, _Size __n, _Function
 
     __reduction_pack<_Rest...> __pack{__reduction_pack_tag(), ::std::forward<_Rest>(__rest)...};
 
-    if constexpr (!typename _Tag::__is_vector{})
-    {
-        for (_Size __i = 0; __i < __n; ++__i, ++__first)
-            __pack.__apply_func(__f, __first, __i);
-    }
-    else
+    if constexpr (typename _Tag::__is_vector{})
     {
         oneapi::dpl::__internal::__brick_walk1(
             __n, [&__pack, __first, __f](_Size __idx) { __pack.__apply_func(__f, __first + __idx, __idx); },
             ::std::true_type{});
     }
+    else
+    {
+        for (_Size __i = 0; __i < __n; ++__i, ++__first)
+            __pack.__apply_func(__f, __first, __i);
+    }
 
     __pack.__finalize(__n);
 }
@@ -196,7 +196,16 @@ __pattern_for_loop_n(_Tag, _ExecutionPolicy&&, _Ip __first, _Size __n, _Function
 
     __reduction_pack<_Rest...> __pack{__reduction_pack_tag(), ::std::forward<_Rest>(__rest)...};
 
-    if constexpr (!typename _Tag::__is_vector{})
+    if constexpr (typename _Tag::__is_vector{})
+    {
+        oneapi::dpl::__internal::__brick_walk1(
+            __n,
+            [&__pack, __first, __f, __stride](_Size __idx) {
+                __pack.__apply_func(__f, __first + __idx * __stride, __idx);
+            },
+            ::std::true_type{});
+    }
+    else
     {
         // Simple loop from 0 to __n is not suitable here as we need to ensure that __first is always
         // <= than the end iterator, even if it's not dereferenced. Some implementation might place
@@ -211,15 +220,6 @@ __pattern_for_loop_n(_Tag, _ExecutionPolicy&&, _Ip __first, _Size __n, _Function
             __pack.__apply_func(__f, __first, __n - 1);
         }
     }
-    else
-    {
-        oneapi::dpl::__internal::__brick_walk1(
-            __n,
-            [&__pack, __first, __f, __stride](_Size __idx) {
-                __pack.__apply_func(__f, __first + __idx * __stride, __idx);
-            },
-            ::std::true_type{});
-    }
 
     __pack.__finalize(__n);
 }

From 0e66dd235c67b683b677e4d9b9c36f9da1bc7793 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 22 Feb 2024 10:58:03 +0100
Subject: [PATCH 422/566] __pattern_for_loop + tag impls + add the new one impl
 of __pattern_for_loop for __parallel_tag<_IsVector> and __single_stride_type

---
 .../experimental/internal/for_loop_impl.h     | 74 ++++++++++++-------
 1 file changed, 47 insertions(+), 27 deletions(-)

diff --git a/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h b/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
index fec71d17642..3fca5cda68e 100644
--- a/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
+++ b/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
@@ -267,36 +267,46 @@ __pattern_for_loop(_Tag __tag, _ExecutionPolicy&& __exec, _Ip __first, _Ip __las
 {
     static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
-    if constexpr (__is_random_access_or_integral_v<_Ip>)
+    if constexpr (!typename _Tag::__is_vector{})
     {
-        oneapi::dpl::__internal::__pattern_for_loop_n(
-            __tag, ::std::forward<_ExecutionPolicy>(__exec), __first,
-            oneapi::dpl::__internal::__calculate_input_sequence_length(__first, __last, __stride), __f, __stride,
-            ::std::forward<_Rest>(__rest)...);
-    }
-    else
-    {
-        __reduction_pack<_Rest...> __pack{__reduction_pack_tag(), ::std::forward<_Rest>(__rest)...};
-
-        // Make sure that our index type is able to hold all the possible values
-        using __index_type = typename __difference<_Ip>::__type;
-        __index_type __ordinal_position = 0;
-
-        if (__stride == 1)
+        if constexpr (__is_random_access_or_integral_v<_Ip>)
         {
-            // Avoid check for i % stride on each iteration for the most common case.
-            for (; __first != __last; ++__first, ++__ordinal_position)
-                __pack.__apply_func(__f, __first, __ordinal_position);
+            oneapi::dpl::__internal::__pattern_for_loop_n(
+                __tag, ::std::forward<_ExecutionPolicy>(__exec), __first,
+                oneapi::dpl::__internal::__calculate_input_sequence_length(__first, __last, __stride), __f, __stride,
+                ::std::forward<_Rest>(__rest)...);
         }
         else
         {
-            __ordinal_position =
-                oneapi::dpl::__internal::__execute_loop_strided(__first, __last, __f, __stride, __pack,
-                                                                // Only passed to deduce the type for internal counter
-                                                                __index_type{});
-        }
+            __reduction_pack<_Rest...> __pack{__reduction_pack_tag(), ::std::forward<_Rest>(__rest)...};
+
+            // Make sure that our index type is able to hold all the possible values
+            using __index_type = typename __difference<_Ip>::__type;
+            __index_type __ordinal_position = 0;
+
+            if (__stride == 1)
+            {
+                // Avoid check for i % stride on each iteration for the most common case.
+                for (; __first != __last; ++__first, ++__ordinal_position)
+                    __pack.__apply_func(__f, __first, __ordinal_position);
+            }
+            else
+            {
+                __ordinal_position = oneapi::dpl::__internal::__execute_loop_strided(
+                    __first, __last, __f, __stride, __pack,
+                    // Only passed to deduce the type for internal counter
+                    __index_type{});
+            }
 
-        __pack.__finalize(__ordinal_position);
+            __pack.__finalize(__ordinal_position);
+        }
+    }
+    else
+    {
+        oneapi::dpl::__internal::__pattern_for_loop_n(
+            __tag, ::std::forward<_ExecutionPolicy>(__exec), __first,
+            oneapi::dpl::__internal::__calculate_input_sequence_length(__first, __last, __stride), __f, __stride,
+            ::std::forward<_Rest>(__rest)...);
     }
 }
 
@@ -362,9 +372,8 @@ __execute_loop_strided(_Ip __first, _Ip __last, _Function __f, _Sp __stride, _Pa
 
 // Sequenced version of for_loop for non-RAI and non-integral types
 template <class _Tag, typename _ExecutionPolicy, typename _Ip, typename _Function, typename... _Rest>
-::std::enable_if_t<!__is_random_access_or_integral_v<_Ip>>
-__pattern_for_loop(_Tag, _ExecutionPolicy&&, _Ip __first, _Ip __last, _Function __f, __single_stride_type,
-                   _Rest&&... __rest) noexcept
+void
+__pattern_for_loop(_Tag, _ExecutionPolicy&& __exec, _Ip __first, _Ip __last, _Function __f, __single_stride_type, _Rest&&... __rest) noexcept
 {
     static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
@@ -424,6 +433,17 @@ __pattern_for_loop_n(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Ip _
     });
 }
 
+template <typename _IsVector, typename _ExecutionPolicy, typename _Ip, typename _Function, typename... _Rest>
+void
+__pattern_for_loop(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _Ip __first, _Ip __last, _Function __f,
+                   __single_stride_type, _Rest&&... __rest)
+{
+    oneapi::dpl::__internal::__pattern_for_loop_n(
+        __tag, ::std::forward<_ExecutionPolicy>(__exec), __first,
+        oneapi::dpl::__internal::__calculate_input_sequence_length(__first, __last, __single_stride_type{}), __f,
+        __single_stride_type{}, ::std::forward<_Rest>(__rest)...);
+}
+
 template <typename _IsVector, typename _ExecutionPolicy, typename _Ip, typename _Size, typename _Function, typename _Sp,
           typename... _Rest>
 void

From f2698ca9e5a9f0670bcbb556c4694a0b3fbdafd1 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 22 Feb 2024 11:50:34 +0100
Subject: [PATCH 423/566] 
 include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h - check by
 static assert host policies only inside __select_backend_for_loop()

---
 include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h b/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
index 3fca5cda68e..5e5b5266a4d 100644
--- a/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
+++ b/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
@@ -498,6 +498,8 @@ template <typename _ExecutionPolicy, typename _Ip>
 constexpr auto
 __select_backend_for_loop()
 {
+    static_assert(oneapi::dpl::__internal::__is_host_execution_policy<::std::decay_t<_ExecutionPolicy>>::value, "for_loop staff can be used with host policies only");
+
     if constexpr (::std::is_integral_v<_Ip>)
     {
         return __select_backend<_ExecutionPolicy>();

From 66b71091a37fdc085cd7ca8fc9bdca5bd3fb1c9d Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 22 Feb 2024 12:07:54 +0100
Subject: [PATCH 424/566] 
 include/oneapi/dpl/pstl/experimental/internal/for_loop.h - add static assert
 for host policies check into for_loop

---
 include/oneapi/dpl/pstl/experimental/internal/for_loop.h | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/include/oneapi/dpl/pstl/experimental/internal/for_loop.h b/include/oneapi/dpl/pstl/experimental/internal/for_loop.h
index b0975209674..5d489f03d26 100644
--- a/include/oneapi/dpl/pstl/experimental/internal/for_loop.h
+++ b/include/oneapi/dpl/pstl/experimental/internal/for_loop.h
@@ -47,6 +47,9 @@ template <typename _ExecutionPolicy, typename _Ip, typename... _Rest>
 void
 for_loop(_ExecutionPolicy&& __exec, type_identity_t<_Ip> __start, _Ip __finish, _Rest&&... __rest)
 {
+    static_assert(oneapi::dpl::__internal::__is_host_execution_policy<::std::decay_t<_ExecutionPolicy>>::value,
+                  "for_loop is implemented for the host policies only");
+
     oneapi::dpl::__internal::__for_loop_repack(::std::forward<_ExecutionPolicy>(__exec), __start, __finish,
                                                oneapi::dpl::__internal::__single_stride_type{},
                                                ::std::forward_as_tuple(::std::forward<_Rest>(__rest)...));

From 01957a948e9cb0ff711bc40c95bfbb940b2c4c0b Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 22 Feb 2024 12:08:19 +0100
Subject: [PATCH 425/566] 
 include/oneapi/dpl/pstl/experimental/internal/for_loop.h - add static assert
 for host policies check into for_loop_n

---
 include/oneapi/dpl/pstl/experimental/internal/for_loop.h | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/include/oneapi/dpl/pstl/experimental/internal/for_loop.h b/include/oneapi/dpl/pstl/experimental/internal/for_loop.h
index 5d489f03d26..b91ec53d4b9 100644
--- a/include/oneapi/dpl/pstl/experimental/internal/for_loop.h
+++ b/include/oneapi/dpl/pstl/experimental/internal/for_loop.h
@@ -67,6 +67,9 @@ template <typename _ExecutionPolicy, typename _Ip, typename _Size, typename... _
 void
 for_loop_n(_ExecutionPolicy&& __exec, _Ip __start, _Size __n, _Rest&&... __rest)
 {
+    static_assert(oneapi::dpl::__internal::__is_host_execution_policy<::std::decay_t<_ExecutionPolicy>>::value,
+                  "for_loop_n is implemented for the host policies only");
+
     oneapi::dpl::__internal::__for_loop_repack_n(::std::forward<_ExecutionPolicy>(__exec), __start, __n,
                                                  oneapi::dpl::__internal::__single_stride_type{},
                                                  ::std::forward_as_tuple(::std::forward<_Rest>(__rest)...));

From fdc0c1ec75f93f8ef8a0322984a23fe934772c1f Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 22 Feb 2024 12:08:05 +0100
Subject: [PATCH 426/566] 
 include/oneapi/dpl/pstl/experimental/internal/for_loop.h - add static assert
 for host policies check into for_loop_strided

---
 include/oneapi/dpl/pstl/experimental/internal/for_loop.h | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/include/oneapi/dpl/pstl/experimental/internal/for_loop.h b/include/oneapi/dpl/pstl/experimental/internal/for_loop.h
index b91ec53d4b9..78238d8670d 100644
--- a/include/oneapi/dpl/pstl/experimental/internal/for_loop.h
+++ b/include/oneapi/dpl/pstl/experimental/internal/for_loop.h
@@ -59,6 +59,9 @@ template <typename _ExecutionPolicy, typename _Ip, typename _Sp, typename... _Re
 void
 for_loop_strided(_ExecutionPolicy&& __exec, type_identity_t<_Ip> __start, _Ip __finish, _Sp __stride, _Rest&&... __rest)
 {
+    static_assert(oneapi::dpl::__internal::__is_host_execution_policy<::std::decay_t<_ExecutionPolicy>>::value,
+                  "for_loop_strided is implemented for the host policies only");
+
     oneapi::dpl::__internal::__for_loop_repack(::std::forward<_ExecutionPolicy>(__exec), __start, __finish, __stride,
                                                ::std::forward_as_tuple(::std::forward<_Rest>(__rest)...));
 }

From baf4a5fab86fb9f89f6f47e5affe9848edf8337e Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 22 Feb 2024 12:08:33 +0100
Subject: [PATCH 427/566] 
 include/oneapi/dpl/pstl/experimental/internal/for_loop.h - add static assert
 for host policies check into for_loop_n_strided

---
 include/oneapi/dpl/pstl/experimental/internal/for_loop.h | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/include/oneapi/dpl/pstl/experimental/internal/for_loop.h b/include/oneapi/dpl/pstl/experimental/internal/for_loop.h
index 78238d8670d..46d05a9505b 100644
--- a/include/oneapi/dpl/pstl/experimental/internal/for_loop.h
+++ b/include/oneapi/dpl/pstl/experimental/internal/for_loop.h
@@ -82,6 +82,9 @@ template <typename _ExecutionPolicy, typename _Ip, typename _Size, typename _Sp,
 void
 for_loop_n_strided(_ExecutionPolicy&& __exec, _Ip __start, _Size __n, _Sp __stride, _Rest&&... __rest)
 {
+    static_assert(oneapi::dpl::__internal::__is_host_execution_policy<::std::decay_t<_ExecutionPolicy>>::value,
+                  "for_loop_n_strided is implemented for the host policies only");
+
     oneapi::dpl::__internal::__for_loop_repack_n(::std::forward<_ExecutionPolicy>(__exec), __start, __n, __stride,
                                                  ::std::forward_as_tuple(::std::forward<_Rest>(__rest)...));
 }

From dc0a722314433bcac19852bee13530b3b4806b41 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 22 Feb 2024 11:51:09 +0100
Subject: [PATCH 428/566] test/parallel_api/experimental/for_loop.pass.cpp -
 fix error in test: for_loop staff exist only for host policies

---
 test/parallel_api/experimental/for_loop.pass.cpp | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/test/parallel_api/experimental/for_loop.pass.cpp b/test/parallel_api/experimental/for_loop.pass.cpp
index 8c59ffe1afd..62330927543 100644
--- a/test/parallel_api/experimental/for_loop.pass.cpp
+++ b/test/parallel_api/experimental/for_loop.pass.cpp
@@ -296,8 +296,8 @@ test_for_loop()
         Sequence<T> in_out(n, Gen<T>());
         Sequence<T> expected = in_out;
 
-        invoke_on_all_policies<>()(test_for_loop_impl(), in_out.begin(), in_out.end(), expected.begin(), expected.end(),
-                                   in_out.size());
+        invoke_on_all_host_policies()(test_for_loop_impl(), in_out.begin(), in_out.end(), expected.begin(),
+                                      expected.end(), in_out.size());
     }
 }
 
@@ -313,8 +313,8 @@ test_for_loop_strided()
         ::std::vector<size_t> strides = {1, 2, 10, n > 1 ? n - 1 : 1, n > 0 ? n : 1, n + 1};
         for (size_t stride : strides)
         {
-            invoke_on_all_policies<>()(test_for_loop_strided_impl(), in_out.begin(), in_out.end(), expected.begin(),
-                                       expected.end(), in_out.size(), stride);
+            invoke_on_all_host_policies()(test_for_loop_strided_impl(), in_out.begin(), in_out.end(), expected.begin(),
+                                          expected.end(), in_out.size(), stride);
         }
     }
 }

From a3f3a773a23bea4a22e5ed03e5d20ff7e48dc8ce Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 22 Feb 2024 11:55:33 +0100
Subject: [PATCH 429/566] 
 test/parallel_api/experimental/for_loop_reduction.pass.cpp - fix error in
 test: for_loop staff exist only for host policies

---
 .../experimental/for_loop_reduction.pass.cpp         | 12 ++++++------
 1 file changed, 6 insertions(+), 6 deletions(-)

diff --git a/test/parallel_api/experimental/for_loop_reduction.pass.cpp b/test/parallel_api/experimental/for_loop_reduction.pass.cpp
index cfd918ff2bd..405ae6ba7c6 100644
--- a/test/parallel_api/experimental/for_loop_reduction.pass.cpp
+++ b/test/parallel_api/experimental/for_loop_reduction.pass.cpp
@@ -77,8 +77,8 @@ test()
     {
         Sequence<T> in_out(n, [](long int k) { return T(k % 5 != 1 ? 3 * k - 7 : 0); });
         Sequence<T> expected = in_out;
-        invoke_on_all_policies<>()(test_body(), in_out.begin(), in_out.end(), expected.begin(), expected.end(),
-                                   in_out.size());
+        invoke_on_all_host_policies()(test_body(), in_out.begin(), in_out.end(), expected.begin(), expected.end(),
+                                      in_out.size());
     }
 }
 
@@ -176,10 +176,10 @@ test_predefined(::std::initializer_list<T> init_list)
     // Just arbitrary numbers
     Sequence<T> in_out = init_list;
     Sequence<T> expected = in_out;
-    invoke_on_all_policies<>()(test_body_predefined(), in_out.begin(), in_out.end(), expected.begin(), expected.end(),
-                               in_out.size());
-    invoke_on_all_policies<>()(test_body_predefined_bits(), in_out.begin(), in_out.end(), expected.begin(),
-                               expected.end(), in_out.size());
+    invoke_on_all_host_policies()(test_body_predefined(), in_out.begin(), in_out.end(), expected.begin(),
+                                  expected.end(), in_out.size());
+    invoke_on_all_host_policies()(test_body_predefined_bits(), in_out.begin(), in_out.end(), expected.begin(),
+                                  expected.end(), in_out.size());
 }
 
 void

From f77826ee15d90d09519be88c98b6734ab7f8ebb4 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 22 Feb 2024 12:04:25 +0100
Subject: [PATCH 430/566] 
 test/parallel_api/experimental/for_loop_induction.pass.cpp - fix error in
 test: for_loop staff exist only for host policies

---
 test/parallel_api/experimental/for_loop_induction.pass.cpp | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/test/parallel_api/experimental/for_loop_induction.pass.cpp b/test/parallel_api/experimental/for_loop_induction.pass.cpp
index 505ceb54e1e..88e003f981c 100644
--- a/test/parallel_api/experimental/for_loop_induction.pass.cpp
+++ b/test/parallel_api/experimental/for_loop_induction.pass.cpp
@@ -153,8 +153,8 @@ test()
     {
         Sequence<T> in_out(n, [](long int k) { return T(k % 5 != 1 ? 3 * k - 7 : 0); });
         Sequence<T> expected = in_out;
-        invoke_on_all_policies<>()(test_body(), in_out.begin(), in_out.end(), expected.begin(), expected.end(),
-                                   in_out.size());
+        invoke_on_all_host_policies()(test_body(), in_out.begin(), in_out.end(), expected.begin(), expected.end(),
+                                      in_out.size());
     }
 }
 

From 9c8286979f790f7abf6e22d49a6e2d870a49178f Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 22 Feb 2024 12:40:46 +0100
Subject: [PATCH 431/566] 
 include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h - remove text
 from static_assert

---
 include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h b/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
index 5e5b5266a4d..ca4a3a284a8 100644
--- a/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
+++ b/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
@@ -498,7 +498,7 @@ template <typename _ExecutionPolicy, typename _Ip>
 constexpr auto
 __select_backend_for_loop()
 {
-    static_assert(oneapi::dpl::__internal::__is_host_execution_policy<::std::decay_t<_ExecutionPolicy>>::value, "for_loop staff can be used with host policies only");
+    static_assert(oneapi::dpl::__internal::__is_host_execution_policy<::std::decay_t<_ExecutionPolicy>>::value);
 
     if constexpr (::std::is_integral_v<_Ip>)
     {

From 5b3889cd2cb9d1f7800031d6be2671410c30dd57 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 22 Feb 2024 14:19:27 +0100
Subject: [PATCH 432/566] Add comments into for_loop tests

---
 test/parallel_api/experimental/for_loop.pass.cpp           | 2 ++
 test/parallel_api/experimental/for_loop_induction.pass.cpp | 2 ++
 test/parallel_api/experimental/for_loop_reduction.pass.cpp | 4 ++++
 3 files changed, 8 insertions(+)

diff --git a/test/parallel_api/experimental/for_loop.pass.cpp b/test/parallel_api/experimental/for_loop.pass.cpp
index 62330927543..725d2bc9747 100644
--- a/test/parallel_api/experimental/for_loop.pass.cpp
+++ b/test/parallel_api/experimental/for_loop.pass.cpp
@@ -296,6 +296,7 @@ test_for_loop()
         Sequence<T> in_out(n, Gen<T>());
         Sequence<T> expected = in_out;
 
+        // for_loop staff is implemented for the host policies only
         invoke_on_all_host_policies()(test_for_loop_impl(), in_out.begin(), in_out.end(), expected.begin(),
                                       expected.end(), in_out.size());
     }
@@ -313,6 +314,7 @@ test_for_loop_strided()
         ::std::vector<size_t> strides = {1, 2, 10, n > 1 ? n - 1 : 1, n > 0 ? n : 1, n + 1};
         for (size_t stride : strides)
         {
+            // for_loop staff is implemented for the host policies only
             invoke_on_all_host_policies()(test_for_loop_strided_impl(), in_out.begin(), in_out.end(), expected.begin(),
                                           expected.end(), in_out.size(), stride);
         }
diff --git a/test/parallel_api/experimental/for_loop_induction.pass.cpp b/test/parallel_api/experimental/for_loop_induction.pass.cpp
index 88e003f981c..e01569c58a1 100644
--- a/test/parallel_api/experimental/for_loop_induction.pass.cpp
+++ b/test/parallel_api/experimental/for_loop_induction.pass.cpp
@@ -153,6 +153,8 @@ test()
     {
         Sequence<T> in_out(n, [](long int k) { return T(k % 5 != 1 ? 3 * k - 7 : 0); });
         Sequence<T> expected = in_out;
+
+        // for_loop staff is implemented for the host policies only
         invoke_on_all_host_policies()(test_body(), in_out.begin(), in_out.end(), expected.begin(), expected.end(),
                                       in_out.size());
     }
diff --git a/test/parallel_api/experimental/for_loop_reduction.pass.cpp b/test/parallel_api/experimental/for_loop_reduction.pass.cpp
index 405ae6ba7c6..7f6b2b9e7f5 100644
--- a/test/parallel_api/experimental/for_loop_reduction.pass.cpp
+++ b/test/parallel_api/experimental/for_loop_reduction.pass.cpp
@@ -77,6 +77,8 @@ test()
     {
         Sequence<T> in_out(n, [](long int k) { return T(k % 5 != 1 ? 3 * k - 7 : 0); });
         Sequence<T> expected = in_out;
+
+        // for_loop staff is implemented for the host policies only
         invoke_on_all_host_policies()(test_body(), in_out.begin(), in_out.end(), expected.begin(), expected.end(),
                                       in_out.size());
     }
@@ -176,6 +178,8 @@ test_predefined(::std::initializer_list<T> init_list)
     // Just arbitrary numbers
     Sequence<T> in_out = init_list;
     Sequence<T> expected = in_out;
+
+    // for_loop staff is implemented for the host policies only
     invoke_on_all_host_policies()(test_body_predefined(), in_out.begin(), in_out.end(), expected.begin(),
                                   expected.end(), in_out.size());
     invoke_on_all_host_policies()(test_body_predefined_bits(), in_out.begin(), in_out.end(), expected.begin(),

From 8486872f0f3fed93e66e0556de16cbc040244027 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 22 Feb 2024 14:20:22 +0100
Subject: [PATCH 433/566] 
 include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h - refactoring
 of __pattern_for_loop(_Tag...

---
 .../experimental/internal/for_loop_impl.h     | 58 ++++++++-----------
 1 file changed, 24 insertions(+), 34 deletions(-)

diff --git a/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h b/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
index ca4a3a284a8..06d3ed73c25 100644
--- a/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
+++ b/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
@@ -267,46 +267,36 @@ __pattern_for_loop(_Tag __tag, _ExecutionPolicy&& __exec, _Ip __first, _Ip __las
 {
     static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
-    if constexpr (!typename _Tag::__is_vector{})
+    if constexpr (typename _Tag::__is_vector{} || __is_random_access_or_integral_v<_Ip>)
     {
-        if constexpr (__is_random_access_or_integral_v<_Ip>)
+        oneapi::dpl::__internal::__pattern_for_loop_n(
+            __tag, ::std::forward<_ExecutionPolicy>(__exec), __first,
+            oneapi::dpl::__internal::__calculate_input_sequence_length(__first, __last, __stride), __f, __stride,
+            ::std::forward<_Rest>(__rest)...);
+    }
+    else
+    {
+        __reduction_pack<_Rest...> __pack{__reduction_pack_tag(), ::std::forward<_Rest>(__rest)...};
+
+        // Make sure that our index type is able to hold all the possible values
+        using __index_type = typename __difference<_Ip>::__type;
+        __index_type __ordinal_position = 0;
+
+        if (__stride == 1)
         {
-            oneapi::dpl::__internal::__pattern_for_loop_n(
-                __tag, ::std::forward<_ExecutionPolicy>(__exec), __first,
-                oneapi::dpl::__internal::__calculate_input_sequence_length(__first, __last, __stride), __f, __stride,
-                ::std::forward<_Rest>(__rest)...);
+            // Avoid check for i % stride on each iteration for the most common case.
+            for (; __first != __last; ++__first, ++__ordinal_position)
+                __pack.__apply_func(__f, __first, __ordinal_position);
         }
         else
         {
-            __reduction_pack<_Rest...> __pack{__reduction_pack_tag(), ::std::forward<_Rest>(__rest)...};
-
-            // Make sure that our index type is able to hold all the possible values
-            using __index_type = typename __difference<_Ip>::__type;
-            __index_type __ordinal_position = 0;
-
-            if (__stride == 1)
-            {
-                // Avoid check for i % stride on each iteration for the most common case.
-                for (; __first != __last; ++__first, ++__ordinal_position)
-                    __pack.__apply_func(__f, __first, __ordinal_position);
-            }
-            else
-            {
-                __ordinal_position = oneapi::dpl::__internal::__execute_loop_strided(
-                    __first, __last, __f, __stride, __pack,
-                    // Only passed to deduce the type for internal counter
-                    __index_type{});
-            }
-
-            __pack.__finalize(__ordinal_position);
+            __ordinal_position = oneapi::dpl::__internal::__execute_loop_strided(
+                __first, __last, __f, __stride, __pack,
+                // Only passed to deduce the type for internal counter
+                __index_type{});
         }
-    }
-    else
-    {
-        oneapi::dpl::__internal::__pattern_for_loop_n(
-            __tag, ::std::forward<_ExecutionPolicy>(__exec), __first,
-            oneapi::dpl::__internal::__calculate_input_sequence_length(__first, __last, __stride), __f, __stride,
-            ::std::forward<_Rest>(__rest)...);
+
+        __pack.__finalize(__ordinal_position);
     }
 }
 

From 67a678f9b5e33268a10c9d54b696dda6837dcf35 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 22 Feb 2024 16:55:28 +0100
Subject: [PATCH 434/566] include/oneapi/dpl/pstl/iterator_defs.h - remove
 extra code: struct is_hetero_iterator

---
 include/oneapi/dpl/pstl/iterator_defs.h | 12 ------------
 1 file changed, 12 deletions(-)

diff --git a/include/oneapi/dpl/pstl/iterator_defs.h b/include/oneapi/dpl/pstl/iterator_defs.h
index 7106ee19a4e..7e0d4077e5e 100644
--- a/include/oneapi/dpl/pstl/iterator_defs.h
+++ b/include/oneapi/dpl/pstl/iterator_defs.h
@@ -58,18 +58,6 @@ inline constexpr bool __is_random_access_iterator_v = __is_random_access_iterato
 template <typename... _IteratorTypes>
 inline constexpr bool __is_forward_iterator_v = __is_forward_iterator<_IteratorTypes...>::value;
 
-// struct for checking if iterator is heterogeneous or not
-// for non-heterogeneous iterators
-template <typename Iter, typename Void = void>
-struct is_hetero_iterator : ::std::false_type
-{
-};
-
-// for heterogeneous iterators
-template <typename Iter>
-struct is_hetero_iterator<Iter, ::std::enable_if_t<Iter::is_hetero::value>> : ::std::true_type
-{
-};
 // struct for checking if iterator should be passed directly to device or not
 // for iterators that should not be passed directly
 template <typename Iter, typename Void = void>

From 118363315dfcd8171d024fceae023c79d6ca7b4d Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 22 Feb 2024 16:57:16 +0100
Subject: [PATCH 435/566] include/oneapi/dpl/pstl/iterator_defs.h - remove
 extra code: struct is_passed_directly

---
 include/oneapi/dpl/pstl/iterator_defs.h | 19 -------------------
 1 file changed, 19 deletions(-)

diff --git a/include/oneapi/dpl/pstl/iterator_defs.h b/include/oneapi/dpl/pstl/iterator_defs.h
index 7e0d4077e5e..dec30db3774 100644
--- a/include/oneapi/dpl/pstl/iterator_defs.h
+++ b/include/oneapi/dpl/pstl/iterator_defs.h
@@ -58,25 +58,6 @@ inline constexpr bool __is_random_access_iterator_v = __is_random_access_iterato
 template <typename... _IteratorTypes>
 inline constexpr bool __is_forward_iterator_v = __is_forward_iterator<_IteratorTypes...>::value;
 
-// struct for checking if iterator should be passed directly to device or not
-// for iterators that should not be passed directly
-template <typename Iter, typename Void = void>
-struct is_passed_directly : ::std::false_type
-{
-};
-
-// for iterators defined as direct pass
-template <typename Iter>
-struct is_passed_directly<Iter, ::std::enable_if_t<Iter::is_passed_directly::value>> : ::std::true_type
-{
-};
-
-// for pointers to objects on device
-template <typename Iter>
-struct is_passed_directly<Iter, ::std::enable_if_t<::std::is_pointer_v<Iter>>> : ::std::true_type
-{
-};
-
 } // namespace __internal
 } // namespace dpl
 } // namespace oneapi

From 1c3a71906a8df297888c5abb9d79621e9fa0d842 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 22 Feb 2024 18:17:31 +0100
Subject: [PATCH 436/566] - added TODO for __ranges::__select_backend

---
 include/oneapi/dpl/pstl/hetero/dpcpp/utils_ranges_sycl.h | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/utils_ranges_sycl.h b/include/oneapi/dpl/pstl/hetero/dpcpp/utils_ranges_sycl.h
index d75fdae8612..b38c8d32e11 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/utils_ranges_sycl.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/utils_ranges_sycl.h
@@ -716,6 +716,9 @@ struct __get_sycl_range
     }
 };
 
+//TODO required correct implementation of this __ranges::__select_backend()
+// 1. There is still not RA ranges checks
+// 2. Obviously, a return tag is not necessarily oneapi::dpl::__internal::__hetero_tag
 template <class _ExecutionPolicy, typename... _Ranges>
 constexpr oneapi::dpl::__internal::__hetero_tag<
     typename oneapi::dpl::__internal::__select_backend_for_hetero_policy_trait<

From cbf0a1546da39c27da068798ec8a080fab66f8fb Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 23 Feb 2024 12:28:09 +0100
Subject: [PATCH 437/566] include/oneapi/dpl/pstl/execution_impl.h - add
 implementation of __is_backend_tag_parallel, __is_backend_tag_parallel_v

---
 include/oneapi/dpl/pstl/execution_impl.h | 17 +++++++++++++++++
 1 file changed, 17 insertions(+)

diff --git a/include/oneapi/dpl/pstl/execution_impl.h b/include/oneapi/dpl/pstl/execution_impl.h
index 51ee4ed9fd1..78494531b69 100644
--- a/include/oneapi/dpl/pstl/execution_impl.h
+++ b/include/oneapi/dpl/pstl/execution_impl.h
@@ -172,6 +172,23 @@ struct __is_backend_tag_parallel_forward<__parallel_forward_tag> : ::std::true_t
 template <class _Tag>
 inline constexpr bool __is_backend_tag_parallel_forward_v = __is_backend_tag_parallel_forward<::std::decay_t<_Tag>>::value;
 
+//----------------------------------------------------------
+// __is_backend_tag_parallel, __is_backend_tag_parallel_v
+//----------------------------------------------------------
+
+template <class _Tag>
+struct __is_backend_tag_parallel : ::std::false_type
+{
+};
+
+template <class _IsVector>
+struct __is_backend_tag_parallel<__parallel_tag<_IsVector>> : ::std::true_type
+{
+};
+
+template <class _Tag>
+inline constexpr bool __is_backend_tag_parallel_v = __is_backend_tag_parallel<::std::decay_t<_Tag>>::value;
+
 } // namespace __internal
 } // namespace dpl
 } // namespace oneapi

From ab5a6ca497c0e4d1c415177f340832948a712b3b Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 23 Feb 2024 12:28:35 +0100
Subject: [PATCH 438/566] include/oneapi/dpl/internal/reduce_by_segment_impl.h
 - fix static_assert condition in reduce_by_segment_impl

---
 include/oneapi/dpl/internal/reduce_by_segment_impl.h | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/include/oneapi/dpl/internal/reduce_by_segment_impl.h b/include/oneapi/dpl/internal/reduce_by_segment_impl.h
index 1a613089e21..75dfe4afc1f 100644
--- a/include/oneapi/dpl/internal/reduce_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/reduce_by_segment_impl.h
@@ -85,7 +85,9 @@ reduce_by_segment_impl(_Tag, Policy&& policy, InputIterator1 first1, InputIterat
                        OutputIterator1 result1, OutputIterator2 result2, BinaryPred binary_pred,
                        BinaryOperator binary_op)
 {
-    static_assert(__internal::__is_backend_tag_serial_v<_Tag> || __internal::__is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__internal::__is_backend_tag_serial_v<_Tag> ||
+                  __internal::__is_backend_tag_parallel_forward_v<_Tag> ||
+                  __internal::__is_backend_tag_parallel_v<_Tag>);
 
     using __backend_tag = typename _Tag::__backend_tag;
 

From 5e7694957fc5713e20c0e1ce2b12f2b3afe4d066 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 23 Feb 2024 12:49:11 +0100
Subject: [PATCH 439/566] Apply GitHUB clang format

---
 .../oneapi/dpl/internal/binary_search_impl.h  | 24 ++++++------
 .../internal/exclusive_scan_by_segment_impl.h |  8 ++--
 .../dpl/internal/scan_by_segment_impl.h       |  9 +++--
 include/oneapi/dpl/pstl/execution_impl.h      |  3 +-
 .../experimental/internal/for_loop_impl.h     | 11 +++---
 .../dpl/pstl/glue_algorithm_ranges_impl.h     |  6 +--
 .../dpl/pstl/glue_numeric_ranges_impl.h       |  3 +-
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 10 +++--
 .../hetero/algorithm_ranges_impl_hetero.h     | 39 +++++++++++--------
 .../pstl/hetero/dpcpp/parallel_backend_sycl.h |  3 +-
 .../dpcpp/parallel_backend_sycl_histogram.h   |  3 +-
 .../dpcpp/parallel_backend_sycl_radix_sort.h  |  4 +-
 include/oneapi/dpl/pstl/numeric_impl.h        |  8 ++--
 .../oneapi/dpl/pstl/parallel_backend_tbb.h    |  3 +-
 14 files changed, 75 insertions(+), 59 deletions(-)

diff --git a/include/oneapi/dpl/internal/binary_search_impl.h b/include/oneapi/dpl/internal/binary_search_impl.h
index 821e5e869b4..18b38dba803 100644
--- a/include/oneapi/dpl/internal/binary_search_impl.h
+++ b/include/oneapi/dpl/internal/binary_search_impl.h
@@ -74,11 +74,11 @@ OutputIterator
 lower_bound_impl(_Tag, Policy&& policy, InputIterator1 start, InputIterator1 end, InputIterator2 value_start,
                  InputIterator2 value_end, OutputIterator result, StrictWeakOrdering comp)
 {
-//#if _ONEDPL_BACKEND_SYCL
-//    static_assert(__internal::__is_backend_tag_serial_v<_Tag> || __internal::__is_backend_tag_parallel_forward_v<_Tag>);
-//#else
+    //#if _ONEDPL_BACKEND_SYCL
+    //    static_assert(__internal::__is_backend_tag_serial_v<_Tag> || __internal::__is_backend_tag_parallel_forward_v<_Tag>);
+    //#else
     static_assert(__internal::__is_backend_tag_v<_Tag>);
-//#endif
+    //#endif
 
     return oneapi::dpl::transform(policy, value_start, value_end, result,
                                   [=](typename ::std::iterator_traits<InputIterator2>::reference val) {
@@ -92,11 +92,11 @@ OutputIterator
 upper_bound_impl(_Tag, Policy&& policy, InputIterator1 start, InputIterator1 end, InputIterator2 value_start,
                  InputIterator2 value_end, OutputIterator result, StrictWeakOrdering comp)
 {
-//#if _ONEDPL_BACKEND_SYCL
-//    static_assert(__internal::__is_backend_tag_serial_v<_Tag> || __internal::__is_backend_tag_parallel_forward_v<_Tag>);
-//#else
+    //#if _ONEDPL_BACKEND_SYCL
+    //    static_assert(__internal::__is_backend_tag_serial_v<_Tag> || __internal::__is_backend_tag_parallel_forward_v<_Tag>);
+    //#else
     static_assert(__internal::__is_backend_tag_v<_Tag>);
-//#endif
+    //#endif
 
     return oneapi::dpl::transform(policy, value_start, value_end, result,
                                   [=](typename ::std::iterator_traits<InputIterator2>::reference val) {
@@ -110,11 +110,11 @@ OutputIterator
 binary_search_impl(_Tag, Policy&& policy, InputIterator1 start, InputIterator1 end, InputIterator2 value_start,
                    InputIterator2 value_end, OutputIterator result, StrictWeakOrdering comp)
 {
-//#if _ONEDPL_BACKEND_SYCL
-//    static_assert(__internal::__is_backend_tag_serial_v<_Tag> || __internal::__is_backend_tag_parallel_forward_v<_Tag>);
-//#else
+    //#if _ONEDPL_BACKEND_SYCL
+    //    static_assert(__internal::__is_backend_tag_serial_v<_Tag> || __internal::__is_backend_tag_parallel_forward_v<_Tag>);
+    //#else
     static_assert(__internal::__is_backend_tag_v<_Tag>);
-//#endif
+    //#endif
 
     return oneapi::dpl::transform(policy, value_start, value_end, result,
                                   [=](typename ::std::iterator_traits<InputIterator2>::reference val) {
diff --git a/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h b/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h
index 2ff180251a0..32553a0b701 100644
--- a/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h
@@ -44,11 +44,11 @@ pattern_exclusive_scan_by_segment(_Tag, Policy&& policy, InputIterator1 first1,
                                   InputIterator2 first2, OutputIterator result, T init, BinaryPredicate binary_pred,
                                   Operator binary_op)
 {
-//#if _ONEDPL_BACKEND_SYCL
-//    static_assert(__internal::__is_backend_tag_serial_v<_Tag> || __internal::__is_backend_tag_parallel_forward_v<_Tag>);
-//#else
+    //#if _ONEDPL_BACKEND_SYCL
+    //    static_assert(__internal::__is_backend_tag_serial_v<_Tag> || __internal::__is_backend_tag_parallel_forward_v<_Tag>);
+    //#else
     static_assert(__internal::__is_backend_tag_v<_Tag>);
-//#endif
+    //#endif
 
     using __backend_tag = typename _Tag::__backend_tag;
 
diff --git a/include/oneapi/dpl/internal/scan_by_segment_impl.h b/include/oneapi/dpl/internal/scan_by_segment_impl.h
index 856493f2d2a..2e2fb87fffa 100644
--- a/include/oneapi/dpl/internal/scan_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/scan_by_segment_impl.h
@@ -109,7 +109,8 @@ struct __sycl_scan_by_segment_impl
     operator()(_ExecutionPolicy&& __exec, _Range1&& __keys, _Range2&& __values, _Range3&& __out_values,
                _BinaryPredicate __binary_pred, _BinaryOperator __binary_op, _T __init, _T __identity)
     {
-        constexpr auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range1, _Range2, _Range3>();
+        constexpr auto __dispatch_tag =
+            oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range1, _Range2, _Range3>();
         using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
 
         using _CustomName = oneapi::dpl::__internal::__policy_kernel_name<_ExecutionPolicy>;
@@ -148,11 +149,13 @@ struct __sycl_scan_by_segment_impl
         ::std::size_t __n_groups = __internal::__dpl_ceiling_div(__n, __wgroup_size * __vals_per_item);
 
         auto __partials =
-            oneapi::dpl::__par_backend_hetero::__buffer<__backend_tag, _ExecutionPolicy, __val_type>(__exec, __n_groups).get_buffer();
+            oneapi::dpl::__par_backend_hetero::__buffer<__backend_tag, _ExecutionPolicy, __val_type>(__exec, __n_groups)
+                .get_buffer();
 
         // the number of segment ends found in each work group
         auto __seg_ends =
-            oneapi::dpl::__par_backend_hetero::__buffer<__backend_tag, _ExecutionPolicy, bool>(__exec, __n_groups).get_buffer();
+            oneapi::dpl::__par_backend_hetero::__buffer<__backend_tag, _ExecutionPolicy, bool>(__exec, __n_groups)
+                .get_buffer();
 
         // 1. Work group reduction
         auto __wg_scan = __exec.queue().submit([&](sycl::handler& __cgh) {
diff --git a/include/oneapi/dpl/pstl/execution_impl.h b/include/oneapi/dpl/pstl/execution_impl.h
index 78494531b69..6a42c61ea59 100644
--- a/include/oneapi/dpl/pstl/execution_impl.h
+++ b/include/oneapi/dpl/pstl/execution_impl.h
@@ -170,7 +170,8 @@ struct __is_backend_tag_parallel_forward<__parallel_forward_tag> : ::std::true_t
 };
 
 template <class _Tag>
-inline constexpr bool __is_backend_tag_parallel_forward_v = __is_backend_tag_parallel_forward<::std::decay_t<_Tag>>::value;
+inline constexpr bool __is_backend_tag_parallel_forward_v =
+    __is_backend_tag_parallel_forward<::std::decay_t<_Tag>>::value;
 
 //----------------------------------------------------------
 // __is_backend_tag_parallel, __is_backend_tag_parallel_v
diff --git a/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h b/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
index 06d3ed73c25..1bc0a309bcd 100644
--- a/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
+++ b/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
@@ -290,10 +290,10 @@ __pattern_for_loop(_Tag __tag, _ExecutionPolicy&& __exec, _Ip __first, _Ip __las
         }
         else
         {
-            __ordinal_position = oneapi::dpl::__internal::__execute_loop_strided(
-                __first, __last, __f, __stride, __pack,
-                // Only passed to deduce the type for internal counter
-                __index_type{});
+            __ordinal_position =
+                oneapi::dpl::__internal::__execute_loop_strided(__first, __last, __f, __stride, __pack,
+                                                                // Only passed to deduce the type for internal counter
+                                                                __index_type{});
         }
 
         __pack.__finalize(__ordinal_position);
@@ -363,7 +363,8 @@ __execute_loop_strided(_Ip __first, _Ip __last, _Function __f, _Sp __stride, _Pa
 // Sequenced version of for_loop for non-RAI and non-integral types
 template <class _Tag, typename _ExecutionPolicy, typename _Ip, typename _Function, typename... _Rest>
 void
-__pattern_for_loop(_Tag, _ExecutionPolicy&& __exec, _Ip __first, _Ip __last, _Function __f, __single_stride_type, _Rest&&... __rest) noexcept
+__pattern_for_loop(_Tag, _ExecutionPolicy&& __exec, _Ip __first, _Ip __last, _Function __f, __single_stride_type,
+                   _Rest&&... __rest) noexcept
 {
     static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
diff --git a/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
index 6f93fc7971c..8efc60f95e5 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
@@ -86,7 +86,8 @@ find_if(_ExecutionPolicy&& __exec, _Range&& __rng, _Predicate __pred)
 {
     constexpr auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range>();
 
-    return oneapi::dpl::__internal::__ranges::__pattern_find_if(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
+    return oneapi::dpl::__internal::__ranges::__pattern_find_if(__dispatch_tag,
+                                                                ::std::forward<_ExecutionPolicy>(__exec),
                                                                 views::all_read(::std::forward<_Range>(__rng)), __pred);
 }
 
@@ -297,8 +298,7 @@ transform(_ExecutionPolicy&& __exec, _Range1&& __rng, _Range2&& __result, _Unary
     constexpr auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range1, _Range2>();
 
     oneapi::dpl::__internal::__ranges::__pattern_walk_n(
-        __dispatch_tag,
-        ::std::forward<_ExecutionPolicy>(__exec), [__op](auto x, auto& z) { z = __op(x); },
+        __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), [__op](auto x, auto& z) { z = __op(x); },
         views::all_read(::std::forward<_Range1>(__rng)), views::all_write(::std::forward<_Range2>(__result)));
 }
 
diff --git a/include/oneapi/dpl/pstl/glue_numeric_ranges_impl.h b/include/oneapi/dpl/pstl/glue_numeric_ranges_impl.h
index c8b4f19b560..162c6906595 100644
--- a/include/oneapi/dpl/pstl/glue_numeric_ranges_impl.h
+++ b/include/oneapi/dpl/pstl/glue_numeric_ranges_impl.h
@@ -63,8 +63,7 @@ template <typename _ExecutionPolicy, typename _Range1, typename _Range2, typenam
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _Tp>
 transform_reduce(_ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&& __rng2, _Tp __init)
 {
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range1, _Range2>();
+    constexpr auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range1, _Range2>();
 
     using _ValueType = oneapi::dpl::__internal::__value_t<_Range1>;
     return oneapi::dpl::__internal::__ranges::__pattern_transform_reduce(
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 6611378952a..ca565f7ba05 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -1034,7 +1034,8 @@ __pattern_remove_if(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec,
 
     using _ValueType = typename ::std::iterator_traits<_Iterator>::value_type;
 
-    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _ValueType> __buf(__exec, __last - __first);
+    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _ValueType> __buf(__exec,
+                                                                                                 __last - __first);
     auto __copy_first = __buf.get();
 
     auto __copy_last = __pattern_copy_if(__tag, __exec, __first, __last, __copy_first, __not_pred<_Predicate>{__pred});
@@ -1055,7 +1056,8 @@ __pattern_unique(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _It
 
     using _ValueType = typename ::std::iterator_traits<_Iterator>::value_type;
 
-    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _ValueType> __buf(__exec, __last - __first);
+    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _ValueType> __buf(__exec,
+                                                                                                 __last - __first);
     auto __copy_first = __buf.get();
     auto __copy_last = __pattern_unique_copy(__tag, __exec, __first, __last, __copy_first, __pred);
 
@@ -1216,8 +1218,8 @@ __pattern_merge(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Ite
 
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator, typename _Compare>
 void
-__pattern_inplace_merge(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __middle,
-                        _Iterator __last, _Compare __comp)
+__pattern_inplace_merge(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator __first,
+                        _Iterator __middle, _Iterator __last, _Compare __comp)
 {
     using _ValueType = typename ::std::iterator_traits<_Iterator>::value_type;
 
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
index 73d93a15086..f9392d34a06 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
@@ -69,7 +69,8 @@ class __swap2_wrapper
 
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Range1, typename _Range2, typename _Function>
 bool
-__pattern_swap(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&& __rng2, _Function __f)
+__pattern_swap(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&& __rng2,
+               _Function __f)
 {
     if (__rng1.size() <= __rng2.size())
     {
@@ -250,8 +251,8 @@ __pattern_search(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Ra
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Range, typename _Size, typename _Tp,
           typename _BinaryPredicate>
 oneapi::dpl::__internal::__difference_t<_Range>
-__pattern_search_n(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Range&& __rng, _Size __count, const _Tp& __value,
-                   _BinaryPredicate __pred)
+__pattern_search_n(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Range&& __rng, _Size __count,
+                   const _Tp& __value, _BinaryPredicate __pred)
 {
     //TODO: To consider definition a kind of special factory "multiple_view" (addition to standard "single_view").
     //The factory "multiple_view" would generate a range of N identical values.
@@ -359,7 +360,8 @@ __pattern_scan_copy(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Range
     _DataAcc __get_data_op;
     _MaskAssigner __add_mask_op;
 
-    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, int32_t> __mask_buf(__exec, __rng1.size());
+    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, int32_t> __mask_buf(__exec,
+                                                                                                   __rng1.size());
 
     auto __res =
         __par_backend_hetero::__parallel_transform_scan_base(
@@ -416,7 +418,7 @@ __pattern_remove_if(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec,
     auto __copy_rng = oneapi::dpl::__ranges::views::all(__buf.get_buffer());
 
     auto __copy_last_id = __ranges::__pattern_copy_if(__tag, __exec, __rng, __copy_rng, __not_pred<_Predicate>{__pred},
-                                            oneapi::dpl::__internal::__pstl_assign());
+                                                      oneapi::dpl::__internal::__pstl_assign());
     auto __copy_rng_truncated = __copy_rng | oneapi::dpl::experimental::ranges::views::take(__copy_last_id);
 
     oneapi::dpl::__internal::__ranges::__pattern_walk_n(__tag, ::std::forward<_ExecutionPolicy>(__exec),
@@ -461,7 +463,8 @@ __pattern_unique(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Ra
 
     oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _ValueType> __buf(__exec, __rng.size());
     auto res_rng = oneapi::dpl::__ranges::views::all(__buf.get_buffer());
-    auto res = __ranges::__pattern_unique_copy(__tag, __exec, __rng, res_rng, __pred, oneapi::dpl::__internal::__pstl_assign());
+    auto res = __ranges::__pattern_unique_copy(__tag, __exec, __rng, res_rng, __pred,
+                                               oneapi::dpl::__internal::__pstl_assign());
 
     __pattern_walk_n(__tag, ::std::forward<_ExecutionPolicy>(__exec), __brick_copy<_ExecutionPolicy>{}, res_rng,
                      ::std::forward<_Range>(__rng));
@@ -669,9 +672,9 @@ class __assign_key2_wrapper
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Range1, typename _Range2, typename _Range3,
           typename _Range4, typename _BinaryPredicate, typename _BinaryOperator>
 oneapi::dpl::__internal::__difference_t<_Range3>
-__pattern_reduce_by_segment(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Range1&& __keys, _Range2&& __values,
-                            _Range3&& __out_keys, _Range4&& __out_values, _BinaryPredicate __binary_pred,
-                            _BinaryOperator __binary_op)
+__pattern_reduce_by_segment(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Range1&& __keys,
+                            _Range2&& __values, _Range3&& __out_keys, _Range4&& __out_values,
+                            _BinaryPredicate __binary_pred, _BinaryOperator __binary_op)
 {
     // The algorithm reduces values in __values where the
     // associated keys for the values are equal to the adjacent key.
@@ -692,14 +695,13 @@ __pattern_reduce_by_segment(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&&
         __brick_copy<_ExecutionPolicy> __copy_range{};
 
         oneapi::dpl::__internal::__ranges::__pattern_walk_n(
-            __tag,
-            oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__copy_keys_wrapper>(__exec),
-            __copy_range, ::std::forward<_Range1>(__keys), ::std::forward<_Range3>(__out_keys));
+            __tag, oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__copy_keys_wrapper>(__exec), __copy_range,
+            ::std::forward<_Range1>(__keys), ::std::forward<_Range3>(__out_keys));
 
         oneapi::dpl::__internal::__ranges::__pattern_walk_n(
             __tag,
-            oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__copy_values_wrapper>
-                (::std::forward<_ExecutionPolicy>(__exec)),
+            oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__copy_values_wrapper>(
+                ::std::forward<_ExecutionPolicy>(__exec)),
             __copy_range, ::std::forward<_Range2>(__values), ::std::forward<_Range4>(__out_values));
 
         return 1;
@@ -712,11 +714,14 @@ __pattern_reduce_by_segment(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&&
     // Round 1: reduce with extra indices added to avoid long segments
     // TODO: At threshold points check if the key is equal to the key at the previous threshold point, indicating a long sequence.
     // Skip a round of copy_if and reduces if there are none.
-    auto __idx = oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, __diff_type>(__exec, __n).get_buffer();
+    auto __idx = oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, __diff_type>(__exec, __n)
+                     .get_buffer();
     auto __tmp_out_keys =
-        oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, __key_type>(__exec, __n).get_buffer();
+        oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, __key_type>(__exec, __n)
+            .get_buffer();
     auto __tmp_out_values =
-        oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, __val_type>(__exec, __n).get_buffer();
+        oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, __val_type>(__exec, __n)
+            .get_buffer();
 
     // Replicating first element of keys view to be able to compare (i-1)-th and (i)-th key with aligned sequences,
     //  dropping the last key for the i-1 sequence.
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
index dd2505c57ec..deb02050628 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
@@ -890,7 +890,8 @@ __parallel_scan_copy(oneapi::dpl::__internal::__device_backend_tag __backend_tag
     _MaskAssigner __add_mask_op;
 
     // temporary buffer to store boolean mask
-    oneapi::dpl::__par_backend_hetero::__buffer<oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy, int32_t>
+    oneapi::dpl::__par_backend_hetero::__buffer<oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy,
+                                                int32_t>
         __mask_buf(__exec, __n);
 
     return __parallel_transform_scan_base(
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h
index 0c3cb721ed9..ec5cd7bdbbf 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h
@@ -429,7 +429,8 @@ struct __histogram_general_private_global_atomics_submitter<__internal::__option
             oneapi::dpl::__internal::__dpl_ceiling_div(__n, __work_group_size * __iters_per_work_item);
 
         auto __private_histograms =
-            oneapi::dpl::__par_backend_hetero::__buffer<__backend_tag, _ExecutionPolicy, _bin_type>(__exec, __segments * __num_bins)
+            oneapi::dpl::__par_backend_hetero::__buffer<__backend_tag, _ExecutionPolicy, _bin_type>(
+                __exec, __segments * __num_bins)
                 .get_buffer();
 
         return __exec.queue().submit([&](auto& __h) {
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_radix_sort.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_radix_sort.h
index 9ae0b9f788c..221031e4ac8 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_radix_sort.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_radix_sort.h
@@ -813,7 +813,9 @@ __parallel_radix_sort(oneapi::dpl::__internal::__device_backend_tag, _ExecutionP
         sycl::buffer<::std::uint32_t, 1> __tmp_buf{sycl::range<1>(__tmp_buf_size)};
 
         // memory for storing values sorted for an iteration
-        oneapi::dpl::__par_backend_hetero::__buffer<oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy, _ValueT> __out_buffer_holder{__exec, __n};
+        oneapi::dpl::__par_backend_hetero::__buffer<oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy,
+                                                    _ValueT>
+            __out_buffer_holder{__exec, __n};
         auto __out_rng = oneapi::dpl::__ranges::all_view<_ValueT, __par_backend_hetero::access_mode::read_write>(
             __out_buffer_holder.get_buffer());
 
diff --git a/include/oneapi/dpl/pstl/numeric_impl.h b/include/oneapi/dpl/pstl/numeric_impl.h
index 7aeac1afc4b..5354a72586a 100644
--- a/include/oneapi/dpl/pstl/numeric_impl.h
+++ b/include/oneapi/dpl/pstl/numeric_impl.h
@@ -327,11 +327,11 @@ _OutputIterator
 __pattern_transform_scan(_Tag __tag, _ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last,
                          _OutputIterator __result, _UnaryOperation __unary_op, _BinaryOperation __binary_op, _Inclusive)
 {
-//#if _ONEDPL_BACKEND_SYCL
-//    static_assert(is_backend_tag_serial_v<_Tag> || is_backend_tag_parallel_forward_v<_Tag>);
-//#else
+    //#if _ONEDPL_BACKEND_SYCL
+    //    static_assert(is_backend_tag_serial_v<_Tag> || is_backend_tag_parallel_forward_v<_Tag>);
+    //#else
     static_assert(__is_backend_tag_v<_Tag>);
-//#endif
+    //#endif
 
     typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _ValueType;
     if (__first != __last)
diff --git a/include/oneapi/dpl/pstl/parallel_backend_tbb.h b/include/oneapi/dpl/pstl/parallel_backend_tbb.h
index 659596d932c..4292f0ec405 100644
--- a/include/oneapi/dpl/pstl/parallel_backend_tbb.h
+++ b/include/oneapi/dpl/pstl/parallel_backend_tbb.h
@@ -1196,7 +1196,8 @@ __parallel_stable_sort(oneapi::dpl::__internal::__tbb_backend_tag, _ExecutionPol
         const _DifferenceType __sort_cut_off = _ONEDPL_STABLE_SORT_CUT_OFF;
         if (__n > __sort_cut_off)
         {
-            __tbb_backend::__buffer<oneapi::dpl::__internal::__tbb_backend_tag, _ExecutionPolicy, _ValueType> __buf(__n);
+            __tbb_backend::__buffer<oneapi::dpl::__internal::__tbb_backend_tag, _ExecutionPolicy, _ValueType> __buf(
+                __n);
             __root_task<__stable_sort_func<_RandomAccessIterator, _ValueType*, _Compare, _LeafSort>> __root{
                 __xs, __xe, __buf.get(), true, __comp, __leaf_sort, __nsort, __xs, __buf.get()};
             __task::spawn_root_and_wait(__root);

From 28283e5374c48f7e3f9d919cd6b4d16bc44163d9 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 23 Feb 2024 13:13:41 +0100
Subject: [PATCH 440/566] include/oneapi/dpl/pstl/numeric_impl.h - fix
 static_assert check in __pattern_transform_scan(_Tag __tag, ...

---
 include/oneapi/dpl/pstl/numeric_impl.h | 7 ++-----
 1 file changed, 2 insertions(+), 5 deletions(-)

diff --git a/include/oneapi/dpl/pstl/numeric_impl.h b/include/oneapi/dpl/pstl/numeric_impl.h
index 5354a72586a..5cd3c23b5d3 100644
--- a/include/oneapi/dpl/pstl/numeric_impl.h
+++ b/include/oneapi/dpl/pstl/numeric_impl.h
@@ -327,11 +327,8 @@ _OutputIterator
 __pattern_transform_scan(_Tag __tag, _ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last,
                          _OutputIterator __result, _UnaryOperation __unary_op, _BinaryOperation __binary_op, _Inclusive)
 {
-    //#if _ONEDPL_BACKEND_SYCL
-    //    static_assert(is_backend_tag_serial_v<_Tag> || is_backend_tag_parallel_forward_v<_Tag>);
-    //#else
-    static_assert(__is_backend_tag_v<_Tag>);
-    //#endif
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag> ||
+                  __is_backend_tag_parallel_v<_Tag>);
 
     typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _ValueType;
     if (__first != __last)

From fcb57cda96950b413bb398aa72a69c1a9934f5ce Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 23 Feb 2024 13:16:45 +0100
Subject: [PATCH 441/566] include/oneapi/dpl/internal/binary_search_impl.h -
 fix static_assert check in lower_bound_impl(_Tag, ...

---
 include/oneapi/dpl/internal/binary_search_impl.h | 4 ----
 1 file changed, 4 deletions(-)

diff --git a/include/oneapi/dpl/internal/binary_search_impl.h b/include/oneapi/dpl/internal/binary_search_impl.h
index 18b38dba803..c940455e997 100644
--- a/include/oneapi/dpl/internal/binary_search_impl.h
+++ b/include/oneapi/dpl/internal/binary_search_impl.h
@@ -74,11 +74,7 @@ OutputIterator
 lower_bound_impl(_Tag, Policy&& policy, InputIterator1 start, InputIterator1 end, InputIterator2 value_start,
                  InputIterator2 value_end, OutputIterator result, StrictWeakOrdering comp)
 {
-    //#if _ONEDPL_BACKEND_SYCL
-    //    static_assert(__internal::__is_backend_tag_serial_v<_Tag> || __internal::__is_backend_tag_parallel_forward_v<_Tag>);
-    //#else
     static_assert(__internal::__is_backend_tag_v<_Tag>);
-    //#endif
 
     return oneapi::dpl::transform(policy, value_start, value_end, result,
                                   [=](typename ::std::iterator_traits<InputIterator2>::reference val) {

From 6f406e6e669ac48d7f02ae89b36ec158d2d15fc4 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 23 Feb 2024 13:17:31 +0100
Subject: [PATCH 442/566] include/oneapi/dpl/internal/binary_search_impl.h -
 fix static_assert check in upper_bound_impl(_Tag, ...

---
 include/oneapi/dpl/internal/binary_search_impl.h | 4 ----
 1 file changed, 4 deletions(-)

diff --git a/include/oneapi/dpl/internal/binary_search_impl.h b/include/oneapi/dpl/internal/binary_search_impl.h
index c940455e997..e8e837a77c9 100644
--- a/include/oneapi/dpl/internal/binary_search_impl.h
+++ b/include/oneapi/dpl/internal/binary_search_impl.h
@@ -88,11 +88,7 @@ OutputIterator
 upper_bound_impl(_Tag, Policy&& policy, InputIterator1 start, InputIterator1 end, InputIterator2 value_start,
                  InputIterator2 value_end, OutputIterator result, StrictWeakOrdering comp)
 {
-    //#if _ONEDPL_BACKEND_SYCL
-    //    static_assert(__internal::__is_backend_tag_serial_v<_Tag> || __internal::__is_backend_tag_parallel_forward_v<_Tag>);
-    //#else
     static_assert(__internal::__is_backend_tag_v<_Tag>);
-    //#endif
 
     return oneapi::dpl::transform(policy, value_start, value_end, result,
                                   [=](typename ::std::iterator_traits<InputIterator2>::reference val) {

From df45ca9af114417464795ad12649d771539c5bb7 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 23 Feb 2024 13:18:33 +0100
Subject: [PATCH 443/566] include/oneapi/dpl/internal/binary_search_impl.h -
 fix static_assert check in binary_search_impl(_Tag, ...

---
 include/oneapi/dpl/internal/binary_search_impl.h | 4 ----
 1 file changed, 4 deletions(-)

diff --git a/include/oneapi/dpl/internal/binary_search_impl.h b/include/oneapi/dpl/internal/binary_search_impl.h
index e8e837a77c9..11e47d01417 100644
--- a/include/oneapi/dpl/internal/binary_search_impl.h
+++ b/include/oneapi/dpl/internal/binary_search_impl.h
@@ -102,11 +102,7 @@ OutputIterator
 binary_search_impl(_Tag, Policy&& policy, InputIterator1 start, InputIterator1 end, InputIterator2 value_start,
                    InputIterator2 value_end, OutputIterator result, StrictWeakOrdering comp)
 {
-    //#if _ONEDPL_BACKEND_SYCL
-    //    static_assert(__internal::__is_backend_tag_serial_v<_Tag> || __internal::__is_backend_tag_parallel_forward_v<_Tag>);
-    //#else
     static_assert(__internal::__is_backend_tag_v<_Tag>);
-    //#endif
 
     return oneapi::dpl::transform(policy, value_start, value_end, result,
                                   [=](typename ::std::iterator_traits<InputIterator2>::reference val) {

From 17cceb0c0cf57536ac43faae56d6d939d2e783d4 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 23 Feb 2024 13:19:28 +0100
Subject: [PATCH 444/566] 
 include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h - fix
 static_assert check in pattern_exclusive_scan_by_segment(_Tag, ...

---
 include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h | 4 ----
 1 file changed, 4 deletions(-)

diff --git a/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h b/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h
index 32553a0b701..b7e15b1123e 100644
--- a/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h
@@ -44,11 +44,7 @@ pattern_exclusive_scan_by_segment(_Tag, Policy&& policy, InputIterator1 first1,
                                   InputIterator2 first2, OutputIterator result, T init, BinaryPredicate binary_pred,
                                   Operator binary_op)
 {
-    //#if _ONEDPL_BACKEND_SYCL
-    //    static_assert(__internal::__is_backend_tag_serial_v<_Tag> || __internal::__is_backend_tag_parallel_forward_v<_Tag>);
-    //#else
     static_assert(__internal::__is_backend_tag_v<_Tag>);
-    //#endif
 
     using __backend_tag = typename _Tag::__backend_tag;
 

From 1d168508ec3b684f6c68e98c308aeebf8c237026 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 23 Feb 2024 13:32:09 +0100
Subject: [PATCH 445/566] Introduce __is_host_backend_tag_v

---
 include/oneapi/dpl/internal/reduce_by_segment_impl.h | 4 +---
 include/oneapi/dpl/pstl/execution_impl.h             | 8 ++++++++
 include/oneapi/dpl/pstl/numeric_impl.h               | 3 +--
 3 files changed, 10 insertions(+), 5 deletions(-)

diff --git a/include/oneapi/dpl/internal/reduce_by_segment_impl.h b/include/oneapi/dpl/internal/reduce_by_segment_impl.h
index 75dfe4afc1f..762bbc70602 100644
--- a/include/oneapi/dpl/internal/reduce_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/reduce_by_segment_impl.h
@@ -85,9 +85,7 @@ reduce_by_segment_impl(_Tag, Policy&& policy, InputIterator1 first1, InputIterat
                        OutputIterator1 result1, OutputIterator2 result2, BinaryPred binary_pred,
                        BinaryOperator binary_op)
 {
-    static_assert(__internal::__is_backend_tag_serial_v<_Tag> ||
-                  __internal::__is_backend_tag_parallel_forward_v<_Tag> ||
-                  __internal::__is_backend_tag_parallel_v<_Tag>);
+    static_assert(__internal::__is_host_backend_tag_v<_Tag>);
 
     using __backend_tag = typename _Tag::__backend_tag;
 
diff --git a/include/oneapi/dpl/pstl/execution_impl.h b/include/oneapi/dpl/pstl/execution_impl.h
index 6a42c61ea59..fad37014d4d 100644
--- a/include/oneapi/dpl/pstl/execution_impl.h
+++ b/include/oneapi/dpl/pstl/execution_impl.h
@@ -190,6 +190,14 @@ struct __is_backend_tag_parallel<__parallel_tag<_IsVector>> : ::std::true_type
 template <class _Tag>
 inline constexpr bool __is_backend_tag_parallel_v = __is_backend_tag_parallel<::std::decay_t<_Tag>>::value;
 
+//----------------------------------------------------------
+// __is_host_backend_tag_v
+//----------------------------------------------------------
+
+template <class _Tag>
+inline constexpr bool __is_host_backend_tag_v =
+    __is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag> || __is_backend_tag_parallel_v<_Tag>;
+
 } // namespace __internal
 } // namespace dpl
 } // namespace oneapi
diff --git a/include/oneapi/dpl/pstl/numeric_impl.h b/include/oneapi/dpl/pstl/numeric_impl.h
index 5cd3c23b5d3..0c64681a51e 100644
--- a/include/oneapi/dpl/pstl/numeric_impl.h
+++ b/include/oneapi/dpl/pstl/numeric_impl.h
@@ -327,8 +327,7 @@ _OutputIterator
 __pattern_transform_scan(_Tag __tag, _ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last,
                          _OutputIterator __result, _UnaryOperation __unary_op, _BinaryOperation __binary_op, _Inclusive)
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag> ||
-                  __is_backend_tag_parallel_v<_Tag>);
+    static_assert(__is_host_backend_tag_v<_Tag>);
 
     typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _ValueType;
     if (__first != __last)

From 81488b660bd06053e5fb4d9b4f0fe20e69ce8a20 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 23 Feb 2024 13:37:22 +0100
Subject: [PATCH 446/566] Replace __is_backend_tag_v check to
 __is_host_backend_tag_v

---
 include/oneapi/dpl/internal/binary_search_impl.h            | 6 +++---
 .../oneapi/dpl/internal/exclusive_scan_by_segment_impl.h    | 2 +-
 .../oneapi/dpl/internal/inclusive_scan_by_segment_impl.h    | 2 +-
 3 files changed, 5 insertions(+), 5 deletions(-)

diff --git a/include/oneapi/dpl/internal/binary_search_impl.h b/include/oneapi/dpl/internal/binary_search_impl.h
index 11e47d01417..00505df5cd6 100644
--- a/include/oneapi/dpl/internal/binary_search_impl.h
+++ b/include/oneapi/dpl/internal/binary_search_impl.h
@@ -74,7 +74,7 @@ OutputIterator
 lower_bound_impl(_Tag, Policy&& policy, InputIterator1 start, InputIterator1 end, InputIterator2 value_start,
                  InputIterator2 value_end, OutputIterator result, StrictWeakOrdering comp)
 {
-    static_assert(__internal::__is_backend_tag_v<_Tag>);
+    static_assert(__internal::__is_host_backend_tag_v<_Tag>);
 
     return oneapi::dpl::transform(policy, value_start, value_end, result,
                                   [=](typename ::std::iterator_traits<InputIterator2>::reference val) {
@@ -88,7 +88,7 @@ OutputIterator
 upper_bound_impl(_Tag, Policy&& policy, InputIterator1 start, InputIterator1 end, InputIterator2 value_start,
                  InputIterator2 value_end, OutputIterator result, StrictWeakOrdering comp)
 {
-    static_assert(__internal::__is_backend_tag_v<_Tag>);
+    static_assert(__internal::__is_host_backend_tag_v<_Tag>);
 
     return oneapi::dpl::transform(policy, value_start, value_end, result,
                                   [=](typename ::std::iterator_traits<InputIterator2>::reference val) {
@@ -102,7 +102,7 @@ OutputIterator
 binary_search_impl(_Tag, Policy&& policy, InputIterator1 start, InputIterator1 end, InputIterator2 value_start,
                    InputIterator2 value_end, OutputIterator result, StrictWeakOrdering comp)
 {
-    static_assert(__internal::__is_backend_tag_v<_Tag>);
+    static_assert(__internal::__is_host_backend_tag_v<_Tag>);
 
     return oneapi::dpl::transform(policy, value_start, value_end, result,
                                   [=](typename ::std::iterator_traits<InputIterator2>::reference val) {
diff --git a/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h b/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h
index b7e15b1123e..c818213b180 100644
--- a/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h
@@ -44,7 +44,7 @@ pattern_exclusive_scan_by_segment(_Tag, Policy&& policy, InputIterator1 first1,
                                   InputIterator2 first2, OutputIterator result, T init, BinaryPredicate binary_pred,
                                   Operator binary_op)
 {
-    static_assert(__internal::__is_backend_tag_v<_Tag>);
+    static_assert(__internal::__is_host_backend_tag_v<_Tag>);
 
     using __backend_tag = typename _Tag::__backend_tag;
 
diff --git a/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h b/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h
index c15715a0823..05c6f6635f7 100644
--- a/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h
@@ -43,7 +43,7 @@ pattern_inclusive_scan_by_segment(_Tag, Policy&& policy, InputIterator1 first1,
                                   InputIterator2 first2, OutputIterator result, BinaryPredicate binary_pred,
                                   BinaryOperator binary_op)
 {
-    static_assert(__internal::__is_backend_tag_v<_Tag>);
+    static_assert(__internal::__is_host_backend_tag_v<_Tag>);
 
     using __backend_tag = typename _Tag::__backend_tag;
 

From 16f727f97a8a5071b25ed4f20aeebdffe8bd5e24 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 23 Feb 2024 13:37:38 +0100
Subject: [PATCH 447/566] Remove __is_backend_tag, __is_backend_tag_v as not
 required anymore

---
 include/oneapi/dpl/pstl/execution_impl.h      | 27 -------------------
 .../pstl/hetero/dpcpp/execution_sycl_defs.h   |  8 ------
 2 files changed, 35 deletions(-)

diff --git a/include/oneapi/dpl/pstl/execution_impl.h b/include/oneapi/dpl/pstl/execution_impl.h
index fad37014d4d..68949c792ab 100644
--- a/include/oneapi/dpl/pstl/execution_impl.h
+++ b/include/oneapi/dpl/pstl/execution_impl.h
@@ -111,33 +111,6 @@ __select_backend()
     return {};
 }
 
-//----------------------------------------------------------
-// __is_backend_tag, __is_backend_tag_v
-//----------------------------------------------------------
-
-template <class _Tag>
-struct __is_backend_tag : ::std::false_type
-{
-};
-
-template <class _IsVector>
-struct __is_backend_tag<__serial_tag<_IsVector>> : ::std::true_type
-{
-};
-
-template <class _IsVector>
-struct __is_backend_tag<__parallel_tag<_IsVector>> : ::std::true_type
-{
-};
-
-template <>
-struct __is_backend_tag<__parallel_forward_tag> : ::std::true_type
-{
-};
-
-template <class _Tag>
-inline constexpr bool __is_backend_tag_v = __is_backend_tag<::std::decay_t<_Tag>>::value;
-
 //----------------------------------------------------------
 // __is_backend_tag_serial, __is_backend_tag_serial_v
 //----------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h b/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h
index 3cdddc16efc..4a63ea9d151 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h
@@ -326,14 +326,6 @@ __select_backend()
     return {}; // return __hetero_tag<__device_backend_tag> or __hetero_tag<__fpga_backend_tag>
 }
 
-template <class _Tag>
-struct __is_backend_tag;
-
-template <typename _BackendTag>
-struct __is_backend_tag<__hetero_tag<_BackendTag>> : ::std::true_type
-{
-};
-
 } // namespace __internal
 
 } // namespace dpl

From 16cf4b29917257b1808f13eab0e8f8a8925eccd6 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 23 Feb 2024 14:32:45 +0100
Subject: [PATCH 448/566] 
 include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h - remove unused param
 `__tag`

---
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 46 +++++++++----------
 1 file changed, 23 insertions(+), 23 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index ca565f7ba05..32bffaf0b29 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -440,7 +440,7 @@ struct __brick_fill_n<_SourceT, _ExecutionPolicy,
 
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator, typename _Compare>
 _Iterator
-__pattern_min_element(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last,
+__pattern_min_element(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last,
                       _Compare __comp)
 {
     if (__first == __last)
@@ -516,8 +516,8 @@ __pattern_min_element(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec
 
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator, typename _Compare>
 ::std::pair<_Iterator, _Iterator>
-__pattern_minmax_element(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator __first,
-                         _Iterator __last, _Compare __comp)
+__pattern_minmax_element(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last,
+                         _Compare __comp)
 {
     if (__first == __last)
         return ::std::make_pair(__first, __first);
@@ -570,7 +570,7 @@ __pattern_minmax_element(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __e
 
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator, typename _BinaryPredicate>
 _Iterator
-__pattern_adjacent_find(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last,
+__pattern_adjacent_find(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last,
                         _BinaryPredicate __predicate, oneapi::dpl::__internal::__or_semantic)
 {
     if (__last - __first < 2)
@@ -598,7 +598,7 @@ __pattern_adjacent_find(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __ex
 
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator, typename _BinaryPredicate>
 _Iterator
-__pattern_adjacent_find(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last,
+__pattern_adjacent_find(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last,
                         _BinaryPredicate __predicate, oneapi::dpl::__internal::__first_semantic)
 {
     if (__last - __first < 2)
@@ -630,7 +630,7 @@ __pattern_adjacent_find(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __ex
 
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator, typename _Predicate>
 typename ::std::iterator_traits<_Iterator>::difference_type
-__pattern_count(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last,
+__pattern_count(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last,
                 _Predicate __predicate)
 {
     if (__first == __last)
@@ -662,7 +662,7 @@ __pattern_count(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Ite
 
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator, typename _Pred>
 bool
-__pattern_any_of(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last,
+__pattern_any_of(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last,
                  _Pred __pred)
 {
     if (__first == __last)
@@ -686,7 +686,7 @@ __pattern_any_of(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _It
 
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2, typename _Pred>
 bool
-__pattern_equal(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator1 __first1, _Iterator1 __last1,
+__pattern_equal(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Iterator1 __first1, _Iterator1 __last1,
                 _Iterator2 __first2, _Iterator2 __last2, _Pred __pred)
 {
     if (__last1 == __first1 || __last2 == __first2 || __last1 - __first1 != __last2 - __first2)
@@ -726,7 +726,7 @@ __pattern_equal(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Ite
 
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator, typename _Pred>
 _Iterator
-__pattern_find_if(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last,
+__pattern_find_if(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last,
                   _Pred __pred)
 {
     if (__first == __last)
@@ -779,8 +779,8 @@ __pattern_find_end(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _
 
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2, typename _Pred>
 _Iterator1
-__pattern_find_first_of(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator1 __first,
-                        _Iterator1 __last, _Iterator2 __s_first, _Iterator2 __s_last, _Pred __pred)
+__pattern_find_first_of(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Iterator1 __first, _Iterator1 __last,
+                        _Iterator2 __s_first, _Iterator2 __s_last, _Pred __pred)
 {
     if (__first == __last || __s_last == __s_first)
         return __last;
@@ -888,7 +888,7 @@ __pattern_search_n(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _
 
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2, typename _Pred>
 ::std::pair<_Iterator1, _Iterator2>
-__pattern_mismatch(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator1 __first1, _Iterator1 __last1,
+__pattern_mismatch(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Iterator1 __first1, _Iterator1 __last1,
                    _Iterator2 __first2, _Iterator2 __last2, _Pred __pred)
 {
     auto __n = ::std::min(__last1 - __first1, __last2 - __first2);
@@ -941,7 +941,7 @@ __pattern_scan_copy(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Itera
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2,
           typename _Predicate>
 _Iterator2
-__pattern_copy_if(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator1 __first, _Iterator1 __last,
+__pattern_copy_if(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Iterator1 __first, _Iterator1 __last,
                   _Iterator2 __result_first, _Predicate __pred)
 {
     using _It1DifferenceType = typename ::std::iterator_traits<_Iterator1>::difference_type;
@@ -1082,8 +1082,8 @@ enum _IsPartitionedReduceType : signed char
 
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator, typename _Predicate>
 bool
-__pattern_is_partitioned(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator __first,
-                         _Iterator __last, _Predicate __predicate)
+__pattern_is_partitioned(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last,
+                         _Predicate __predicate)
 {
     if (__last - __first < 2)
         return true;
@@ -1133,7 +1133,7 @@ struct __is_heap_check
 
 template <typename _BackendTag, typename _ExecutionPolicy, typename _RandomAccessIterator, typename _Compare>
 _RandomAccessIterator
-__pattern_is_heap_until(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
+__pattern_is_heap_until(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
                         _RandomAccessIterator __last, _Compare __comp)
 {
     if (__last - __first < 2)
@@ -1151,7 +1151,7 @@ __pattern_is_heap_until(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __ex
 
 template <typename _BackendTag, typename _ExecutionPolicy, typename _RandomAccessIterator, typename _Compare>
 bool
-__pattern_is_heap(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
+__pattern_is_heap(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
                   _RandomAccessIterator __last, _Compare __comp)
 {
     if (__last - __first < 2)
@@ -1251,8 +1251,8 @@ __pattern_inplace_merge(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __ex
 //------------------------------------------------------------------------
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator, typename _Compare, typename _Proj>
 void
-__stable_sort_with_projection(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator __first,
-                              _Iterator __last, _Compare __comp, _Proj __proj)
+__stable_sort_with_projection(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last,
+                              _Compare __comp, _Proj __proj)
 {
     if (__last - __first < 2)
         return;
@@ -1352,7 +1352,7 @@ __pattern_partition(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec,
 
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator1, typename _Iterator2, typename _Compare>
 bool
-__pattern_lexicographical_compare(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator1 __first1,
+__pattern_lexicographical_compare(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Iterator1 __first1,
                                   _Iterator1 __last1, _Iterator2 __first2, _Iterator2 __last2, _Compare __comp)
 {
     //trivial pre-checks
@@ -1401,7 +1401,7 @@ __pattern_lexicographical_compare(__hetero_tag<_BackendTag> __tag, _ExecutionPol
 template <typename _BackendTag, typename _ExecutionPolicy, typename _ForwardIterator1, typename _ForwardIterator2,
           typename _Compare>
 bool
-__pattern_includes(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _ForwardIterator1 __first1,
+__pattern_includes(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _ForwardIterator1 __first1,
                    _ForwardIterator1 __last1, _ForwardIterator2 __first2, _ForwardIterator2 __last2, _Compare __comp)
 {
     //according to the spec
@@ -1432,7 +1432,7 @@ __pattern_includes(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _
 
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator, typename _Compare>
 void
-__pattern_partial_sort(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __mid,
+__pattern_partial_sort(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __mid,
                        _Iterator __last, _Compare __comp)
 {
     if (__last - __first < 2)
@@ -1583,7 +1583,7 @@ __pattern_reverse(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Iterato
 
 template <typename _BackendTag, typename _ExecutionPolicy, typename _BidirectionalIterator, typename _ForwardIterator>
 _ForwardIterator
-__pattern_reverse_copy(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _BidirectionalIterator __first,
+__pattern_reverse_copy(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _BidirectionalIterator __first,
                        _BidirectionalIterator __last, _ForwardIterator __result)
 {
     auto __n = __last - __first;

From 8533f7fe004d2cd883621280b43910d053fd1676 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 23 Feb 2024 14:33:06 +0100
Subject: [PATCH 449/566] include/oneapi/dpl/pstl/numeric_impl.h - remove
 unused param `__tag`

---
 include/oneapi/dpl/pstl/numeric_impl.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/include/oneapi/dpl/pstl/numeric_impl.h b/include/oneapi/dpl/pstl/numeric_impl.h
index 0c64681a51e..2b8102dfe7a 100644
--- a/include/oneapi/dpl/pstl/numeric_impl.h
+++ b/include/oneapi/dpl/pstl/numeric_impl.h
@@ -77,7 +77,7 @@ __pattern_transform_reduce(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1,
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
           class _Tp, class _BinaryOperation1, class _BinaryOperation2>
 _Tp
-__pattern_transform_reduce(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1,
+__pattern_transform_reduce(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1,
                            _RandomAccessIterator1 __last1, _RandomAccessIterator2 __first2, _Tp __init,
                            _BinaryOperation1 __binary_op1, _BinaryOperation2 __binary_op2)
 {

From f881b2e1b04d2b2588c5a59211326c4e9da89084 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 23 Feb 2024 14:33:37 +0100
Subject: [PATCH 450/566] include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
 - remove unused param `__tag`

---
 include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
index ff089c083d3..0083299169c 100644
--- a/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
@@ -40,7 +40,7 @@ namespace __internal
 template <typename _BackendTag, typename _ExecutionPolicy, typename _RandomAccessIterator1,
           typename _RandomAccessIterator2, typename _Tp, typename _BinaryOperation1, typename _BinaryOperation2>
 _Tp
-__pattern_transform_reduce(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1,
+__pattern_transform_reduce(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1,
                            _RandomAccessIterator1 __last1, _RandomAccessIterator2 __first2, _Tp __init,
                            _BinaryOperation1 __binary_op1, _BinaryOperation2 __binary_op2)
 {
@@ -73,7 +73,7 @@ __pattern_transform_reduce(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& _
 template <typename _BackendTag, typename _ExecutionPolicy, typename _ForwardIterator, typename _Tp,
           typename _BinaryOperation, typename _UnaryOperation>
 _Tp
-__pattern_transform_reduce(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _ForwardIterator __first,
+__pattern_transform_reduce(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _ForwardIterator __first,
                            _ForwardIterator __last, _Tp __init, _BinaryOperation __binary_op,
                            _UnaryOperation __unary_op)
 {

From f13e4732bb5c88c32542cd19252c9ff4a5556f5c Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 23 Feb 2024 14:47:33 +0100
Subject: [PATCH 451/566] 
 include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h - fix
 __histogram_general_private_global_atomics_submitter : __select_backend call
 inside not required

---
 .../hetero/dpcpp/parallel_backend_sycl_histogram.h   | 12 +++++-------
 1 file changed, 5 insertions(+), 7 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h
index ec5cd7bdbbf..05864be358d 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h
@@ -399,10 +399,10 @@ __histogram_general_local_atomics(oneapi::dpl::__internal::__device_backend_tag,
         ::std::forward<_Range2>(__bins), __binhash_manager);
 }
 
-template <typename _KernelName>
+template <typename _BackendTag, typename _KernelName>
 struct __histogram_general_private_global_atomics_submitter;
 
-template <typename... _KernelName>
+template <typename _BackendTag, typename... _KernelName>
 struct __histogram_general_private_global_atomics_submitter<__internal::__optional_kernel_name<_KernelName...>>
 {
     template <typename _ExecutionPolicy, typename _Range1, typename _Range2, typename _BinHashMgr>
@@ -411,9 +411,6 @@ struct __histogram_general_private_global_atomics_submitter<__internal::__option
                ::std::uint16_t __work_group_size, _Range1&& __input, _Range2&& __bins,
                const _BinHashMgr& __binhash_manager)
     {
-        constexpr auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range1, _Range2>();
-        using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
         const ::std::size_t __n = __input.size();
         const ::std::size_t __num_bins = __bins.size();
         using _bin_type = oneapi::dpl::__internal::__value_t<_Range2>;
@@ -429,7 +426,7 @@ struct __histogram_general_private_global_atomics_submitter<__internal::__option
             oneapi::dpl::__internal::__dpl_ceiling_div(__n, __work_group_size * __iters_per_work_item);
 
         auto __private_histograms =
-            oneapi::dpl::__par_backend_hetero::__buffer<__backend_tag, _ExecutionPolicy, _bin_type>(
+            oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _bin_type>(
                 __exec, __segments * __num_bins)
                 .get_buffer();
 
@@ -491,7 +488,8 @@ __histogram_general_private_global_atomics(oneapi::dpl::__internal::__device_bac
     using _global_atomics_name = oneapi::dpl::__par_backend_hetero::__internal::__kernel_name_provider<
         __histo_kernel_private_glocal_atomics<_kernel_base_name>>;
 
-    return __histogram_general_private_global_atomics_submitter<_global_atomics_name>()(
+    return __histogram_general_private_global_atomics_submitter<oneapi::dpl::__internal::__device_backend_tag,
+                                                                _global_atomics_name>()(
         ::std::forward<_ExecutionPolicy>(__exec), __init_event, __min_iters_per_work_item, __work_group_size,
         ::std::forward<_Range1>(__input), ::std::forward<_Range2>(__bins), __binhash_manager);
 }

From c547583fd1a5b4b3e85e0d49699f67965a6bb87c Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 23 Feb 2024 14:47:53 +0100
Subject: [PATCH 452/566] 
 include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h - fix
 __parallel_sort_submitter : __select_backend call inside not required

---
 .../pstl/hetero/dpcpp/parallel_backend_sycl.h | 23 ++++++++++---------
 1 file changed, 12 insertions(+), 11 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
index deb02050628..d10a959ffe9 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
@@ -1565,11 +1565,13 @@ struct __leaf_sort_kernel
 };
 
 // Please see the comment for __parallel_for_submitter for optional kernel name explanation
-template <typename _IdType, typename _LeafSortName, typename _GlobalSortName, typename _CopyBackName>
+template <typename _BackendTag, typename _IdType, typename _LeafSortName, typename _GlobalSortName,
+          typename _CopyBackName>
 struct __parallel_sort_submitter;
 
-template <typename _IdType, typename... _LeafSortName, typename... _GlobalSortName, typename... _CopyBackName>
-struct __parallel_sort_submitter<_IdType, __internal::__optional_kernel_name<_LeafSortName...>,
+template <typename _BackendTag, typename _IdType, typename... _LeafSortName, typename... _GlobalSortName,
+          typename... _CopyBackName>
+struct __parallel_sort_submitter<_BackendTag, _IdType, __internal::__optional_kernel_name<_LeafSortName...>,
                                  __internal::__optional_kernel_name<_GlobalSortName...>,
                                  __internal::__optional_kernel_name<_CopyBackName...>>
 {
@@ -1577,9 +1579,6 @@ struct __parallel_sort_submitter<_IdType, __internal::__optional_kernel_name<_Le
     auto
     operator()(_ExecutionPolicy&& __exec, _Range&& __rng, _Compare __comp) const
     {
-        constexpr auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range>();
-        using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
         using _Tp = oneapi::dpl::__internal::__value_t<_Range>;
         using _Size = oneapi::dpl::__internal::__difference_t<_Range>;
 
@@ -1601,7 +1600,7 @@ struct __parallel_sort_submitter<_IdType, __internal::__optional_kernel_name<_Le
         });
 
         // 2. Merge sorting
-        oneapi::dpl::__par_backend_hetero::__buffer<__backend_tag, _ExecutionPolicy, _Tp> __temp_buf(__exec, __n);
+        oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _Tp> __temp_buf(__exec, __n);
         auto __temp = __temp_buf.get_buffer();
         bool __data_in_temp = false;
         _IdType __n_sorted = __leaf;
@@ -1685,8 +1684,9 @@ __parallel_sort_impl(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPo
             oneapi::dpl::__par_backend_hetero::__internal::__kernel_name_provider<__sort_global_kernel<_CustomName, _wi_index_type>>;
         using _CopyBackKernel =
             oneapi::dpl::__par_backend_hetero::__internal::__kernel_name_provider<__sort_copy_back_kernel<_CustomName, _wi_index_type>>;
-        return __parallel_sort_submitter<_wi_index_type, _LeafSortKernel, _GlobalSortKernel, _CopyBackKernel>()(
-            ::std::forward<_ExecutionPolicy>(__exec), ::std::forward<_Range>(__rng), __comp);
+        return __parallel_sort_submitter<oneapi::dpl::__internal::__device_backend_tag, _wi_index_type, _LeafSortKernel,
+                                         _GlobalSortKernel, _CopyBackKernel>()(::std::forward<_ExecutionPolicy>(__exec),
+                                                                               ::std::forward<_Range>(__rng), __comp);
     }
     else
     {
@@ -1697,8 +1697,9 @@ __parallel_sort_impl(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPo
             oneapi::dpl::__par_backend_hetero::__internal::__kernel_name_provider<__sort_global_kernel<_CustomName, _wi_index_type>>;
         using _CopyBackKernel =
             oneapi::dpl::__par_backend_hetero::__internal::__kernel_name_provider<__sort_copy_back_kernel<_CustomName, _wi_index_type>>;
-        return __parallel_sort_submitter<_wi_index_type, _LeafSortKernel, _GlobalSortKernel, _CopyBackKernel>()(
-            ::std::forward<_ExecutionPolicy>(__exec), ::std::forward<_Range>(__rng), __comp);
+        return __parallel_sort_submitter<oneapi::dpl::__internal::__device_backend_tag, _wi_index_type, _LeafSortKernel,
+                                         _GlobalSortKernel, _CopyBackKernel>()(::std::forward<_ExecutionPolicy>(__exec),
+                                                                               ::std::forward<_Range>(__rng), __comp);
     }
 }
 

From c11f1a077e248a7f837e21c52a42f8a0c651cf22 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 23 Feb 2024 14:51:30 +0100
Subject: [PATCH 453/566] Fix
 __histogram_general_private_global_atomics_submitter : __select_backend call
 inside not required

---
 .../pstl/hetero/dpcpp/parallel_backend_sycl.h    | 16 +++++++---------
 .../dpcpp/parallel_backend_sycl_histogram.h      |  3 ++-
 2 files changed, 9 insertions(+), 10 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
index d10a959ffe9..f966b334493 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
@@ -1704,27 +1704,24 @@ __parallel_sort_impl(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPo
 }
 
 // Please see the comment for __parallel_for_submitter for optional kernel name explanation
-template <typename _GlobalSortName, typename _CopyBackName>
+template <typename _BackendTag, typename _GlobalSortName, typename _CopyBackName>
 struct __parallel_partial_sort_submitter;
 
-template <typename... _GlobalSortName, typename... _CopyBackName>
-struct __parallel_partial_sort_submitter<__internal::__optional_kernel_name<_GlobalSortName...>,
+template <typename _BackendTag, typename... _GlobalSortName, typename... _CopyBackName>
+struct __parallel_partial_sort_submitter<_BackendTag, __internal::__optional_kernel_name<_GlobalSortName...>,
                                          __internal::__optional_kernel_name<_CopyBackName...>>
 {
     template <typename _ExecutionPolicy, typename _Range, typename _Merge, typename _Compare>
     auto
     operator()(_ExecutionPolicy&& __exec, _Range&& __rng, _Merge __merge, _Compare __comp) const
     {
-        constexpr auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range>();
-        using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
         using _Tp = oneapi::dpl::__internal::__value_t<_Range>;
         using _Size = oneapi::dpl::__internal::__difference_t<_Range>;
 
         _Size __n = __rng.size();
         assert(__n > 1);
 
-        oneapi::dpl::__par_backend_hetero::__buffer<__backend_tag, _ExecutionPolicy, _Tp> __temp_buf(__exec, __n);
+        oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _Tp> __temp_buf(__exec, __n);
         auto __temp = __temp_buf.get_buffer();
         _PRINT_INFO_IN_DEBUG_MODE(__exec);
 
@@ -1790,8 +1787,9 @@ __parallel_partial_sort_impl(oneapi::dpl::__internal::__device_backend_tag, _Exe
     using _CopyBackKernel =
         oneapi::dpl::__par_backend_hetero::__internal::__kernel_name_provider<__sort_copy_back_kernel<_CustomName>>;
 
-    return __parallel_partial_sort_submitter<_GlobalSortKernel, _CopyBackKernel>()(
-        ::std::forward<_ExecutionPolicy>(__exec), ::std::forward<_Range>(__rng), __merge, __comp);
+    return __parallel_partial_sort_submitter<oneapi::dpl::__internal::__device_backend_tag, _GlobalSortKernel,
+                                             _CopyBackKernel>()(::std::forward<_ExecutionPolicy>(__exec),
+                                                                ::std::forward<_Range>(__rng), __merge, __comp);
 }
 
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h
index 05864be358d..083bffcefc0 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h
@@ -403,7 +403,8 @@ template <typename _BackendTag, typename _KernelName>
 struct __histogram_general_private_global_atomics_submitter;
 
 template <typename _BackendTag, typename... _KernelName>
-struct __histogram_general_private_global_atomics_submitter<__internal::__optional_kernel_name<_KernelName...>>
+struct __histogram_general_private_global_atomics_submitter<_BackendTag,
+                                                            __internal::__optional_kernel_name<_KernelName...>>
 {
     template <typename _ExecutionPolicy, typename _Range1, typename _Range2, typename _BinHashMgr>
     auto

From 650b239d56a40cbfce665dbadc8408e80c188366 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 23 Feb 2024 16:00:19 +0100
Subject: [PATCH 454/566] include/oneapi/dpl/internal/scan_by_segment_impl.h -
 fix __sycl_scan_by_segment_impl : __select_backend call inside not required

---
 .../oneapi/dpl/internal/scan_by_segment_impl.h   | 16 ++++++----------
 1 file changed, 6 insertions(+), 10 deletions(-)

diff --git a/include/oneapi/dpl/internal/scan_by_segment_impl.h b/include/oneapi/dpl/internal/scan_by_segment_impl.h
index 77848128f98..f897c1ec899 100644
--- a/include/oneapi/dpl/internal/scan_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/scan_by_segment_impl.h
@@ -94,7 +94,7 @@ class __seg_scan_wg_kernel;
 template <bool __is_inclusive, typename... Name>
 class __seg_scan_prefix_kernel;
 
-template <bool __is_inclusive>
+template <typename _BackendTag, bool __is_inclusive>
 struct __sycl_scan_by_segment_impl
 {
     template <typename... _Name>
@@ -109,10 +109,6 @@ struct __sycl_scan_by_segment_impl
     operator()(_ExecutionPolicy&& __exec, _Range1&& __keys, _Range2&& __values, _Range3&& __out_values,
                _BinaryPredicate __binary_pred, _BinaryOperator __binary_op, _T __init, _T __identity)
     {
-        constexpr auto __dispatch_tag =
-            oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range1, _Range2, _Range3>();
-        using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
-
         using _CustomName = oneapi::dpl::__internal::__policy_kernel_name<_ExecutionPolicy>;
 
         using _SegScanWgKernel = oneapi::dpl::__par_backend_hetero::__internal::__kernel_name_generator<
@@ -149,12 +145,12 @@ struct __sycl_scan_by_segment_impl
         ::std::size_t __n_groups = __internal::__dpl_ceiling_div(__n, __wgroup_size * __vals_per_item);
 
         auto __partials =
-            oneapi::dpl::__par_backend_hetero::__buffer<__backend_tag, _ExecutionPolicy, __val_type>(__exec, __n_groups)
+            oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, __val_type>(__exec, __n_groups)
                 .get_buffer();
 
         // the number of segment ends found in each work group
         auto __seg_ends =
-            oneapi::dpl::__par_backend_hetero::__buffer<__backend_tag, _ExecutionPolicy, bool>(__exec, __n_groups)
+            oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, bool>(__exec, __n_groups)
                 .get_buffer();
 
         // 1. Work group reduction
@@ -396,9 +392,9 @@ __scan_by_segment_impl_common(__internal::__hetero_tag<_BackendTag>, Policy&& po
 
     constexpr iter_value_t identity = unseq_backend::__known_identity<Operator, iter_value_t>;
 
-    __sycl_scan_by_segment_impl<Inclusive::value>()(::std::forward<Policy>(policy), key_buf.all_view(),
-                                                    value_buf.all_view(), value_output_buf.all_view(), binary_pred,
-                                                    binary_op, init, identity);
+    __sycl_scan_by_segment_impl<_BackendTag, Inclusive::value>()(::std::forward<Policy>(policy), key_buf.all_view(),
+                                                                 value_buf.all_view(), value_output_buf.all_view(),
+                                                                 binary_pred, binary_op, init, identity);
     return result + n;
 }
 

From e87aac5894e263606b714b6e22b4e4e6c98d3df8 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Mon, 26 Feb 2024 10:06:54 +0100
Subject: [PATCH 455/566] include/oneapi/dpl/pstl/algorithm_fwd.h - fix error:
 remove class _IsVector from template params in forwards declaration

---
 include/oneapi/dpl/pstl/algorithm_fwd.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index 7eba977aba1..547dacb13f7 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -71,7 +71,7 @@ template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Fun
 void
 __pattern_walk1(_Tag, _ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _Function) noexcept;
 
-template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Function>
+template <class _ExecutionPolicy, class _RandomAccessIterator, class _Function>
 void
 __pattern_walk1(__parallel_forward_tag, _ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _Function);
 

From ad2fa560c68af870ce46c58013ee3b9a3edd28ac Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <unknown>
Date: Mon, 26 Feb 2024 18:27:45 +0100
Subject: [PATCH 456/566] include/oneapi/dpl/pstl/algorithm_impl.h - additional
 static asserts

---
 include/oneapi/dpl/pstl/algorithm_impl.h | 8 ++++++++
 1 file changed, 8 insertions(+)

diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index d6430ead636..4c95e590eff 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -561,6 +561,8 @@ _ForwardIterator2
 __pattern_walk2_transform_if(_Tag __tag, _ExecutionPolicy&& __exec, _ForwardIterator1 __first1,
                              _ForwardIterator1 __last1, _ForwardIterator2 __first2, _Function __func) noexcept
 {
+    static_assert(__is_host_backend_tag_v<_Tag>);
+
     return __pattern_walk2(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __func);
 }
 
@@ -571,6 +573,8 @@ __pattern_walk3_transform_if(_Tag __tag, _ExecutionPolicy&& __exec, _ForwardIter
                              _ForwardIterator1 __last1, _ForwardIterator2 __first2, _ForwardIterator3 __first3,
                              _Function __func) noexcept
 {
+    static_assert(__is_host_backend_tag_v<_Tag>);
+
     return __pattern_walk3(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __first3,
                            __func);
 }
@@ -4120,6 +4124,8 @@ _ForwardIterator2
 __pattern_swap(_Tag __tag, _ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
                _ForwardIterator2 __first2, _Function __f)
 {
+    static_assert(__is_host_backend_tag_v<_Tag>);
+
     return __pattern_walk2(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __f);
 }
 
@@ -4252,6 +4258,8 @@ __pattern_shift_right(_Tag __tag, _ExecutionPolicy&& __exec, _BidirectionalItera
                       _BidirectionalIterator __last,
                       typename ::std::iterator_traits<_BidirectionalIterator>::difference_type __n)
 {
+    static_assert(__is_host_backend_tag_v<_Tag>);
+
     using _ReverseIterator = typename ::std::reverse_iterator<_BidirectionalIterator>;
 
     auto __res = oneapi::dpl::__internal::__pattern_shift_left(

From 52a6c3312dbdaaf6f0a8a750ddfa6af00ea14934 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <unknown>
Date: Mon, 26 Feb 2024 18:43:37 +0100
Subject: [PATCH 457/566] include/oneapi/dpl/pstl/execution_impl.h - remove
 extra comments

---
 include/oneapi/dpl/pstl/execution_impl.h | 6 ------
 1 file changed, 6 deletions(-)

diff --git a/include/oneapi/dpl/pstl/execution_impl.h b/include/oneapi/dpl/pstl/execution_impl.h
index 68949c792ab..85f9f3efc5e 100644
--- a/include/oneapi/dpl/pstl/execution_impl.h
+++ b/include/oneapi/dpl/pstl/execution_impl.h
@@ -48,8 +48,6 @@ template <class _IsVector>
 struct __serial_tag
 {
     using __is_vector = _IsVector;
-    // backend tag can be change depending on
-    // TBB availability in the environment
     using __backend_tag = __par_backend_tag;
 };
 
@@ -57,16 +55,12 @@ template <class _IsVector>
 struct __parallel_tag
 {
     using __is_vector = _IsVector;
-    // backend tag can be change depending on
-    // TBB availability in the environment
     using __backend_tag = __par_backend_tag;
 };
 
 struct __parallel_forward_tag
 {
     using __is_vector = ::std::false_type;
-    // backend tag can be change depending on
-    // TBB availability in the environment
     using __backend_tag = __par_backend_tag;
 };
 

From 00ca18c5ccc6f783d1b49292a269ffa7bcdfc479 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 27 Feb 2024 10:13:10 +0100
Subject: [PATCH 458/566] 
 include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h - remove old
 implementation of __parallel_for_submitter::operator() without tag

---
 .../pstl/hetero/dpcpp/parallel_backend_sycl.h  | 18 ------------------
 1 file changed, 18 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
index f966b334493..1e1c1f453a2 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
@@ -223,24 +223,6 @@ struct __parallel_for_submitter;
 template <typename... _Name>
 struct __parallel_for_submitter<__internal::__optional_kernel_name<_Name...>>
 {
-    template <typename _ExecutionPolicy, typename _Fp, typename _Index, typename... _Ranges>
-    auto
-    operator()(_ExecutionPolicy&& __exec, _Fp __brick, _Index __count, _Ranges&&... __rngs) const
-    {
-        assert(oneapi::dpl::__ranges::__get_first_range_size(__rngs...) > 0);
-        _PRINT_INFO_IN_DEBUG_MODE(__exec);
-        auto __event = __exec.queue().submit([&__rngs..., &__brick, __count](sycl::handler& __cgh) {
-            //get an access to data under SYCL buffer:
-            oneapi::dpl::__ranges::__require_access(__cgh, __rngs...);
-
-            __cgh.parallel_for<_Name...>(sycl::range</*dim=*/1>(__count), [=](sycl::item</*dim=*/1> __item_id) {
-                auto __idx = __item_id.get_linear_id();
-                __brick(__idx, __rngs...);
-            });
-        });
-        return __future(__event);
-    }
-
     template <typename _ExecutionPolicy, typename _Fp, typename _Index, typename... _Ranges>
     auto
     operator()(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&& __exec, _Fp __brick, _Index __count,

From 3f092198a6016b0bb87b7d600a20065ca339d8f3 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 27 Feb 2024 10:15:54 +0100
Subject: [PATCH 459/566] 
 include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h - remove extra
 changes

---
 .../dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h     | 11 +++++------
 1 file changed, 5 insertions(+), 6 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
index 1e1c1f453a2..d2a935a8463 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
@@ -225,8 +225,7 @@ struct __parallel_for_submitter<__internal::__optional_kernel_name<_Name...>>
 {
     template <typename _ExecutionPolicy, typename _Fp, typename _Index, typename... _Ranges>
     auto
-    operator()(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&& __exec, _Fp __brick, _Index __count,
-               _Ranges&&... __rngs) const
+    operator()(_ExecutionPolicy&& __exec, _Fp __brick, _Index __count, _Ranges&&... __rngs) const
     {
         assert(oneapi::dpl::__ranges::__get_first_range_size(__rngs...) > 0);
         _PRINT_INFO_IN_DEBUG_MODE(__exec);
@@ -247,14 +246,14 @@ struct __parallel_for_submitter<__internal::__optional_kernel_name<_Name...>>
 //for some algorithms happens that size of processing range is n, but amount of iterations is n/2.
 template <typename _ExecutionPolicy, typename _Fp, typename _Index, typename... _Ranges>
 auto
-__parallel_for(oneapi::dpl::__internal::__device_backend_tag __backend_tag, _ExecutionPolicy&& __exec, _Fp __brick,
-               _Index __count, _Ranges&&... __rngs)
+__parallel_for(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy&& __exec, _Fp __brick, _Index __count,
+               _Ranges&&... __rngs)
 {
     using _CustomName = oneapi::dpl::__internal::__policy_kernel_name<_ExecutionPolicy>;
     using _ForKernel = oneapi::dpl::__par_backend_hetero::__internal::__kernel_name_provider<_CustomName>;
 
-    return __parallel_for_submitter<_ForKernel>()(__backend_tag, ::std::forward<_ExecutionPolicy>(__exec), __brick,
-                                                  __count, ::std::forward<_Ranges>(__rngs)...);
+    return __parallel_for_submitter<_ForKernel>()(::std::forward<_ExecutionPolicy>(__exec), __brick, __count,
+                                                  ::std::forward<_Ranges>(__rngs)...);
 }
 
 //------------------------------------------------------------------------

From 4d84c6f208195eab1cec5d222702ec5051c85814 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 27 Feb 2024 20:21:07 +0100
Subject: [PATCH 460/566] include/oneapi/dpl/pstl/iterator_defs.h - add '::'
 before std namespace

---
 include/oneapi/dpl/pstl/iterator_defs.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/include/oneapi/dpl/pstl/iterator_defs.h b/include/oneapi/dpl/pstl/iterator_defs.h
index dec30db3774..f757770f28f 100644
--- a/include/oneapi/dpl/pstl/iterator_defs.h
+++ b/include/oneapi/dpl/pstl/iterator_defs.h
@@ -32,7 +32,7 @@ namespace __internal
 template <typename _IteratorTag, typename... _IteratorTypes>
 auto
 __is_iterator_of(int) -> decltype(
-    std::conjunction<::std::is_base_of<
+    ::std::conjunction<::std::is_base_of<
         _IteratorTag, typename ::std::iterator_traits<::std::decay_t<_IteratorTypes>>::iterator_category>...>{});
 
 template <typename... _IteratorTypes>

From e31aa1bba8dbe2b462df81551c585713619416dd Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 27 Feb 2024 18:08:38 +0100
Subject: [PATCH 461/566] include/oneapi/dpl/pstl/execution_impl.h - make
 __select_backend as non-constexpr function with params again for
 compatibility with up-stream code

---
 include/oneapi/dpl/pstl/execution_impl.h | 32 +++++++++---------------
 1 file changed, 12 insertions(+), 20 deletions(-)

diff --git a/include/oneapi/dpl/pstl/execution_impl.h b/include/oneapi/dpl/pstl/execution_impl.h
index 85f9f3efc5e..fe3ae6bc52c 100644
--- a/include/oneapi/dpl/pstl/execution_impl.h
+++ b/include/oneapi/dpl/pstl/execution_impl.h
@@ -69,38 +69,30 @@ using __tag_type = ::std::conditional_t<
     __internal::__is_random_access_iterator_v<_IteratorTypes...>, __parallel_tag<_IsVector>,
     ::std::conditional_t<__is_forward_iterator_v<_IteratorTypes...>, __parallel_forward_tag, __serial_tag<_IsVector>>>;
 
-template <class _ExecutionPolicy, class... _IteratorTypes>
-constexpr ::std::enable_if_t<
-    ::std::is_same_v<::std::decay_t<_ExecutionPolicy>, oneapi::dpl::execution::sequenced_policy>,
-    __serial_tag<std::false_type>>
-__select_backend()
+template <class... _IteratorTypes>
+__serial_tag<std::false_type>
+__select_backend(oneapi::dpl::execution::sequenced_policy, _IteratorTypes&&...)
 {
     return {};
 }
 
-template <class _ExecutionPolicy, class... _IteratorTypes>
-constexpr ::std::enable_if_t<
-    ::std::is_same_v<::std::decay_t<_ExecutionPolicy>, oneapi::dpl::execution::unsequenced_policy>,
-    __serial_tag<__internal::__is_random_access_iterator<_IteratorTypes...>>>
-__select_backend()
+template <class... _IteratorTypes>
+__serial_tag<__internal::__is_random_access_iterator<_IteratorTypes...>>
+__select_backend(oneapi::dpl::execution::unsequenced_policy, _IteratorTypes&&...)
 {
     return {};
 }
 
-template <class _ExecutionPolicy, class... _IteratorTypes>
-constexpr ::std::enable_if_t<
-    ::std::is_same_v<::std::decay_t<_ExecutionPolicy>, oneapi::dpl::execution::parallel_policy>,
-    __tag_type<std::false_type, _IteratorTypes...>>
-__select_backend()
+template <class... _IteratorTypes>
+__tag_type<std::false_type, _IteratorTypes...>
+__select_backend(oneapi::dpl::execution::parallel_policy, _IteratorTypes&&...)
 {
     return {};
 }
 
-template <class _ExecutionPolicy, class... _IteratorTypes>
-constexpr ::std::enable_if_t<
-    ::std::is_same_v<::std::decay_t<_ExecutionPolicy>, oneapi::dpl::execution::parallel_unsequenced_policy>,
-    __tag_type<__internal::__is_random_access_iterator<_IteratorTypes...>, _IteratorTypes...>>
-__select_backend()
+template <class... _IteratorTypes>
+__tag_type<__internal::__is_random_access_iterator<_IteratorTypes...>, _IteratorTypes...>
+__select_backend(oneapi::dpl::execution::parallel_unsequenced_policy, _IteratorTypes&&...)
 {
     return {};
 }

From c5a882b05e2814198812e3cbdd6870f2a906db70 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 27 Feb 2024 18:12:30 +0100
Subject: [PATCH 462/566] 
 include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h - make
 __select_backend as non-constexpr function with params again for
 compatibility with up-stream code

---
 .../pstl/hetero/dpcpp/execution_sycl_defs.h   | 30 +++++++------------
 1 file changed, 10 insertions(+), 20 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h b/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h
index 4a63ea9d151..01589e1ff20 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h
@@ -296,35 +296,25 @@ struct __device_backend_tag
 {
 };
 
-template <typename HeteroPolicy>
-struct __select_backend_for_hetero_policy_trait;
-
-template <typename KernelName>
-struct __select_backend_for_hetero_policy_trait<execution::device_policy<KernelName>>
+template <class... _IteratorTypes, typename _KernelName>
+::std::enable_if_t<__is_random_access_iterator_v<_IteratorTypes...>, __hetero_tag<__device_backend_tag>>
+__select_backend(const execution::device_policy<_KernelName>&, _IteratorTypes&&...)
 {
-    using __backend_tag = __device_backend_tag;
-};
+    return {};
+}
 
 #if _ONEDPL_FPGA_DEVICE
 struct __fpga_backend_tag : __device_backend_tag
 {
 };
 
-template <unsigned int factor, typename KernelName>
-struct __select_backend_for_hetero_policy_trait<execution::fpga_policy<factor, KernelName>>
-{
-    using __backend_tag = __fpga_backend_tag;
-};
-#endif
-
-template <class _ExecutionPolicy, class... _IteratorTypes>
-constexpr ::std::enable_if_t<
-    __is_random_access_iterator_v<_IteratorTypes...>,
-    __hetero_tag<typename __select_backend_for_hetero_policy_trait<::std::decay_t<_ExecutionPolicy>>::__backend_tag>>
-__select_backend()
+template <class... _IteratorTypes, unsigned int _Factor, typename _KernelName>
+::std::enable_if_t<__is_random_access_iterator_v<_IteratorTypes...>, __hetero_tag<__fpga_backend_tag>>
+__select_backend(const execution::fpga_policy<_Factor, _KernelName>&, _IteratorTypes&&...)
 {
-    return {}; // return __hetero_tag<__device_backend_tag> or __hetero_tag<__fpga_backend_tag>
+    return {};
 }
+#endif
 
 } // namespace __internal
 

From 8a914bb996c3cd7ccb548a8c3448ec6b4af4de83 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 27 Feb 2024 18:12:43 +0100
Subject: [PATCH 463/566] 
 include/oneapi/dpl/pstl/hetero/dpcpp/utils_ranges_sycl.h - make
 __select_backend as non-constexpr function with params again for
 compatibility with up-stream code

---
 .../dpl/pstl/hetero/dpcpp/utils_ranges_sycl.h | 22 ++++++++++++++-----
 1 file changed, 16 insertions(+), 6 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/utils_ranges_sycl.h b/include/oneapi/dpl/pstl/hetero/dpcpp/utils_ranges_sycl.h
index b38c8d32e11..cdb43159d15 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/utils_ranges_sycl.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/utils_ranges_sycl.h
@@ -719,15 +719,25 @@ struct __get_sycl_range
 //TODO required correct implementation of this __ranges::__select_backend()
 // 1. There is still not RA ranges checks
 // 2. Obviously, a return tag is not necessarily oneapi::dpl::__internal::__hetero_tag
-template <class _ExecutionPolicy, typename... _Ranges>
-constexpr oneapi::dpl::__internal::__hetero_tag<
-    typename oneapi::dpl::__internal::__select_backend_for_hetero_policy_trait<
-        ::std::decay_t<_ExecutionPolicy>>::__backend_tag>
-__select_backend()
+template <typename _KernelName, typename... _Ranges>
+oneapi::dpl::__internal::__hetero_tag<oneapi::dpl::__internal::__device_backend_tag>
+__select_backend(const execution::device_policy<_KernelName>&, _Ranges&&...)
 {
-    return {}; // return __hetero_tag<__device_backend_tag> or __hetero_tag<__fpga_backend_tag>
+    return {};
 }
 
+#if _ONEDPL_FPGA_DEVICE
+//TODO required correct implementation of this __ranges::__select_backend()
+// 1. There is still not RA ranges checks
+// 2. Obviously, a return tag is not necessarily oneapi::dpl::__internal::__hetero_tag
+template <unsigned int _Factor, typename _KernelName, typename... _Ranges>
+oneapi::dpl::__internal::__hetero_tag<oneapi::dpl::__internal::__fpga_backend_tag>
+__select_backend(const execution::fpga_policy<_Factor, _KernelName>&, _Ranges&&...)
+{
+    return {};
+}
+#endif
+
 } // namespace __ranges
 } // namespace dpl
 } // namespace oneapi

From 68539ba91ff8523bacbd7f3ffbe39b77970c2dc6 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 27 Feb 2024 18:21:58 +0100
Subject: [PATCH 464/566] 
 include/oneapi/dpl/internal/async_impl/glue_async_impl.h - rewrite calls of
 __select_backend() function with run-time params

---
 .../dpl/internal/async_impl/glue_async_impl.h | 46 +++++++------------
 1 file changed, 17 insertions(+), 29 deletions(-)

diff --git a/include/oneapi/dpl/internal/async_impl/glue_async_impl.h b/include/oneapi/dpl/internal/async_impl/glue_async_impl.h
index 586360f83af..06c441fb81d 100644
--- a/include/oneapi/dpl/internal/async_impl/glue_async_impl.h
+++ b/include/oneapi/dpl/internal/async_impl/glue_async_impl.h
@@ -43,8 +43,7 @@ auto
 transform_async(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _ForwardIterator1 __last,
                 _ForwardIterator2 __result, _UnaryOperation __op, _Events&&... __dependencies)
 {
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first, __result);
 
     wait_for_all(::std::forward<_Events>(__dependencies)...);
     auto ret_val = oneapi::dpl::__internal::__pattern_walk2_async(
@@ -62,8 +61,7 @@ transform_async(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardI
                 _ForwardIterator2 __first2, _ForwardIterator __result, _BinaryOperation __op,
                 _Events&&... __dependencies)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1,
-                                                                              _ForwardIterator2, _ForwardIterator>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first1, __first2, __result);
 
     wait_for_all(::std::forward<_Events>(__dependencies)...);
     auto ret_val = oneapi::dpl::__internal::__pattern_walk3_async(
@@ -79,8 +77,7 @@ auto
 copy_async(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _ForwardIterator1 __last, _ForwardIterator2 __result,
            _Events&&... __dependencies)
 {
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first, __result);
 
     wait_for_all(::std::forward<_Events>(__dependencies)...);
     auto ret_val = oneapi::dpl::__internal::__pattern_walk2_brick_async(
@@ -102,7 +99,7 @@ sort_async(_ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last, _Comp
     auto __keep = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read_write, _Iterator>();
     auto __buf = __keep(__first, __last);
 
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _Iterator>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first);
     using __backend_tag = typename decltype(__dispatch_tag)::__backend_tag;
 
     return __par_backend_hetero::__parallel_stable_sort(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
@@ -127,7 +124,7 @@ auto
 for_each_async(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last, _Function __f,
                _Events&&... __dependencies)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first);
 
     wait_for_all(::std::forward<_Events>(__dependencies)...);
     auto ret_val = oneapi::dpl::__internal::__pattern_walk1_async(
@@ -144,7 +141,7 @@ auto
 reduce_async(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last, _Tp __init,
              _BinaryOperation __binary_op, _Events&&... __dependencies)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first);
 
     wait_for_all(::std::forward<_Events>(__dependencies)...);
     auto ret_val = oneapi::dpl::__internal::__pattern_transform_reduce_async(
@@ -181,7 +178,7 @@ auto
 fill_async(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last, const _Tp& __value,
            _Events&&... __dependencies)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first);
 
     wait_for_all(::std::forward<_Events>(__dependencies)...);
     return oneapi::dpl::__internal::__pattern_fill_async(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
@@ -198,8 +195,7 @@ auto
 transform_reduce_async(_ExecutionPolicy&& __exec, _ForwardIt1 __first1, _ForwardIt1 __last1, _ForwardIt2 __first2,
                        _T __init, _BinaryOp1 __binary_op1, _BinaryOp2 __binary_op2, _Events&&... __dependencies)
 {
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIt1, _ForwardIt2>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first1, __first2);
 
     wait_for_all(::std::forward<_Events>(__dependencies)...);
     return oneapi::dpl::__internal::__pattern_transform_reduce_async(
@@ -214,7 +210,7 @@ auto
 transform_reduce_async(_ExecutionPolicy&& __exec, _ForwardIt __first, _ForwardIt __last, _T __init,
                        _BinaryOp __binary_op, _UnaryOp __unary_op, _Events&&... __dependencies)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIt>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first);
 
     wait_for_all(::std::forward<_Events>(__dependencies)...);
     return oneapi::dpl::__internal::__pattern_transform_reduce_async(
@@ -241,8 +237,7 @@ auto
 inclusive_scan_async(_ExecutionPolicy&& __exec, _ForwardIt1 __first1, _ForwardIt1 __last1, _ForwardIt2 __first2,
                      _Events&&... __dependencies)
 {
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIt1, _ForwardIt2>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first1, __first2);
 
     using _ValueType = typename ::std::iterator_traits<_ForwardIt1>::value_type;
     wait_for_all(::std::forward<_Events>(__dependencies)...);
@@ -258,8 +253,7 @@ auto
 inclusive_scan_async(_ExecutionPolicy&& __exec, _ForwardIt1 __first1, _ForwardIt1 __last1, _ForwardIt2 __first2,
                      _BinaryOperation __binary_op, _Events&&... __dependencies)
 {
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIt1, _ForwardIt2>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first1, __first2);
 
     wait_for_all(::std::forward<_Events>(__dependencies)...);
     return oneapi::dpl::__internal::__pattern_transform_scan_async(
@@ -275,8 +269,7 @@ auto
 inclusive_scan_async(_ExecutionPolicy&& __exec, _ForwardIt1 __first1, _ForwardIt1 __last1, _ForwardIt2 __first2,
                      _BinaryOperation __binary_op, _T __init, _Events&&... __dependencies)
 {
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIt1, _ForwardIt2>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first1, __first2);
 
     wait_for_all(::std::forward<_Events>(__dependencies)...);
     return oneapi::dpl::__internal::__pattern_transform_scan_async(
@@ -290,8 +283,7 @@ auto
 exclusive_scan_async(_ExecutionPolicy&& __exec, _ForwardIt1 __first1, _ForwardIt1 __last1, _ForwardIt2 __first2,
                      _T __init, _Events&&... __dependencies)
 {
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIt1, _ForwardIt2>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first1, __first2);
 
     wait_for_all(::std::forward<_Events>(__dependencies)...);
     return oneapi::dpl::__internal::__pattern_transform_scan_async(
@@ -307,8 +299,7 @@ auto
 exclusive_scan_async(_ExecutionPolicy&& __exec, _ForwardIt1 __first1, _ForwardIt1 __last1, _ForwardIt2 __first2,
                      _T __init, _BinaryOperation __binary_op, _Events&&... __dependencies)
 {
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIt1, _ForwardIt2>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first1, __first2);
 
     wait_for_all(::std::forward<_Events>(__dependencies)...);
     return oneapi::dpl::__internal::__pattern_transform_scan_async(
@@ -324,8 +315,7 @@ transform_exclusive_scan_async(_ExecutionPolicy&& __exec, _ForwardIt1 __first1,
                                _ForwardIt2 __first2, _T __init, _BinaryOperation __binary_op,
                                _UnaryOperation __unary_op, _Events&&... __dependencies)
 {
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIt1, _ForwardIt2>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first1, __first2);
 
     wait_for_all(::std::forward<_Events>(__dependencies)...);
     return oneapi::dpl::__internal::__pattern_transform_scan_async(__dispatch_tag,
@@ -342,8 +332,7 @@ transform_inclusive_scan_async(_ExecutionPolicy&& __exec, _ForwardIt1 __first1,
                                _ForwardIt2 __first2, _BinaryOperation __binary_op, _UnaryOperation __unary_op,
                                _Events&&... __dependencies)
 {
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIt1, _ForwardIt2>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first1, __first2);
 
     wait_for_all(::std::forward<_Events>(__dependencies)...);
     return oneapi::dpl::__internal::__pattern_transform_scan_async(
@@ -360,8 +349,7 @@ transform_inclusive_scan_async(_ExecutionPolicy&& __exec, _ForwardIt1 __first1,
                                _ForwardIt2 __first2, _BinaryOperation __binary_op, _UnaryOperation __unary_op,
                                _T __init, _Events&&... __dependencies)
 {
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIt1, _ForwardIt2>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first1, __first2);
 
     wait_for_all(::std::forward<_Events>(__dependencies)...);
     return oneapi::dpl::__internal::__pattern_transform_scan_async(__dispatch_tag,

From f415b62831c7236857669af3c10cce9fe755918d Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 27 Feb 2024 18:26:11 +0100
Subject: [PATCH 465/566] include/oneapi/dpl/internal/binary_search_impl.h -
 rewrite calls of __select_backend() function with run-time params

---
 .../oneapi/dpl/internal/binary_search_impl.h   | 18 ++++++------------
 1 file changed, 6 insertions(+), 12 deletions(-)

diff --git a/include/oneapi/dpl/internal/binary_search_impl.h b/include/oneapi/dpl/internal/binary_search_impl.h
index 00505df5cd6..4a2d7be1391 100644
--- a/include/oneapi/dpl/internal/binary_search_impl.h
+++ b/include/oneapi/dpl/internal/binary_search_impl.h
@@ -216,8 +216,7 @@ oneapi::dpl::__internal::__enable_if_execution_policy<Policy, OutputIterator>
 lower_bound(Policy&& policy, InputIterator1 start, InputIterator1 end, InputIterator2 value_start,
             InputIterator2 value_end, OutputIterator result)
 {
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<Policy, InputIterator1, InputIterator2, OutputIterator>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(policy, start, value_start, result);
 
     return internal::lower_bound_impl(__dispatch_tag, ::std::forward<Policy>(policy), start, end, value_start,
                                       value_end, result, oneapi::dpl::__internal::__pstl_less());
@@ -229,8 +228,7 @@ oneapi::dpl::__internal::__enable_if_execution_policy<Policy, OutputIterator>
 lower_bound(Policy&& policy, InputIterator1 start, InputIterator1 end, InputIterator2 value_start,
             InputIterator2 value_end, OutputIterator result, StrictWeakOrdering comp)
 {
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<Policy, InputIterator1, InputIterator2, OutputIterator>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(policy, start, value_start, result);
 
     return internal::lower_bound_impl(__dispatch_tag, ::std::forward<Policy>(policy), start, end, value_start,
                                       value_end, result, comp);
@@ -244,8 +242,7 @@ oneapi::dpl::__internal::__enable_if_execution_policy<Policy, OutputIterator>
 upper_bound(Policy&& policy, InputIterator1 start, InputIterator1 end, InputIterator2 value_start,
             InputIterator2 value_end, OutputIterator result)
 {
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<Policy, InputIterator1, InputIterator2, OutputIterator>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(policy, start, value_start, result);
 
     return internal::upper_bound_impl(__dispatch_tag, ::std::forward<Policy>(policy), start, end, value_start,
                                       value_end, result, oneapi::dpl::__internal::__pstl_less());
@@ -257,8 +254,7 @@ oneapi::dpl::__internal::__enable_if_execution_policy<Policy, OutputIterator>
 upper_bound(Policy&& policy, InputIterator1 start, InputIterator1 end, InputIterator2 value_start,
             InputIterator2 value_end, OutputIterator result, StrictWeakOrdering comp)
 {
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<Policy, InputIterator1, InputIterator2, OutputIterator>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(policy, start, value_start, result);
 
     return internal::upper_bound_impl(__dispatch_tag, ::std::forward<Policy>(policy), start, end, value_start,
                                       value_end, result, comp);
@@ -273,8 +269,7 @@ oneapi::dpl::__internal::__enable_if_execution_policy<Policy, OutputIterator>
 binary_search(Policy&& policy, InputIterator1 start, InputIterator1 end, InputIterator2 value_start,
               InputIterator2 value_end, OutputIterator result)
 {
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<Policy, InputIterator1, InputIterator2, OutputIterator>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(policy, start, value_start, result);
 
     return internal::binary_search_impl(__dispatch_tag, ::std::forward<Policy>(policy), start, end, value_start,
                                         value_end, result, oneapi::dpl::__internal::__pstl_less());
@@ -286,8 +281,7 @@ oneapi::dpl::__internal::__enable_if_execution_policy<Policy, OutputIterator>
 binary_search(Policy&& policy, InputIterator1 start, InputIterator1 end, InputIterator2 value_start,
               InputIterator2 value_end, OutputIterator result, StrictWeakOrdering comp)
 {
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<Policy, InputIterator1, InputIterator2, OutputIterator>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(policy, start, value_start, result);
 
     return internal::binary_search_impl(__dispatch_tag, ::std::forward<Policy>(policy), start, end, value_start,
                                         value_end, result, comp);

From b5e55039f6f36c88ab2d3b2303e9136f85a157d2 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 27 Feb 2024 18:26:53 +0100
Subject: [PATCH 466/566] 
 include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h - rewrite calls
 of __select_backend() function with run-time params

---
 include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)

diff --git a/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h b/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h
index c818213b180..211046642ce 100644
--- a/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h
@@ -189,8 +189,7 @@ oneapi::dpl::__internal::__enable_if_execution_policy<Policy, OutputIterator>
 exclusive_scan_by_segment(Policy&& policy, InputIterator1 first1, InputIterator1 last1, InputIterator2 first2,
                           OutputIterator result, T init, BinaryPredicate binary_pred, Operator binary_op)
 {
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<Policy, InputIterator1, InputIterator2, OutputIterator>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(policy, first1, first2, result);
 
     return internal::pattern_exclusive_scan_by_segment(__dispatch_tag, ::std::forward<Policy>(policy), first1, last1,
                                                        first2, result, init, binary_pred, binary_op);

From 0b348947da366786c17822313713cbc67e5cc921 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 27 Feb 2024 18:27:35 +0100
Subject: [PATCH 467/566] 
 include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h - rewrite calls
 of __select_backend() function with run-time params

---
 include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)

diff --git a/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h b/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h
index 05c6f6635f7..95dc8fa62db 100644
--- a/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h
@@ -152,8 +152,7 @@ oneapi::dpl::__internal::__enable_if_execution_policy<Policy, OutputIterator>
 inclusive_scan_by_segment(Policy&& policy, InputIterator1 first1, InputIterator1 last1, InputIterator2 first2,
                           OutputIterator result, BinaryPredicate binary_pred, BinaryOperator binary_op)
 {
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<Policy, InputIterator1, InputIterator2, OutputIterator>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(policy, first1, first2, result);
 
     return internal::pattern_inclusive_scan_by_segment(__dispatch_tag, ::std::forward<Policy>(policy), first1, last1,
                                                        first2, result, binary_pred, binary_op);

From 237d2fad6c37d0762651852b39a27c9c0a53ab5e Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 27 Feb 2024 18:28:26 +0100
Subject: [PATCH 468/566] include/oneapi/dpl/internal/reduce_by_segment_impl.h
 - rewrite calls of __select_backend() function with run-time params

---
 include/oneapi/dpl/internal/reduce_by_segment_impl.h | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)

diff --git a/include/oneapi/dpl/internal/reduce_by_segment_impl.h b/include/oneapi/dpl/internal/reduce_by_segment_impl.h
index 762bbc70602..5cda04d6127 100644
--- a/include/oneapi/dpl/internal/reduce_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/reduce_by_segment_impl.h
@@ -632,8 +632,7 @@ oneapi::dpl::__internal::__enable_if_execution_policy<Policy, ::std::pair<Output
 reduce_by_segment(Policy&& policy, InputIterator1 first1, InputIterator1 last1, InputIterator2 first2,
                   OutputIterator1 result1, OutputIterator2 result2, BinaryPred binary_pred, BinaryOperator binary_op)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<Policy, InputIterator1, InputIterator2,
-                                                                              OutputIterator1, OutputIterator2>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(policy, first1, first2, result1, result2);
 
     return internal::reduce_by_segment_impl(__dispatch_tag, ::std::forward<Policy>(policy), first1, last1, first2,
                                             result1, result2, binary_pred, binary_op);

From 3511276567fa64e3cb028e12c665b8e30fd29e98 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 27 Feb 2024 18:31:34 +0100
Subject: [PATCH 469/566] 
 include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h - rewrite calls
 of __select_backend() function with run-time params (__select_backend +
 __select_backend_for_loop)

---
 .../dpl/pstl/experimental/internal/for_loop_impl.h   | 12 ++++++------
 1 file changed, 6 insertions(+), 6 deletions(-)

diff --git a/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h b/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
index 1bc0a309bcd..a6a6952738a 100644
--- a/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
+++ b/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
@@ -486,18 +486,18 @@ __pattern_for_loop(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _
 
 // Special versions for for_loop: handles both iterators and integral types(treated as random access iterators)
 template <typename _ExecutionPolicy, typename _Ip>
-constexpr auto
-__select_backend_for_loop()
+auto
+__select_backend_for_loop(_ExecutionPolicy&& exec, _Ip __start)
 {
     static_assert(oneapi::dpl::__internal::__is_host_execution_policy<::std::decay_t<_ExecutionPolicy>>::value);
 
     if constexpr (::std::is_integral_v<_Ip>)
     {
-        return __select_backend<_ExecutionPolicy>();
+        return __select_backend(exec);
     }
     else
     {
-        return __select_backend<_ExecutionPolicy, _Ip>();
+        return __select_backend(exec, __start);
     }
 }
 
@@ -507,7 +507,7 @@ void
 __for_loop_impl(_ExecutionPolicy&& __exec, _Ip __start, _Ip __finish, _Fp&& __f, _Sp __stride,
                 ::std::tuple<_Rest...>&& __t, ::std::index_sequence<_Is...>)
 {
-    constexpr auto __dispatch_tag = __select_backend_for_loop<_ExecutionPolicy, _Ip>();
+    const auto __dispatch_tag = __select_backend_for_loop(__exec, __start);
 
     oneapi::dpl::__internal::__pattern_for_loop(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __start,
                                                 __finish, __f, __stride, ::std::get<_Is>(::std::move(__t))...);
@@ -519,7 +519,7 @@ void
 __for_loop_n_impl(_ExecutionPolicy&& __exec, _Ip __start, _Size __n, _Fp&& __f, _Sp __stride,
                   ::std::tuple<_Rest...>&& __t, ::std::index_sequence<_Is...>)
 {
-    constexpr auto __dispatch_tag = __select_backend_for_loop<_ExecutionPolicy, _Ip>();
+    const auto __dispatch_tag = __select_backend_for_loop(__exec, __start);
 
     oneapi::dpl::__internal::__pattern_for_loop_n(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __start,
                                                   __n, __f, __stride, ::std::get<_Is>(::std::move(__t))...);

From 7c2cb0c3c5746adfcbb76b374b37e12213f21e2f Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 27 Feb 2024 18:33:10 +0100
Subject: [PATCH 470/566] include/oneapi/dpl/pstl/histogram_impl.h - rewrite
 calls of __select_backend() function with run-time params

---
 include/oneapi/dpl/pstl/histogram_impl.h | 7 ++-----
 1 file changed, 2 insertions(+), 5 deletions(-)

diff --git a/include/oneapi/dpl/pstl/histogram_impl.h b/include/oneapi/dpl/pstl/histogram_impl.h
index 8e75800e2af..885600f0288 100644
--- a/include/oneapi/dpl/pstl/histogram_impl.h
+++ b/include/oneapi/dpl/pstl/histogram_impl.h
@@ -52,8 +52,7 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _RandomA
 histogram(_ExecutionPolicy&& exec, _RandomAccessIterator1 first, _RandomAccessIterator1 last, _Size num_bins,
           _ValueType first_bin_min_val, _ValueType last_bin_max_val, _RandomAccessIterator2 histogram_first)
 {
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(exec, first, histogram_first);
 
     oneapi::dpl::__internal::__pattern_histogram(
         __dispatch_tag, ::std::forward<_ExecutionPolicy>(exec), first, last, num_bins,
@@ -69,9 +68,7 @@ histogram(_ExecutionPolicy&& exec, _RandomAccessIterator1 first, _RandomAccessIt
           _RandomAccessIterator2 boundary_first, _RandomAccessIterator2 boundary_last,
           _RandomAccessIterator3 histogram_first)
 {
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2,
-                                                  _RandomAccessIterator3>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(exec, first, boundary_first, histogram_first);
 
     ::std::ptrdiff_t num_bins = boundary_last - boundary_first - 1;
     oneapi::dpl::__internal::__pattern_histogram(

From 0919e372763c82181ec3e5aa7a8add57127e70d2 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 27 Feb 2024 18:35:57 +0100
Subject: [PATCH 471/566] include/oneapi/dpl/pstl/glue_numeric_ranges_impl.h -
 rewrite calls of __select_backend() function with run-time params

---
 include/oneapi/dpl/pstl/glue_numeric_ranges_impl.h | 12 ++++++------
 1 file changed, 6 insertions(+), 6 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_numeric_ranges_impl.h b/include/oneapi/dpl/pstl/glue_numeric_ranges_impl.h
index 162c6906595..521ebee46b7 100644
--- a/include/oneapi/dpl/pstl/glue_numeric_ranges_impl.h
+++ b/include/oneapi/dpl/pstl/glue_numeric_ranges_impl.h
@@ -63,7 +63,7 @@ template <typename _ExecutionPolicy, typename _Range1, typename _Range2, typenam
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _Tp>
 transform_reduce(_ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&& __rng2, _Tp __init)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range1, _Range2>();
+    const auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend(__exec, __rng1, __rng2);
 
     using _ValueType = oneapi::dpl::__internal::__value_t<_Range1>;
     return oneapi::dpl::__internal::__ranges::__pattern_transform_reduce(
@@ -78,7 +78,7 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _Tp>
 transform_reduce(_ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&& __rng2, _Tp __init,
                  _BinaryOperation1 __binary_op1, _BinaryOperation2 __binary_op2)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range1, _Range2>();
+    const auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend(__exec, __rng1, __rng2);
 
     return oneapi::dpl::__internal::__ranges::__pattern_transform_reduce(
         __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), views::all_read(::std::forward<_Range1>(__rng1)),
@@ -90,7 +90,7 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _Tp>
 transform_reduce(_ExecutionPolicy&& __exec, _Range&& __rng, _Tp __init, _BinaryOperation __binary_op,
                  _UnaryOperation __unary_op)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range>();
+    const auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend(__exec, __rng);
 
     return oneapi::dpl::__internal::__ranges::__pattern_transform_reduce(
         __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), views::all_read(::std::forward<_Range>(__rng)),
@@ -160,7 +160,7 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy,
 transform_exclusive_scan(_ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&& __rng2, _Tp __init,
                          _BinaryOperation __binary_op, _UnaryOperation __unary_op)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range1, _Range2>();
+    const auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend(__exec, __rng1, __rng2);
 
     return oneapi::dpl::__internal::__ranges::__pattern_transform_scan(
         __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), views::all_read(::std::forward<_Range1>(__rng1)),
@@ -177,7 +177,7 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy,
 transform_inclusive_scan(_ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&& __rng2, _BinaryOperation __binary_op,
                          _UnaryOperation __unary_op, _Tp __init)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range1, _Range2>();
+    const auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend(__exec, __rng1, __rng2);
 
     return oneapi::dpl::__internal::__ranges::__pattern_transform_scan(
         __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), views::all_read(::std::forward<_Range1>(__rng1)),
@@ -192,7 +192,7 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy,
 transform_inclusive_scan(_ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&& __rng2, _BinaryOperation __binary_op,
                          _UnaryOperation __unary_op)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range1, _Range2>();
+    const auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend(__exec, __rng1, __rng2);
 
     return oneapi::dpl::__internal::__ranges::__pattern_transform_scan(
         __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), views::all_read(::std::forward<_Range1>(__rng1)),

From 44a29b8d5f6652a564cf855389fd11b8c0228bb4 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 27 Feb 2024 18:39:16 +0100
Subject: [PATCH 472/566] include/oneapi/dpl/pstl/glue_numeric_impl.h - rewrite
 calls of __select_backend() function with run-time params

---
 include/oneapi/dpl/pstl/glue_numeric_impl.h | 20 +++++++-------------
 1 file changed, 7 insertions(+), 13 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_numeric_impl.h b/include/oneapi/dpl/pstl/glue_numeric_impl.h
index d6a98e6529f..17ed09d0ca4 100644
--- a/include/oneapi/dpl/pstl/glue_numeric_impl.h
+++ b/include/oneapi/dpl/pstl/glue_numeric_impl.h
@@ -71,8 +71,7 @@ transform_reduce(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _Forward
 {
     typedef typename ::std::iterator_traits<_ForwardIterator1>::value_type _InputType;
 
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first1, __first2);
 
     return oneapi::dpl::__internal::__pattern_transform_reduce(
         __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __init,
@@ -85,8 +84,7 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _Tp>
 transform_reduce(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
                  _ForwardIterator2 __first2, _Tp __init, _BinaryOperation1 __binary_op1, _BinaryOperation2 __binary_op2)
 {
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first1, __first2);
 
     return oneapi::dpl::__internal::__pattern_transform_reduce(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
                                                                __first1, __last1, __first2, __init, __binary_op1,
@@ -98,7 +96,7 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _Tp>
 transform_reduce(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last, _Tp __init,
                  _BinaryOperation __binary_op, _UnaryOperation __unary_op)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first);
 
     return oneapi::dpl::__internal::__pattern_transform_reduce(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
                                                                __first, __last, __init, __binary_op, __unary_op);
@@ -225,8 +223,7 @@ transform_exclusive_scan(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _
                          _ForwardIterator2 __result, _Tp __init, _BinaryOperation __binary_op,
                          _UnaryOperation __unary_op)
 {
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first, __result);
 
     return oneapi::dpl::__internal::__pattern_transform_scan(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
                                                              __first, __last, __result, __unary_op, __init, __binary_op,
@@ -242,8 +239,7 @@ transform_inclusive_scan(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _
                          _ForwardIterator2 __result, _BinaryOperation __binary_op, _UnaryOperation __unary_op,
                          _Tp __init)
 {
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first, __result);
 
     return oneapi::dpl::__internal::__pattern_transform_scan(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
                                                              __first, __last, __result, __unary_op, __init, __binary_op,
@@ -256,8 +252,7 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _Forward
 transform_inclusive_scan(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _ForwardIterator1 __last,
                          _ForwardIterator2 __result, _BinaryOperation __binary_op, _UnaryOperation __unary_op)
 {
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first, __result);
 
     return oneapi::dpl::__internal::__pattern_transform_scan(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
                                                              __first, __last, __result, __unary_op, __binary_op,
@@ -274,8 +269,7 @@ adjacent_difference(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _Forwa
     if (__first == __last)
         return __d_first;
 
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first, __d_first);
 
     return oneapi::dpl::__internal::__pattern_adjacent_difference(
         __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __d_first, __op);

From e01bd896127fc966cc00cc015d48d35897ea764b Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 27 Feb 2024 18:44:05 +0100
Subject: [PATCH 473/566] include/oneapi/dpl/pstl/glue_memory_impl.h - rewrite
 calls of __select_backend() function with run-time params

---
 include/oneapi/dpl/pstl/glue_memory_impl.h | 30 +++++++++-------------
 1 file changed, 12 insertions(+), 18 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_memory_impl.h b/include/oneapi/dpl/pstl/glue_memory_impl.h
index 9058b0dc07c..60f02af73d0 100644
--- a/include/oneapi/dpl/pstl/glue_memory_impl.h
+++ b/include/oneapi/dpl/pstl/glue_memory_impl.h
@@ -45,8 +45,7 @@ uninitialized_copy(_ExecutionPolicy&& __exec, _InputIterator __first, _InputIter
     typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _ValueType2;
     typedef ::std::decay_t<_ExecutionPolicy> _DecayedExecutionPolicy;
 
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _InputIterator, _ForwardIterator>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first, __result);
 
     if constexpr (::std::is_trivial_v<_ValueType1> && ::std::is_trivial_v<_ValueType2>)
     {
@@ -70,8 +69,7 @@ uninitialized_copy_n(_ExecutionPolicy&& __exec, _InputIterator __first, _Size __
     typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _ValueType2;
     typedef ::std::decay_t<_ExecutionPolicy> _DecayedExecutionPolicy;
 
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _InputIterator, _ForwardIterator>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first, __result);
 
     if constexpr (::std::is_trivial_v<_ValueType1> && ::std::is_trivial_v<_ValueType2>)
     {
@@ -97,8 +95,7 @@ uninitialized_move(_ExecutionPolicy&& __exec, _InputIterator __first, _InputIter
     typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _ValueType2;
     typedef ::std::decay_t<_ExecutionPolicy> _DecayedExecutionPolicy;
 
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _InputIterator, _ForwardIterator>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first, __result);
 
     if constexpr (::std::is_trivial_v<_ValueType1> && ::std::is_trivial_v<_ValueType2>)
     {
@@ -122,8 +119,7 @@ uninitialized_move_n(_ExecutionPolicy&& __exec, _InputIterator __first, _Size __
     typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _ValueType2;
     typedef ::std::decay_t<_ExecutionPolicy> _DecayedExecutionPolicy;
 
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _InputIterator, _ForwardIterator>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first, __result);
 
     if constexpr (::std::is_trivial_v<_ValueType1> && ::std::is_trivial_v<_ValueType2>)
     {
@@ -148,7 +144,7 @@ uninitialized_fill(_ExecutionPolicy&& __exec, _ForwardIterator __first, _Forward
     typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _ValueType;
     typedef ::std::decay_t<_ExecutionPolicy> _DecayedExecutionPolicy;
 
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first);
 
     if constexpr (::std::is_arithmetic_v<_ValueType>)
     {
@@ -171,7 +167,7 @@ uninitialized_fill_n(_ExecutionPolicy&& __exec, _ForwardIterator __first, _Size
     typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _ValueType;
     typedef ::std::decay_t<_ExecutionPolicy> _DecayedExecutionPolicy;
 
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first);
 
     if constexpr (::std::is_arithmetic_v<_ValueType>)
     {
@@ -225,8 +221,7 @@ destroy(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __
             _ExecutionPolicy;
 #endif // _PSTL_ICPX_OMP_SIMD_DESTROY_WINDOWS_BROKEN || _ONEDPL_ICPX_OMP_SIMD_DESTROY_WINDOWS_BROKEN
 
-        constexpr auto __dispatch_tag =
-            oneapi::dpl::__internal::__select_backend<_ExecutionPolicyDest, _ForwardIterator>();
+        const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first);
 
         oneapi::dpl::__internal::__pattern_walk1(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first,
                                                  __last, [](_ReferenceType __val) { __val.~_ValueType(); });
@@ -253,8 +248,7 @@ destroy_n(_ExecutionPolicy&& __exec, _ForwardIterator __first, _Size __n)
             _ExecutionPolicy;
 #endif // _PSTL_ICPX_OMP_SIMD_DESTROY_WINDOWS_BROKEN || _ONEDPL_ICPX_OMP_SIMD_DESTROY_WINDOWS_BROKEN
 
-        constexpr auto __dispatch_tag =
-            oneapi::dpl::__internal::__select_backend<_ExecutionPolicyDest, _ForwardIterator>();
+        const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first);
 
         return oneapi::dpl::__internal::__pattern_walk1_n(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
                                                           __first, __n,
@@ -273,7 +267,7 @@ uninitialized_default_construct(_ExecutionPolicy&& __exec, _ForwardIterator __fi
 
     if constexpr (!::std::is_trivial_v<_ValueType>)
     {
-        constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+        const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first);
 
         oneapi::dpl::__internal::__pattern_walk1(
             __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
@@ -294,7 +288,7 @@ uninitialized_default_construct_n(_ExecutionPolicy&& __exec, _ForwardIterator __
     }
     else
     {
-        constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+        const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first);
 
         return oneapi::dpl::__internal::__pattern_walk1_n(
             __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __n,
@@ -311,7 +305,7 @@ uninitialized_value_construct(_ExecutionPolicy&& __exec, _ForwardIterator __firs
     typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _ValueType;
     typedef ::std::decay_t<_ExecutionPolicy> _DecayedExecutionPolicy;
 
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first);
 
     if constexpr (::std::is_trivial_v<_ValueType>)
     {
@@ -334,7 +328,7 @@ uninitialized_value_construct_n(_ExecutionPolicy&& __exec, _ForwardIterator __fi
     typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _ValueType;
     typedef ::std::decay_t<_ExecutionPolicy> _DecayedExecutionPolicy;
 
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first);
 
     if constexpr (::std::is_trivial_v<_ValueType>)
     {

From c0fd0297587d92276d6fec372024cffbb1c77b67 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 27 Feb 2024 18:49:33 +0100
Subject: [PATCH 474/566] include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
 - rewrite calls of __select_backend() function with run-time params

---
 .../dpl/pstl/glue_algorithm_ranges_impl.h     | 60 +++++++++----------
 1 file changed, 29 insertions(+), 31 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
index 8efc60f95e5..b27140869ff 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
@@ -39,7 +39,7 @@ template <typename _ExecutionPolicy, typename _Range, typename _Predicate>
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, bool>
 any_of(_ExecutionPolicy&& __exec, _Range&& __rng, _Predicate __pred)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range>();
+    const auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend(__exec, __rng);
 
     return oneapi::dpl::__internal::__ranges::__pattern_any_of(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
                                                                views::all_read(::std::forward<_Range>(__rng)), __pred);
@@ -72,7 +72,7 @@ template <typename _ExecutionPolicy, typename _Range, typename _Function>
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy>
 for_each(_ExecutionPolicy&& __exec, _Range&& __rng, _Function __f)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range>();
+    const auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend(__exec, __rng);
 
     oneapi::dpl::__internal::__ranges::__pattern_walk_n(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __f,
                                                         views::all(::std::forward<_Range>(__rng)));
@@ -84,7 +84,7 @@ template <typename _ExecutionPolicy, typename _Range, typename _Predicate>
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, oneapi::dpl::__internal::__difference_t<_Range>>
 find_if(_ExecutionPolicy&& __exec, _Range&& __rng, _Predicate __pred)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range>();
+    const auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend(__exec, __rng);
 
     return oneapi::dpl::__internal::__ranges::__pattern_find_if(__dispatch_tag,
                                                                 ::std::forward<_ExecutionPolicy>(__exec),
@@ -118,7 +118,7 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy,
                                                       oneapi::dpl::__internal::__difference_t<_Range1>>
 find_end(_ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&& __rng2, _BinaryPredicate __pred)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range1, _Range2>();
+    const auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend(__exec, __rng1, __rng2);
 
     return oneapi::dpl::__internal::__ranges::__pattern_find_end(
         __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), views::all_read(::std::forward<_Range1>(__rng1)),
@@ -141,7 +141,7 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy,
                                                       oneapi::dpl::__internal::__difference_t<_Range1>>
 find_first_of(_ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&& __rng2, _BinaryPredicate __pred)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range1, _Range2>();
+    const auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend(__exec, __rng1, __rng2);
 
     return oneapi::dpl::__internal::__ranges::__pattern_find_first_of(
         __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), views::all_read(::std::forward<_Range1>(__rng1)),
@@ -163,7 +163,7 @@ template <typename _ExecutionPolicy, typename _Range, typename _BinaryPredicate>
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, oneapi::dpl::__internal::__difference_t<_Range>>
 adjacent_find(_ExecutionPolicy&& __exec, _Range&& __rng, _BinaryPredicate __pred)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range>();
+    const auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend(__exec, __rng);
 
     return oneapi::dpl::__internal::__ranges::__pattern_adjacent_find(
         __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), views::all_read(::std::forward<_Range>(__rng)),
@@ -185,7 +185,7 @@ template <typename _ExecutionPolicy, typename _Range, typename _Predicate>
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, oneapi::dpl::__internal::__difference_t<_Range>>
 count_if(_ExecutionPolicy&& __exec, _Range&& __rng, _Predicate __pred)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range>();
+    const auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend(__exec, __rng);
 
     return oneapi::dpl::__internal::__ranges::__pattern_count(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
                                                               views::all_read(::std::forward<_Range>(__rng)), __pred);
@@ -208,7 +208,7 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy,
                                                       oneapi::dpl::__internal::__difference_t<_Range1>>
 search(_ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&& __rng2, _BinaryPredicate __pred)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range1, _Range2>();
+    const auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend(__exec, __rng1, __rng2);
 
     return oneapi::dpl::__internal::__ranges::__pattern_search(
         __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), views::all_read(::std::forward<_Range1>(__rng1)),
@@ -228,7 +228,7 @@ template <typename _ExecutionPolicy, typename _Range, typename _Size, typename _
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, oneapi::dpl::__internal::__difference_t<_Range>>
 search_n(_ExecutionPolicy&& __exec, _Range&& __rng, _Size __count, const _Tp& __value, _BinaryPredicate __pred)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range>();
+    const auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend(__exec, __rng);
 
     return oneapi::dpl::__internal::__ranges::__pattern_search_n(
         __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), views::all_read(::std::forward<_Range>(__rng)),
@@ -249,7 +249,7 @@ template <typename _ExecutionPolicy, typename _Range1, typename _Range2>
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy>
 copy(_ExecutionPolicy&& __exec, _Range1&& __rng, _Range2&& __result)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range1, _Range2>();
+    const auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend(__exec, __rng, __result);
 
     oneapi::dpl::__internal::__ranges::__pattern_walk_n(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
                                                         oneapi::dpl::__internal::__brick_copy<_ExecutionPolicy>{},
@@ -262,7 +262,7 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy,
                                                       oneapi::dpl::__internal::__difference_t<_Range2>>
 copy_if(_ExecutionPolicy&& __exec, _Range1&& __rng, _Range2&& __result, _Predicate __pred)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range1, _Range2>();
+    const auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend(__exec, __rng, __result);
 
     return oneapi::dpl::__internal::__ranges::__pattern_copy_if(
         __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), views::all_read(::std::forward<_Range1>(__rng)),
@@ -276,7 +276,7 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy,
                                                       oneapi::dpl::__internal::__difference_t<_Range1>>
 swap_ranges(_ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&& __rng2)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range1, _Range2>();
+    const auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend(__exec, __rng1, __rng2);
 
     using _ReferenceType1 = oneapi::dpl::__internal::__value_t<_Range1>&;
     using _ReferenceType2 = oneapi::dpl::__internal::__value_t<_Range2>&;
@@ -295,7 +295,7 @@ template <typename _ExecutionPolicy, typename _Range1, typename _Range2, typenam
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy>
 transform(_ExecutionPolicy&& __exec, _Range1&& __rng, _Range2&& __result, _UnaryOperation __op)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range1, _Range2>();
+    const auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend(__exec, __rng, __result);
 
     oneapi::dpl::__internal::__ranges::__pattern_walk_n(
         __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), [__op](auto x, auto& z) { z = __op(x); },
@@ -306,8 +306,7 @@ template <typename _ExecutionPolicy, typename _Range1, typename _Range2, typenam
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy>
 transform(_ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&& __rng2, _Range3&& __result, _BinaryOperation __op)
 {
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range1, _Range2, _Range3>();
+    const auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend(__exec, __rng1, __rng2, __result);
 
     oneapi::dpl::__internal::__ranges::__pattern_walk_n(
         __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), [__op](auto x, auto y, auto& z) { z = __op(x, y); },
@@ -321,7 +320,7 @@ template <typename _ExecutionPolicy, typename _Range, typename _UnaryPredicate>
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, oneapi::dpl::__internal::__difference_t<_Range>>
 remove_if(_ExecutionPolicy&& __exec, _Range&& __rng, _UnaryPredicate __pred)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range>();
+    const auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend(__exec, __rng);
 
     return oneapi::dpl::__internal::__ranges::__pattern_remove_if(
         __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), views::all(::std::forward<_Range>(__rng)), __pred);
@@ -365,7 +364,7 @@ template <typename _ExecutionPolicy, typename _Range, typename _BinaryPredicate>
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, oneapi::dpl::__internal::__difference_t<_Range>>
 unique(_ExecutionPolicy&& __exec, _Range&& __rng, _BinaryPredicate __pred)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range>();
+    const auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend(__exec, __rng);
 
     return oneapi::dpl::__internal::__ranges::__pattern_unique(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
                                                                views::all(::std::forward<_Range>(__rng)), __pred);
@@ -384,7 +383,7 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy,
                                                       oneapi::dpl::__internal::__difference_t<_Range2>>
 unique_copy(_ExecutionPolicy&& __exec, _Range1&& __rng, _Range2&& __result, _BinaryPredicate __pred)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range1, _Range2>();
+    const auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend(__exec, __rng, __result);
 
     return oneapi::dpl::__internal::__ranges::__pattern_unique_copy(
         __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), views::all_read(::std::forward<_Range1>(__rng)),
@@ -443,7 +442,7 @@ template <typename _ExecutionPolicy, typename _Range, typename _UnaryPredicate,
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy>
 replace_if(_ExecutionPolicy&& __exec, _Range&& __rng, _UnaryPredicate __pred, const _Tp& __new_value)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range>();
+    const auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend(__exec, __rng);
 
     oneapi::dpl::__internal::__ranges::__pattern_walk_n(
         __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
@@ -470,7 +469,7 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy,
 replace_copy_if(_ExecutionPolicy&& __exec, _Range1&& __rng, _Range2&& __result, _UnaryPredicate __pred,
                 const _Tp& __new_value)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range1, _Range2>();
+    const auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend(__exec, __rng, __result);
 
     auto __src = views::all_read(::std::forward<_Range1>(__rng));
     oneapi::dpl::__internal::__ranges::__pattern_walk_n(
@@ -504,7 +503,7 @@ template <typename _ExecutionPolicy, typename _Range, typename _Compare, typenam
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy>
 sort(_ExecutionPolicy&& __exec, _Range&& __rng, _Compare __comp, _Proj __proj)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range>();
+    const auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend(__exec, __rng);
 
     oneapi::dpl::__internal::__ranges::__pattern_sort(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
                                                       views::all(::std::forward<_Range>(__rng)), __comp, __proj);
@@ -541,7 +540,7 @@ template <typename _ExecutionPolicy, typename _Range, typename _Compare>
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, oneapi::dpl::__internal::__difference_t<_Range>>
 is_sorted_until(_ExecutionPolicy&& __exec, _Range&& __rng, _Compare __comp)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range>();
+    const auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend(__exec, __rng);
 
     auto __view = views::all_read(::std::forward<_Range>(__rng));
     const auto __res = oneapi::dpl::__internal::__ranges::__pattern_adjacent_find(
@@ -563,7 +562,7 @@ template <typename _ExecutionPolicy, typename _Range, typename _Compare>
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, bool>
 is_sorted(_ExecutionPolicy&& __exec, _Range&& __rng, _Compare __comp)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range>();
+    const auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend(__exec, __rng);
 
     auto __view = views::all_read(::std::forward<_Range>(__rng));
     return oneapi::dpl::__internal::__ranges::__pattern_adjacent_find(
@@ -586,7 +585,7 @@ template <typename _ExecutionPolicy, typename _Range1, typename _Range2, typenam
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, bool>
 equal(_ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&& __rng2, _BinaryPredicate __p)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range1, _Range2>();
+    const auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend(__exec, __rng1, __rng2);
 
     return oneapi::dpl::__internal::__ranges::__pattern_equal(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
                                                               views::all_read(::std::forward<_Range1>(__rng1)),
@@ -607,7 +606,7 @@ template <typename _ExecutionPolicy, typename _Range1, typename _Range2>
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy>
 move(_ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&& __rng2)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range1, _Range2>();
+    const auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend(__exec, __rng1, __rng2);
 
     using _DecayedExecutionPolicy = ::std::decay_t<_ExecutionPolicy>;
 
@@ -624,8 +623,7 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy,
                                                       oneapi::dpl::__internal::__difference_t<_Range3>>
 merge(_ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&& __rng2, _Range3&& __rng3, _Compare __comp)
 {
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range1, _Range2, _Range3>();
+    const auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend(__exec, __rng1, __rng2, __rng3);
 
     return oneapi::dpl::__internal::__ranges::__pattern_merge(
         __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), views::all_read(::std::forward<_Range1>(__rng1)),
@@ -648,7 +646,7 @@ template <typename _ExecutionPolicy, typename _Range, typename _Compare>
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, oneapi::dpl::__internal::__difference_t<_Range>>
 min_element(_ExecutionPolicy&& __exec, _Range&& __rng, _Compare __comp)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range>();
+    const auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend(__exec, __rng);
 
     return oneapi::dpl::__internal::__ranges::__pattern_min_element(
         __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), views::all_read(::std::forward<_Range>(__rng)),
@@ -686,7 +684,7 @@ oneapi::dpl::__internal::__enable_if_execution_policy<
     ::std::pair<oneapi::dpl::__internal::__difference_t<_Range>, oneapi::dpl::__internal::__difference_t<_Range>>>
 minmax_element(_ExecutionPolicy&& __exec, _Range&& __rng, _Compare __comp)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range>();
+    const auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend(__exec, __rng);
 
     return oneapi::dpl::__internal::__ranges::__pattern_minmax_element(
         __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), views::all_read(::std::forward<_Range>(__rng)),
@@ -710,8 +708,8 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy,
 reduce_by_segment(_ExecutionPolicy&& __exec, _Range1&& __keys, _Range2&& __values, _Range3&& __out_keys,
                   _Range4&& __out_values, _BinaryPredicate __binary_pred, _BinaryOperator __binary_op)
 {
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__ranges::__select_backend<_ExecutionPolicy, _Range1, _Range2, _Range3>();
+    const auto __dispatch_tag =
+        oneapi::dpl::__ranges::__select_backend(__exec, __keys, __values, __out_keys, __out_values);
 
     return oneapi::dpl::__internal::__ranges::__pattern_reduce_by_segment(
         __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), views::all_read(::std::forward<_Range1>(__keys)),

From f0b5db69a34ca17749bfcb6eaa956a4613d5cc4c Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 27 Feb 2024 18:54:48 +0100
Subject: [PATCH 475/566] include/oneapi/dpl/pstl/glue_algorithm_impl.h -
 rewrite calls of __select_backend() function with run-time params

---
 include/oneapi/dpl/pstl/glue_algorithm_impl.h | 167 +++++++-----------
 1 file changed, 63 insertions(+), 104 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_algorithm_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
index abb6d6f6c7f..c74ad3ced5d 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
@@ -43,7 +43,7 @@ template <class _ExecutionPolicy, class _ForwardIterator, class _Predicate>
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, bool>
 any_of(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last, _Predicate __pred)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first);
 
     return oneapi::dpl::__internal::__pattern_any_of(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first,
                                                      __last, __pred);
@@ -75,7 +75,7 @@ template <class _ExecutionPolicy, class _ForwardIterator, class _Function>
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy>
 for_each(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last, _Function __f)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first);
 
     oneapi::dpl::__internal::__pattern_walk1(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
                                              __f);
@@ -85,7 +85,7 @@ template <class _ExecutionPolicy, class _ForwardIterator, class _Size, class _Fu
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _ForwardIterator>
 for_each_n(_ExecutionPolicy&& __exec, _ForwardIterator __first, _Size __n, _Function __f)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first);
 
     return oneapi::dpl::__internal::__pattern_walk1_n(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first,
                                                       __n, __f);
@@ -97,7 +97,7 @@ template <class _ExecutionPolicy, class _ForwardIterator, class _Predicate>
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _ForwardIterator>
 find_if(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last, _Predicate __pred)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first);
 
     return oneapi::dpl::__internal::__pattern_find_if(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first,
                                                       __last, __pred);
@@ -129,8 +129,7 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _Forward
 find_end(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _ForwardIterator1 __last, _ForwardIterator2 __s_first,
          _ForwardIterator2 __s_last, _BinaryPredicate __pred)
 {
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first, __s_first);
 
     return oneapi::dpl::__internal::__pattern_find_end(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
                                                        __first, __last, __s_first, __s_last, __pred);
@@ -151,8 +150,7 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _Forward
 find_first_of(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _ForwardIterator1 __last,
               _ForwardIterator2 __s_first, _ForwardIterator2 __s_last, _BinaryPredicate __pred)
 {
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first, __s_first);
 
     return oneapi::dpl::__internal::__pattern_find_first_of(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
                                                             __first, __last, __s_first, __s_last, __pred);
@@ -174,7 +172,7 @@ adjacent_find(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardItera
 {
     typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _ValueType;
 
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first);
 
     return oneapi::dpl::__internal::__pattern_adjacent_find(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
                                                             __first, __last, ::std::equal_to<_ValueType>(),
@@ -185,7 +183,7 @@ template <class _ExecutionPolicy, class _ForwardIterator, class _BinaryPredicate
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _ForwardIterator>
 adjacent_find(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last, _BinaryPredicate __pred)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first);
 
     return oneapi::dpl::__internal::__pattern_adjacent_find(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
                                                             __first, __last, __pred,
@@ -202,7 +200,7 @@ oneapi::dpl::__internal::__enable_if_execution_policy<
     _ExecutionPolicy, typename ::std::iterator_traits<_ForwardIterator>::difference_type>
 count(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last, const _Tp& __value)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first);
 
     return oneapi::dpl::__internal::__pattern_count(
         __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
@@ -215,7 +213,7 @@ oneapi::dpl::__internal::__enable_if_execution_policy<
     _ExecutionPolicy, typename ::std::iterator_traits<_ForwardIterator>::difference_type>
 count_if(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last, _Predicate __pred)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first);
 
     return oneapi::dpl::__internal::__pattern_count(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first,
                                                     __last, __pred);
@@ -228,8 +226,7 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _Forward
 search(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _ForwardIterator1 __last, _ForwardIterator2 __s_first,
        _ForwardIterator2 __s_last, _BinaryPredicate __pred)
 {
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first, __s_first);
 
     return oneapi::dpl::__internal::__pattern_search(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first,
                                                      __last, __s_first, __s_last, __pred);
@@ -249,7 +246,7 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _Forward
 search_n(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last, _Size __count,
          const _Tp& __value, _BinaryPredicate __pred)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first);
 
     return oneapi::dpl::__internal::__pattern_search_n(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
                                                        __first, __last, __count, __value, __pred);
@@ -270,8 +267,7 @@ template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterato
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _ForwardIterator2>
 copy(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _ForwardIterator1 __last, _ForwardIterator2 __result)
 {
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first, __result);
 
     return oneapi::dpl::__internal::__pattern_walk2_brick(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
                                                           __first, __last, __result,
@@ -284,8 +280,7 @@ copy_n(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _Size __n, _Forward
 {
     using _DecayedExecutionPolicy = ::std::decay_t<_ExecutionPolicy>;
 
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first, __result);
 
     return oneapi::dpl::__internal::__pattern_walk2_brick_n(
         __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __n, __result,
@@ -297,8 +292,7 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _Forward
 copy_if(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _ForwardIterator1 __last, _ForwardIterator2 __result,
         _Predicate __pred)
 {
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first, __result);
 
     return oneapi::dpl::__internal::__pattern_copy_if(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first,
                                                       __last, __result, __pred);
@@ -314,8 +308,7 @@ swap_ranges(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardItera
     typedef typename ::std::iterator_traits<_ForwardIterator1>::reference _ReferenceType1;
     typedef typename ::std::iterator_traits<_ForwardIterator2>::reference _ReferenceType2;
 
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first1, __first2);
 
     return oneapi::dpl::__internal::__pattern_swap(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first1,
                                                    __last1, __first2, [](_ReferenceType1 __x, _ReferenceType2 __y) {
@@ -331,8 +324,7 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _Forward
 transform(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _ForwardIterator1 __last, _ForwardIterator2 __result,
           _UnaryOperation __op)
 {
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first, __result);
 
     return oneapi::dpl::__internal::__pattern_walk2(
         __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result,
@@ -346,8 +338,7 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _Forward
 transform(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 __last1, _ForwardIterator2 __first2,
           _ForwardIterator __result, _BinaryOperation __op)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1,
-                                                                              _ForwardIterator2, _ForwardIterator>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first1, __first2, __result);
 
     return oneapi::dpl::__internal::__pattern_walk3(
         __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __result,
@@ -362,8 +353,7 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _Forward
 transform_if(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _ForwardIterator1 __last, _ForwardIterator2 __result,
              _UnaryOperation __op, _UnaryPredicate __pred)
 {
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first, __result);
 
     return oneapi::dpl::__internal::__pattern_walk2_transform_if(
         __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result,
@@ -377,8 +367,7 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _Forward
 transform_if(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
              _ForwardIterator2 __first2, _ForwardIterator3 __result, _BinaryOperation __op, _BinaryPredicate __pred)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1,
-                                                                              _ForwardIterator2, _ForwardIterator3>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first1, __first2, __result);
 
     return oneapi::dpl::__internal::__pattern_walk3_transform_if(
         __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __result,
@@ -393,7 +382,7 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy>
 replace_if(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last, _UnaryPredicate __pred,
            const _Tp& __new_value)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first);
 
     oneapi::dpl::__internal::__pattern_walk1(
         __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
@@ -419,8 +408,7 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _Forward
 replace_copy_if(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _ForwardIterator1 __last,
                 _ForwardIterator2 __result, _UnaryPredicate __pred, const _Tp& __new_value)
 {
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first, __result);
 
     return oneapi::dpl::__internal::__pattern_walk2(
         __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result,
@@ -450,7 +438,7 @@ template <class _ExecutionPolicy, class _ForwardIterator, class _Tp>
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy>
 fill(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last, const _Tp& __value)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first);
 
     oneapi::dpl::__internal::__pattern_fill(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
                                             __value);
@@ -463,7 +451,7 @@ fill_n(_ExecutionPolicy&& __exec, _ForwardIterator __first, _Size __count, const
     if (__count <= 0)
         return __first;
 
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first);
 
     return oneapi::dpl::__internal::__pattern_fill_n(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first,
                                                      __count, __value);
@@ -474,7 +462,7 @@ template <class _ExecutionPolicy, class _ForwardIterator, class _Generator>
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy>
 generate(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last, _Generator __g)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first);
 
     oneapi::dpl::__internal::__pattern_generate(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first,
                                                 __last, __g);
@@ -487,7 +475,7 @@ generate_n(_ExecutionPolicy&& __exec, _ForwardIterator __first, _Size __count, _
     if (__count <= 0)
         return __first;
 
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first);
 
     return oneapi::dpl::__internal::__pattern_generate_n(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
                                                          __first, __count, __g);
@@ -521,7 +509,7 @@ template <class _ExecutionPolicy, class _ForwardIterator, class _UnaryPredicate>
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _ForwardIterator>
 remove_if(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last, _UnaryPredicate __pred)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first);
 
     return oneapi::dpl::__internal::__pattern_remove_if(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
                                                         __first, __last, __pred);
@@ -543,7 +531,7 @@ template <class _ExecutionPolicy, class _ForwardIterator, class _BinaryPredicate
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _ForwardIterator>
 unique(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last, _BinaryPredicate __pred)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first);
 
     return oneapi::dpl::__internal::__pattern_unique(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first,
                                                      __last, __pred);
@@ -562,8 +550,7 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _Forward
 unique_copy(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _ForwardIterator1 __last, _ForwardIterator2 __result,
             _BinaryPredicate __pred)
 {
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first, __result);
 
     return oneapi::dpl::__internal::__pattern_unique_copy(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
                                                           __first, __last, __result, __pred);
@@ -583,8 +570,7 @@ template <class _ExecutionPolicy, class _BidirectionalIterator>
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy>
 reverse(_ExecutionPolicy&& __exec, _BidirectionalIterator __first, _BidirectionalIterator __last)
 {
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _BidirectionalIterator>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first);
 
     oneapi::dpl::__internal::__pattern_reverse(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first,
                                                __last);
@@ -595,8 +581,7 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _Forward
 reverse_copy(_ExecutionPolicy&& __exec, _BidirectionalIterator __first, _BidirectionalIterator __last,
              _ForwardIterator __d_first)
 {
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _BidirectionalIterator, _ForwardIterator>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first, __d_first);
 
     return oneapi::dpl::__internal::__pattern_reverse_copy(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
                                                            __first, __last, __d_first);
@@ -608,7 +593,7 @@ template <class _ExecutionPolicy, class _ForwardIterator>
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _ForwardIterator>
 rotate(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __middle, _ForwardIterator __last)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first);
 
     return oneapi::dpl::__internal::__pattern_rotate(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first,
                                                      __middle, __last);
@@ -619,8 +604,7 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _Forward
 rotate_copy(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _ForwardIterator1 __middle, _ForwardIterator1 __last,
             _ForwardIterator2 __result)
 {
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first, __result);
 
     return oneapi::dpl::__internal::__pattern_rotate_copy(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
                                                           __first, __middle, __last, __result);
@@ -632,7 +616,7 @@ template <class _ExecutionPolicy, class _ForwardIterator, class _UnaryPredicate>
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, bool>
 is_partitioned(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last, _UnaryPredicate __pred)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first);
 
     return oneapi::dpl::__internal::__pattern_is_partitioned(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
                                                              __first, __last, __pred);
@@ -642,7 +626,7 @@ template <class _ExecutionPolicy, class _ForwardIterator, class _UnaryPredicate>
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _ForwardIterator>
 partition(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last, _UnaryPredicate __pred)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first);
 
     return oneapi::dpl::__internal::__pattern_partition(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
                                                         __first, __last, __pred);
@@ -653,8 +637,7 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _Bidirec
 stable_partition(_ExecutionPolicy&& __exec, _BidirectionalIterator __first, _BidirectionalIterator __last,
                  _UnaryPredicate __pred)
 {
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _BidirectionalIterator>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first);
 
     return oneapi::dpl::__internal::__pattern_stable_partition(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
                                                                __first, __last, __pred);
@@ -667,8 +650,7 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy,
 partition_copy(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last,
                _ForwardIterator1 __out_true, _ForwardIterator2 __out_false, _UnaryPredicate __pred)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator,
-                                                                              _ForwardIterator1, _ForwardIterator2>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first, __out_true, __out_false);
 
     return oneapi::dpl::__internal::__pattern_partition_copy(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
                                                              __first, __last, __out_true, __out_false, __pred);
@@ -680,8 +662,7 @@ template <class _ExecutionPolicy, class _RandomAccessIterator, class _Compare>
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy>
 sort(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIterator __last, _Compare __comp)
 {
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first);
 
     typedef typename ::std::iterator_traits<_RandomAccessIterator>::value_type _InputType;
 
@@ -703,8 +684,7 @@ template <class _ExecutionPolicy, class _RandomAccessIterator, class _Compare>
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy>
 stable_sort(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIterator __last, _Compare __comp)
 {
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first);
 
     oneapi::dpl::__internal::__pattern_stable_sort(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first,
                                                    __last, __comp);
@@ -726,8 +706,7 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy>
 sort_by_key(_ExecutionPolicy&& __exec, _RandomAccessIterator1 __keys_first, _RandomAccessIterator1 __keys_last,
             _RandomAccessIterator2 __values_first, _Compare __comp)
 {
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator1, _RandomAccessIterator2>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __keys_first, __values_first);
 
     oneapi::dpl::__internal::__pattern_sort_by_key(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
                                                    __keys_first, __keys_last, __values_first, __comp);
@@ -750,8 +729,7 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy,
 mismatch(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 __last1, _ForwardIterator2 __first2,
          _ForwardIterator2 __last2, _BinaryPredicate __pred)
 {
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first1, __first2);
 
     return oneapi::dpl::__internal::__pattern_mismatch(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
                                                        __first1, __last1, __first2, __last2, __pred);
@@ -795,9 +773,7 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, bool>
 equal(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 __last1, _ForwardIterator2 __first2,
       _BinaryPredicate __p)
 {
-    // TODO is it correct that we check _ForwardIterator2 in __select_backend ?
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first1, __first2);
 
     return oneapi::dpl::__internal::__pattern_equal(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first1,
                                                     __last1, __first2, __p);
@@ -816,9 +792,7 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, bool>
 equal(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 __last1, _ForwardIterator2 __first2,
       _ForwardIterator2 __last2, _BinaryPredicate __p)
 {
-    // TODO is it correct that we check _ForwardIterator2 in __select_backend ?
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first1, __first2);
 
     return oneapi::dpl::__internal::__pattern_equal(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first1,
                                                     __last1, __first2, __last2, __p);
@@ -840,8 +814,7 @@ move(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _ForwardIterator1 __l
 {
     using _DecayedExecutionPolicy = ::std::decay_t<_ExecutionPolicy>;
 
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first, __d_first);
 
     return oneapi::dpl::__internal::__pattern_walk2_brick(
         __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __d_first,
@@ -855,8 +828,7 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy>
 partial_sort(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIterator __middle,
              _RandomAccessIterator __last, _Compare __comp)
 {
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first);
 
     oneapi::dpl::__internal::__pattern_partial_sort(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first,
                                                     __middle, __last, __comp);
@@ -878,8 +850,7 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _RandomA
 partial_sort_copy(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last,
                   _RandomAccessIterator __d_first, _RandomAccessIterator __d_last, _Compare __comp)
 {
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator, _RandomAccessIterator>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first, __d_first);
 
     return oneapi::dpl::__internal::__pattern_partial_sort_copy(
         __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __d_first, __d_last, __comp);
@@ -899,7 +870,7 @@ template <class _ExecutionPolicy, class _ForwardIterator, class _Compare>
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _ForwardIterator>
 is_sorted_until(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last, _Compare __comp)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first);
 
     const _ForwardIterator __res = oneapi::dpl::__internal::__pattern_adjacent_find(
         __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
@@ -919,7 +890,7 @@ template <class _ExecutionPolicy, class _ForwardIterator, class _Compare>
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, bool>
 is_sorted(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last, _Compare __comp)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first);
 
     return oneapi::dpl::__internal::__pattern_adjacent_find(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
                                                             __first, __last,
@@ -942,8 +913,7 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _Forward
 merge(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 __last1, _ForwardIterator2 __first2,
       _ForwardIterator2 __last2, _ForwardIterator __d_first, _Compare __comp)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1,
-                                                                              _ForwardIterator2, _ForwardIterator>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first1, __first2, __d_first);
 
     return oneapi::dpl::__internal::__pattern_merge(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first1,
                                                     __last1, __first2, __last2, __d_first, __comp);
@@ -963,8 +933,7 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy>
 inplace_merge(_ExecutionPolicy&& __exec, _BidirectionalIterator __first, _BidirectionalIterator __middle,
               _BidirectionalIterator __last, _Compare __comp)
 {
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _BidirectionalIterator>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first);
 
     oneapi::dpl::__internal::__pattern_inplace_merge(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first,
                                                      __middle, __last, __comp);
@@ -986,8 +955,7 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, bool>
 includes(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 __last1, _ForwardIterator2 __first2,
          _ForwardIterator2 __last2, _Compare __comp)
 {
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first1, __first2);
 
     return oneapi::dpl::__internal::__pattern_includes(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
                                                        __first1, __last1, __first2, __last2, __comp);
@@ -1010,8 +978,7 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _Forward
 set_union(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 __last1, _ForwardIterator2 __first2,
           _ForwardIterator2 __last2, _ForwardIterator __result, _Compare __comp)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1,
-                                                                              _ForwardIterator2, _ForwardIterator>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first1, __first2, __result);
 
     return oneapi::dpl::__internal::__pattern_set_union(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
                                                         __first1, __last1, __first2, __last2, __result, __comp);
@@ -1034,8 +1001,7 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _Forward
 set_intersection(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
                  _ForwardIterator2 __first2, _ForwardIterator2 __last2, _ForwardIterator __result, _Compare __comp)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1,
-                                                                              _ForwardIterator2, _ForwardIterator>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first1, __first2, __result);
 
     return oneapi::dpl::__internal::__pattern_set_intersection(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
                                                                __first1, __last1, __first2, __last2, __result, __comp);
@@ -1058,8 +1024,7 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _Forward
 set_difference(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
                _ForwardIterator2 __first2, _ForwardIterator2 __last2, _ForwardIterator __result, _Compare __comp)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1,
-                                                                              _ForwardIterator2, _ForwardIterator>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first1, __first2, __result);
 
     return oneapi::dpl::__internal::__pattern_set_difference(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
                                                              __first1, __last1, __first2, __last2, __result, __comp);
@@ -1083,8 +1048,7 @@ set_symmetric_difference(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1,
                          _ForwardIterator2 __first2, _ForwardIterator2 __last2, _ForwardIterator __result,
                          _Compare __comp)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1,
-                                                                              _ForwardIterator2, _ForwardIterator>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first1, __first2, __result);
 
     return oneapi::dpl::__internal::__pattern_set_symmetric_difference(
         __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2, __result,
@@ -1105,8 +1069,7 @@ template <class _ExecutionPolicy, class _RandomAccessIterator, class _Compare>
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _RandomAccessIterator>
 is_heap_until(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIterator __last, _Compare __comp)
 {
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first);
 
     return oneapi::dpl::__internal::__pattern_is_heap_until(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
                                                             __first, __last, __comp);
@@ -1124,8 +1087,7 @@ template <class _ExecutionPolicy, class _RandomAccessIterator, class _Compare>
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, bool>
 is_heap(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIterator __last, _Compare __comp)
 {
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first);
 
     return oneapi::dpl::__internal::__pattern_is_heap(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first,
                                                       __last, __comp);
@@ -1145,7 +1107,7 @@ template <class _ExecutionPolicy, class _ForwardIterator, class _Compare>
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _ForwardIterator>
 min_element(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last, _Compare __comp)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first);
 
     return oneapi::dpl::__internal::__pattern_min_element(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
                                                           __first, __last, __comp);
@@ -1180,7 +1142,7 @@ template <class _ExecutionPolicy, class _ForwardIterator, class _Compare>
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, ::std::pair<_ForwardIterator, _ForwardIterator>>
 minmax_element(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last, _Compare __comp)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first);
 
     return oneapi::dpl::__internal::__pattern_minmax_element(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
                                                              __first, __last, __comp);
@@ -1201,8 +1163,7 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy>
 nth_element(_ExecutionPolicy&& __exec, _RandomAccessIterator __first, _RandomAccessIterator __nth,
             _RandomAccessIterator __last, _Compare __comp)
 {
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _RandomAccessIterator>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first);
 
     oneapi::dpl::__internal::__pattern_nth_element(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first,
                                                    __nth, __last, __comp);
@@ -1224,8 +1185,7 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, bool>
 lexicographical_compare(_ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
                         _ForwardIterator2 __first2, _ForwardIterator2 __last2, _Compare __comp)
 {
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator1, _ForwardIterator2>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first1, __first2);
 
     return oneapi::dpl::__internal::__pattern_lexicographical_compare(
         __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __last2, __comp);
@@ -1247,7 +1207,7 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _Forward
 shift_left(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last,
            typename ::std::iterator_traits<_ForwardIterator>::difference_type __n)
 {
-    constexpr auto __dispatch_tag = oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _ForwardIterator>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first);
 
     return oneapi::dpl::__internal::__pattern_shift_left(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
                                                          __first, __last, __n);
@@ -1260,8 +1220,7 @@ oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _Bidirec
 shift_right(_ExecutionPolicy&& __exec, _BidirectionalIterator __first, _BidirectionalIterator __last,
             typename ::std::iterator_traits<_BidirectionalIterator>::difference_type __n)
 {
-    constexpr auto __dispatch_tag =
-        oneapi::dpl::__internal::__select_backend<_ExecutionPolicy, _BidirectionalIterator>();
+    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first);
 
     return oneapi::dpl::__internal::__pattern_shift_right(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
                                                           __first, __last, __n);

From 3c8ab1c458f01982deb2bb7cd7967ad356f09f8c Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 28 Feb 2024 14:55:02 +0100
Subject: [PATCH 476/566] include/oneapi/dpl/pstl/glue_memory_impl.h - fix
 compile error introduced in previous commit "Merge branch
 'dev/skopienko/tag_dispatching_make_select_backend_non_constexpr_again' into
 dev/skopienko/tag_dispatching"

---
 include/oneapi/dpl/pstl/glue_memory_impl.h | 49 +++++++++-------------
 1 file changed, 20 insertions(+), 29 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_memory_impl.h b/include/oneapi/dpl/pstl/glue_memory_impl.h
index 60f02af73d0..6f12099a224 100644
--- a/include/oneapi/dpl/pstl/glue_memory_impl.h
+++ b/include/oneapi/dpl/pstl/glue_memory_impl.h
@@ -184,24 +184,29 @@ uninitialized_fill_n(_ExecutionPolicy&& __exec, _ForwardIterator __first, _Size
 }
 
 #if (_PSTL_ICPX_OMP_SIMD_DESTROY_WINDOWS_BROKEN || _ONEDPL_ICPX_OMP_SIMD_DESTROY_WINDOWS_BROKEN)
-template <class _ExecutionPolicy>
-struct EvaDestroylWorkaroundPolicy
+
+// oneapi::dpl::execution::parallel_unsequenced_policy -> oneapi::dpl::execution::parallel_policy
+auto
+get_destroy_dest_policy(oneapi::dpl::execution::parallel_unsequenced_policy)
 {
-    using ExecutionPolicy = _ExecutionPolicy;
-};
+    return oneapi::dpl::execution::parallel_policy{};
+}
 
-template <>
-struct EvaDestroylWorkaroundPolicy<oneapi::dpl::execution::parallel_unsequenced_policy>
+// oneapi::dpl::execution::unsequenced_policy -> oneapi::dpl::execution::sequenced_policy
+auto get_destroy_dest_policy(oneapi::dpl::execution::unsequenced_policy)
 {
-    using ExecutionPolicy = oneapi::dpl::execution::parallel_policy;
-};
+    return oneapi::dpl::execution::sequenced_policy{};
+}
 
-template <>
-struct EvaDestroylWorkaroundPolicy<oneapi::dpl::execution::unsequenced_policy>
+#endif // (_PSTL_ICPX_OMP_SIMD_DESTROY_WINDOWS_BROKEN || _ONEDPL_ICPX_OMP_SIMD_DESTROY_WINDOWS_BROKEN)
+
+// 1:1 for all other policies
+template <typename _ExecutionPolicy>
+auto
+get_destroy_dest_policy(_ExecutionPolicy&& __exec)
 {
-    using ExecutionPolicy = oneapi::dpl::execution::sequenced_policy;
-};
-#endif // _PSTL_ICPX_OMP_SIMD_DESTROY_WINDOWS_BROKEN || _ONEDPL_ICPX_OMP_SIMD_DESTROY_WINDOWS_BROKEN
+    return __exec;
+}
 
 // [specialized.destroy]
 
@@ -214,14 +219,7 @@ destroy(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __
 
     if constexpr (!::std::is_trivially_destructible_v<_ValueType>)
     {
-        using _ExecutionPolicyDest =
-#if (_PSTL_ICPX_OMP_SIMD_DESTROY_WINDOWS_BROKEN || _ONEDPL_ICPX_OMP_SIMD_DESTROY_WINDOWS_BROKEN)
-            typename EvaDestroylWorkaroundPolicy<::std::decay_t<_ExecutionPolicy>>::ExecutionPolicy;
-#else
-            _ExecutionPolicy;
-#endif // _PSTL_ICPX_OMP_SIMD_DESTROY_WINDOWS_BROKEN || _ONEDPL_ICPX_OMP_SIMD_DESTROY_WINDOWS_BROKEN
-
-        const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first);
+        const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(get_destroy_dest_policy(__exec), __first);
 
         oneapi::dpl::__internal::__pattern_walk1(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first,
                                                  __last, [](_ReferenceType __val) { __val.~_ValueType(); });
@@ -241,14 +239,7 @@ destroy_n(_ExecutionPolicy&& __exec, _ForwardIterator __first, _Size __n)
     }
     else
     {
-        using _ExecutionPolicyDest =
-#if (_PSTL_ICPX_OMP_SIMD_DESTROY_WINDOWS_BROKEN || _ONEDPL_ICPX_OMP_SIMD_DESTROY_WINDOWS_BROKEN)
-            typename EvaDestroylWorkaroundPolicy<::std::decay_t<_ExecutionPolicy>>::ExecutionPolicy;
-#else
-            _ExecutionPolicy;
-#endif // _PSTL_ICPX_OMP_SIMD_DESTROY_WINDOWS_BROKEN || _ONEDPL_ICPX_OMP_SIMD_DESTROY_WINDOWS_BROKEN
-
-        const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first);
+        const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(get_destroy_dest_policy(__exec), __first);
 
         return oneapi::dpl::__internal::__pattern_walk1_n(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
                                                           __first, __n,

From 02b61f1391fc15b6c4dd006208fc83cab5c2c840 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 28 Feb 2024 16:12:03 +0100
Subject: [PATCH 477/566] Apply GitHUB clang format

---
 include/oneapi/dpl/pstl/glue_memory_impl.h | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/include/oneapi/dpl/pstl/glue_memory_impl.h b/include/oneapi/dpl/pstl/glue_memory_impl.h
index 6f12099a224..adf7f65ca18 100644
--- a/include/oneapi/dpl/pstl/glue_memory_impl.h
+++ b/include/oneapi/dpl/pstl/glue_memory_impl.h
@@ -193,7 +193,8 @@ get_destroy_dest_policy(oneapi::dpl::execution::parallel_unsequenced_policy)
 }
 
 // oneapi::dpl::execution::unsequenced_policy -> oneapi::dpl::execution::sequenced_policy
-auto get_destroy_dest_policy(oneapi::dpl::execution::unsequenced_policy)
+auto
+get_destroy_dest_policy(oneapi::dpl::execution::unsequenced_policy)
 {
     return oneapi::dpl::execution::sequenced_policy{};
 }

From 366163bd513aeafaccdbf733f6fe885b9e707e6a Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 29 Feb 2024 09:25:57 +0100
Subject: [PATCH 478/566] include/oneapi/dpl/internal/scan_by_segment_impl.h -
 fix review comment: Why __BackendTag is not a part of operator()? Looks
 inconsistent with everything else.

---
 include/oneapi/dpl/internal/scan_by_segment_impl.h | 12 ++++++------
 1 file changed, 6 insertions(+), 6 deletions(-)

diff --git a/include/oneapi/dpl/internal/scan_by_segment_impl.h b/include/oneapi/dpl/internal/scan_by_segment_impl.h
index f897c1ec899..85903a93fdf 100644
--- a/include/oneapi/dpl/internal/scan_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/scan_by_segment_impl.h
@@ -94,7 +94,7 @@ class __seg_scan_wg_kernel;
 template <bool __is_inclusive, typename... Name>
 class __seg_scan_prefix_kernel;
 
-template <typename _BackendTag, bool __is_inclusive>
+template <bool __is_inclusive>
 struct __sycl_scan_by_segment_impl
 {
     template <typename... _Name>
@@ -103,10 +103,10 @@ struct __sycl_scan_by_segment_impl
     template <typename... _Name>
     using _SegScanPrefixPhase = __seg_scan_prefix_kernel<__is_inclusive, _Name...>;
 
-    template <typename _ExecutionPolicy, typename _Range1, typename _Range2, typename _Range3,
+    template <typename _BackendTag, typename _ExecutionPolicy, typename _Range1, typename _Range2, typename _Range3,
               typename _BinaryPredicate, typename _BinaryOperator, typename _T>
     void
-    operator()(_ExecutionPolicy&& __exec, _Range1&& __keys, _Range2&& __values, _Range3&& __out_values,
+    operator()(_BackendTag, _ExecutionPolicy&& __exec, _Range1&& __keys, _Range2&& __values, _Range3&& __out_values,
                _BinaryPredicate __binary_pred, _BinaryOperator __binary_op, _T __init, _T __identity)
     {
         using _CustomName = oneapi::dpl::__internal::__policy_kernel_name<_ExecutionPolicy>;
@@ -392,9 +392,9 @@ __scan_by_segment_impl_common(__internal::__hetero_tag<_BackendTag>, Policy&& po
 
     constexpr iter_value_t identity = unseq_backend::__known_identity<Operator, iter_value_t>;
 
-    __sycl_scan_by_segment_impl<_BackendTag, Inclusive::value>()(::std::forward<Policy>(policy), key_buf.all_view(),
-                                                                 value_buf.all_view(), value_output_buf.all_view(),
-                                                                 binary_pred, binary_op, init, identity);
+    __sycl_scan_by_segment_impl<Inclusive::value>()(_BackendTag{}, ::std::forward<Policy>(policy), key_buf.all_view(),
+                                                    value_buf.all_view(), value_output_buf.all_view(), binary_pred,
+                                                    binary_op, init, identity);
     return result + n;
 }
 

From eb95d71c89e044ab7ca55f0ed94f213e3d89b554 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 29 Feb 2024 09:31:42 +0100
Subject: [PATCH 479/566] include/oneapi/dpl/pstl/algorithm_impl.h - fix review
 comment: ::std::move(::std::forward<_ReferenceType1>(__val)) looks really
 weird.

---
 include/oneapi/dpl/pstl/algorithm_impl.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 4c95e590eff..859979bb370 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -1131,7 +1131,7 @@ struct __brick_move<_ExecutionPolicy, oneapi::dpl::__internal::__enable_if_host_
     void
     operator()(_ReferenceType1&& __val, _ReferenceType2&& __result) const
     {
-        ::std::forward<_ReferenceType2>(__result) = ::std::move(::std::forward<_ReferenceType1>(__val));
+        __result = ::std::move(__val);
     }
 };
 

From 730d1ba43bfced67851ecbc81ca25db5d841e1ca Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 29 Feb 2024 09:44:11 +0100
Subject: [PATCH 480/566] include/oneapi/dpl/pstl/algorithm_impl.h - fix review
 comment: where is the code from the left side? Was it moved to the brick?

---
 include/oneapi/dpl/pstl/algorithm_impl.h | 46 +++++++++++++++++++++---
 1 file changed, 42 insertions(+), 4 deletions(-)

diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 859979bb370..8b1d26989a9 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -4106,13 +4106,51 @@ __pattern_lexicographical_compare(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __
                                                        typename _Tag::__is_vector{});
 }
 
-template <class _IsVector, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Compare>
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
+          class _Compare>
 bool
-__pattern_lexicographical_compare(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _ForwardIterator1 __first1,
-                                  _ForwardIterator1 __last1, _ForwardIterator2 __first2, _ForwardIterator2 __last2,
+__pattern_lexicographical_compare(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec,
+                                  _RandomAccessIterator1 __first1, _RandomAccessIterator1 __last1,
+                                  _RandomAccessIterator2 __first2, _RandomAccessIterator2 __last2,
                                   _Compare __comp) noexcept
 {
-    return __internal::__brick_lexicographical_compare(__first1, __last1, __first2, __last2, __comp, _IsVector{});
+    if (__first2 == __last2)
+    { // if second sequence is empty
+        return false;
+    }
+    else if (__first1 == __last1)
+    { // if first sequence is empty
+        return true;
+    }
+    else
+    {
+        typedef typename ::std::iterator_traits<_RandomAccessIterator1>::reference _RefType1;
+        typedef typename ::std::iterator_traits<_RandomAccessIterator2>::reference _RefType2;
+        --__last1;
+        --__last2;
+        auto __n = ::std::min(__last1 - __first1, __last2 - __first2);
+        auto __result = __internal::__parallel_find(
+            __tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __first1 + __n,
+            [__first1, __first2, &__comp](_RandomAccessIterator1 __i, _RandomAccessIterator1 __j) {
+                return __internal::__brick_mismatch(
+                           __i, __j, __first2 + (__i - __first1), __first2 + (__j - __first1),
+                           [&__comp](const _RefType1 __x, const _RefType2 __y) {
+                               return !__comp(__x, __y) && !__comp(__y, __x);
+                           },
+                           _IsVector{})
+                    .first;
+            },
+            ::std::true_type{});
+
+        if (__result == __last1 && __first2 + (__result - __first1) != __last2)
+        { // if first sequence shorter than second
+            return !__comp(*(__first2 + (__result - __first1)), *__result);
+        }
+        else
+        { // if second sequence shorter than first or both have the same number of elements
+            return __comp(*__result, *(__first2 + (__result - __first1)));
+        }
+    }
 }
 
 //------------------------------------------------------------------------

From dfc6f89a68cdf176fcd6296879c22f67af69c8c8 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 29 Feb 2024 10:51:31 +0100
Subject: [PATCH 481/566] include/oneapi/dpl/pstl/glue_memory_impl.h - fix
 review comment: In case we would not have a broken configuration we will not
 have get_destroy_dest_policy at all.

---
 include/oneapi/dpl/pstl/glue_memory_impl.h | 32 +++++++++++++++-------
 1 file changed, 22 insertions(+), 10 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_memory_impl.h b/include/oneapi/dpl/pstl/glue_memory_impl.h
index adf7f65ca18..ba109b9da70 100644
--- a/include/oneapi/dpl/pstl/glue_memory_impl.h
+++ b/include/oneapi/dpl/pstl/glue_memory_impl.h
@@ -186,29 +186,29 @@ uninitialized_fill_n(_ExecutionPolicy&& __exec, _ForwardIterator __first, _Size
 #if (_PSTL_ICPX_OMP_SIMD_DESTROY_WINDOWS_BROKEN || _ONEDPL_ICPX_OMP_SIMD_DESTROY_WINDOWS_BROKEN)
 
 // oneapi::dpl::execution::parallel_unsequenced_policy -> oneapi::dpl::execution::parallel_policy
-auto
+oneapi::dpl::execution::parallel_policy
 get_destroy_dest_policy(oneapi::dpl::execution::parallel_unsequenced_policy)
 {
-    return oneapi::dpl::execution::parallel_policy{};
+    return {};
 }
 
 // oneapi::dpl::execution::unsequenced_policy -> oneapi::dpl::execution::sequenced_policy
-auto
+oneapi::dpl::execution::sequenced_policy
 get_destroy_dest_policy(oneapi::dpl::execution::unsequenced_policy)
 {
-    return oneapi::dpl::execution::sequenced_policy{};
+    return {};
 }
 
-#endif // (_PSTL_ICPX_OMP_SIMD_DESTROY_WINDOWS_BROKEN || _ONEDPL_ICPX_OMP_SIMD_DESTROY_WINDOWS_BROKEN)
-
 // 1:1 for all other policies
 template <typename _ExecutionPolicy>
-auto
+_ExecutionPolicy&&
 get_destroy_dest_policy(_ExecutionPolicy&& __exec)
 {
-    return __exec;
+    return ::std::forward<_ExecutionPolicy>(__exec);
 }
 
+#endif // (_PSTL_ICPX_OMP_SIMD_DESTROY_WINDOWS_BROKEN || _ONEDPL_ICPX_OMP_SIMD_DESTROY_WINDOWS_BROKEN)
+
 // [specialized.destroy]
 
 template <class _ExecutionPolicy, class _ForwardIterator>
@@ -220,7 +220,13 @@ destroy(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __
 
     if constexpr (!::std::is_trivially_destructible_v<_ValueType>)
     {
-        const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(get_destroy_dest_policy(__exec), __first);
+        const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(
+#if (_PSTL_ICPX_OMP_SIMD_DESTROY_WINDOWS_BROKEN || _ONEDPL_ICPX_OMP_SIMD_DESTROY_WINDOWS_BROKEN)
+            get_destroy_dest_policy(__exec),
+#else
+            __exec,
+#endif
+            __first);
 
         oneapi::dpl::__internal::__pattern_walk1(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first,
                                                  __last, [](_ReferenceType __val) { __val.~_ValueType(); });
@@ -240,7 +246,13 @@ destroy_n(_ExecutionPolicy&& __exec, _ForwardIterator __first, _Size __n)
     }
     else
     {
-        const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(get_destroy_dest_policy(__exec), __first);
+        const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(
+#if (_PSTL_ICPX_OMP_SIMD_DESTROY_WINDOWS_BROKEN || _ONEDPL_ICPX_OMP_SIMD_DESTROY_WINDOWS_BROKEN)
+            get_destroy_dest_policy(__exec),
+#else
+            __exec,
+#endif
+            __first);
 
         return oneapi::dpl::__internal::__pattern_walk1_n(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
                                                           __first, __n,

From d5f94bdd65e7830f25ead740b2007290088ab9af Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 29 Feb 2024 10:53:59 +0100
Subject: [PATCH 482/566] include/oneapi/dpl/pstl/glue_memory_impl.h - fix
 review comment: rename get_destroy_dest_policy to get_unvectorized_policy

---
 include/oneapi/dpl/pstl/glue_memory_impl.h | 23 +++++++++-------------
 1 file changed, 9 insertions(+), 14 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_memory_impl.h b/include/oneapi/dpl/pstl/glue_memory_impl.h
index ba109b9da70..9f68203278c 100644
--- a/include/oneapi/dpl/pstl/glue_memory_impl.h
+++ b/include/oneapi/dpl/pstl/glue_memory_impl.h
@@ -185,24 +185,21 @@ uninitialized_fill_n(_ExecutionPolicy&& __exec, _ForwardIterator __first, _Size
 
 #if (_PSTL_ICPX_OMP_SIMD_DESTROY_WINDOWS_BROKEN || _ONEDPL_ICPX_OMP_SIMD_DESTROY_WINDOWS_BROKEN)
 
-// oneapi::dpl::execution::parallel_unsequenced_policy -> oneapi::dpl::execution::parallel_policy
 oneapi::dpl::execution::parallel_policy
-get_destroy_dest_policy(oneapi::dpl::execution::parallel_unsequenced_policy)
+get_unvectorized_policy(oneapi::dpl::execution::parallel_unsequenced_policy)
 {
     return {};
 }
 
-// oneapi::dpl::execution::unsequenced_policy -> oneapi::dpl::execution::sequenced_policy
 oneapi::dpl::execution::sequenced_policy
-get_destroy_dest_policy(oneapi::dpl::execution::unsequenced_policy)
+get_unvectorized_policy(oneapi::dpl::execution::unsequenced_policy)
 {
     return {};
 }
 
-// 1:1 for all other policies
 template <typename _ExecutionPolicy>
 _ExecutionPolicy&&
-get_destroy_dest_policy(_ExecutionPolicy&& __exec)
+get_unvectorized_policy(_ExecutionPolicy&& __exec)
 {
     return ::std::forward<_ExecutionPolicy>(__exec);
 }
@@ -220,13 +217,12 @@ destroy(_ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __
 
     if constexpr (!::std::is_trivially_destructible_v<_ValueType>)
     {
-        const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(
+        const auto __dispatch_tag =
 #if (_PSTL_ICPX_OMP_SIMD_DESTROY_WINDOWS_BROKEN || _ONEDPL_ICPX_OMP_SIMD_DESTROY_WINDOWS_BROKEN)
-            get_destroy_dest_policy(__exec),
+            oneapi::dpl::__internal::__select_backend(get_unvectorized_policy(__exec), __first);
 #else
-            __exec,
+            oneapi::dpl::__internal::__select_backend(__exec, __first);
 #endif
-            __first);
 
         oneapi::dpl::__internal::__pattern_walk1(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first,
                                                  __last, [](_ReferenceType __val) { __val.~_ValueType(); });
@@ -246,13 +242,12 @@ destroy_n(_ExecutionPolicy&& __exec, _ForwardIterator __first, _Size __n)
     }
     else
     {
-        const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(
+        const auto __dispatch_tag =
 #if (_PSTL_ICPX_OMP_SIMD_DESTROY_WINDOWS_BROKEN || _ONEDPL_ICPX_OMP_SIMD_DESTROY_WINDOWS_BROKEN)
-            get_destroy_dest_policy(__exec),
+            oneapi::dpl::__internal::__select_backend(get_unvectorized_policy(__exec), __first);
 #else
-            __exec,
+            oneapi::dpl::__internal::__select_backend(__exec, __first);
 #endif
-            __first);
 
         return oneapi::dpl::__internal::__pattern_walk1_n(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
                                                           __first, __n,

From 19331f37576bf2e8506a38be4c45a658a699c492 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 29 Feb 2024 17:05:20 +0100
Subject: [PATCH 483/566] struct __brick_copy_n with tags - impls + calls

---
 include/oneapi/dpl/pstl/algorithm_fwd.h                | 2 +-
 include/oneapi/dpl/pstl/algorithm_impl.h               | 6 ++++--
 include/oneapi/dpl/pstl/glue_algorithm_impl.h          | 2 +-
 include/oneapi/dpl/pstl/glue_memory_impl.h             | 6 ++++--
 include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h | 6 +++---
 5 files changed, 13 insertions(+), 9 deletions(-)

diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index 547dacb13f7..434d92cb9b3 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -406,7 +406,7 @@ __pattern_search_n(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessI
 // copy_n
 //------------------------------------------------------------------------
 
-template <typename _ExecutionPolicy, typename = void>
+template <class _Tag, typename _ExecutionPolicy>
 struct __brick_copy_n;
 
 template <typename _ExecutionPolicy, typename = void>
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 8b1d26989a9..2514ea05545 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -1048,9 +1048,11 @@ __pattern_search_n(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _
 // clear that doing so is worth the trouble and extra layers of call chain.
 // Sometimes a little duplication for sake of regularity is better than the alternative.
 
-template <typename _ExecutionPolicy>
-struct __brick_copy_n<_ExecutionPolicy, oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>>
+template <class _Tag, typename _ExecutionPolicy>
+struct __brick_copy_n
 {
+    static_assert(__is_host_backend_tag_v<_Tag>);
+
     template <typename _RandomAccessIterator1, typename _Size, typename _RandomAccessIterator2>
     _RandomAccessIterator2
     operator()(_RandomAccessIterator1 __first, _Size __n, _RandomAccessIterator2 __result,
diff --git a/include/oneapi/dpl/pstl/glue_algorithm_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
index c74ad3ced5d..91b618c11a7 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
@@ -284,7 +284,7 @@ copy_n(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _Size __n, _Forward
 
     return oneapi::dpl::__internal::__pattern_walk2_brick_n(
         __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __n, __result,
-        oneapi::dpl::__internal::__brick_copy_n<_DecayedExecutionPolicy>{});
+        oneapi::dpl::__internal::__brick_copy_n<::std::decay_t<decltype(__dispatch_tag)>, _DecayedExecutionPolicy>{});
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Predicate>
diff --git a/include/oneapi/dpl/pstl/glue_memory_impl.h b/include/oneapi/dpl/pstl/glue_memory_impl.h
index 9f68203278c..4ba9fe1be47 100644
--- a/include/oneapi/dpl/pstl/glue_memory_impl.h
+++ b/include/oneapi/dpl/pstl/glue_memory_impl.h
@@ -75,7 +75,8 @@ uninitialized_copy_n(_ExecutionPolicy&& __exec, _InputIterator __first, _Size __
     {
         return oneapi::dpl::__internal::__pattern_walk2_brick_n(
             __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __n, __result,
-            oneapi::dpl::__internal::__brick_copy_n<_DecayedExecutionPolicy>{});
+            oneapi::dpl::__internal::__brick_copy_n<::std::decay_t<decltype(__dispatch_tag)>,
+                                                    _DecayedExecutionPolicy>{});
     }
     else
     {
@@ -125,7 +126,8 @@ uninitialized_move_n(_ExecutionPolicy&& __exec, _InputIterator __first, _Size __
     {
         return oneapi::dpl::__internal::__pattern_walk2_brick_n(
             __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __n, __result,
-            oneapi::dpl::__internal::__brick_copy_n<_DecayedExecutionPolicy>{});
+            oneapi::dpl::__internal::__brick_copy_n<::std::decay_t<decltype(__dispatch_tag)>,
+                                                    _DecayedExecutionPolicy>{});
     }
     else
     {
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 32bffaf0b29..2ab06e80540 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -375,8 +375,8 @@ __pattern_generate_n(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec,
 // brick_copy, brick_move
 //------------------------------------------------------------------------
 
-template <typename _ExecutionPolicy>
-struct __brick_copy_n<_ExecutionPolicy, oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy>>
+template <typename _BackendTag, typename _ExecutionPolicy>
+struct __brick_copy_n<__hetero_tag<_BackendTag>, _ExecutionPolicy>
 {
     template <typename _SourceT, typename _TargetT>
     void
@@ -390,7 +390,7 @@ template <typename _ExecutionPolicy>
 struct __brick_copy<_ExecutionPolicy, oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy>>
 {
     template <typename _SourceT, typename _TargetT>
-    oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy>
+    void
     operator()(_SourceT&& __source, _TargetT&& __target) const
     {
         __target = ::std::forward<_SourceT>(__source);

From 2f68546da9d07ae5887c6c1850838268c22936c4 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 29 Feb 2024 17:25:33 +0100
Subject: [PATCH 484/566] struct __brick_copy, __brick_rotate_copy with tags -
 impls + calls

---
 .../dpl/internal/async_impl/glue_async_impl.h |  2 +-
 include/oneapi/dpl/pstl/algorithm_fwd.h       | 14 +++---
 include/oneapi/dpl/pstl/algorithm_impl.h      | 44 ++++++++++---------
 include/oneapi/dpl/pstl/glue_algorithm_impl.h |  6 +--
 .../dpl/pstl/glue_algorithm_ranges_impl.h     |  8 ++--
 include/oneapi/dpl/pstl/glue_memory_impl.h    |  4 +-
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 35 +++++++++------
 .../hetero/algorithm_ranges_impl_hetero.h     | 20 +++++----
 .../dpl/pstl/hetero/numeric_impl_hetero.h     |  5 ++-
 9 files changed, 75 insertions(+), 63 deletions(-)

diff --git a/include/oneapi/dpl/internal/async_impl/glue_async_impl.h b/include/oneapi/dpl/internal/async_impl/glue_async_impl.h
index 06c441fb81d..194f6a84f08 100644
--- a/include/oneapi/dpl/internal/async_impl/glue_async_impl.h
+++ b/include/oneapi/dpl/internal/async_impl/glue_async_impl.h
@@ -82,7 +82,7 @@ copy_async(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _ForwardIterato
     wait_for_all(::std::forward<_Events>(__dependencies)...);
     auto ret_val = oneapi::dpl::__internal::__pattern_walk2_brick_async(
         __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result,
-        oneapi::dpl::__internal::__brick_copy<_ExecutionPolicy>{});
+        oneapi::dpl::__internal::__brick_copy<::std::decay_t<decltype(__dispatch_tag)>, _ExecutionPolicy>{});
     return ret_val;
 }
 
diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index 434d92cb9b3..81cf8e02f73 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -409,7 +409,7 @@ __pattern_search_n(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessI
 template <class _Tag, typename _ExecutionPolicy>
 struct __brick_copy_n;
 
-template <typename _ExecutionPolicy, typename = void>
+template <class _Tag, typename _ExecutionPolicy>
 struct __brick_copy;
 
 template <typename _ExecutionPolicy, typename = void>
@@ -632,14 +632,12 @@ __pattern_rotate(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIte
 // rotate_copy
 //------------------------------------------------------------------------
 
-template <class _ForwardIterator, class _OutputIterator>
-_OutputIterator __brick_rotate_copy(_ForwardIterator, _ForwardIterator, _ForwardIterator, _OutputIterator,
-                                    /*__is_vector=*/::std::false_type) noexcept;
+template <class _Tag, class _ForwardIterator, class _OutputIterator>
+_OutputIterator __brick_rotate_copy(_Tag, _ForwardIterator, _ForwardIterator, _ForwardIterator, _OutputIterator) noexcept;
 
-template <class _RandomAccessIterator, class _OutputIterator>
-_OutputIterator __brick_rotate_copy(_RandomAccessIterator, _RandomAccessIterator, _RandomAccessIterator,
-                                    _OutputIterator,
-                                    /*__is_vector=*/::std::true_type) noexcept;
+template <class _IsVector, class _RandomAccessIterator, class _OutputIterator>
+_OutputIterator __brick_rotate_copy(__parallel_tag<_IsVector>, _RandomAccessIterator, _RandomAccessIterator,
+                                    _RandomAccessIterator, _OutputIterator) noexcept;
 
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _OutputIterator>
 _OutputIterator
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 2514ea05545..a4a081f4d19 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -1075,9 +1075,10 @@ struct __brick_copy_n
 // copy
 //------------------------------------------------------------------------
 
-template <typename _ExecutionPolicy>
-struct __brick_copy<_ExecutionPolicy, oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>>
+template <class _Tag, typename _ExecutionPolicy>
+struct __brick_copy
 {
+    static_assert(__is_host_backend_tag_v<_Tag>);
 
     template <typename _RandomAccessIterator1, typename _RandomAccessIterator2>
     _RandomAccessIterator2
@@ -1913,33 +1914,36 @@ __pattern_rotate(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAc
 // rotate_copy
 //------------------------------------------------------------------------
 
-template <class _ExecutionPolicy, class _ForwardIterator, class _OutputIterator>
+template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _OutputIterator>
 _OutputIterator
-__brick_rotate_copy(_ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __middle, _ForwardIterator __last,
-                    _OutputIterator __result, /*__is_vector=*/::std::false_type) noexcept
+__brick_rotate_copy(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __middle, _ForwardIterator __last,
+                    _OutputIterator __result) noexcept
 {
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+
     return ::std::rotate_copy(__first, __middle, __last, __result);
 }
 
-template <class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2>
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2>
 _RandomAccessIterator2
-__brick_rotate_copy(_ExecutionPolicy&&, _RandomAccessIterator1 __first, _RandomAccessIterator1 __middle,
-                    _RandomAccessIterator1 __last, _RandomAccessIterator2 __result,
-                    /*__is_vector=*/::std::true_type) noexcept
+__brick_rotate_copy(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator1 __first,
+                    _RandomAccessIterator1 __middle,
+                    _RandomAccessIterator1 __last, _RandomAccessIterator2 __result) noexcept
 {
-    _RandomAccessIterator2 __res = __brick_copy<_ExecutionPolicy>{}(__middle, __last, __result, ::std::true_type());
-    return __internal::__brick_copy<_ExecutionPolicy>{}(__first, __middle, __res, ::std::true_type());
+    _RandomAccessIterator2 __res =
+        __brick_copy<__parallel_tag<_IsVector>, _ExecutionPolicy>{}(__middle, __last, __result);
+    return __internal::__brick_copy<__parallel_tag<_IsVector>, _ExecutionPolicy>{}(__first, __middle, __res);
 }
 
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _OutputIterator>
 _OutputIterator
-__pattern_rotate_copy(_Tag, _ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __middle,
+__pattern_rotate_copy(_Tag __tag, _ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __middle,
                       _ForwardIterator __last, _OutputIterator __result) noexcept
 {
     static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
-    return __internal::__brick_rotate_copy(::std::forward<_ExecutionPolicy>(__exec), __first, __middle, __last,
-                                           __result, typename _Tag::__is_vector{});
+    return __internal::__brick_rotate_copy(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __middle, __last,
+                                           __result);
 }
 
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2>
@@ -1952,7 +1956,7 @@ __pattern_rotate_copy(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Ran
     __par_backend::__parallel_for(
         __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
         [__first, __last, __middle, __result](_RandomAccessIterator1 __b, _RandomAccessIterator1 __e) {
-            __internal::__brick_copy<_ExecutionPolicy> __copy{};
+            __internal::__brick_copy<__parallel_tag<_IsVector>, _ExecutionPolicy> __copy{};
             if (__b > __middle)
             {
                 __copy(__b, __e, __result + (__b - __middle), _IsVector{});
@@ -2573,7 +2577,7 @@ __pattern_partial_sort_copy(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec
                     _RandomAccessIterator1 __j1 = __first + (__j - __d_first);
 
                     // 1. Copy elements from input to output
-                    __brick_copy<_ExecutionPolicy>{}(__i1, __j1, __i, _IsVector{});
+                    __brick_copy<__parallel_tag<_IsVector>, _ExecutionPolicy>{}(__i1, __j1, __i, _IsVector{});
                     // 2. Sort elements in output sequence
                     ::std::sort(__i, __j, __comp);
                 },
@@ -3293,7 +3297,7 @@ __parallel_set_union_op(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __ex
     const auto __n1 = __last1 - __first1;
     const auto __n2 = __last2 - __first2;
 
-    __brick_copy<_ExecutionPolicy> __copy_range{};
+    __brick_copy<__parallel_tag<_IsVector>, _ExecutionPolicy> __copy_range{};
 
     // {1} {}: parallel copying just first sequence
     if (__n2 == 0)
@@ -3617,21 +3621,21 @@ __pattern_set_difference(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __e
     // {1} \ {}: parallel copying just first sequence
     if (__n2 == 0)
         return __pattern_walk2_brick(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __result,
-                                     __internal::__brick_copy<_ExecutionPolicy>{});
+                                     __internal::__brick_copy<__parallel_tag<_IsVector>, _ExecutionPolicy>{});
 
     // testing  whether the sequences are intersected
     _RandomAccessIterator1 __left_bound_seq_1 = ::std::lower_bound(__first1, __last1, *__first2, __comp);
     //{1} < {2}: seq 2 is wholly greater than seq 1, so, parallel copying just first sequence
     if (__left_bound_seq_1 == __last1)
         return __pattern_walk2_brick(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __result,
-                                     __internal::__brick_copy<_ExecutionPolicy>{});
+                                     __internal::__brick_copy<__parallel_tag<_IsVector>, _ExecutionPolicy>{});
 
     // testing  whether the sequences are intersected
     _RandomAccessIterator2 __left_bound_seq_2 = ::std::lower_bound(__first2, __last2, *__first1, __comp);
     //{2} < {1}: seq 1 is wholly greater than seq 2, so, parallel copying just first sequence
     if (__left_bound_seq_2 == __last2)
         return __internal::__pattern_walk2_brick(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1,
-                                                 __result, __brick_copy<_ExecutionPolicy>{});
+                                                 __result, __brick_copy<__parallel_tag<_IsVector>, _ExecutionPolicy>{});
 
     if (__n1 + __n2 > __set_algo_cut_off)
         return __parallel_set_op(
diff --git a/include/oneapi/dpl/pstl/glue_algorithm_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
index 91b618c11a7..b2e582164e1 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
@@ -269,9 +269,9 @@ copy(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _ForwardIterator1 __l
 {
     const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first, __result);
 
-    return oneapi::dpl::__internal::__pattern_walk2_brick(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
-                                                          __first, __last, __result,
-                                                          oneapi::dpl::__internal::__brick_copy<_ExecutionPolicy>{});
+    return oneapi::dpl::__internal::__pattern_walk2_brick(
+        __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result,
+        oneapi::dpl::__internal::__brick_copy<::std::decay_t<decltype(__dispatch_tag)>, _ExecutionPolicy>{});
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator1, class _Size, class _ForwardIterator2>
diff --git a/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
index b27140869ff..ec07cc7f7cd 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
@@ -251,10 +251,10 @@ copy(_ExecutionPolicy&& __exec, _Range1&& __rng, _Range2&& __result)
 {
     const auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend(__exec, __rng, __result);
 
-    oneapi::dpl::__internal::__ranges::__pattern_walk_n(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
-                                                        oneapi::dpl::__internal::__brick_copy<_ExecutionPolicy>{},
-                                                        views::all_read(::std::forward<_Range1>(__rng)),
-                                                        views::all_write(::std::forward<_Range2>(__result)));
+    oneapi::dpl::__internal::__ranges::__pattern_walk_n(
+        __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
+        oneapi::dpl::__internal::__brick_copy<::std::decay_t<decltype(__dispatch_tag)>, _ExecutionPolicy>{},
+        views::all_read(::std::forward<_Range1>(__rng)), views::all_write(::std::forward<_Range2>(__result)));
 }
 
 template <typename _ExecutionPolicy, typename _Range1, typename _Range2, typename _Predicate>
diff --git a/include/oneapi/dpl/pstl/glue_memory_impl.h b/include/oneapi/dpl/pstl/glue_memory_impl.h
index 4ba9fe1be47..8278f84620c 100644
--- a/include/oneapi/dpl/pstl/glue_memory_impl.h
+++ b/include/oneapi/dpl/pstl/glue_memory_impl.h
@@ -51,7 +51,7 @@ uninitialized_copy(_ExecutionPolicy&& __exec, _InputIterator __first, _InputIter
     {
         return oneapi::dpl::__internal::__pattern_walk2_brick(
             __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result,
-            oneapi::dpl::__internal::__brick_copy<_DecayedExecutionPolicy>{});
+            oneapi::dpl::__internal::__brick_copy<::std::decay_t<decltype(__dispatch_tag)>, _DecayedExecutionPolicy>{});
     }
     else
     {
@@ -102,7 +102,7 @@ uninitialized_move(_ExecutionPolicy&& __exec, _InputIterator __first, _InputIter
     {
         return oneapi::dpl::__internal::__pattern_walk2_brick(
             __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result,
-            oneapi::dpl::__internal::__brick_copy<_DecayedExecutionPolicy>{});
+            oneapi::dpl::__internal::__brick_copy<::std::decay_t<decltype(__dispatch_tag)>, _DecayedExecutionPolicy>{});
     }
     else
     {
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 2ab06e80540..6757e0b04d4 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -386,8 +386,8 @@ struct __brick_copy_n<__hetero_tag<_BackendTag>, _ExecutionPolicy>
     }
 };
 
-template <typename _ExecutionPolicy>
-struct __brick_copy<_ExecutionPolicy, oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy>>
+template <typename _BackendTag, typename _ExecutionPolicy>
+struct __brick_copy<__hetero_tag<_BackendTag>, _ExecutionPolicy>
 {
     template <typename _SourceT, typename _TargetT>
     void
@@ -1043,7 +1043,7 @@ __pattern_remove_if(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec,
     //TODO: optimize copy back depending on Iterator, i.e. set_final_data for host iterator/pointer
     return __pattern_walk2(
         __tag, __par_backend_hetero::make_wrapped_policy<copy_back_wrapper>(::std::forward<_ExecutionPolicy>(__exec)),
-        __copy_first, __copy_last, __first, __brick_copy<_ExecutionPolicy>{});
+        __copy_first, __copy_last, __first, __brick_copy<__hetero_tag<_BackendTag>, _ExecutionPolicy>{});
 }
 
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator, typename _BinaryPredicate>
@@ -1065,7 +1065,7 @@ __pattern_unique(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _It
     return __pattern_walk2<_BackendTag, /*_IsSync=*/::std::true_type, __par_backend_hetero::access_mode::read_write,
                            __par_backend_hetero::access_mode::read_write>(
         __tag, __par_backend_hetero::make_wrapped_policy<copy_back_wrapper>(::std::forward<_ExecutionPolicy>(__exec)),
-        __copy_first, __copy_last, __first, __brick_copy<_ExecutionPolicy>{});
+        __copy_first, __copy_last, __first, __brick_copy<__hetero_tag<_BackendTag>, _ExecutionPolicy>{});
 }
 
 //------------------------------------------------------------------------
@@ -1188,13 +1188,15 @@ __pattern_merge(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Ite
             __tag,
             oneapi::dpl::__par_backend_hetero::make_wrapped_policy<copy_back_wrapper>(
                 ::std::forward<_ExecutionPolicy>(__exec)),
-            __first2, __last2, __d_first, oneapi::dpl::__internal::__brick_copy<_ExecutionPolicy>{});
+            __first2, __last2, __d_first,
+            oneapi::dpl::__internal::__brick_copy<__hetero_tag<_BackendTag>, _ExecutionPolicy>{});
     else if (__n2 == 0)
         oneapi::dpl::__internal::__pattern_walk2_brick(
             __tag,
             oneapi::dpl::__par_backend_hetero::make_wrapped_policy<copy_back_wrapper2>(
                 ::std::forward<_ExecutionPolicy>(__exec)),
-            __first1, __last1, __d_first, oneapi::dpl::__internal::__brick_copy<_ExecutionPolicy>{});
+            __first1, __last1, __d_first,
+            oneapi::dpl::__internal::__brick_copy<__hetero_tag<_BackendTag>, _ExecutionPolicy>{});
     else
     {
         auto __keep1 = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _Iterator1>();
@@ -1498,7 +1500,7 @@ __pattern_partial_sort_copy(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&&
         // full sort on them.
         auto __out_end = __pattern_walk2<_BackendTag, /*_IsSync=*/::std::false_type>(
             __tag, __par_backend_hetero::make_wrapped_policy<__initial_copy_1>(__exec), __first, __last, __out_first,
-            __brick_copy<_ExecutionPolicy>{});
+            __brick_copy<__hetero_tag<_BackendTag>, _ExecutionPolicy>{});
 
         // Use regular sort as partial_sort isn't required to be stable
         __pattern_sort(
@@ -1520,7 +1522,7 @@ __pattern_partial_sort_copy(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&&
 
         auto __buf_last = __pattern_walk2<_BackendTag, /*_IsSync=*/::std::false_type>(
             __tag, __par_backend_hetero::make_wrapped_policy<__initial_copy_2>(__exec), __first, __last, __buf_first,
-            __brick_copy<_ExecutionPolicy>{});
+            __brick_copy<__hetero_tag<_BackendTag>, _ExecutionPolicy>{});
 
         auto __buf_mid = __buf_first + __out_size;
 
@@ -1532,7 +1534,7 @@ __pattern_partial_sort_copy(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&&
 
         return __pattern_walk2(
             __tag, __par_backend_hetero::make_wrapped_policy<__copy_back>(::std::forward<_ExecutionPolicy>(__exec)),
-            __buf_first, __buf_mid, __out_first, __brick_copy<_ExecutionPolicy>{});
+            __buf_first, __buf_mid, __out_first, __brick_copy<__hetero_tag<_BackendTag>, _ExecutionPolicy>{});
     }
 }
 
@@ -1788,7 +1790,8 @@ __pattern_set_difference(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __e
             __tag,
             oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__set_difference_copy_case_1>(
                 ::std::forward<_ExecutionPolicy>(__exec)),
-            __first1, __last1, __result, oneapi::dpl::__internal::__brick_copy<_ExecutionPolicy>{});
+            __first1, __last1, __result,
+            oneapi::dpl::__internal::__brick_copy<__hetero_tag<_BackendTag>, _ExecutionPolicy>{});
     }
 
     return __pattern_hetero_set_op(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2,
@@ -1823,7 +1826,8 @@ __pattern_set_union(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec,
             __tag,
             oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__set_union_copy_case_1>(
                 ::std::forward<_ExecutionPolicy>(__exec)),
-            __first2, __last2, __result, oneapi::dpl::__internal::__brick_copy<_ExecutionPolicy>{});
+            __first2, __last2, __result,
+            oneapi::dpl::__internal::__brick_copy<__hetero_tag<_BackendTag>, _ExecutionPolicy>{});
     }
 
     //{2} is empty
@@ -1833,7 +1837,8 @@ __pattern_set_union(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec,
             __tag,
             oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__set_union_copy_case_2>(
                 ::std::forward<_ExecutionPolicy>(__exec)),
-            __first1, __last1, __result, oneapi::dpl::__internal::__brick_copy<_ExecutionPolicy>{});
+            __first1, __last1, __result,
+            oneapi::dpl::__internal::__brick_copy<__hetero_tag<_BackendTag>, _ExecutionPolicy>{});
     }
 
     typedef typename ::std::iterator_traits<_OutputIterator>::value_type _ValueType;
@@ -1902,7 +1907,8 @@ __pattern_set_symmetric_difference(__hetero_tag<_BackendTag> __tag, _ExecutionPo
             __tag,
             oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__set_symmetric_difference_copy_case_1>(
                 ::std::forward<_ExecutionPolicy>(__exec)),
-            __first2, __last2, __result, oneapi::dpl::__internal::__brick_copy<_ExecutionPolicy>{});
+            __first2, __last2, __result,
+            oneapi::dpl::__internal::__brick_copy<__hetero_tag<_BackendTag>, _ExecutionPolicy>{});
     }
 
     //{2} is empty
@@ -1912,7 +1918,8 @@ __pattern_set_symmetric_difference(__hetero_tag<_BackendTag> __tag, _ExecutionPo
             __tag,
             oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__set_symmetric_difference_copy_case_2>(
                 ::std::forward<_ExecutionPolicy>(__exec)),
-            __first1, __last1, __result, oneapi::dpl::__internal::__brick_copy<_ExecutionPolicy>{});
+            __first1, __last1, __result,
+            oneapi::dpl::__internal::__brick_copy<__hetero_tag<_BackendTag>, _ExecutionPolicy>{});
     }
 
     typedef typename ::std::iterator_traits<_OutputIterator>::value_type _ValueType;
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
index f9392d34a06..227ff9761b5 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
@@ -421,9 +421,10 @@ __pattern_remove_if(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec,
                                                       oneapi::dpl::__internal::__pstl_assign());
     auto __copy_rng_truncated = __copy_rng | oneapi::dpl::experimental::ranges::views::take(__copy_last_id);
 
-    oneapi::dpl::__internal::__ranges::__pattern_walk_n(__tag, ::std::forward<_ExecutionPolicy>(__exec),
-                                                        oneapi::dpl::__internal::__brick_copy<_ExecutionPolicy>{},
-                                                        __copy_rng_truncated, ::std::forward<_Range>(__rng));
+    oneapi::dpl::__internal::__ranges::__pattern_walk_n(
+        __tag, ::std::forward<_ExecutionPolicy>(__exec),
+        oneapi::dpl::__internal::__brick_copy<__hetero_tag<_BackendTag>, _ExecutionPolicy>{}, __copy_rng_truncated,
+        ::std::forward<_Range>(__rng));
 
     return __copy_last_id;
 }
@@ -466,7 +467,8 @@ __pattern_unique(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Ra
     auto res = __ranges::__pattern_unique_copy(__tag, __exec, __rng, res_rng, __pred,
                                                oneapi::dpl::__internal::__pstl_assign());
 
-    __pattern_walk_n(__tag, ::std::forward<_ExecutionPolicy>(__exec), __brick_copy<_ExecutionPolicy>{}, res_rng,
+    __pattern_walk_n(__tag, ::std::forward<_ExecutionPolicy>(__exec),
+                     __brick_copy<__hetero_tag<_BackendTag>, _ExecutionPolicy>{}, res_rng,
                      ::std::forward<_Range>(__rng));
     return res;
 }
@@ -504,8 +506,8 @@ __pattern_merge(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Ran
             __tag,
             oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__copy1_wrapper>(
                 ::std::forward<_ExecutionPolicy>(__exec)),
-            oneapi::dpl::__internal::__brick_copy<_ExecutionPolicy>{}, ::std::forward<_Range2>(__rng2),
-            ::std::forward<_Range3>(__rng3));
+            oneapi::dpl::__internal::__brick_copy<__hetero_tag<_BackendTag>, _ExecutionPolicy>{},
+            ::std::forward<_Range2>(__rng2), ::std::forward<_Range3>(__rng3));
     }
     else if (__n2 == 0)
     {
@@ -513,8 +515,8 @@ __pattern_merge(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Ran
             __tag,
             oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__copy2_wrapper>(
                 ::std::forward<_ExecutionPolicy>(__exec)),
-            oneapi::dpl::__internal::__brick_copy<_ExecutionPolicy>{}, ::std::forward<_Range1>(__rng1),
-            ::std::forward<_Range3>(__rng3));
+            oneapi::dpl::__internal::__brick_copy<__hetero_tag<_BackendTag>, _ExecutionPolicy>{},
+            ::std::forward<_Range1>(__rng1), ::std::forward<_Range3>(__rng3));
     }
     else
     {
@@ -692,7 +694,7 @@ __pattern_reduce_by_segment(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&&
 
     if (__n == 1)
     {
-        __brick_copy<_ExecutionPolicy> __copy_range{};
+        __brick_copy<__hetero_tag<_BackendTag>, _ExecutionPolicy> __copy_range{};
 
         oneapi::dpl::__internal::__ranges::__pattern_walk_n(
             __tag, oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__copy_keys_wrapper>(__exec), __copy_range,
diff --git a/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
index 0083299169c..c2b75712b6e 100644
--- a/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
@@ -248,8 +248,9 @@ __pattern_adjacent_difference(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&
             auto __wrapped_policy = __par_backend_hetero::make_wrapped_policy<adjacent_difference_wrapper>(
                 ::std::forward<_ExecutionPolicy>(__exec));
 
-            __internal::__pattern_walk2_brick(__tag, __wrapped_policy, __first, __last, __d_first,
-                                              __internal::__brick_copy<decltype(__wrapped_policy)>{});
+            __internal::__pattern_walk2_brick(
+                __tag, __wrapped_policy, __first, __last, __d_first,
+                __internal::__brick_copy<__hetero_tag<_BackendTag>, decltype(__wrapped_policy)>{});
 
             return __d_last;
         });

From 03d6b4b2c51ec8c3cbe5c924b9ead5eae1b21b3c Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 29 Feb 2024 17:34:21 +0100
Subject: [PATCH 485/566] struct __brick_move with tags - impls + calls

---
 include/oneapi/dpl/pstl/algorithm_fwd.h         |  2 +-
 include/oneapi/dpl/pstl/algorithm_impl.h        | 17 +++++++++--------
 include/oneapi/dpl/pstl/glue_algorithm_impl.h   |  2 +-
 .../dpl/pstl/glue_algorithm_ranges_impl.h       |  2 +-
 .../dpl/pstl/hetero/algorithm_impl_hetero.h     | 15 ++++++++-------
 .../dpl/pstl/hetero/numeric_impl_hetero.h       |  5 +++--
 6 files changed, 23 insertions(+), 20 deletions(-)

diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index 81cf8e02f73..524f10df622 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -412,7 +412,7 @@ struct __brick_copy_n;
 template <class _Tag, typename _ExecutionPolicy>
 struct __brick_copy;
 
-template <typename _ExecutionPolicy, typename = void>
+template <class _Tag, typename _ExecutionPolicy>
 struct __brick_move;
 
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index a4a081f4d19..5ad8ae9ae1f 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -1109,9 +1109,10 @@ struct __brick_copy
 // move
 //------------------------------------------------------------------------
 
-template <typename _ExecutionPolicy>
-struct __brick_move<_ExecutionPolicy, oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>>
+template <class _Tag, typename _ExecutionPolicy>
+struct __brick_move
 {
+    static_assert(__is_host_backend_tag_v<_Tag>);
 
     template <typename _RandomAccessIterator1, typename _RandomAccessIterator2>
     _RandomAccessIterator2
@@ -1869,7 +1870,7 @@ __pattern_rotate(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAc
 
             __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __middle,
                                           [__last, __middle](_RandomAccessIterator __b, _RandomAccessIterator __e) {
-                                              __internal::__brick_move<_ExecutionPolicy>{}(
+                                              __internal::__brick_move<__parallel_tag<_IsVector>, _ExecutionPolicy>{}(
                                                   __b, __e, __b + (__last - __middle), _IsVector{});
                                           });
 
@@ -1895,7 +1896,7 @@ __pattern_rotate(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAc
 
             __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __middle, __last,
                                           [__first, __middle](_RandomAccessIterator __b, _RandomAccessIterator __e) {
-                                              __internal::__brick_move<_ExecutionPolicy>{}(
+                                              __internal::__brick_move<__parallel_tag<_IsVector>, _ExecutionPolicy>{}(
                                                   __b, __e, __first + (__b - __middle), _IsVector{});
                                           });
 
@@ -4275,8 +4276,8 @@ __pattern_shift_left(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Forw
     {
         __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __n, __size,
                                       [__first, __n](_DiffType __i, _DiffType __j) {
-                                          __brick_move<_ExecutionPolicy>{}(__first + __i, __first + __j,
-                                                                           __first + __i - __n, _IsVector{});
+                                          __brick_move<__parallel_tag<_IsVector>, _ExecutionPolicy>{}(
+                                              __first + __i, __first + __j, __first + __i - __n, _IsVector{});
                                       });
     }
     else //2. n < size/2; there is not enough memory to parallel copying; doing parallel copying by n elements
@@ -4287,8 +4288,8 @@ __pattern_shift_left(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Forw
             auto __end = ::std::min(__k + __n, __size);
             __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __k, __end,
                                           [__first, __n](_DiffType __i, _DiffType __j) {
-                                              __brick_move<_ExecutionPolicy>{}(__first + __i, __first + __j,
-                                                                               __first + __i - __n, _IsVector{});
+                                              __brick_move<__parallel_tag<_IsVector>, _ExecutionPolicy>{}(
+                                                  __first + __i, __first + __j, __first + __i - __n, _IsVector{});
                                           });
         }
     }
diff --git a/include/oneapi/dpl/pstl/glue_algorithm_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
index b2e582164e1..fa9bbb43051 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
@@ -818,7 +818,7 @@ move(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _ForwardIterator1 __l
 
     return oneapi::dpl::__internal::__pattern_walk2_brick(
         __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __d_first,
-        oneapi::dpl::__internal::__brick_move<_DecayedExecutionPolicy>{});
+        oneapi::dpl::__internal::__brick_move<::std::decay_t<decltype(__dispatch_tag)>, _DecayedExecutionPolicy>{});
 }
 
 // [partial.sort]
diff --git a/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
index ec07cc7f7cd..3796812f620 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
@@ -612,7 +612,7 @@ move(_ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&& __rng2)
 
     oneapi::dpl::__internal::__ranges::__pattern_walk_n(
         __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
-        oneapi::dpl::__internal::__brick_move<_DecayedExecutionPolicy>{},
+        oneapi::dpl::__internal::__brick_move<::std::decay_t<decltype(__dispatch_tag)>, _DecayedExecutionPolicy>{},
         views::all_read(::std::forward<_Range1>(__rng1)), views::all_write(::std::forward<_Range2>(__rng2)));
 }
 
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 6757e0b04d4..9168bb31568 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -397,8 +397,8 @@ struct __brick_copy<__hetero_tag<_BackendTag>, _ExecutionPolicy>
     }
 };
 
-template <typename _ExecutionPolicy>
-struct __brick_move<_ExecutionPolicy, oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy>>
+template <typename _BackendTag, typename _ExecutionPolicy>
+struct __brick_move<__hetero_tag<_BackendTag>, _ExecutionPolicy>
 {
     template <typename _SourceT, typename _TargetT>
     oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy>
@@ -1245,7 +1245,7 @@ __pattern_inplace_merge(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __ex
     //TODO: optimize copy back depending on Iterator, i.e. set_final_data for host iterator/pointer
     __pattern_walk2(
         __tag, __par_backend_hetero::make_wrapped_policy<copy_back_wrapper>(::std::forward<_ExecutionPolicy>(__exec)),
-        __copy_first, __copy_last, __first, __brick_move<_ExecutionPolicy>{});
+        __copy_first, __copy_last, __first, __brick_move<__hetero_tag<_BackendTag>, _ExecutionPolicy>{});
 }
 
 //------------------------------------------------------------------------
@@ -1330,11 +1330,12 @@ __pattern_stable_partition(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& _
     //TODO: optimize copy back if possible (inplace, decrease number of submits)
     __pattern_walk2<_BackendTag, /*_IsSync=*/::std::false_type>(
         __tag, __par_backend_hetero::make_wrapped_policy<copy_back_wrapper>(__exec), __true_result, copy_result.first,
-        __first, __brick_move<_ExecutionPolicy>{});
+        __first, __brick_move<__hetero_tag<_BackendTag>, _ExecutionPolicy>{});
 
     __pattern_walk2(
         __tag, __par_backend_hetero::make_wrapped_policy<copy_back_wrapper2>(::std::forward<_ExecutionPolicy>(__exec)),
-        __false_result, copy_result.second, __first + true_count, __brick_move<_ExecutionPolicy>{});
+        __false_result, copy_result.second, __first + true_count,
+        __brick_move<__hetero_tag<_BackendTag>, _ExecutionPolicy>{});
 
     return __first + true_count;
 }
@@ -1644,7 +1645,7 @@ __pattern_rotate(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Iterator
         unseq_backend::__rotate_copy<typename ::std::iterator_traits<_Iterator>::difference_type>{__n, __shift}, __n,
         __buf.all_view(), __temp_rng);
 
-    using _Function = __brick_move<_ExecutionPolicy>;
+    using _Function = __brick_move<__hetero_tag<_BackendTag>, _ExecutionPolicy>;
     auto __brick = unseq_backend::walk_n<_ExecutionPolicy, _Function>{_Function{}};
 
     oneapi::dpl::__par_backend_hetero::__parallel_for(_BackendTag{}, ::std::forward<_ExecutionPolicy>(__exec), __brick,
@@ -1974,7 +1975,7 @@ __pattern_shift_left(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Rang
     //1. n >= size/2; 'size - _n' parallel copying
     if (__n >= __mid)
     {
-        using _Function = __brick_move<_ExecutionPolicy>;
+        using _Function = __brick_move<__hetero_tag<_BackendTag>, _ExecutionPolicy>;
         auto __brick = oneapi::dpl::unseq_backend::walk_n<_ExecutionPolicy, _Function>{_Function{}};
 
         //TODO: to consider use just "read" access mode for a source range and just "write" - for a destination range.
diff --git a/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
index c2b75712b6e..a313a944125 100644
--- a/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
@@ -177,8 +177,9 @@ __pattern_transform_scan_base(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&
             .wait();
 
         // Move data from temporary buffer into results
-        oneapi::dpl::__internal::__pattern_walk2_brick(__tag, ::std::move(__policy), __first_tmp, __last_tmp, __result,
-                                                       oneapi::dpl::__internal::__brick_move<_NewExecutionPolicy>{});
+        oneapi::dpl::__internal::__pattern_walk2_brick(
+            __tag, ::std::move(__policy), __first_tmp, __last_tmp, __result,
+            oneapi::dpl::__internal::__brick_move<__hetero_tag<_BackendTag>, _NewExecutionPolicy>{});
 
         //TODO: optimize copy back depending on Iterator, i.e. set_final_data for host iterator/pointer
     }

From 74a2357231acea8bd0c00b6ef629ba76ae082c24 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 29 Feb 2024 17:38:43 +0100
Subject: [PATCH 486/566] struct __brick_move_destroy with tags - impls + calls

---
 include/oneapi/dpl/pstl/algorithm_impl.h | 30 +++++++++++-------------
 1 file changed, 14 insertions(+), 16 deletions(-)

diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 5ad8ae9ae1f..8c65bd01436 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -1139,13 +1139,11 @@ struct __brick_move
     }
 };
 
-template <typename _ExecutionPolicy, typename _Dummy = void>
-struct __brick_move_destroy;
-
-template <typename _ExecutionPolicy>
-struct __brick_move_destroy<_ExecutionPolicy,
-                            oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>>
+template <class _Tag, typename _ExecutionPolicy>
+struct __brick_move_destroy
 {
+    static_assert(__is_host_backend_tag_v<_Tag>);
+
     template <typename _RandomAccessIterator1, typename _RandomAccessIterator2>
     _RandomAccessIterator2
     operator()(_RandomAccessIterator1 __first, _RandomAccessIterator1 __last, _RandomAccessIterator2 __result,
@@ -1518,8 +1516,8 @@ __remove_elements(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Forward
         // 3. Elements from result are moved to [first, last)
         __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __result,
                                       __result + __m, [__result, __first](_Tp* __i, _Tp* __j) {
-                                          __brick_move_destroy<_ExecutionPolicy>{}(__i, __j, __first + (__i - __result),
-                                                                                   _IsVector{});
+                                          __brick_move_destroy<__parallel_tag<_IsVector>, _ExecutionPolicy>{}(
+                                              __i, __j, __first + (__i - __result), _IsVector{});
                                       });
         return __first + __m;
     });
@@ -1876,7 +1874,7 @@ __pattern_rotate(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAc
 
             __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __result,
                                           __result + (__n - __m), [__first, __result](_Tp* __b, _Tp* __e) {
-                                              __brick_move_destroy<_ExecutionPolicy>{}(
+                                              __brick_move_destroy<__parallel_tag<_IsVector>, _ExecutionPolicy>{}(
                                                   __b, __e, __first + (__b - __result), _IsVector{});
                                           });
 
@@ -1902,7 +1900,7 @@ __pattern_rotate(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAc
 
             __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __result,
                                           __result + __m, [__n, __m, __first, __result](_Tp* __b, _Tp* __e) {
-                                              __brick_move_destroy<_ExecutionPolicy>{}(
+                                              __brick_move_destroy<__parallel_tag<_IsVector>, _ExecutionPolicy>{}(
                                                   __b, __e, __first + ((__n - __m) + (__b - __result)), _IsVector{});
                                           });
 
@@ -2614,7 +2612,7 @@ __pattern_partial_sort_copy(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec
             // 3. Move elements from temporary buffer to output
             __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __r, __r + __n2,
                                           [__r, __d_first](_T1* __i, _T1* __j) {
-                                              __brick_move_destroy<_ExecutionPolicy>{}(
+                                              __brick_move_destroy<__parallel_tag<_IsVector>, _ExecutionPolicy>{}(
                                                   __i, __j, __d_first + (__i - __r), _IsVector{});
                                           });
 
@@ -3107,8 +3105,8 @@ __pattern_inplace_merge(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _R
             });
         __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __r, __r + __n,
                                       [__r, __first](_Tp* __i, _Tp* __j) {
-                                          __brick_move_destroy<_ExecutionPolicy>{}(__i, __j, __first + (__i - __r),
-                                                                                   _IsVector{});
+                                          __brick_move_destroy<__parallel_tag<_IsVector>, _ExecutionPolicy>{}(
+                                              __i, __j, __first + (__i - __r), _IsVector{});
                                       });
     });
 }
@@ -3224,9 +3222,9 @@ __parallel_set_op(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Forward
         _DifferenceType __m{};
         auto __scan = [=](_DifferenceType, _DifferenceType, const _SetRange& __s) { // Scan
             if (!__s.empty())
-                __brick_move_destroy<_ExecutionPolicy>{}(__tmp_memory + __s.__buf_pos,
-                                                         __tmp_memory + (__s.__buf_pos + __s.__len),
-                                                         __result + __s.__pos, _IsVector{});
+                __brick_move_destroy<__parallel_tag<_IsVector>, _ExecutionPolicy>{}(
+                    __tmp_memory + __s.__buf_pos, __tmp_memory + (__s.__buf_pos + __s.__len), __result + __s.__pos,
+                    _IsVector{});
         };
         __par_backend::__parallel_strict_scan(
             __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __n1, _SetRange{0, 0, 0}, //-1, 0},

From c52f153235854e4f2bbb815a5d2dbf03360b135e Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 1 Mar 2024 09:48:59 +0100
Subject: [PATCH 487/566] struct __brick_fill with tags - impls + calls

---
 include/oneapi/dpl/pstl/algorithm_fwd.h              |  2 +-
 include/oneapi/dpl/pstl/algorithm_impl.h             | 12 +++++++-----
 include/oneapi/dpl/pstl/glue_memory_impl.h           |  7 +++++--
 .../oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h   |  5 ++---
 4 files changed, 15 insertions(+), 11 deletions(-)

diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index 524f10df622..bf36f11059c 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -852,7 +852,7 @@ __pattern_nth_element(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAcce
 //------------------------------------------------------------------------
 // fill, fill_n
 //------------------------------------------------------------------------
-template <typename _Tp, typename _ExecutionPolicy, typename = void>
+template <class _Tag, typename _Tp, typename _ExecutionPolicy>
 struct __brick_fill;
 
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Tp>
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 8c65bd01436..ab3288ed61f 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -2759,9 +2759,11 @@ __pattern_nth_element(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec
 //------------------------------------------------------------------------
 // fill, fill_n
 //------------------------------------------------------------------------
-template <typename _Tp, typename _ExecutionPolicy>
-struct __brick_fill<_Tp, _ExecutionPolicy, oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>>
+template <class _Tag, typename _Tp, typename _ExecutionPolicy>
+struct __brick_fill
 {
+    static_assert(__is_host_backend_tag_v<_Tag>);
+
     const _Tp& __value;
 
     template <typename _RandomAccessIterator>
@@ -2787,7 +2789,7 @@ __pattern_fill(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardItera
 {
     static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
-    __internal::__brick_fill<_Tp, _ExecutionPolicy>{__value}(__first, __last, typename _Tag::__is_vector{});
+    __internal::__brick_fill<_Tag, _Tp, _ExecutionPolicy>{__value}(__first, __last, typename _Tag::__is_vector{});
 }
 
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Tp>
@@ -2800,8 +2802,8 @@ __pattern_fill(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAcce
     return __internal::__except_handler([&__exec, __first, __last, &__value]() {
         __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
                                       [&__value](_RandomAccessIterator __begin, _RandomAccessIterator __end) {
-                                          __internal::__brick_fill<_Tp, _ExecutionPolicy>{__value}(__begin, __end,
-                                                                                                   _IsVector{});
+                                          __internal::__brick_fill<__parallel_tag<_IsVector>, _Tp, _ExecutionPolicy>{
+                                              __value}(__begin, __end, _IsVector{});
                                       });
         return __last;
     });
diff --git a/include/oneapi/dpl/pstl/glue_memory_impl.h b/include/oneapi/dpl/pstl/glue_memory_impl.h
index 8278f84620c..91a2c81575b 100644
--- a/include/oneapi/dpl/pstl/glue_memory_impl.h
+++ b/include/oneapi/dpl/pstl/glue_memory_impl.h
@@ -152,7 +152,9 @@ uninitialized_fill(_ExecutionPolicy&& __exec, _ForwardIterator __first, _Forward
     {
         oneapi::dpl::__internal::__pattern_walk_brick(
             __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
-            oneapi::dpl::__internal::__brick_fill<_ValueType, _DecayedExecutionPolicy>{_ValueType(__value)});
+            oneapi::dpl::__internal::__brick_fill<::std::decay_t<decltype(__dispatch_tag)>, _ValueType,
+                                                  _DecayedExecutionPolicy>{
+                _ValueType(__value)});
     }
     else
     {
@@ -312,7 +314,8 @@ uninitialized_value_construct(_ExecutionPolicy&& __exec, _ForwardIterator __firs
     {
         oneapi::dpl::__internal::__pattern_walk_brick(
             __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
-            oneapi::dpl::__internal::__brick_fill<_ValueType, _DecayedExecutionPolicy>{_ValueType()});
+            oneapi::dpl::__internal::__brick_fill<::std::decay_t<decltype(__dispatch_tag)>, _ValueType,
+                                                  _DecayedExecutionPolicy>{_ValueType()});
     }
     else
     {
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 9168bb31568..104b9091f4a 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -408,9 +408,8 @@ struct __brick_move<__hetero_tag<_BackendTag>, _ExecutionPolicy>
     }
 };
 
-template <typename _SourceT, typename _ExecutionPolicy>
-struct __brick_fill<_SourceT, _ExecutionPolicy,
-                    oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy>>
+template <typename _BackendTag, typename _SourceT, typename _ExecutionPolicy>
+struct __brick_fill<__hetero_tag<_BackendTag>, _SourceT, _ExecutionPolicy>
 {
     _SourceT __value;
     template <typename _TargetT>

From ecca122eb99bf7b43651dee6ce3a928f167b02fc Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 1 Mar 2024 09:48:33 +0100
Subject: [PATCH 488/566] struct __brick_fill_n with tags - impls + calls

---
 include/oneapi/dpl/pstl/algorithm_fwd.h                |  2 +-
 include/oneapi/dpl/pstl/algorithm_impl.h               | 10 ++++++----
 include/oneapi/dpl/pstl/glue_memory_impl.h             |  6 ++++--
 include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h |  5 ++---
 4 files changed, 13 insertions(+), 10 deletions(-)

diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index bf36f11059c..6cb4d735710 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -863,7 +863,7 @@ template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator,
 _RandomAccessIterator
 __pattern_fill(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, const _Tp&);
 
-template <typename _Tp, typename _ExecutionPolicy, typename = void>
+template <class _Tag, typename _Tp, typename _ExecutionPolicy>
 struct __brick_fill_n;
 
 template <class _Tag, class _ExecutionPolicy, class _OutputIterator, class _Size, class _Tp>
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index ab3288ed61f..2741f47bc1a 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -2809,10 +2809,11 @@ __pattern_fill(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAcce
     });
 }
 
-template <typename _Tp, typename _ExecutionPolicy>
-struct __brick_fill_n<_Tp, _ExecutionPolicy,
-                      oneapi::dpl::__internal::__enable_if_host_execution_policy<_ExecutionPolicy>>
+template <class _Tag, typename _Tp, typename _ExecutionPolicy>
+struct __brick_fill_n
 {
+    static_assert(__is_host_backend_tag_v<_Tag>);
+
     const _Tp& __value;
 
     template <typename _RandomAccessIterator, typename _Size>
@@ -2838,7 +2839,8 @@ __pattern_fill_n(_Tag, _ExecutionPolicy&&, _OutputIterator __first, _Size __coun
 {
     static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
-    return __internal::__brick_fill_n<_Tp, _ExecutionPolicy>{__value}(__first, __count, typename _Tag::__is_vector{});
+    return __internal::__brick_fill_n<_Tag, _Tp, _ExecutionPolicy>{__value}(__first, __count,
+                                                                            typename _Tag::__is_vector{});
 }
 
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Size, class _Tp>
diff --git a/include/oneapi/dpl/pstl/glue_memory_impl.h b/include/oneapi/dpl/pstl/glue_memory_impl.h
index 91a2c81575b..ec264c5805a 100644
--- a/include/oneapi/dpl/pstl/glue_memory_impl.h
+++ b/include/oneapi/dpl/pstl/glue_memory_impl.h
@@ -177,7 +177,8 @@ uninitialized_fill_n(_ExecutionPolicy&& __exec, _ForwardIterator __first, _Size
     {
         return oneapi::dpl::__internal::__pattern_walk_brick_n(
             __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __n,
-            oneapi::dpl::__internal::__brick_fill_n<_ValueType, _DecayedExecutionPolicy>{_ValueType(__value)});
+            oneapi::dpl::__internal::__brick_fill_n<::std::decay_t<decltype(__dispatch_tag)>, _ValueType,
+                                                    _DecayedExecutionPolicy>{_ValueType(__value)});
     }
     else
     {
@@ -338,7 +339,8 @@ uninitialized_value_construct_n(_ExecutionPolicy&& __exec, _ForwardIterator __fi
     {
         return oneapi::dpl::__internal::__pattern_walk_brick_n(
             __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __n,
-            oneapi::dpl::__internal::__brick_fill_n<_ValueType, _DecayedExecutionPolicy>{_ValueType()});
+            oneapi::dpl::__internal::__brick_fill_n<::std::decay_t<decltype(__dispatch_tag)>, _ValueType,
+                                                    _DecayedExecutionPolicy>{_ValueType()});
     }
     else
     {
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 104b9091f4a..b68c90bd61c 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -420,9 +420,8 @@ struct __brick_fill<__hetero_tag<_BackendTag>, _SourceT, _ExecutionPolicy>
     }
 };
 
-template <typename _SourceT, typename _ExecutionPolicy>
-struct __brick_fill_n<_SourceT, _ExecutionPolicy,
-                      oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy>>
+template <typename _BackendTag, typename _SourceT, typename _ExecutionPolicy>
+struct __brick_fill_n<__hetero_tag<_BackendTag>, _SourceT, _ExecutionPolicy>
 {
     _SourceT __value;
     template <typename _TargetT>

From baf38db4cb3c2db71f84fb38a495a0663f86797a Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 29 Feb 2024 17:54:55 +0100
Subject: [PATCH 489/566] include/oneapi/dpl/pstl/execution_defs.h - remove
 __enable_if_host_execution_policy as not required anymore

---
 include/oneapi/dpl/pstl/execution_defs.h | 4 ----
 1 file changed, 4 deletions(-)

diff --git a/include/oneapi/dpl/pstl/execution_defs.h b/include/oneapi/dpl/pstl/execution_defs.h
index 98b99f6cce5..26287ccbf6e 100644
--- a/include/oneapi/dpl/pstl/execution_defs.h
+++ b/include/oneapi/dpl/pstl/execution_defs.h
@@ -112,10 +112,6 @@ template <class _ExecPolicy, class _T = void>
 using __enable_if_execution_policy =
     ::std::enable_if_t<oneapi::dpl::execution::is_execution_policy_v<::std::decay_t<_ExecPolicy>>, _T>;
 
-template <class _ExecPolicy, class _T = void>
-using __enable_if_host_execution_policy =
-    ::std::enable_if_t<__is_host_execution_policy<::std::decay_t<_ExecPolicy>>::value, _T>;
-
 template <typename _ExecPolicy, typename _T>
 struct __ref_or_copy_impl
 {

From 29d519fe43d926749d14dec8c14bbb4c03b7d792 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 1 Mar 2024 11:59:25 +0100
Subject: [PATCH 490/566] struct __brick_copy_n - remove _ExecutionPolicy from
 template params

---
 include/oneapi/dpl/pstl/algorithm_fwd.h                | 2 +-
 include/oneapi/dpl/pstl/algorithm_impl.h               | 2 +-
 include/oneapi/dpl/pstl/glue_algorithm_impl.h          | 2 +-
 include/oneapi/dpl/pstl/glue_memory_impl.h             | 6 ++----
 include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h | 4 ++--
 5 files changed, 7 insertions(+), 9 deletions(-)

diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index 6cb4d735710..16dc7a6e522 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -406,7 +406,7 @@ __pattern_search_n(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessI
 // copy_n
 //------------------------------------------------------------------------
 
-template <class _Tag, typename _ExecutionPolicy>
+template <class _Tag>
 struct __brick_copy_n;
 
 template <class _Tag, typename _ExecutionPolicy>
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 2741f47bc1a..e8b850212c3 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -1048,7 +1048,7 @@ __pattern_search_n(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _
 // clear that doing so is worth the trouble and extra layers of call chain.
 // Sometimes a little duplication for sake of regularity is better than the alternative.
 
-template <class _Tag, typename _ExecutionPolicy>
+template <class _Tag>
 struct __brick_copy_n
 {
     static_assert(__is_host_backend_tag_v<_Tag>);
diff --git a/include/oneapi/dpl/pstl/glue_algorithm_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
index fa9bbb43051..2fa0151b8ae 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
@@ -284,7 +284,7 @@ copy_n(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _Size __n, _Forward
 
     return oneapi::dpl::__internal::__pattern_walk2_brick_n(
         __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __n, __result,
-        oneapi::dpl::__internal::__brick_copy_n<::std::decay_t<decltype(__dispatch_tag)>, _DecayedExecutionPolicy>{});
+        oneapi::dpl::__internal::__brick_copy_n<::std::decay_t<decltype(__dispatch_tag)>>{});
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Predicate>
diff --git a/include/oneapi/dpl/pstl/glue_memory_impl.h b/include/oneapi/dpl/pstl/glue_memory_impl.h
index ec264c5805a..995b9f10a0d 100644
--- a/include/oneapi/dpl/pstl/glue_memory_impl.h
+++ b/include/oneapi/dpl/pstl/glue_memory_impl.h
@@ -75,8 +75,7 @@ uninitialized_copy_n(_ExecutionPolicy&& __exec, _InputIterator __first, _Size __
     {
         return oneapi::dpl::__internal::__pattern_walk2_brick_n(
             __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __n, __result,
-            oneapi::dpl::__internal::__brick_copy_n<::std::decay_t<decltype(__dispatch_tag)>,
-                                                    _DecayedExecutionPolicy>{});
+            oneapi::dpl::__internal::__brick_copy_n<::std::decay_t<decltype(__dispatch_tag)>>{});
     }
     else
     {
@@ -126,8 +125,7 @@ uninitialized_move_n(_ExecutionPolicy&& __exec, _InputIterator __first, _Size __
     {
         return oneapi::dpl::__internal::__pattern_walk2_brick_n(
             __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __n, __result,
-            oneapi::dpl::__internal::__brick_copy_n<::std::decay_t<decltype(__dispatch_tag)>,
-                                                    _DecayedExecutionPolicy>{});
+            oneapi::dpl::__internal::__brick_copy_n<::std::decay_t<decltype(__dispatch_tag)>>{});
     }
     else
     {
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index b68c90bd61c..1d9560fd8d3 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -375,8 +375,8 @@ __pattern_generate_n(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec,
 // brick_copy, brick_move
 //------------------------------------------------------------------------
 
-template <typename _BackendTag, typename _ExecutionPolicy>
-struct __brick_copy_n<__hetero_tag<_BackendTag>, _ExecutionPolicy>
+template <typename _BackendTag>
+struct __brick_copy_n<__hetero_tag<_BackendTag>>
 {
     template <typename _SourceT, typename _TargetT>
     void

From c94d8e2414bf5fac259f063728fd4508baf743a8 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 1 Mar 2024 12:03:45 +0100
Subject: [PATCH 491/566] struct __brick_copy - remove _ExecutionPolicy from
 template params

---
 .../dpl/internal/async_impl/glue_async_impl.h |  2 +-
 include/oneapi/dpl/pstl/algorithm_fwd.h       |  2 +-
 include/oneapi/dpl/pstl/algorithm_impl.h      | 19 ++++++-------
 include/oneapi/dpl/pstl/glue_algorithm_impl.h |  2 +-
 .../dpl/pstl/glue_algorithm_ranges_impl.h     |  2 +-
 include/oneapi/dpl/pstl/glue_memory_impl.h    |  4 +--
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 28 +++++++++----------
 .../hetero/algorithm_ranges_impl_hetero.h     | 10 +++----
 .../dpl/pstl/hetero/numeric_impl_hetero.h     |  2 +-
 9 files changed, 35 insertions(+), 36 deletions(-)

diff --git a/include/oneapi/dpl/internal/async_impl/glue_async_impl.h b/include/oneapi/dpl/internal/async_impl/glue_async_impl.h
index 194f6a84f08..4dd32e3a012 100644
--- a/include/oneapi/dpl/internal/async_impl/glue_async_impl.h
+++ b/include/oneapi/dpl/internal/async_impl/glue_async_impl.h
@@ -82,7 +82,7 @@ copy_async(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _ForwardIterato
     wait_for_all(::std::forward<_Events>(__dependencies)...);
     auto ret_val = oneapi::dpl::__internal::__pattern_walk2_brick_async(
         __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result,
-        oneapi::dpl::__internal::__brick_copy<::std::decay_t<decltype(__dispatch_tag)>, _ExecutionPolicy>{});
+        oneapi::dpl::__internal::__brick_copy<::std::decay_t<decltype(__dispatch_tag)>>{});
     return ret_val;
 }
 
diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index 16dc7a6e522..d9e5cb46e81 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -409,7 +409,7 @@ __pattern_search_n(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessI
 template <class _Tag>
 struct __brick_copy_n;
 
-template <class _Tag, typename _ExecutionPolicy>
+template <class _Tag>
 struct __brick_copy;
 
 template <class _Tag, typename _ExecutionPolicy>
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index e8b850212c3..85bda3e634a 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -1075,7 +1075,7 @@ struct __brick_copy_n
 // copy
 //------------------------------------------------------------------------
 
-template <class _Tag, typename _ExecutionPolicy>
+template <class _Tag>
 struct __brick_copy
 {
     static_assert(__is_host_backend_tag_v<_Tag>);
@@ -1929,9 +1929,8 @@ __brick_rotate_copy(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccess
                     _RandomAccessIterator1 __middle,
                     _RandomAccessIterator1 __last, _RandomAccessIterator2 __result) noexcept
 {
-    _RandomAccessIterator2 __res =
-        __brick_copy<__parallel_tag<_IsVector>, _ExecutionPolicy>{}(__middle, __last, __result);
-    return __internal::__brick_copy<__parallel_tag<_IsVector>, _ExecutionPolicy>{}(__first, __middle, __res);
+    _RandomAccessIterator2 __res = __brick_copy<__parallel_tag<_IsVector>>{}(__middle, __last, __result);
+    return __internal::__brick_copy<__parallel_tag<_IsVector>>{}(__first, __middle, __res);
 }
 
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _OutputIterator>
@@ -1955,7 +1954,7 @@ __pattern_rotate_copy(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Ran
     __par_backend::__parallel_for(
         __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
         [__first, __last, __middle, __result](_RandomAccessIterator1 __b, _RandomAccessIterator1 __e) {
-            __internal::__brick_copy<__parallel_tag<_IsVector>, _ExecutionPolicy> __copy{};
+            __internal::__brick_copy<__parallel_tag<_IsVector>> __copy{};
             if (__b > __middle)
             {
                 __copy(__b, __e, __result + (__b - __middle), _IsVector{});
@@ -2576,7 +2575,7 @@ __pattern_partial_sort_copy(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec
                     _RandomAccessIterator1 __j1 = __first + (__j - __d_first);
 
                     // 1. Copy elements from input to output
-                    __brick_copy<__parallel_tag<_IsVector>, _ExecutionPolicy>{}(__i1, __j1, __i, _IsVector{});
+                    __brick_copy<__parallel_tag<_IsVector>>{}(__i1, __j1, __i, _IsVector{});
                     // 2. Sort elements in output sequence
                     ::std::sort(__i, __j, __comp);
                 },
@@ -3300,7 +3299,7 @@ __parallel_set_union_op(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __ex
     const auto __n1 = __last1 - __first1;
     const auto __n2 = __last2 - __first2;
 
-    __brick_copy<__parallel_tag<_IsVector>, _ExecutionPolicy> __copy_range{};
+    __brick_copy<__parallel_tag<_IsVector>> __copy_range{};
 
     // {1} {}: parallel copying just first sequence
     if (__n2 == 0)
@@ -3624,21 +3623,21 @@ __pattern_set_difference(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __e
     // {1} \ {}: parallel copying just first sequence
     if (__n2 == 0)
         return __pattern_walk2_brick(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __result,
-                                     __internal::__brick_copy<__parallel_tag<_IsVector>, _ExecutionPolicy>{});
+                                     __internal::__brick_copy<__parallel_tag<_IsVector>>{});
 
     // testing  whether the sequences are intersected
     _RandomAccessIterator1 __left_bound_seq_1 = ::std::lower_bound(__first1, __last1, *__first2, __comp);
     //{1} < {2}: seq 2 is wholly greater than seq 1, so, parallel copying just first sequence
     if (__left_bound_seq_1 == __last1)
         return __pattern_walk2_brick(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __result,
-                                     __internal::__brick_copy<__parallel_tag<_IsVector>, _ExecutionPolicy>{});
+                                     __internal::__brick_copy<__parallel_tag<_IsVector>>{});
 
     // testing  whether the sequences are intersected
     _RandomAccessIterator2 __left_bound_seq_2 = ::std::lower_bound(__first2, __last2, *__first1, __comp);
     //{2} < {1}: seq 1 is wholly greater than seq 2, so, parallel copying just first sequence
     if (__left_bound_seq_2 == __last2)
         return __internal::__pattern_walk2_brick(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1,
-                                                 __result, __brick_copy<__parallel_tag<_IsVector>, _ExecutionPolicy>{});
+                                                 __result, __brick_copy<__parallel_tag<_IsVector>>{});
 
     if (__n1 + __n2 > __set_algo_cut_off)
         return __parallel_set_op(
diff --git a/include/oneapi/dpl/pstl/glue_algorithm_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
index 2fa0151b8ae..02a1c0d71c6 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
@@ -271,7 +271,7 @@ copy(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _ForwardIterator1 __l
 
     return oneapi::dpl::__internal::__pattern_walk2_brick(
         __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result,
-        oneapi::dpl::__internal::__brick_copy<::std::decay_t<decltype(__dispatch_tag)>, _ExecutionPolicy>{});
+        oneapi::dpl::__internal::__brick_copy<::std::decay_t<decltype(__dispatch_tag)>>{});
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator1, class _Size, class _ForwardIterator2>
diff --git a/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
index 3796812f620..095fb1e20ae 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
@@ -253,7 +253,7 @@ copy(_ExecutionPolicy&& __exec, _Range1&& __rng, _Range2&& __result)
 
     oneapi::dpl::__internal::__ranges::__pattern_walk_n(
         __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
-        oneapi::dpl::__internal::__brick_copy<::std::decay_t<decltype(__dispatch_tag)>, _ExecutionPolicy>{},
+        oneapi::dpl::__internal::__brick_copy<::std::decay_t<decltype(__dispatch_tag)>>{},
         views::all_read(::std::forward<_Range1>(__rng)), views::all_write(::std::forward<_Range2>(__result)));
 }
 
diff --git a/include/oneapi/dpl/pstl/glue_memory_impl.h b/include/oneapi/dpl/pstl/glue_memory_impl.h
index 995b9f10a0d..860057753eb 100644
--- a/include/oneapi/dpl/pstl/glue_memory_impl.h
+++ b/include/oneapi/dpl/pstl/glue_memory_impl.h
@@ -51,7 +51,7 @@ uninitialized_copy(_ExecutionPolicy&& __exec, _InputIterator __first, _InputIter
     {
         return oneapi::dpl::__internal::__pattern_walk2_brick(
             __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result,
-            oneapi::dpl::__internal::__brick_copy<::std::decay_t<decltype(__dispatch_tag)>, _DecayedExecutionPolicy>{});
+            oneapi::dpl::__internal::__brick_copy<::std::decay_t<decltype(__dispatch_tag)>>{});
     }
     else
     {
@@ -101,7 +101,7 @@ uninitialized_move(_ExecutionPolicy&& __exec, _InputIterator __first, _InputIter
     {
         return oneapi::dpl::__internal::__pattern_walk2_brick(
             __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result,
-            oneapi::dpl::__internal::__brick_copy<::std::decay_t<decltype(__dispatch_tag)>, _DecayedExecutionPolicy>{});
+            oneapi::dpl::__internal::__brick_copy<::std::decay_t<decltype(__dispatch_tag)>>{});
     }
     else
     {
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 1d9560fd8d3..2f28843cdfe 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -386,8 +386,8 @@ struct __brick_copy_n<__hetero_tag<_BackendTag>>
     }
 };
 
-template <typename _BackendTag, typename _ExecutionPolicy>
-struct __brick_copy<__hetero_tag<_BackendTag>, _ExecutionPolicy>
+template <typename _BackendTag>
+struct __brick_copy<__hetero_tag<_BackendTag>>
 {
     template <typename _SourceT, typename _TargetT>
     void
@@ -1041,7 +1041,7 @@ __pattern_remove_if(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec,
     //TODO: optimize copy back depending on Iterator, i.e. set_final_data for host iterator/pointer
     return __pattern_walk2(
         __tag, __par_backend_hetero::make_wrapped_policy<copy_back_wrapper>(::std::forward<_ExecutionPolicy>(__exec)),
-        __copy_first, __copy_last, __first, __brick_copy<__hetero_tag<_BackendTag>, _ExecutionPolicy>{});
+        __copy_first, __copy_last, __first, __brick_copy<__hetero_tag<_BackendTag>>{});
 }
 
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator, typename _BinaryPredicate>
@@ -1063,7 +1063,7 @@ __pattern_unique(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _It
     return __pattern_walk2<_BackendTag, /*_IsSync=*/::std::true_type, __par_backend_hetero::access_mode::read_write,
                            __par_backend_hetero::access_mode::read_write>(
         __tag, __par_backend_hetero::make_wrapped_policy<copy_back_wrapper>(::std::forward<_ExecutionPolicy>(__exec)),
-        __copy_first, __copy_last, __first, __brick_copy<__hetero_tag<_BackendTag>, _ExecutionPolicy>{});
+        __copy_first, __copy_last, __first, __brick_copy<__hetero_tag<_BackendTag>>{});
 }
 
 //------------------------------------------------------------------------
@@ -1187,14 +1187,14 @@ __pattern_merge(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Ite
             oneapi::dpl::__par_backend_hetero::make_wrapped_policy<copy_back_wrapper>(
                 ::std::forward<_ExecutionPolicy>(__exec)),
             __first2, __last2, __d_first,
-            oneapi::dpl::__internal::__brick_copy<__hetero_tag<_BackendTag>, _ExecutionPolicy>{});
+            oneapi::dpl::__internal::__brick_copy<__hetero_tag<_BackendTag>>{});
     else if (__n2 == 0)
         oneapi::dpl::__internal::__pattern_walk2_brick(
             __tag,
             oneapi::dpl::__par_backend_hetero::make_wrapped_policy<copy_back_wrapper2>(
                 ::std::forward<_ExecutionPolicy>(__exec)),
             __first1, __last1, __d_first,
-            oneapi::dpl::__internal::__brick_copy<__hetero_tag<_BackendTag>, _ExecutionPolicy>{});
+            oneapi::dpl::__internal::__brick_copy<__hetero_tag<_BackendTag>>{});
     else
     {
         auto __keep1 = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _Iterator1>();
@@ -1499,7 +1499,7 @@ __pattern_partial_sort_copy(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&&
         // full sort on them.
         auto __out_end = __pattern_walk2<_BackendTag, /*_IsSync=*/::std::false_type>(
             __tag, __par_backend_hetero::make_wrapped_policy<__initial_copy_1>(__exec), __first, __last, __out_first,
-            __brick_copy<__hetero_tag<_BackendTag>, _ExecutionPolicy>{});
+            __brick_copy<__hetero_tag<_BackendTag>>{});
 
         // Use regular sort as partial_sort isn't required to be stable
         __pattern_sort(
@@ -1521,7 +1521,7 @@ __pattern_partial_sort_copy(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&&
 
         auto __buf_last = __pattern_walk2<_BackendTag, /*_IsSync=*/::std::false_type>(
             __tag, __par_backend_hetero::make_wrapped_policy<__initial_copy_2>(__exec), __first, __last, __buf_first,
-            __brick_copy<__hetero_tag<_BackendTag>, _ExecutionPolicy>{});
+            __brick_copy<__hetero_tag<_BackendTag>>{});
 
         auto __buf_mid = __buf_first + __out_size;
 
@@ -1533,7 +1533,7 @@ __pattern_partial_sort_copy(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&&
 
         return __pattern_walk2(
             __tag, __par_backend_hetero::make_wrapped_policy<__copy_back>(::std::forward<_ExecutionPolicy>(__exec)),
-            __buf_first, __buf_mid, __out_first, __brick_copy<__hetero_tag<_BackendTag>, _ExecutionPolicy>{});
+            __buf_first, __buf_mid, __out_first, __brick_copy<__hetero_tag<_BackendTag>>{});
     }
 }
 
@@ -1790,7 +1790,7 @@ __pattern_set_difference(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __e
             oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__set_difference_copy_case_1>(
                 ::std::forward<_ExecutionPolicy>(__exec)),
             __first1, __last1, __result,
-            oneapi::dpl::__internal::__brick_copy<__hetero_tag<_BackendTag>, _ExecutionPolicy>{});
+            oneapi::dpl::__internal::__brick_copy<__hetero_tag<_BackendTag>>{});
     }
 
     return __pattern_hetero_set_op(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2,
@@ -1826,7 +1826,7 @@ __pattern_set_union(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec,
             oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__set_union_copy_case_1>(
                 ::std::forward<_ExecutionPolicy>(__exec)),
             __first2, __last2, __result,
-            oneapi::dpl::__internal::__brick_copy<__hetero_tag<_BackendTag>, _ExecutionPolicy>{});
+            oneapi::dpl::__internal::__brick_copy<__hetero_tag<_BackendTag>>{});
     }
 
     //{2} is empty
@@ -1837,7 +1837,7 @@ __pattern_set_union(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec,
             oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__set_union_copy_case_2>(
                 ::std::forward<_ExecutionPolicy>(__exec)),
             __first1, __last1, __result,
-            oneapi::dpl::__internal::__brick_copy<__hetero_tag<_BackendTag>, _ExecutionPolicy>{});
+            oneapi::dpl::__internal::__brick_copy<__hetero_tag<_BackendTag>>{});
     }
 
     typedef typename ::std::iterator_traits<_OutputIterator>::value_type _ValueType;
@@ -1907,7 +1907,7 @@ __pattern_set_symmetric_difference(__hetero_tag<_BackendTag> __tag, _ExecutionPo
             oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__set_symmetric_difference_copy_case_1>(
                 ::std::forward<_ExecutionPolicy>(__exec)),
             __first2, __last2, __result,
-            oneapi::dpl::__internal::__brick_copy<__hetero_tag<_BackendTag>, _ExecutionPolicy>{});
+            oneapi::dpl::__internal::__brick_copy<__hetero_tag<_BackendTag>>{});
     }
 
     //{2} is empty
@@ -1918,7 +1918,7 @@ __pattern_set_symmetric_difference(__hetero_tag<_BackendTag> __tag, _ExecutionPo
             oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__set_symmetric_difference_copy_case_2>(
                 ::std::forward<_ExecutionPolicy>(__exec)),
             __first1, __last1, __result,
-            oneapi::dpl::__internal::__brick_copy<__hetero_tag<_BackendTag>, _ExecutionPolicy>{});
+            oneapi::dpl::__internal::__brick_copy<__hetero_tag<_BackendTag>>{});
     }
 
     typedef typename ::std::iterator_traits<_OutputIterator>::value_type _ValueType;
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
index 227ff9761b5..0e9e4fcc7d4 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
@@ -423,7 +423,7 @@ __pattern_remove_if(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec,
 
     oneapi::dpl::__internal::__ranges::__pattern_walk_n(
         __tag, ::std::forward<_ExecutionPolicy>(__exec),
-        oneapi::dpl::__internal::__brick_copy<__hetero_tag<_BackendTag>, _ExecutionPolicy>{}, __copy_rng_truncated,
+        oneapi::dpl::__internal::__brick_copy<__hetero_tag<_BackendTag>>{}, __copy_rng_truncated,
         ::std::forward<_Range>(__rng));
 
     return __copy_last_id;
@@ -468,7 +468,7 @@ __pattern_unique(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Ra
                                                oneapi::dpl::__internal::__pstl_assign());
 
     __pattern_walk_n(__tag, ::std::forward<_ExecutionPolicy>(__exec),
-                     __brick_copy<__hetero_tag<_BackendTag>, _ExecutionPolicy>{}, res_rng,
+                     __brick_copy<__hetero_tag<_BackendTag>>{}, res_rng,
                      ::std::forward<_Range>(__rng));
     return res;
 }
@@ -506,7 +506,7 @@ __pattern_merge(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Ran
             __tag,
             oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__copy1_wrapper>(
                 ::std::forward<_ExecutionPolicy>(__exec)),
-            oneapi::dpl::__internal::__brick_copy<__hetero_tag<_BackendTag>, _ExecutionPolicy>{},
+            oneapi::dpl::__internal::__brick_copy<__hetero_tag<_BackendTag>>{},
             ::std::forward<_Range2>(__rng2), ::std::forward<_Range3>(__rng3));
     }
     else if (__n2 == 0)
@@ -515,7 +515,7 @@ __pattern_merge(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Ran
             __tag,
             oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__copy2_wrapper>(
                 ::std::forward<_ExecutionPolicy>(__exec)),
-            oneapi::dpl::__internal::__brick_copy<__hetero_tag<_BackendTag>, _ExecutionPolicy>{},
+            oneapi::dpl::__internal::__brick_copy<__hetero_tag<_BackendTag>>{},
             ::std::forward<_Range1>(__rng1), ::std::forward<_Range3>(__rng3));
     }
     else
@@ -694,7 +694,7 @@ __pattern_reduce_by_segment(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&&
 
     if (__n == 1)
     {
-        __brick_copy<__hetero_tag<_BackendTag>, _ExecutionPolicy> __copy_range{};
+        __brick_copy<__hetero_tag<_BackendTag>> __copy_range{};
 
         oneapi::dpl::__internal::__ranges::__pattern_walk_n(
             __tag, oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__copy_keys_wrapper>(__exec), __copy_range,
diff --git a/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
index a313a944125..f26e8533efa 100644
--- a/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
@@ -251,7 +251,7 @@ __pattern_adjacent_difference(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&
 
             __internal::__pattern_walk2_brick(
                 __tag, __wrapped_policy, __first, __last, __d_first,
-                __internal::__brick_copy<__hetero_tag<_BackendTag>, decltype(__wrapped_policy)>{});
+                __internal::__brick_copy<__hetero_tag<_BackendTag>>{});
 
             return __d_last;
         });

From d0acc451888125756b532d22aae786e6e16908f0 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 1 Mar 2024 12:06:17 +0100
Subject: [PATCH 492/566] struct __brick_move - remove _ExecutionPolicy from
 template params

---
 include/oneapi/dpl/pstl/algorithm_fwd.h          |  2 +-
 include/oneapi/dpl/pstl/algorithm_impl.h         | 10 +++++-----
 include/oneapi/dpl/pstl/glue_algorithm_impl.h    |  2 +-
 .../oneapi/dpl/pstl/glue_algorithm_ranges_impl.h |  2 +-
 .../dpl/pstl/hetero/algorithm_impl_hetero.h      | 16 ++++++++--------
 .../oneapi/dpl/pstl/hetero/numeric_impl_hetero.h |  2 +-
 6 files changed, 17 insertions(+), 17 deletions(-)

diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index d9e5cb46e81..5e9e7d4b4dd 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -412,7 +412,7 @@ struct __brick_copy_n;
 template <class _Tag>
 struct __brick_copy;
 
-template <class _Tag, typename _ExecutionPolicy>
+template <class _Tag>
 struct __brick_move;
 
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 85bda3e634a..4132368eaaa 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -1109,7 +1109,7 @@ struct __brick_copy
 // move
 //------------------------------------------------------------------------
 
-template <class _Tag, typename _ExecutionPolicy>
+template <class _Tag>
 struct __brick_move
 {
     static_assert(__is_host_backend_tag_v<_Tag>);
@@ -1868,7 +1868,7 @@ __pattern_rotate(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAc
 
             __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __middle,
                                           [__last, __middle](_RandomAccessIterator __b, _RandomAccessIterator __e) {
-                                              __internal::__brick_move<__parallel_tag<_IsVector>, _ExecutionPolicy>{}(
+                                              __internal::__brick_move<__parallel_tag<_IsVector>>{}(
                                                   __b, __e, __b + (__last - __middle), _IsVector{});
                                           });
 
@@ -1894,7 +1894,7 @@ __pattern_rotate(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAc
 
             __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __middle, __last,
                                           [__first, __middle](_RandomAccessIterator __b, _RandomAccessIterator __e) {
-                                              __internal::__brick_move<__parallel_tag<_IsVector>, _ExecutionPolicy>{}(
+                                              __internal::__brick_move<__parallel_tag<_IsVector>>{}(
                                                   __b, __e, __first + (__b - __middle), _IsVector{});
                                           });
 
@@ -4277,7 +4277,7 @@ __pattern_shift_left(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Forw
     {
         __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __n, __size,
                                       [__first, __n](_DiffType __i, _DiffType __j) {
-                                          __brick_move<__parallel_tag<_IsVector>, _ExecutionPolicy>{}(
+                                          __brick_move<__parallel_tag<_IsVector>>{}(
                                               __first + __i, __first + __j, __first + __i - __n, _IsVector{});
                                       });
     }
@@ -4289,7 +4289,7 @@ __pattern_shift_left(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Forw
             auto __end = ::std::min(__k + __n, __size);
             __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __k, __end,
                                           [__first, __n](_DiffType __i, _DiffType __j) {
-                                              __brick_move<__parallel_tag<_IsVector>, _ExecutionPolicy>{}(
+                                              __brick_move<__parallel_tag<_IsVector>>{}(
                                                   __first + __i, __first + __j, __first + __i - __n, _IsVector{});
                                           });
         }
diff --git a/include/oneapi/dpl/pstl/glue_algorithm_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
index 02a1c0d71c6..35bdfac332b 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
@@ -818,7 +818,7 @@ move(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _ForwardIterator1 __l
 
     return oneapi::dpl::__internal::__pattern_walk2_brick(
         __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __d_first,
-        oneapi::dpl::__internal::__brick_move<::std::decay_t<decltype(__dispatch_tag)>, _DecayedExecutionPolicy>{});
+        oneapi::dpl::__internal::__brick_move<::std::decay_t<decltype(__dispatch_tag)>>{});
 }
 
 // [partial.sort]
diff --git a/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
index 095fb1e20ae..32f178b5153 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
@@ -612,7 +612,7 @@ move(_ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&& __rng2)
 
     oneapi::dpl::__internal::__ranges::__pattern_walk_n(
         __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
-        oneapi::dpl::__internal::__brick_move<::std::decay_t<decltype(__dispatch_tag)>, _DecayedExecutionPolicy>{},
+        oneapi::dpl::__internal::__brick_move<::std::decay_t<decltype(__dispatch_tag)>>{},
         views::all_read(::std::forward<_Range1>(__rng1)), views::all_write(::std::forward<_Range2>(__rng2)));
 }
 
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 2f28843cdfe..74756589ccb 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -397,11 +397,11 @@ struct __brick_copy<__hetero_tag<_BackendTag>>
     }
 };
 
-template <typename _BackendTag, typename _ExecutionPolicy>
-struct __brick_move<__hetero_tag<_BackendTag>, _ExecutionPolicy>
+template <typename _BackendTag>
+struct __brick_move<__hetero_tag<_BackendTag>>
 {
     template <typename _SourceT, typename _TargetT>
-    oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy>
+    void
     operator()(_SourceT&& __source, _TargetT&& __target) const
     {
         __target = ::std::move(__source);
@@ -1243,7 +1243,7 @@ __pattern_inplace_merge(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __ex
     //TODO: optimize copy back depending on Iterator, i.e. set_final_data for host iterator/pointer
     __pattern_walk2(
         __tag, __par_backend_hetero::make_wrapped_policy<copy_back_wrapper>(::std::forward<_ExecutionPolicy>(__exec)),
-        __copy_first, __copy_last, __first, __brick_move<__hetero_tag<_BackendTag>, _ExecutionPolicy>{});
+        __copy_first, __copy_last, __first, __brick_move<__hetero_tag<_BackendTag>>{});
 }
 
 //------------------------------------------------------------------------
@@ -1328,12 +1328,12 @@ __pattern_stable_partition(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& _
     //TODO: optimize copy back if possible (inplace, decrease number of submits)
     __pattern_walk2<_BackendTag, /*_IsSync=*/::std::false_type>(
         __tag, __par_backend_hetero::make_wrapped_policy<copy_back_wrapper>(__exec), __true_result, copy_result.first,
-        __first, __brick_move<__hetero_tag<_BackendTag>, _ExecutionPolicy>{});
+        __first, __brick_move<__hetero_tag<_BackendTag>>{});
 
     __pattern_walk2(
         __tag, __par_backend_hetero::make_wrapped_policy<copy_back_wrapper2>(::std::forward<_ExecutionPolicy>(__exec)),
         __false_result, copy_result.second, __first + true_count,
-        __brick_move<__hetero_tag<_BackendTag>, _ExecutionPolicy>{});
+        __brick_move<__hetero_tag<_BackendTag>>{});
 
     return __first + true_count;
 }
@@ -1643,7 +1643,7 @@ __pattern_rotate(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Iterator
         unseq_backend::__rotate_copy<typename ::std::iterator_traits<_Iterator>::difference_type>{__n, __shift}, __n,
         __buf.all_view(), __temp_rng);
 
-    using _Function = __brick_move<__hetero_tag<_BackendTag>, _ExecutionPolicy>;
+    using _Function = __brick_move<__hetero_tag<_BackendTag>>;
     auto __brick = unseq_backend::walk_n<_ExecutionPolicy, _Function>{_Function{}};
 
     oneapi::dpl::__par_backend_hetero::__parallel_for(_BackendTag{}, ::std::forward<_ExecutionPolicy>(__exec), __brick,
@@ -1973,7 +1973,7 @@ __pattern_shift_left(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Rang
     //1. n >= size/2; 'size - _n' parallel copying
     if (__n >= __mid)
     {
-        using _Function = __brick_move<__hetero_tag<_BackendTag>, _ExecutionPolicy>;
+        using _Function = __brick_move<__hetero_tag<_BackendTag>>;
         auto __brick = oneapi::dpl::unseq_backend::walk_n<_ExecutionPolicy, _Function>{_Function{}};
 
         //TODO: to consider use just "read" access mode for a source range and just "write" - for a destination range.
diff --git a/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
index f26e8533efa..336e51ecd29 100644
--- a/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
@@ -179,7 +179,7 @@ __pattern_transform_scan_base(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&
         // Move data from temporary buffer into results
         oneapi::dpl::__internal::__pattern_walk2_brick(
             __tag, ::std::move(__policy), __first_tmp, __last_tmp, __result,
-            oneapi::dpl::__internal::__brick_move<__hetero_tag<_BackendTag>, _NewExecutionPolicy>{});
+            oneapi::dpl::__internal::__brick_move<__hetero_tag<_BackendTag>>{});
 
         //TODO: optimize copy back depending on Iterator, i.e. set_final_data for host iterator/pointer
     }

From 448f8050f2463fac91de1abbe69ba7abab42f6e8 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 1 Mar 2024 12:07:42 +0100
Subject: [PATCH 493/566] struct __brick_move_destroy - remove _ExecutionPolicy
 from template params

---
 include/oneapi/dpl/pstl/algorithm_impl.h | 14 +++++++-------
 1 file changed, 7 insertions(+), 7 deletions(-)

diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 4132368eaaa..7ca9f45d763 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -1139,7 +1139,7 @@ struct __brick_move
     }
 };
 
-template <class _Tag, typename _ExecutionPolicy>
+template <class _Tag>
 struct __brick_move_destroy
 {
     static_assert(__is_host_backend_tag_v<_Tag>);
@@ -1516,7 +1516,7 @@ __remove_elements(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Forward
         // 3. Elements from result are moved to [first, last)
         __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __result,
                                       __result + __m, [__result, __first](_Tp* __i, _Tp* __j) {
-                                          __brick_move_destroy<__parallel_tag<_IsVector>, _ExecutionPolicy>{}(
+                                          __brick_move_destroy<__parallel_tag<_IsVector>>{}(
                                               __i, __j, __first + (__i - __result), _IsVector{});
                                       });
         return __first + __m;
@@ -1874,7 +1874,7 @@ __pattern_rotate(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAc
 
             __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __result,
                                           __result + (__n - __m), [__first, __result](_Tp* __b, _Tp* __e) {
-                                              __brick_move_destroy<__parallel_tag<_IsVector>, _ExecutionPolicy>{}(
+                                              __brick_move_destroy<__parallel_tag<_IsVector>>{}(
                                                   __b, __e, __first + (__b - __result), _IsVector{});
                                           });
 
@@ -1900,7 +1900,7 @@ __pattern_rotate(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAc
 
             __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __result,
                                           __result + __m, [__n, __m, __first, __result](_Tp* __b, _Tp* __e) {
-                                              __brick_move_destroy<__parallel_tag<_IsVector>, _ExecutionPolicy>{}(
+                                              __brick_move_destroy<__parallel_tag<_IsVector>>{}(
                                                   __b, __e, __first + ((__n - __m) + (__b - __result)), _IsVector{});
                                           });
 
@@ -2611,7 +2611,7 @@ __pattern_partial_sort_copy(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec
             // 3. Move elements from temporary buffer to output
             __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __r, __r + __n2,
                                           [__r, __d_first](_T1* __i, _T1* __j) {
-                                              __brick_move_destroy<__parallel_tag<_IsVector>, _ExecutionPolicy>{}(
+                                              __brick_move_destroy<__parallel_tag<_IsVector>>{}(
                                                   __i, __j, __d_first + (__i - __r), _IsVector{});
                                           });
 
@@ -3108,7 +3108,7 @@ __pattern_inplace_merge(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _R
             });
         __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __r, __r + __n,
                                       [__r, __first](_Tp* __i, _Tp* __j) {
-                                          __brick_move_destroy<__parallel_tag<_IsVector>, _ExecutionPolicy>{}(
+                                          __brick_move_destroy<__parallel_tag<_IsVector>>{}(
                                               __i, __j, __first + (__i - __r), _IsVector{});
                                       });
     });
@@ -3225,7 +3225,7 @@ __parallel_set_op(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Forward
         _DifferenceType __m{};
         auto __scan = [=](_DifferenceType, _DifferenceType, const _SetRange& __s) { // Scan
             if (!__s.empty())
-                __brick_move_destroy<__parallel_tag<_IsVector>, _ExecutionPolicy>{}(
+                __brick_move_destroy<__parallel_tag<_IsVector>>{}(
                     __tmp_memory + __s.__buf_pos, __tmp_memory + (__s.__buf_pos + __s.__len), __result + __s.__pos,
                     _IsVector{});
         };

From 4659c025820713300b5beaf04dc778cadd555c6f Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 1 Mar 2024 12:09:25 +0100
Subject: [PATCH 494/566] struct __brick_fill - remove _ExecutionPolicy from
 template params

---
 include/oneapi/dpl/pstl/algorithm_fwd.h                | 2 +-
 include/oneapi/dpl/pstl/algorithm_impl.h               | 6 +++---
 include/oneapi/dpl/pstl/glue_memory_impl.h             | 6 ++----
 include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h | 6 +++---
 4 files changed, 9 insertions(+), 11 deletions(-)

diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index 5e9e7d4b4dd..b1e592805c3 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -852,7 +852,7 @@ __pattern_nth_element(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAcce
 //------------------------------------------------------------------------
 // fill, fill_n
 //------------------------------------------------------------------------
-template <class _Tag, typename _Tp, typename _ExecutionPolicy>
+template <class _Tag, typename _Tp>
 struct __brick_fill;
 
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Tp>
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 7ca9f45d763..bf5aa7a8b02 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -2758,7 +2758,7 @@ __pattern_nth_element(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec
 //------------------------------------------------------------------------
 // fill, fill_n
 //------------------------------------------------------------------------
-template <class _Tag, typename _Tp, typename _ExecutionPolicy>
+template <class _Tag, typename _Tp>
 struct __brick_fill
 {
     static_assert(__is_host_backend_tag_v<_Tag>);
@@ -2788,7 +2788,7 @@ __pattern_fill(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardItera
 {
     static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
-    __internal::__brick_fill<_Tag, _Tp, _ExecutionPolicy>{__value}(__first, __last, typename _Tag::__is_vector{});
+    __internal::__brick_fill<_Tag, _Tp>{__value}(__first, __last, typename _Tag::__is_vector{});
 }
 
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Tp>
@@ -2801,7 +2801,7 @@ __pattern_fill(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAcce
     return __internal::__except_handler([&__exec, __first, __last, &__value]() {
         __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
                                       [&__value](_RandomAccessIterator __begin, _RandomAccessIterator __end) {
-                                          __internal::__brick_fill<__parallel_tag<_IsVector>, _Tp, _ExecutionPolicy>{
+                                          __internal::__brick_fill<__parallel_tag<_IsVector>, _Tp>{
                                               __value}(__begin, __end, _IsVector{});
                                       });
         return __last;
diff --git a/include/oneapi/dpl/pstl/glue_memory_impl.h b/include/oneapi/dpl/pstl/glue_memory_impl.h
index 860057753eb..243ef51d1b1 100644
--- a/include/oneapi/dpl/pstl/glue_memory_impl.h
+++ b/include/oneapi/dpl/pstl/glue_memory_impl.h
@@ -150,8 +150,7 @@ uninitialized_fill(_ExecutionPolicy&& __exec, _ForwardIterator __first, _Forward
     {
         oneapi::dpl::__internal::__pattern_walk_brick(
             __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
-            oneapi::dpl::__internal::__brick_fill<::std::decay_t<decltype(__dispatch_tag)>, _ValueType,
-                                                  _DecayedExecutionPolicy>{
+            oneapi::dpl::__internal::__brick_fill<::std::decay_t<decltype(__dispatch_tag)>, _ValueType>{
                 _ValueType(__value)});
     }
     else
@@ -313,8 +312,7 @@ uninitialized_value_construct(_ExecutionPolicy&& __exec, _ForwardIterator __firs
     {
         oneapi::dpl::__internal::__pattern_walk_brick(
             __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
-            oneapi::dpl::__internal::__brick_fill<::std::decay_t<decltype(__dispatch_tag)>, _ValueType,
-                                                  _DecayedExecutionPolicy>{_ValueType()});
+            oneapi::dpl::__internal::__brick_fill<::std::decay_t<decltype(__dispatch_tag)>, _ValueType>{_ValueType()});
     }
     else
     {
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 74756589ccb..80ee79587bc 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -408,12 +408,12 @@ struct __brick_move<__hetero_tag<_BackendTag>>
     }
 };
 
-template <typename _BackendTag, typename _SourceT, typename _ExecutionPolicy>
-struct __brick_fill<__hetero_tag<_BackendTag>, _SourceT, _ExecutionPolicy>
+template <typename _BackendTag, typename _SourceT>
+struct __brick_fill<__hetero_tag<_BackendTag>, _SourceT>
 {
     _SourceT __value;
     template <typename _TargetT>
-    oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy>
+    void
     operator()(_TargetT& __target) const
     {
         __target = __value;

From d5f85bc2b4c8d974f957127b53f469f21440a457 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 1 Mar 2024 12:12:51 +0100
Subject: [PATCH 495/566] struct __brick_fill_n - remove _ExecutionPolicy from
 template params

---
 include/oneapi/dpl/pstl/algorithm_fwd.h                | 2 +-
 include/oneapi/dpl/pstl/algorithm_impl.h               | 5 ++---
 include/oneapi/dpl/pstl/glue_memory_impl.h             | 8 ++++----
 include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h | 6 +++---
 4 files changed, 10 insertions(+), 11 deletions(-)

diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index b1e592805c3..76aebfef030 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -863,7 +863,7 @@ template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator,
 _RandomAccessIterator
 __pattern_fill(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, const _Tp&);
 
-template <class _Tag, typename _Tp, typename _ExecutionPolicy>
+template <class _Tag, typename _Tp>
 struct __brick_fill_n;
 
 template <class _Tag, class _ExecutionPolicy, class _OutputIterator, class _Size, class _Tp>
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index bf5aa7a8b02..14d9bcf915a 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -2808,7 +2808,7 @@ __pattern_fill(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAcce
     });
 }
 
-template <class _Tag, typename _Tp, typename _ExecutionPolicy>
+template <class _Tag, typename _Tp>
 struct __brick_fill_n
 {
     static_assert(__is_host_backend_tag_v<_Tag>);
@@ -2838,8 +2838,7 @@ __pattern_fill_n(_Tag, _ExecutionPolicy&&, _OutputIterator __first, _Size __coun
 {
     static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
-    return __internal::__brick_fill_n<_Tag, _Tp, _ExecutionPolicy>{__value}(__first, __count,
-                                                                            typename _Tag::__is_vector{});
+    return __internal::__brick_fill_n<_Tag, _Tp>{__value}(__first, __count, typename _Tag::__is_vector{});
 }
 
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Size, class _Tp>
diff --git a/include/oneapi/dpl/pstl/glue_memory_impl.h b/include/oneapi/dpl/pstl/glue_memory_impl.h
index 243ef51d1b1..edd93953626 100644
--- a/include/oneapi/dpl/pstl/glue_memory_impl.h
+++ b/include/oneapi/dpl/pstl/glue_memory_impl.h
@@ -174,8 +174,8 @@ uninitialized_fill_n(_ExecutionPolicy&& __exec, _ForwardIterator __first, _Size
     {
         return oneapi::dpl::__internal::__pattern_walk_brick_n(
             __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __n,
-            oneapi::dpl::__internal::__brick_fill_n<::std::decay_t<decltype(__dispatch_tag)>, _ValueType,
-                                                    _DecayedExecutionPolicy>{_ValueType(__value)});
+            oneapi::dpl::__internal::__brick_fill_n<::std::decay_t<decltype(__dispatch_tag)>, _ValueType>{
+                _ValueType(__value)});
     }
     else
     {
@@ -335,8 +335,8 @@ uninitialized_value_construct_n(_ExecutionPolicy&& __exec, _ForwardIterator __fi
     {
         return oneapi::dpl::__internal::__pattern_walk_brick_n(
             __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __n,
-            oneapi::dpl::__internal::__brick_fill_n<::std::decay_t<decltype(__dispatch_tag)>, _ValueType,
-                                                    _DecayedExecutionPolicy>{_ValueType()});
+            oneapi::dpl::__internal::__brick_fill_n<::std::decay_t<decltype(__dispatch_tag)>, _ValueType>{
+                _ValueType()});
     }
     else
     {
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 80ee79587bc..e4cfdee6cda 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -420,12 +420,12 @@ struct __brick_fill<__hetero_tag<_BackendTag>, _SourceT>
     }
 };
 
-template <typename _BackendTag, typename _SourceT, typename _ExecutionPolicy>
-struct __brick_fill_n<__hetero_tag<_BackendTag>, _SourceT, _ExecutionPolicy>
+template <typename _BackendTag, typename _SourceT>
+struct __brick_fill_n<__hetero_tag<_BackendTag>, _SourceT>
 {
     _SourceT __value;
     template <typename _TargetT>
-    oneapi::dpl::__internal::__enable_if_hetero_execution_policy<_ExecutionPolicy>
+    void
     operator()(_TargetT& __target) const
     {
         __target = __value;

From 0df1a0e54fba7d780a6440ef33df65fd3cd467ec Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 1 Mar 2024 12:18:37 +0100
Subject: [PATCH 496/566] include/oneapi/dpl/pstl/algorithm_impl.h - remove
 __is_vector param from host implementation of __brick_fill_n::operator()

---
 include/oneapi/dpl/pstl/algorithm_impl.h | 26 +++++++++++-------------
 1 file changed, 12 insertions(+), 14 deletions(-)

diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 14d9bcf915a..3bd612c5f63 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -242,7 +242,7 @@ __pattern_walk_brick_n(_Tag, _ExecutionPolicy&& __exec, _ForwardIterator __first
 {
     static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
-    return __brick(__first, __n, typename _Tag::__is_vector{});
+    return __brick(__first, __n,);
 }
 
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Size, class _Brick>
@@ -255,7 +255,7 @@ __pattern_walk_brick_n(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Ra
     return __internal::__except_handler([&]() {
         __par_backend::__parallel_for(
             __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __first + __n,
-            [__brick](_RandomAccessIterator __i, _RandomAccessIterator __j) { __brick(__i, __j - __i, _IsVector{}); });
+            [__brick](_RandomAccessIterator __i, _RandomAccessIterator __j) { __brick(__i, __j - __i); });
         return __first + __n;
     });
 }
@@ -2817,18 +2817,16 @@ struct __brick_fill_n
 
     template <typename _RandomAccessIterator, typename _Size>
     _RandomAccessIterator
-    operator()(_RandomAccessIterator __first, _Size __count,
-               /* __is_vector = */ ::std::true_type) const noexcept
-    {
-        return __unseq_backend::__simd_fill_n(__first, __count, __value);
-    }
-
-    template <typename _OutputIterator, typename _Size>
-    _OutputIterator
-    operator()(_OutputIterator __first, _Size __count,
-               /* __is_vector = */ ::std::false_type) const noexcept
+    operator()(_RandomAccessIterator __first, _Size __count) const noexcept
     {
-        return ::std::fill_n(__first, __count, __value);
+        if constexpr (typename _Tag::__is_vector{})
+        {
+            return __unseq_backend::__simd_fill_n(__first, __count, __value);
+        }
+        else
+        {
+            return ::std::fill_n(__first, __count, __value);
+        }
     }
 };
 
@@ -2838,7 +2836,7 @@ __pattern_fill_n(_Tag, _ExecutionPolicy&&, _OutputIterator __first, _Size __coun
 {
     static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
-    return __internal::__brick_fill_n<_Tag, _Tp>{__value}(__first, __count, typename _Tag::__is_vector{});
+    return __internal::__brick_fill_n<_Tag, _Tp>{__value}(__first, __count);
 }
 
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Size, class _Tp>

From fdbb0099548c3b7bac08ba8e966e4ccd1c583d9c Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 1 Mar 2024 12:19:54 +0100
Subject: [PATCH 497/566] Revert "include/oneapi/dpl/pstl/algorithm_impl.h -
 remove __is_vector param from host implementation of
 __brick_fill_n::operator()"

This reverts commit 0df1a0e54fba7d780a6440ef33df65fd3cd467ec.
---
 include/oneapi/dpl/pstl/algorithm_impl.h | 26 +++++++++++++-----------
 1 file changed, 14 insertions(+), 12 deletions(-)

diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 3bd612c5f63..14d9bcf915a 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -242,7 +242,7 @@ __pattern_walk_brick_n(_Tag, _ExecutionPolicy&& __exec, _ForwardIterator __first
 {
     static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
-    return __brick(__first, __n,);
+    return __brick(__first, __n, typename _Tag::__is_vector{});
 }
 
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Size, class _Brick>
@@ -255,7 +255,7 @@ __pattern_walk_brick_n(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Ra
     return __internal::__except_handler([&]() {
         __par_backend::__parallel_for(
             __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __first + __n,
-            [__brick](_RandomAccessIterator __i, _RandomAccessIterator __j) { __brick(__i, __j - __i); });
+            [__brick](_RandomAccessIterator __i, _RandomAccessIterator __j) { __brick(__i, __j - __i, _IsVector{}); });
         return __first + __n;
     });
 }
@@ -2817,16 +2817,18 @@ struct __brick_fill_n
 
     template <typename _RandomAccessIterator, typename _Size>
     _RandomAccessIterator
-    operator()(_RandomAccessIterator __first, _Size __count) const noexcept
+    operator()(_RandomAccessIterator __first, _Size __count,
+               /* __is_vector = */ ::std::true_type) const noexcept
     {
-        if constexpr (typename _Tag::__is_vector{})
-        {
-            return __unseq_backend::__simd_fill_n(__first, __count, __value);
-        }
-        else
-        {
-            return ::std::fill_n(__first, __count, __value);
-        }
+        return __unseq_backend::__simd_fill_n(__first, __count, __value);
+    }
+
+    template <typename _OutputIterator, typename _Size>
+    _OutputIterator
+    operator()(_OutputIterator __first, _Size __count,
+               /* __is_vector = */ ::std::false_type) const noexcept
+    {
+        return ::std::fill_n(__first, __count, __value);
     }
 };
 
@@ -2836,7 +2838,7 @@ __pattern_fill_n(_Tag, _ExecutionPolicy&&, _OutputIterator __first, _Size __coun
 {
     static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
-    return __internal::__brick_fill_n<_Tag, _Tp>{__value}(__first, __count);
+    return __internal::__brick_fill_n<_Tag, _Tp>{__value}(__first, __count, typename _Tag::__is_vector{});
 }
 
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Size, class _Tp>

From f1dab7f730c11ebe48d3a90890b8778344d85142 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 1 Mar 2024 13:13:57 +0100
Subject: [PATCH 498/566] Apply GitHUB clang format

---
 include/oneapi/dpl/pstl/algorithm_fwd.h       |  3 ++-
 include/oneapi/dpl/pstl/algorithm_impl.h      | 22 ++++++++---------
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 24 +++++++------------
 .../hetero/algorithm_ranges_impl_hetero.h     | 13 +++++-----
 .../dpl/pstl/hetero/numeric_impl_hetero.h     |  5 ++--
 5 files changed, 29 insertions(+), 38 deletions(-)

diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index 76aebfef030..56dc6b54d6a 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -633,7 +633,8 @@ __pattern_rotate(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIte
 //------------------------------------------------------------------------
 
 template <class _Tag, class _ForwardIterator, class _OutputIterator>
-_OutputIterator __brick_rotate_copy(_Tag, _ForwardIterator, _ForwardIterator, _ForwardIterator, _OutputIterator) noexcept;
+_OutputIterator __brick_rotate_copy(_Tag, _ForwardIterator, _ForwardIterator, _ForwardIterator,
+                                    _OutputIterator) noexcept;
 
 template <class _IsVector, class _RandomAccessIterator, class _OutputIterator>
 _OutputIterator __brick_rotate_copy(__parallel_tag<_IsVector>, _RandomAccessIterator, _RandomAccessIterator,
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 14d9bcf915a..04410b3998d 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -1915,8 +1915,8 @@ __pattern_rotate(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAc
 
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _OutputIterator>
 _OutputIterator
-__brick_rotate_copy(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __middle, _ForwardIterator __last,
-                    _OutputIterator __result) noexcept
+__brick_rotate_copy(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __middle,
+                    _ForwardIterator __last, _OutputIterator __result) noexcept
 {
     static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
@@ -1926,8 +1926,8 @@ __brick_rotate_copy(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _Forward
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2>
 _RandomAccessIterator2
 __brick_rotate_copy(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator1 __first,
-                    _RandomAccessIterator1 __middle,
-                    _RandomAccessIterator1 __last, _RandomAccessIterator2 __result) noexcept
+                    _RandomAccessIterator1 __middle, _RandomAccessIterator1 __last,
+                    _RandomAccessIterator2 __result) noexcept
 {
     _RandomAccessIterator2 __res = __brick_copy<__parallel_tag<_IsVector>>{}(__middle, __last, __result);
     return __internal::__brick_copy<__parallel_tag<_IsVector>>{}(__first, __middle, __res);
@@ -2801,8 +2801,8 @@ __pattern_fill(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAcce
     return __internal::__except_handler([&__exec, __first, __last, &__value]() {
         __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
                                       [&__value](_RandomAccessIterator __begin, _RandomAccessIterator __end) {
-                                          __internal::__brick_fill<__parallel_tag<_IsVector>, _Tp>{
-                                              __value}(__begin, __end, _IsVector{});
+                                          __internal::__brick_fill<__parallel_tag<_IsVector>, _Tp>{__value}(
+                                              __begin, __end, _IsVector{});
                                       });
         return __last;
     });
@@ -3224,9 +3224,9 @@ __parallel_set_op(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Forward
         _DifferenceType __m{};
         auto __scan = [=](_DifferenceType, _DifferenceType, const _SetRange& __s) { // Scan
             if (!__s.empty())
-                __brick_move_destroy<__parallel_tag<_IsVector>>{}(
-                    __tmp_memory + __s.__buf_pos, __tmp_memory + (__s.__buf_pos + __s.__len), __result + __s.__pos,
-                    _IsVector{});
+                __brick_move_destroy<__parallel_tag<_IsVector>>{}(__tmp_memory + __s.__buf_pos,
+                                                                  __tmp_memory + (__s.__buf_pos + __s.__len),
+                                                                  __result + __s.__pos, _IsVector{});
         };
         __par_backend::__parallel_strict_scan(
             __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __n1, _SetRange{0, 0, 0}, //-1, 0},
@@ -4276,8 +4276,8 @@ __pattern_shift_left(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Forw
     {
         __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __n, __size,
                                       [__first, __n](_DiffType __i, _DiffType __j) {
-                                          __brick_move<__parallel_tag<_IsVector>>{}(
-                                              __first + __i, __first + __j, __first + __i - __n, _IsVector{});
+                                          __brick_move<__parallel_tag<_IsVector>>{}(__first + __i, __first + __j,
+                                                                                    __first + __i - __n, _IsVector{});
                                       });
     }
     else //2. n < size/2; there is not enough memory to parallel copying; doing parallel copying by n elements
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index e4cfdee6cda..6a42ad15792 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -1186,15 +1186,13 @@ __pattern_merge(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Ite
             __tag,
             oneapi::dpl::__par_backend_hetero::make_wrapped_policy<copy_back_wrapper>(
                 ::std::forward<_ExecutionPolicy>(__exec)),
-            __first2, __last2, __d_first,
-            oneapi::dpl::__internal::__brick_copy<__hetero_tag<_BackendTag>>{});
+            __first2, __last2, __d_first, oneapi::dpl::__internal::__brick_copy<__hetero_tag<_BackendTag>>{});
     else if (__n2 == 0)
         oneapi::dpl::__internal::__pattern_walk2_brick(
             __tag,
             oneapi::dpl::__par_backend_hetero::make_wrapped_policy<copy_back_wrapper2>(
                 ::std::forward<_ExecutionPolicy>(__exec)),
-            __first1, __last1, __d_first,
-            oneapi::dpl::__internal::__brick_copy<__hetero_tag<_BackendTag>>{});
+            __first1, __last1, __d_first, oneapi::dpl::__internal::__brick_copy<__hetero_tag<_BackendTag>>{});
     else
     {
         auto __keep1 = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _Iterator1>();
@@ -1332,8 +1330,7 @@ __pattern_stable_partition(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& _
 
     __pattern_walk2(
         __tag, __par_backend_hetero::make_wrapped_policy<copy_back_wrapper2>(::std::forward<_ExecutionPolicy>(__exec)),
-        __false_result, copy_result.second, __first + true_count,
-        __brick_move<__hetero_tag<_BackendTag>>{});
+        __false_result, copy_result.second, __first + true_count, __brick_move<__hetero_tag<_BackendTag>>{});
 
     return __first + true_count;
 }
@@ -1789,8 +1786,7 @@ __pattern_set_difference(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __e
             __tag,
             oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__set_difference_copy_case_1>(
                 ::std::forward<_ExecutionPolicy>(__exec)),
-            __first1, __last1, __result,
-            oneapi::dpl::__internal::__brick_copy<__hetero_tag<_BackendTag>>{});
+            __first1, __last1, __result, oneapi::dpl::__internal::__brick_copy<__hetero_tag<_BackendTag>>{});
     }
 
     return __pattern_hetero_set_op(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2,
@@ -1825,8 +1821,7 @@ __pattern_set_union(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec,
             __tag,
             oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__set_union_copy_case_1>(
                 ::std::forward<_ExecutionPolicy>(__exec)),
-            __first2, __last2, __result,
-            oneapi::dpl::__internal::__brick_copy<__hetero_tag<_BackendTag>>{});
+            __first2, __last2, __result, oneapi::dpl::__internal::__brick_copy<__hetero_tag<_BackendTag>>{});
     }
 
     //{2} is empty
@@ -1836,8 +1831,7 @@ __pattern_set_union(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec,
             __tag,
             oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__set_union_copy_case_2>(
                 ::std::forward<_ExecutionPolicy>(__exec)),
-            __first1, __last1, __result,
-            oneapi::dpl::__internal::__brick_copy<__hetero_tag<_BackendTag>>{});
+            __first1, __last1, __result, oneapi::dpl::__internal::__brick_copy<__hetero_tag<_BackendTag>>{});
     }
 
     typedef typename ::std::iterator_traits<_OutputIterator>::value_type _ValueType;
@@ -1906,8 +1900,7 @@ __pattern_set_symmetric_difference(__hetero_tag<_BackendTag> __tag, _ExecutionPo
             __tag,
             oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__set_symmetric_difference_copy_case_1>(
                 ::std::forward<_ExecutionPolicy>(__exec)),
-            __first2, __last2, __result,
-            oneapi::dpl::__internal::__brick_copy<__hetero_tag<_BackendTag>>{});
+            __first2, __last2, __result, oneapi::dpl::__internal::__brick_copy<__hetero_tag<_BackendTag>>{});
     }
 
     //{2} is empty
@@ -1917,8 +1910,7 @@ __pattern_set_symmetric_difference(__hetero_tag<_BackendTag> __tag, _ExecutionPo
             __tag,
             oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__set_symmetric_difference_copy_case_2>(
                 ::std::forward<_ExecutionPolicy>(__exec)),
-            __first1, __last1, __result,
-            oneapi::dpl::__internal::__brick_copy<__hetero_tag<_BackendTag>>{});
+            __first1, __last1, __result, oneapi::dpl::__internal::__brick_copy<__hetero_tag<_BackendTag>>{});
     }
 
     typedef typename ::std::iterator_traits<_OutputIterator>::value_type _ValueType;
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
index 0e9e4fcc7d4..45863f75615 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
@@ -467,9 +467,8 @@ __pattern_unique(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Ra
     auto res = __ranges::__pattern_unique_copy(__tag, __exec, __rng, res_rng, __pred,
                                                oneapi::dpl::__internal::__pstl_assign());
 
-    __pattern_walk_n(__tag, ::std::forward<_ExecutionPolicy>(__exec),
-                     __brick_copy<__hetero_tag<_BackendTag>>{}, res_rng,
-                     ::std::forward<_Range>(__rng));
+    __pattern_walk_n(__tag, ::std::forward<_ExecutionPolicy>(__exec), __brick_copy<__hetero_tag<_BackendTag>>{},
+                     res_rng, ::std::forward<_Range>(__rng));
     return res;
 }
 
@@ -506,8 +505,8 @@ __pattern_merge(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Ran
             __tag,
             oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__copy1_wrapper>(
                 ::std::forward<_ExecutionPolicy>(__exec)),
-            oneapi::dpl::__internal::__brick_copy<__hetero_tag<_BackendTag>>{},
-            ::std::forward<_Range2>(__rng2), ::std::forward<_Range3>(__rng3));
+            oneapi::dpl::__internal::__brick_copy<__hetero_tag<_BackendTag>>{}, ::std::forward<_Range2>(__rng2),
+            ::std::forward<_Range3>(__rng3));
     }
     else if (__n2 == 0)
     {
@@ -515,8 +514,8 @@ __pattern_merge(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Ran
             __tag,
             oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__copy2_wrapper>(
                 ::std::forward<_ExecutionPolicy>(__exec)),
-            oneapi::dpl::__internal::__brick_copy<__hetero_tag<_BackendTag>>{},
-            ::std::forward<_Range1>(__rng1), ::std::forward<_Range3>(__rng3));
+            oneapi::dpl::__internal::__brick_copy<__hetero_tag<_BackendTag>>{}, ::std::forward<_Range1>(__rng1),
+            ::std::forward<_Range3>(__rng3));
     }
     else
     {
diff --git a/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
index 336e51ecd29..216c84820f2 100644
--- a/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
@@ -249,9 +249,8 @@ __pattern_adjacent_difference(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&
             auto __wrapped_policy = __par_backend_hetero::make_wrapped_policy<adjacent_difference_wrapper>(
                 ::std::forward<_ExecutionPolicy>(__exec));
 
-            __internal::__pattern_walk2_brick(
-                __tag, __wrapped_policy, __first, __last, __d_first,
-                __internal::__brick_copy<__hetero_tag<_BackendTag>>{});
+            __internal::__pattern_walk2_brick(__tag, __wrapped_policy, __first, __last, __d_first,
+                                              __internal::__brick_copy<__hetero_tag<_BackendTag>>{});
 
             return __d_last;
         });

From 4682a90909cbc52170827826226a907f0d5b2d6d Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 1 Mar 2024 13:30:17 +0100
Subject: [PATCH 499/566] Fix review comment: It's very weird when the trait
 and _v shortcut might produce different values.

---
 .../oneapi/dpl/internal/binary_search_impl.h  |   6 +-
 .../internal/exclusive_scan_by_segment_impl.h |   2 +-
 .../internal/inclusive_scan_by_segment_impl.h |   2 +-
 .../dpl/internal/reduce_by_segment_impl.h     |   2 +-
 include/oneapi/dpl/pstl/algorithm_impl.h      | 132 +++++++++---------
 include/oneapi/dpl/pstl/execution_impl.h      |   7 +-
 .../experimental/internal/for_loop_impl.h     |   8 +-
 include/oneapi/dpl/pstl/histogram_impl.h      |   2 +-
 include/oneapi/dpl/pstl/numeric_impl.h        |  10 +-
 9 files changed, 85 insertions(+), 86 deletions(-)

diff --git a/include/oneapi/dpl/internal/binary_search_impl.h b/include/oneapi/dpl/internal/binary_search_impl.h
index 4a2d7be1391..92ffd046be9 100644
--- a/include/oneapi/dpl/internal/binary_search_impl.h
+++ b/include/oneapi/dpl/internal/binary_search_impl.h
@@ -74,7 +74,7 @@ OutputIterator
 lower_bound_impl(_Tag, Policy&& policy, InputIterator1 start, InputIterator1 end, InputIterator2 value_start,
                  InputIterator2 value_end, OutputIterator result, StrictWeakOrdering comp)
 {
-    static_assert(__internal::__is_host_backend_tag_v<_Tag>);
+    static_assert(__internal::__is_host_backend_tag_v<::std::decay_t<_Tag>>);
 
     return oneapi::dpl::transform(policy, value_start, value_end, result,
                                   [=](typename ::std::iterator_traits<InputIterator2>::reference val) {
@@ -88,7 +88,7 @@ OutputIterator
 upper_bound_impl(_Tag, Policy&& policy, InputIterator1 start, InputIterator1 end, InputIterator2 value_start,
                  InputIterator2 value_end, OutputIterator result, StrictWeakOrdering comp)
 {
-    static_assert(__internal::__is_host_backend_tag_v<_Tag>);
+    static_assert(__internal::__is_host_backend_tag_v<::std::decay_t<_Tag>>);
 
     return oneapi::dpl::transform(policy, value_start, value_end, result,
                                   [=](typename ::std::iterator_traits<InputIterator2>::reference val) {
@@ -102,7 +102,7 @@ OutputIterator
 binary_search_impl(_Tag, Policy&& policy, InputIterator1 start, InputIterator1 end, InputIterator2 value_start,
                    InputIterator2 value_end, OutputIterator result, StrictWeakOrdering comp)
 {
-    static_assert(__internal::__is_host_backend_tag_v<_Tag>);
+    static_assert(__internal::__is_host_backend_tag_v<::std::decay_t<_Tag>>);
 
     return oneapi::dpl::transform(policy, value_start, value_end, result,
                                   [=](typename ::std::iterator_traits<InputIterator2>::reference val) {
diff --git a/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h b/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h
index 211046642ce..a2c0eaebc4b 100644
--- a/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h
@@ -44,7 +44,7 @@ pattern_exclusive_scan_by_segment(_Tag, Policy&& policy, InputIterator1 first1,
                                   InputIterator2 first2, OutputIterator result, T init, BinaryPredicate binary_pred,
                                   Operator binary_op)
 {
-    static_assert(__internal::__is_host_backend_tag_v<_Tag>);
+    static_assert(__internal::__is_host_backend_tag_v<::std::decay_t<_Tag>>);
 
     using __backend_tag = typename _Tag::__backend_tag;
 
diff --git a/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h b/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h
index 95dc8fa62db..f4d2b8d6a33 100644
--- a/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h
@@ -43,7 +43,7 @@ pattern_inclusive_scan_by_segment(_Tag, Policy&& policy, InputIterator1 first1,
                                   InputIterator2 first2, OutputIterator result, BinaryPredicate binary_pred,
                                   BinaryOperator binary_op)
 {
-    static_assert(__internal::__is_host_backend_tag_v<_Tag>);
+    static_assert(__internal::__is_host_backend_tag_v<::std::decay_t<_Tag>>);
 
     using __backend_tag = typename _Tag::__backend_tag;
 
diff --git a/include/oneapi/dpl/internal/reduce_by_segment_impl.h b/include/oneapi/dpl/internal/reduce_by_segment_impl.h
index 5cda04d6127..37a7aad1379 100644
--- a/include/oneapi/dpl/internal/reduce_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/reduce_by_segment_impl.h
@@ -85,7 +85,7 @@ reduce_by_segment_impl(_Tag, Policy&& policy, InputIterator1 first1, InputIterat
                        OutputIterator1 result1, OutputIterator2 result2, BinaryPred binary_pred,
                        BinaryOperator binary_op)
 {
-    static_assert(__internal::__is_host_backend_tag_v<_Tag>);
+    static_assert(__internal::__is_host_backend_tag_v<::std::decay_t<_Tag>>);
 
     using __backend_tag = typename _Tag::__backend_tag;
 
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 04410b3998d..ca36a8429e1 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -67,7 +67,7 @@ template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Pre
 bool
 __pattern_any_of(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _Pred __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_any_of(__first, __last, __pred, typename _Tag::__is_vector{});
 }
@@ -139,7 +139,7 @@ template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Fun
 void
 __pattern_walk1(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _Function __f) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>>);
 
     __internal::__brick_walk1(__first, __last, __f, typename _Tag::__is_vector{});
 }
@@ -179,7 +179,7 @@ void
 __pattern_walk_brick(_Tag, _ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last,
                      _Brick __brick) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     __brick(__first, __last, typename _Tag::__is_vector{});
 }
@@ -221,7 +221,7 @@ template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Siz
 _ForwardIterator
 __pattern_walk1_n(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _Size __n, _Function __f) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_walk1_n(__first, __n, __f, typename _Tag::__is_vector{});
 }
@@ -240,7 +240,7 @@ template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Siz
 _ForwardIterator
 __pattern_walk_brick_n(_Tag, _ExecutionPolicy&& __exec, _ForwardIterator __first, _Size __n, _Brick __brick) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __brick(__first, __n, typename _Tag::__is_vector{});
 }
@@ -307,7 +307,7 @@ _ForwardIterator2
 __pattern_walk2(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
                 _ForwardIterator2 __first2, _Function __f) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_walk2(__first1, __last1, __first2, __f, typename _Tag::__is_vector{});
 }
@@ -364,7 +364,7 @@ _ForwardIterator2
 __pattern_walk2_n(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _Size __n, _ForwardIterator2 __first2,
                   _Function __f) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_walk2_n(__first1, __n, __first2, __f, typename _Tag::__is_vector{});
 }
@@ -384,7 +384,7 @@ _ForwardIterator2
 __pattern_walk2_brick(_Tag, _ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
                       _ForwardIterator2 __first2, _Brick __brick) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>>);
 
     return __brick(__first1, __last1, __first2, typename _Tag::__is_vector{});
 }
@@ -461,7 +461,7 @@ _ForwardIterator2
 __pattern_walk2_brick_n(_Tag, _ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _Size __n,
                         _ForwardIterator2 __first2, _Brick __brick) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __brick(__first1, __n, __first2, typename _Tag::__is_vector{});
 }
@@ -495,7 +495,7 @@ _ForwardIterator3
 __pattern_walk3(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
                 _ForwardIterator2 __first2, _ForwardIterator3 __first3, _Function __f) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_walk3(__first1, __last1, __first2, __first3, __f, typename _Tag::__is_vector{});
 }
@@ -561,7 +561,7 @@ _ForwardIterator2
 __pattern_walk2_transform_if(_Tag __tag, _ExecutionPolicy&& __exec, _ForwardIterator1 __first1,
                              _ForwardIterator1 __last1, _ForwardIterator2 __first2, _Function __func) noexcept
 {
-    static_assert(__is_host_backend_tag_v<_Tag>);
+    static_assert(__is_host_backend_tag_v<::std::decay_t<_Tag>>);
 
     return __pattern_walk2(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __func);
 }
@@ -573,7 +573,7 @@ __pattern_walk3_transform_if(_Tag __tag, _ExecutionPolicy&& __exec, _ForwardIter
                              _ForwardIterator1 __last1, _ForwardIterator2 __first2, _ForwardIterator3 __first3,
                              _Function __func) noexcept
 {
-    static_assert(__is_host_backend_tag_v<_Tag>);
+    static_assert(__is_host_backend_tag_v<::std::decay_t<_Tag>>);
 
     return __pattern_walk3(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __first3,
                            __func);
@@ -608,7 +608,7 @@ bool
 __pattern_equal(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
                 _ForwardIterator2 __first2, _ForwardIterator2 __last2, _BinaryPredicate __p) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_equal(__first1, __last1, __first2, __last2, __p, typename _Tag::__is_vector{});
 }
@@ -659,7 +659,7 @@ bool
 __pattern_equal(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
                 _ForwardIterator2 __first2, _BinaryPredicate __p) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_equal(__first1, __last1, __first2, __p, typename _Tag::__is_vector{});
 }
@@ -706,7 +706,7 @@ _ForwardIterator
 __pattern_find_if(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                   _Predicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_find_if(__first, __last, __pred, typename _Tag::__is_vector{});
 }
@@ -849,7 +849,7 @@ _ForwardIterator1
 __pattern_find_end(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first, _ForwardIterator1 __last,
                    _ForwardIterator2 __s_first, _ForwardIterator2 __s_last, _BinaryPredicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_find_end(__first, __last, __s_first, __s_last, __pred, typename _Tag::__is_vector{});
 }
@@ -905,7 +905,7 @@ _ForwardIterator1
 __pattern_find_first_of(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first, _ForwardIterator1 __last,
                         _ForwardIterator2 __s_first, _ForwardIterator2 __s_last, _BinaryPredicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_find_first_of(__first, __last, __s_first, __s_last, __pred,
                                              typename _Tag::__is_vector{});
@@ -954,7 +954,7 @@ _ForwardIterator1
 __pattern_search(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first, _ForwardIterator1 __last,
                  _ForwardIterator2 __s_first, _ForwardIterator2 __s_last, _BinaryPredicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_search(__first, __last, __s_first, __s_last, __pred, typename _Tag::__is_vector{});
 }
@@ -1010,7 +1010,7 @@ _ForwardIterator
 __pattern_search_n(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _Size __count,
                    const _Tp& __value, _BinaryPredicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_search_n(__first, __last, __count, __value, __pred, typename _Tag::__is_vector{});
 }
@@ -1051,7 +1051,7 @@ __pattern_search_n(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _
 template <class _Tag>
 struct __brick_copy_n
 {
-    static_assert(__is_host_backend_tag_v<_Tag>);
+    static_assert(__is_host_backend_tag_v<::std::decay_t<_Tag>>);
 
     template <typename _RandomAccessIterator1, typename _Size, typename _RandomAccessIterator2>
     _RandomAccessIterator2
@@ -1078,7 +1078,7 @@ struct __brick_copy_n
 template <class _Tag>
 struct __brick_copy
 {
-    static_assert(__is_host_backend_tag_v<_Tag>);
+    static_assert(__is_host_backend_tag_v<::std::decay_t<_Tag>>);
 
     template <typename _RandomAccessIterator1, typename _RandomAccessIterator2>
     _RandomAccessIterator2
@@ -1112,7 +1112,7 @@ struct __brick_copy
 template <class _Tag>
 struct __brick_move
 {
-    static_assert(__is_host_backend_tag_v<_Tag>);
+    static_assert(__is_host_backend_tag_v<::std::decay_t<_Tag>>);
 
     template <typename _RandomAccessIterator1, typename _RandomAccessIterator2>
     _RandomAccessIterator2
@@ -1142,7 +1142,7 @@ struct __brick_move
 template <class _Tag>
 struct __brick_move_destroy
 {
-    static_assert(__is_host_backend_tag_v<_Tag>);
+    static_assert(__is_host_backend_tag_v<::std::decay_t<_Tag>>);
 
     template <typename _RandomAccessIterator1, typename _RandomAccessIterator2>
     _RandomAccessIterator2
@@ -1315,7 +1315,7 @@ _OutputIterator
 __pattern_copy_if(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _OutputIterator __result,
                   _UnaryPredicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_copy_if(__first, __last, __result, __pred, typename _Tag::__is_vector{});
 }
@@ -1380,7 +1380,7 @@ template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Pre
 typename ::std::iterator_traits<_ForwardIterator>::difference_type
 __pattern_count(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _Predicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_count(__first, __last, __pred, typename _Tag::__is_vector{});
 }
@@ -1434,7 +1434,7 @@ _ForwardIterator
 __pattern_unique(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                  _BinaryPredicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_unique(__first, __last, __pred, typename _Tag::__is_vector{});
 }
@@ -1577,7 +1577,7 @@ _OutputIterator
 __pattern_unique_copy(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                       _OutputIterator __result, _BinaryPredicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_unique_copy(__first, __last, __result, __pred, typename _Tag::__is_vector{});
 }
@@ -1713,7 +1713,7 @@ template <class _Tag, class _ExecutionPolicy, class _BidirectionalIterator>
 void
 __pattern_reverse(_Tag, _ExecutionPolicy&&, _BidirectionalIterator __first, _BidirectionalIterator __last) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     __internal::__brick_reverse(__first, __last, typename _Tag::__is_vector{});
 }
@@ -1761,7 +1761,7 @@ _OutputIterator
 __pattern_reverse_copy(_Tag, _ExecutionPolicy&&, _BidirectionalIterator __first, _BidirectionalIterator __last,
                        _OutputIterator __d_first) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_reverse_copy(__first, __last, __d_first, typename _Tag::__is_vector{});
 }
@@ -1840,7 +1840,7 @@ _ForwardIterator
 __pattern_rotate(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __middle,
                  _ForwardIterator __last) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_rotate(__first, __middle, __last, typename _Tag::__is_vector{});
 }
@@ -1918,7 +1918,7 @@ _OutputIterator
 __brick_rotate_copy(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __middle,
                     _ForwardIterator __last, _OutputIterator __result) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return ::std::rotate_copy(__first, __middle, __last, __result);
 }
@@ -1938,7 +1938,7 @@ _OutputIterator
 __pattern_rotate_copy(_Tag __tag, _ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __middle,
                       _ForwardIterator __last, _OutputIterator __result) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_rotate_copy(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __middle, __last,
                                            __result);
@@ -2020,7 +2020,7 @@ bool
 __pattern_is_partitioned(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                          _UnaryPredicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_is_partitioned(__first, __last, __pred, typename _Tag::__is_vector{});
 }
@@ -2150,7 +2150,7 @@ _ForwardIterator
 __pattern_partition(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                     _UnaryPredicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_partition(__first, __last, __pred, typename _Tag::__is_vector{});
 }
@@ -2250,7 +2250,7 @@ _BidirectionalIterator
 __pattern_stable_partition(_Tag, _ExecutionPolicy&&, _BidirectionalIterator __first, _BidirectionalIterator __last,
                            _UnaryPredicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_stable_partition(__first, __last, __pred, typename _Tag::__is_vector{});
 }
@@ -2340,7 +2340,7 @@ ::std::pair<_OutputIterator1, _OutputIterator2>
 __pattern_partition_copy(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                          _OutputIterator1 __out_true, _OutputIterator2 __out_false, _UnaryPredicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_partition_copy(__first, __last, __out_true, __out_false, __pred,
                                               typename _Tag::__is_vector{});
@@ -2396,7 +2396,7 @@ void
 __pattern_sort(_Tag, _ExecutionPolicy&&, _RandomAccessIterator __first, _RandomAccessIterator __last, _Compare __comp,
                _IsMoveConstructible) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     ::std::sort(__first, __last, __comp);
 }
@@ -2428,7 +2428,7 @@ void
 __pattern_stable_sort(_Tag, _ExecutionPolicy&&, _RandomAccessIterator __first, _RandomAccessIterator __last,
                       _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     ::std::stable_sort(__first, __last, __comp);
 }
@@ -2461,7 +2461,7 @@ __pattern_sort_by_key(_Tag, _ExecutionPolicy&& __exec, _RandomAccessIterator1 __
                       _RandomAccessIterator1 __keys_last, _RandomAccessIterator2 __values_first,
                       _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     auto __beg = oneapi::dpl::make_zip_iterator(__keys_first, __values_first);
     auto __end = __beg + (__keys_last - __keys_first);
@@ -2507,7 +2507,7 @@ void
 __pattern_partial_sort(_Tag, _ExecutionPolicy&&, _RandomAccessIterator __first, _RandomAccessIterator __middle,
                        _RandomAccessIterator __last, _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     ::std::partial_sort(__first, __middle, __last, __comp);
 }
@@ -2545,7 +2545,7 @@ _RandomAccessIterator
 __pattern_partial_sort_copy(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                             _RandomAccessIterator __d_first, _RandomAccessIterator __d_last, _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return ::std::partial_sort_copy(__first, __last, __d_first, __d_last, __comp);
 }
@@ -2649,7 +2649,7 @@ _ForwardIterator
 __pattern_adjacent_find(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                         _BinaryPredicate __pred, _Semantic) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_adjacent_find(__first, __last, __pred, typename _Tag::__is_vector{}, _Semantic::value);
 }
@@ -2709,7 +2709,7 @@ void
 __pattern_nth_element(_Tag, _ExecutionPolicy&&, _RandomAccessIterator __first, _RandomAccessIterator __nth,
                       _RandomAccessIterator __last, _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     ::std::nth_element(__first, __nth, __last, __comp);
 }
@@ -2761,7 +2761,7 @@ __pattern_nth_element(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec
 template <class _Tag, typename _Tp>
 struct __brick_fill
 {
-    static_assert(__is_host_backend_tag_v<_Tag>);
+    static_assert(__is_host_backend_tag_v<::std::decay_t<_Tag>>);
 
     const _Tp& __value;
 
@@ -2786,7 +2786,7 @@ template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Tp>
 void
 __pattern_fill(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, const _Tp& __value) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     __internal::__brick_fill<_Tag, _Tp>{__value}(__first, __last, typename _Tag::__is_vector{});
 }
@@ -2811,7 +2811,7 @@ __pattern_fill(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAcce
 template <class _Tag, typename _Tp>
 struct __brick_fill_n
 {
-    static_assert(__is_host_backend_tag_v<_Tag>);
+    static_assert(__is_host_backend_tag_v<::std::decay_t<_Tag>>);
 
     const _Tp& __value;
 
@@ -2836,7 +2836,7 @@ template <class _Tag, class _ExecutionPolicy, class _OutputIterator, class _Size
 _OutputIterator
 __pattern_fill_n(_Tag, _ExecutionPolicy&&, _OutputIterator __first, _Size __count, const _Tp& __value) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_fill_n<_Tag, _Tp>{__value}(__first, __count, typename _Tag::__is_vector{});
 }
@@ -2873,7 +2873,7 @@ template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Gen
 void
 __pattern_generate(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _Generator __g) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     __internal::__brick_generate(__first, __last, __g, typename _Tag::__is_vector{});
 }
@@ -2913,7 +2913,7 @@ template <class _Tag, class _ExecutionPolicy, class _OutputIterator, class _Size
 _OutputIterator
 __pattern_generate_n(_Tag, _ExecutionPolicy&&, _OutputIterator __first, _Size __count, _Generator __g) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_generate_n(__first, __count, __g, typename _Tag::__is_vector{});
 }
@@ -2958,7 +2958,7 @@ _ForwardIterator
 __pattern_remove_if(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                     _UnaryPredicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_remove_if(__first, __last, __pred, typename _Tag::__is_vector{});
 }
@@ -3014,7 +3014,7 @@ __pattern_merge(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIt
                 _ForwardIterator2 __first2, _ForwardIterator2 __last2, _OutputIterator __d_first,
                 _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_merge(__first1, __last1, __first2, __last2, __d_first, __comp,
                                      typename _Tag::__is_vector{});
@@ -3063,7 +3063,7 @@ void
 __pattern_inplace_merge(_Tag, _ExecutionPolicy&&, _BidirectionalIterator __first, _BidirectionalIterator __middle,
                         _BidirectionalIterator __last, _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     __internal::__brick_inplace_merge(__first, __middle, __last, __comp, typename _Tag::__is_vector{});
 }
@@ -3122,7 +3122,7 @@ bool
 __pattern_includes(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
                    _ForwardIterator2 __first2, _ForwardIterator2 __last2, _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return ::std::includes(__first1, __last1, __first2, __last2, __comp);
 }
@@ -3437,7 +3437,7 @@ __pattern_set_union(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _Forwa
                     _ForwardIterator2 __first2, _ForwardIterator2 __last2, _OutputIterator __result,
                     _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_set_union(__first1, __last1, __first2, __last2, __result, __comp,
                                          typename _Tag::__is_vector{});
@@ -3498,7 +3498,7 @@ __pattern_set_intersection(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1,
                            _ForwardIterator2 __first2, _ForwardIterator2 __last2, _OutputIterator __result,
                            _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_set_intersection(__first1, __last1, __first2, __last2, __result, __comp,
                                                 typename _Tag::__is_vector{});
@@ -3596,7 +3596,7 @@ __pattern_set_difference(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _
                          _ForwardIterator2 __first2, _ForwardIterator2 __last2, _OutputIterator __result,
                          _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_set_difference(__first1, __last1, __first2, __last2, __result, __comp,
                                               typename _Tag::__is_vector{});
@@ -3683,7 +3683,7 @@ __pattern_set_symmetric_difference(_Tag, _ExecutionPolicy&&, _ForwardIterator1 _
                                    _ForwardIterator2 __first2, _ForwardIterator2 __last2, _OutputIterator __result,
                                    _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_set_symmetric_difference(__first1, __last1, __first2, __last2, __result, __comp,
                                                         typename _Tag::__is_vector{});
@@ -3742,7 +3742,7 @@ _RandomAccessIterator
 __pattern_is_heap_until(_Tag, _ExecutionPolicy&&, _RandomAccessIterator __first, _RandomAccessIterator __last,
                         _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_is_heap_until(__first, __last, __comp, typename _Tag::__is_vector{});
 }
@@ -3830,7 +3830,7 @@ bool
 __pattern_is_heap(_Tag, _ExecutionPolicy&&, _RandomAccessIterator __first, _RandomAccessIterator __last,
                   _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_is_heap(__first, __last, __comp, typename _Tag::__is_vector{});
 }
@@ -3878,7 +3878,7 @@ _ForwardIterator
 __pattern_min_element(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                       _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_min_element(__first, __last, __comp, typename _Tag::__is_vector{});
 }
@@ -3945,7 +3945,7 @@ ::std::pair<_ForwardIterator, _ForwardIterator>
 __pattern_minmax_element(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                          _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_minmax_element(__first, __last, __comp, typename _Tag::__is_vector{});
 }
@@ -4026,7 +4026,7 @@ ::std::pair<_ForwardIterator1, _ForwardIterator2>
 __pattern_mismatch(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
                    _ForwardIterator2 __first2, _ForwardIterator2 __last2, _Predicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_mismatch(__first1, __last1, __first2, __last2, __pred, typename _Tag::__is_vector{});
 }
@@ -4107,7 +4107,7 @@ bool
 __pattern_lexicographical_compare(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
                                   _ForwardIterator2 __first2, _ForwardIterator2 __last2, _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_lexicographical_compare(__first1, __last1, __first2, __last2, __comp,
                                                        typename _Tag::__is_vector{});
@@ -4169,7 +4169,7 @@ _ForwardIterator2
 __pattern_swap(_Tag __tag, _ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
                _ForwardIterator2 __first2, _Function __f)
 {
-    static_assert(__is_host_backend_tag_v<_Tag>);
+    static_assert(__is_host_backend_tag_v<::std::decay_t<_Tag>>);
 
     return __pattern_walk2(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __f);
 }
@@ -4247,7 +4247,7 @@ _ForwardIterator
 __pattern_shift_left(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                      typename ::std::iterator_traits<_ForwardIterator>::difference_type __n) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __brick_shift_left(__first, __last, __n, typename _Tag::__is_vector{});
 }
@@ -4303,7 +4303,7 @@ __pattern_shift_right(_Tag __tag, _ExecutionPolicy&& __exec, _BidirectionalItera
                       _BidirectionalIterator __last,
                       typename ::std::iterator_traits<_BidirectionalIterator>::difference_type __n)
 {
-    static_assert(__is_host_backend_tag_v<_Tag>);
+    static_assert(__is_host_backend_tag_v<::std::decay_t<_Tag>>);
 
     using _ReverseIterator = typename ::std::reverse_iterator<_BidirectionalIterator>;
 
diff --git a/include/oneapi/dpl/pstl/execution_impl.h b/include/oneapi/dpl/pstl/execution_impl.h
index fe3ae6bc52c..b99114ebfc1 100644
--- a/include/oneapi/dpl/pstl/execution_impl.h
+++ b/include/oneapi/dpl/pstl/execution_impl.h
@@ -112,7 +112,7 @@ struct __is_backend_tag_serial<__serial_tag<_IsVector>> : ::std::true_type
 };
 
 template <class _Tag>
-inline constexpr bool __is_backend_tag_serial_v = __is_backend_tag_serial<::std::decay_t<_Tag>>::value;
+inline constexpr bool __is_backend_tag_serial_v = __is_backend_tag_serial<_Tag>::value;
 
 //----------------------------------------------------------
 // __is_backend_tag_parallel_forward, __is_backend_tag_parallel_forward_v
@@ -129,8 +129,7 @@ struct __is_backend_tag_parallel_forward<__parallel_forward_tag> : ::std::true_t
 };
 
 template <class _Tag>
-inline constexpr bool __is_backend_tag_parallel_forward_v =
-    __is_backend_tag_parallel_forward<::std::decay_t<_Tag>>::value;
+inline constexpr bool __is_backend_tag_parallel_forward_v = __is_backend_tag_parallel_forward<_Tag>::value;
 
 //----------------------------------------------------------
 // __is_backend_tag_parallel, __is_backend_tag_parallel_v
@@ -147,7 +146,7 @@ struct __is_backend_tag_parallel<__parallel_tag<_IsVector>> : ::std::true_type
 };
 
 template <class _Tag>
-inline constexpr bool __is_backend_tag_parallel_v = __is_backend_tag_parallel<::std::decay_t<_Tag>>::value;
+inline constexpr bool __is_backend_tag_parallel_v = __is_backend_tag_parallel<_Tag>::value;
 
 //----------------------------------------------------------
 // __is_host_backend_tag_v
diff --git a/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h b/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
index a6a6952738a..53d1e355c8f 100644
--- a/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
+++ b/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
@@ -167,7 +167,7 @@ void
 __pattern_for_loop_n(_Tag, _ExecutionPolicy&&, _Ip __first, _Size __n, _Function __f, __single_stride_type,
                      _Rest&&... __rest) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     __reduction_pack<_Rest...> __pack{__reduction_pack_tag(), ::std::forward<_Rest>(__rest)...};
 
@@ -192,7 +192,7 @@ void
 __pattern_for_loop_n(_Tag, _ExecutionPolicy&&, _Ip __first, _Size __n, _Function __f, _Sp __stride,
                      _Rest&&... __rest) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     __reduction_pack<_Rest...> __pack{__reduction_pack_tag(), ::std::forward<_Rest>(__rest)...};
 
@@ -265,7 +265,7 @@ void
 __pattern_for_loop(_Tag __tag, _ExecutionPolicy&& __exec, _Ip __first, _Ip __last, _Function __f, _Sp __stride,
                    _Rest&&... __rest) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     if constexpr (typename _Tag::__is_vector{} || __is_random_access_or_integral_v<_Ip>)
     {
@@ -366,7 +366,7 @@ void
 __pattern_for_loop(_Tag, _ExecutionPolicy&& __exec, _Ip __first, _Ip __last, _Function __f, __single_stride_type,
                    _Rest&&... __rest) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     __reduction_pack<_Rest...> __pack{__reduction_pack_tag(), ::std::forward<_Rest>(__rest)...};
 
diff --git a/include/oneapi/dpl/pstl/histogram_impl.h b/include/oneapi/dpl/pstl/histogram_impl.h
index 885600f0288..07b359b938e 100644
--- a/include/oneapi/dpl/pstl/histogram_impl.h
+++ b/include/oneapi/dpl/pstl/histogram_impl.h
@@ -38,7 +38,7 @@ void
 __pattern_histogram(_Tag, _ExecutionPolicy&& exec, _RandomAccessIterator1 __first, _RandomAccessIterator1 __last,
                     _Size __num_bins, _IdxHashFunc __func, _RandomAccessIterator2 __histogram_first)
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     static_assert(sizeof(_Size) == 0 /*false*/,
                   "Histogram API is currently unsupported for policies other than device execution policies");
diff --git a/include/oneapi/dpl/pstl/numeric_impl.h b/include/oneapi/dpl/pstl/numeric_impl.h
index 2b8102dfe7a..bffb70c06a7 100644
--- a/include/oneapi/dpl/pstl/numeric_impl.h
+++ b/include/oneapi/dpl/pstl/numeric_impl.h
@@ -68,7 +68,7 @@ __pattern_transform_reduce(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1,
                            _ForwardIterator2 __first2, _Tp __init, _BinaryOperation1 __binary_op1,
                            _BinaryOperation2 __binary_op2) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __brick_transform_reduce(__first1, __last1, __first2, __init, __binary_op1, __binary_op2,
                                     typename _Tag::__is_vector{});
@@ -133,7 +133,7 @@ _Tp
 __pattern_transform_reduce(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _Tp __init,
                            _BinaryOperation __binary_op, _UnaryOperation __unary_op) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_transform_reduce(__first, __last, __init, __binary_op, __unary_op,
                                                 typename _Tag::__is_vector{});
@@ -243,7 +243,7 @@ __pattern_transform_scan(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _Fo
                          _OutputIterator __result, _UnaryOperation __unary_op, _Tp __init, _BinaryOperation __binary_op,
                          _Inclusive) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_transform_scan(__first, __last, __result, __unary_op, __init, __binary_op, _Inclusive(),
                                               typename _Tag::__is_vector{})
@@ -327,7 +327,7 @@ _OutputIterator
 __pattern_transform_scan(_Tag __tag, _ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last,
                          _OutputIterator __result, _UnaryOperation __unary_op, _BinaryOperation __binary_op, _Inclusive)
 {
-    static_assert(__is_host_backend_tag_v<_Tag>);
+    static_assert(__is_host_backend_tag_v<::std::decay_t<_Tag>>);
 
     typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _ValueType;
     if (__first != __last)
@@ -379,7 +379,7 @@ _OutputIterator
 __pattern_adjacent_difference(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                               _OutputIterator __d_first, _BinaryOperation __op) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_adjacent_difference(__first, __last, __d_first, __op, typename _Tag::__is_vector{});
 }

From fbd0f4ffef6e0d2501b0dd71ea604c2b9d0c9f38 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 1 Mar 2024 14:06:23 +0100
Subject: [PATCH 500/566] Apply GitHUB clang format

---
 include/oneapi/dpl/pstl/algorithm_impl.h      | 156 ++++++++++++------
 .../experimental/internal/for_loop_impl.h     |  12 +-
 include/oneapi/dpl/pstl/histogram_impl.h      |   3 +-
 include/oneapi/dpl/pstl/numeric_impl.h        |  12 +-
 4 files changed, 122 insertions(+), 61 deletions(-)

diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index ca36a8429e1..ee475895a70 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -67,7 +67,8 @@ template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Pre
 bool
 __pattern_any_of(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _Pred __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
+                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_any_of(__first, __last, __pred, typename _Tag::__is_vector{});
 }
@@ -179,7 +180,8 @@ void
 __pattern_walk_brick(_Tag, _ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last,
                      _Brick __brick) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
+                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     __brick(__first, __last, typename _Tag::__is_vector{});
 }
@@ -221,7 +223,8 @@ template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Siz
 _ForwardIterator
 __pattern_walk1_n(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _Size __n, _Function __f) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
+                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_walk1_n(__first, __n, __f, typename _Tag::__is_vector{});
 }
@@ -240,7 +243,8 @@ template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Siz
 _ForwardIterator
 __pattern_walk_brick_n(_Tag, _ExecutionPolicy&& __exec, _ForwardIterator __first, _Size __n, _Brick __brick) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
+                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __brick(__first, __n, typename _Tag::__is_vector{});
 }
@@ -364,7 +368,8 @@ _ForwardIterator2
 __pattern_walk2_n(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _Size __n, _ForwardIterator2 __first2,
                   _Function __f) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
+                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_walk2_n(__first1, __n, __first2, __f, typename _Tag::__is_vector{});
 }
@@ -461,7 +466,8 @@ _ForwardIterator2
 __pattern_walk2_brick_n(_Tag, _ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _Size __n,
                         _ForwardIterator2 __first2, _Brick __brick) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
+                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __brick(__first1, __n, __first2, typename _Tag::__is_vector{});
 }
@@ -608,7 +614,8 @@ bool
 __pattern_equal(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
                 _ForwardIterator2 __first2, _ForwardIterator2 __last2, _BinaryPredicate __p) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
+                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_equal(__first1, __last1, __first2, __last2, __p, typename _Tag::__is_vector{});
 }
@@ -659,7 +666,8 @@ bool
 __pattern_equal(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
                 _ForwardIterator2 __first2, _BinaryPredicate __p) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
+                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_equal(__first1, __last1, __first2, __p, typename _Tag::__is_vector{});
 }
@@ -706,7 +714,8 @@ _ForwardIterator
 __pattern_find_if(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                   _Predicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
+                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_find_if(__first, __last, __pred, typename _Tag::__is_vector{});
 }
@@ -849,7 +858,8 @@ _ForwardIterator1
 __pattern_find_end(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first, _ForwardIterator1 __last,
                    _ForwardIterator2 __s_first, _ForwardIterator2 __s_last, _BinaryPredicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
+                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_find_end(__first, __last, __s_first, __s_last, __pred, typename _Tag::__is_vector{});
 }
@@ -905,7 +915,8 @@ _ForwardIterator1
 __pattern_find_first_of(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first, _ForwardIterator1 __last,
                         _ForwardIterator2 __s_first, _ForwardIterator2 __s_last, _BinaryPredicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
+                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_find_first_of(__first, __last, __s_first, __s_last, __pred,
                                              typename _Tag::__is_vector{});
@@ -954,7 +965,8 @@ _ForwardIterator1
 __pattern_search(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first, _ForwardIterator1 __last,
                  _ForwardIterator2 __s_first, _ForwardIterator2 __s_last, _BinaryPredicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
+                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_search(__first, __last, __s_first, __s_last, __pred, typename _Tag::__is_vector{});
 }
@@ -1010,7 +1022,8 @@ _ForwardIterator
 __pattern_search_n(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _Size __count,
                    const _Tp& __value, _BinaryPredicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
+                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_search_n(__first, __last, __count, __value, __pred, typename _Tag::__is_vector{});
 }
@@ -1315,7 +1328,8 @@ _OutputIterator
 __pattern_copy_if(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _OutputIterator __result,
                   _UnaryPredicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
+                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_copy_if(__first, __last, __result, __pred, typename _Tag::__is_vector{});
 }
@@ -1380,7 +1394,8 @@ template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Pre
 typename ::std::iterator_traits<_ForwardIterator>::difference_type
 __pattern_count(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _Predicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
+                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_count(__first, __last, __pred, typename _Tag::__is_vector{});
 }
@@ -1434,7 +1449,8 @@ _ForwardIterator
 __pattern_unique(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                  _BinaryPredicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
+                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_unique(__first, __last, __pred, typename _Tag::__is_vector{});
 }
@@ -1577,7 +1593,8 @@ _OutputIterator
 __pattern_unique_copy(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                       _OutputIterator __result, _BinaryPredicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
+                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_unique_copy(__first, __last, __result, __pred, typename _Tag::__is_vector{});
 }
@@ -1713,7 +1730,8 @@ template <class _Tag, class _ExecutionPolicy, class _BidirectionalIterator>
 void
 __pattern_reverse(_Tag, _ExecutionPolicy&&, _BidirectionalIterator __first, _BidirectionalIterator __last) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
+                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     __internal::__brick_reverse(__first, __last, typename _Tag::__is_vector{});
 }
@@ -1761,7 +1779,8 @@ _OutputIterator
 __pattern_reverse_copy(_Tag, _ExecutionPolicy&&, _BidirectionalIterator __first, _BidirectionalIterator __last,
                        _OutputIterator __d_first) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
+                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_reverse_copy(__first, __last, __d_first, typename _Tag::__is_vector{});
 }
@@ -1840,7 +1859,8 @@ _ForwardIterator
 __pattern_rotate(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __middle,
                  _ForwardIterator __last) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
+                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_rotate(__first, __middle, __last, typename _Tag::__is_vector{});
 }
@@ -1918,7 +1938,8 @@ _OutputIterator
 __brick_rotate_copy(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __middle,
                     _ForwardIterator __last, _OutputIterator __result) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
+                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return ::std::rotate_copy(__first, __middle, __last, __result);
 }
@@ -1938,7 +1959,8 @@ _OutputIterator
 __pattern_rotate_copy(_Tag __tag, _ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __middle,
                       _ForwardIterator __last, _OutputIterator __result) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
+                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_rotate_copy(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __middle, __last,
                                            __result);
@@ -2020,7 +2042,8 @@ bool
 __pattern_is_partitioned(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                          _UnaryPredicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
+                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_is_partitioned(__first, __last, __pred, typename _Tag::__is_vector{});
 }
@@ -2150,7 +2173,8 @@ _ForwardIterator
 __pattern_partition(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                     _UnaryPredicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
+                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_partition(__first, __last, __pred, typename _Tag::__is_vector{});
 }
@@ -2250,7 +2274,8 @@ _BidirectionalIterator
 __pattern_stable_partition(_Tag, _ExecutionPolicy&&, _BidirectionalIterator __first, _BidirectionalIterator __last,
                            _UnaryPredicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
+                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_stable_partition(__first, __last, __pred, typename _Tag::__is_vector{});
 }
@@ -2340,7 +2365,8 @@ ::std::pair<_OutputIterator1, _OutputIterator2>
 __pattern_partition_copy(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                          _OutputIterator1 __out_true, _OutputIterator2 __out_false, _UnaryPredicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
+                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_partition_copy(__first, __last, __out_true, __out_false, __pred,
                                               typename _Tag::__is_vector{});
@@ -2396,7 +2422,8 @@ void
 __pattern_sort(_Tag, _ExecutionPolicy&&, _RandomAccessIterator __first, _RandomAccessIterator __last, _Compare __comp,
                _IsMoveConstructible) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
+                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     ::std::sort(__first, __last, __comp);
 }
@@ -2428,7 +2455,8 @@ void
 __pattern_stable_sort(_Tag, _ExecutionPolicy&&, _RandomAccessIterator __first, _RandomAccessIterator __last,
                       _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
+                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     ::std::stable_sort(__first, __last, __comp);
 }
@@ -2461,7 +2489,8 @@ __pattern_sort_by_key(_Tag, _ExecutionPolicy&& __exec, _RandomAccessIterator1 __
                       _RandomAccessIterator1 __keys_last, _RandomAccessIterator2 __values_first,
                       _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
+                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     auto __beg = oneapi::dpl::make_zip_iterator(__keys_first, __values_first);
     auto __end = __beg + (__keys_last - __keys_first);
@@ -2507,7 +2536,8 @@ void
 __pattern_partial_sort(_Tag, _ExecutionPolicy&&, _RandomAccessIterator __first, _RandomAccessIterator __middle,
                        _RandomAccessIterator __last, _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
+                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     ::std::partial_sort(__first, __middle, __last, __comp);
 }
@@ -2545,7 +2575,8 @@ _RandomAccessIterator
 __pattern_partial_sort_copy(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                             _RandomAccessIterator __d_first, _RandomAccessIterator __d_last, _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
+                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return ::std::partial_sort_copy(__first, __last, __d_first, __d_last, __comp);
 }
@@ -2649,7 +2680,8 @@ _ForwardIterator
 __pattern_adjacent_find(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                         _BinaryPredicate __pred, _Semantic) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
+                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_adjacent_find(__first, __last, __pred, typename _Tag::__is_vector{}, _Semantic::value);
 }
@@ -2709,7 +2741,8 @@ void
 __pattern_nth_element(_Tag, _ExecutionPolicy&&, _RandomAccessIterator __first, _RandomAccessIterator __nth,
                       _RandomAccessIterator __last, _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
+                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     ::std::nth_element(__first, __nth, __last, __comp);
 }
@@ -2786,7 +2819,8 @@ template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Tp>
 void
 __pattern_fill(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, const _Tp& __value) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
+                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     __internal::__brick_fill<_Tag, _Tp>{__value}(__first, __last, typename _Tag::__is_vector{});
 }
@@ -2836,7 +2870,8 @@ template <class _Tag, class _ExecutionPolicy, class _OutputIterator, class _Size
 _OutputIterator
 __pattern_fill_n(_Tag, _ExecutionPolicy&&, _OutputIterator __first, _Size __count, const _Tp& __value) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
+                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_fill_n<_Tag, _Tp>{__value}(__first, __count, typename _Tag::__is_vector{});
 }
@@ -2873,7 +2908,8 @@ template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Gen
 void
 __pattern_generate(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _Generator __g) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
+                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     __internal::__brick_generate(__first, __last, __g, typename _Tag::__is_vector{});
 }
@@ -2913,7 +2949,8 @@ template <class _Tag, class _ExecutionPolicy, class _OutputIterator, class _Size
 _OutputIterator
 __pattern_generate_n(_Tag, _ExecutionPolicy&&, _OutputIterator __first, _Size __count, _Generator __g) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
+                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_generate_n(__first, __count, __g, typename _Tag::__is_vector{});
 }
@@ -2958,7 +2995,8 @@ _ForwardIterator
 __pattern_remove_if(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                     _UnaryPredicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
+                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_remove_if(__first, __last, __pred, typename _Tag::__is_vector{});
 }
@@ -3014,7 +3052,8 @@ __pattern_merge(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIt
                 _ForwardIterator2 __first2, _ForwardIterator2 __last2, _OutputIterator __d_first,
                 _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
+                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_merge(__first1, __last1, __first2, __last2, __d_first, __comp,
                                      typename _Tag::__is_vector{});
@@ -3063,7 +3102,8 @@ void
 __pattern_inplace_merge(_Tag, _ExecutionPolicy&&, _BidirectionalIterator __first, _BidirectionalIterator __middle,
                         _BidirectionalIterator __last, _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
+                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     __internal::__brick_inplace_merge(__first, __middle, __last, __comp, typename _Tag::__is_vector{});
 }
@@ -3122,7 +3162,8 @@ bool
 __pattern_includes(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
                    _ForwardIterator2 __first2, _ForwardIterator2 __last2, _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
+                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return ::std::includes(__first1, __last1, __first2, __last2, __comp);
 }
@@ -3437,7 +3478,8 @@ __pattern_set_union(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _Forwa
                     _ForwardIterator2 __first2, _ForwardIterator2 __last2, _OutputIterator __result,
                     _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
+                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_set_union(__first1, __last1, __first2, __last2, __result, __comp,
                                          typename _Tag::__is_vector{});
@@ -3498,7 +3540,8 @@ __pattern_set_intersection(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1,
                            _ForwardIterator2 __first2, _ForwardIterator2 __last2, _OutputIterator __result,
                            _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
+                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_set_intersection(__first1, __last1, __first2, __last2, __result, __comp,
                                                 typename _Tag::__is_vector{});
@@ -3596,7 +3639,8 @@ __pattern_set_difference(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _
                          _ForwardIterator2 __first2, _ForwardIterator2 __last2, _OutputIterator __result,
                          _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
+                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_set_difference(__first1, __last1, __first2, __last2, __result, __comp,
                                               typename _Tag::__is_vector{});
@@ -3683,7 +3727,8 @@ __pattern_set_symmetric_difference(_Tag, _ExecutionPolicy&&, _ForwardIterator1 _
                                    _ForwardIterator2 __first2, _ForwardIterator2 __last2, _OutputIterator __result,
                                    _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
+                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_set_symmetric_difference(__first1, __last1, __first2, __last2, __result, __comp,
                                                         typename _Tag::__is_vector{});
@@ -3742,7 +3787,8 @@ _RandomAccessIterator
 __pattern_is_heap_until(_Tag, _ExecutionPolicy&&, _RandomAccessIterator __first, _RandomAccessIterator __last,
                         _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
+                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_is_heap_until(__first, __last, __comp, typename _Tag::__is_vector{});
 }
@@ -3830,7 +3876,8 @@ bool
 __pattern_is_heap(_Tag, _ExecutionPolicy&&, _RandomAccessIterator __first, _RandomAccessIterator __last,
                   _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
+                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_is_heap(__first, __last, __comp, typename _Tag::__is_vector{});
 }
@@ -3878,7 +3925,8 @@ _ForwardIterator
 __pattern_min_element(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                       _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
+                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_min_element(__first, __last, __comp, typename _Tag::__is_vector{});
 }
@@ -3945,7 +3993,8 @@ ::std::pair<_ForwardIterator, _ForwardIterator>
 __pattern_minmax_element(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                          _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
+                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_minmax_element(__first, __last, __comp, typename _Tag::__is_vector{});
 }
@@ -4026,7 +4075,8 @@ ::std::pair<_ForwardIterator1, _ForwardIterator2>
 __pattern_mismatch(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
                    _ForwardIterator2 __first2, _ForwardIterator2 __last2, _Predicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
+                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_mismatch(__first1, __last1, __first2, __last2, __pred, typename _Tag::__is_vector{});
 }
@@ -4107,7 +4157,8 @@ bool
 __pattern_lexicographical_compare(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
                                   _ForwardIterator2 __first2, _ForwardIterator2 __last2, _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
+                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_lexicographical_compare(__first1, __last1, __first2, __last2, __comp,
                                                        typename _Tag::__is_vector{});
@@ -4247,7 +4298,8 @@ _ForwardIterator
 __pattern_shift_left(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                      typename ::std::iterator_traits<_ForwardIterator>::difference_type __n) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
+                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __brick_shift_left(__first, __last, __n, typename _Tag::__is_vector{});
 }
diff --git a/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h b/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
index 53d1e355c8f..88d08abb3a1 100644
--- a/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
+++ b/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
@@ -167,7 +167,8 @@ void
 __pattern_for_loop_n(_Tag, _ExecutionPolicy&&, _Ip __first, _Size __n, _Function __f, __single_stride_type,
                      _Rest&&... __rest) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
+                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     __reduction_pack<_Rest...> __pack{__reduction_pack_tag(), ::std::forward<_Rest>(__rest)...};
 
@@ -192,7 +193,8 @@ void
 __pattern_for_loop_n(_Tag, _ExecutionPolicy&&, _Ip __first, _Size __n, _Function __f, _Sp __stride,
                      _Rest&&... __rest) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
+                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     __reduction_pack<_Rest...> __pack{__reduction_pack_tag(), ::std::forward<_Rest>(__rest)...};
 
@@ -265,7 +267,8 @@ void
 __pattern_for_loop(_Tag __tag, _ExecutionPolicy&& __exec, _Ip __first, _Ip __last, _Function __f, _Sp __stride,
                    _Rest&&... __rest) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
+                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     if constexpr (typename _Tag::__is_vector{} || __is_random_access_or_integral_v<_Ip>)
     {
@@ -366,7 +369,8 @@ void
 __pattern_for_loop(_Tag, _ExecutionPolicy&& __exec, _Ip __first, _Ip __last, _Function __f, __single_stride_type,
                    _Rest&&... __rest) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
+                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     __reduction_pack<_Rest...> __pack{__reduction_pack_tag(), ::std::forward<_Rest>(__rest)...};
 
diff --git a/include/oneapi/dpl/pstl/histogram_impl.h b/include/oneapi/dpl/pstl/histogram_impl.h
index 07b359b938e..5d27df53749 100644
--- a/include/oneapi/dpl/pstl/histogram_impl.h
+++ b/include/oneapi/dpl/pstl/histogram_impl.h
@@ -38,7 +38,8 @@ void
 __pattern_histogram(_Tag, _ExecutionPolicy&& exec, _RandomAccessIterator1 __first, _RandomAccessIterator1 __last,
                     _Size __num_bins, _IdxHashFunc __func, _RandomAccessIterator2 __histogram_first)
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
+                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     static_assert(sizeof(_Size) == 0 /*false*/,
                   "Histogram API is currently unsupported for policies other than device execution policies");
diff --git a/include/oneapi/dpl/pstl/numeric_impl.h b/include/oneapi/dpl/pstl/numeric_impl.h
index bffb70c06a7..b03ee25340a 100644
--- a/include/oneapi/dpl/pstl/numeric_impl.h
+++ b/include/oneapi/dpl/pstl/numeric_impl.h
@@ -68,7 +68,8 @@ __pattern_transform_reduce(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1,
                            _ForwardIterator2 __first2, _Tp __init, _BinaryOperation1 __binary_op1,
                            _BinaryOperation2 __binary_op2) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
+                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __brick_transform_reduce(__first1, __last1, __first2, __init, __binary_op1, __binary_op2,
                                     typename _Tag::__is_vector{});
@@ -133,7 +134,8 @@ _Tp
 __pattern_transform_reduce(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _Tp __init,
                            _BinaryOperation __binary_op, _UnaryOperation __unary_op) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
+                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_transform_reduce(__first, __last, __init, __binary_op, __unary_op,
                                                 typename _Tag::__is_vector{});
@@ -243,7 +245,8 @@ __pattern_transform_scan(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _Fo
                          _OutputIterator __result, _UnaryOperation __unary_op, _Tp __init, _BinaryOperation __binary_op,
                          _Inclusive) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
+                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_transform_scan(__first, __last, __result, __unary_op, __init, __binary_op, _Inclusive(),
                                               typename _Tag::__is_vector{})
@@ -379,7 +382,8 @@ _OutputIterator
 __pattern_adjacent_difference(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                               _OutputIterator __d_first, _BinaryOperation __op) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> || __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
+                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
 
     return __internal::__brick_adjacent_difference(__first, __last, __d_first, __op, typename _Tag::__is_vector{});
 }

From 90af81960c4f9c38248a051032d939477dff6a2a Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 1 Mar 2024 14:31:00 +0100
Subject: [PATCH 501/566] 
 include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_utils.h - rename
 __buffer_impl to __buffer + _ExecutionPolicy ->
 oneapi::dpl::__internal::__device_backend_tag

---
 .../hetero/dpcpp/parallel_backend_sycl_utils.h   | 16 +++++++++-------
 1 file changed, 9 insertions(+), 7 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_utils.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_utils.h
index dc3a53305a8..fcf9a915211 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_utils.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_utils.h
@@ -376,9 +376,12 @@ struct __local_buffer<sycl::buffer<::std::tuple<_T...>, __dim, _AllocT>>
     using type = sycl::buffer<oneapi::dpl::__internal::tuple<_T...>, __dim, _AllocT>;
 };
 
+template <typename _BackendTag, typename _Tp>
+class __buffer_impl;
+
 // impl for sycl::buffer<...>
-template <typename _ExecutionPolicy, typename _T>
-class __buffer_impl
+template <typename _T>
+class __buffer_impl<oneapi::dpl::__internal::__device_backend_tag, _T>
 {
   private:
     using __container_t = typename __local_buffer<sycl::buffer<_T>>::type;
@@ -386,9 +389,9 @@ class __buffer_impl
     __container_t __container;
 
   public:
-    static_assert(::std::is_same_v<_ExecutionPolicy, ::std::decay_t<_ExecutionPolicy>>);
 
-    __buffer_impl(_ExecutionPolicy /*__exec*/, ::std::size_t __n_elements) : __container{sycl::range<1>(__n_elements)}
+    __buffer_impl(oneapi::dpl::__internal::__device_backend_tag, ::std::size_t __n_elements)
+        : __container{sycl::range<1>(__n_elements)}
     {
     }
 
@@ -448,9 +451,8 @@ struct __memobj_traits<_T*>
 
 } // namespace __internal
 
-template <class _Tag, typename _ExecutionPolicy, typename _T>
-using __buffer = ::std::enable_if_t<::std::is_base_of<oneapi::dpl::__internal::__device_backend_tag, _Tag>::value,
-                                    __internal::__buffer_impl<::std::decay_t<_ExecutionPolicy>, _T>>;
+template <typename _T>
+using __buffer = __internal::__buffer_impl<oneapi::dpl::__internal::__device_backend_tag, _T>;
 
 template <typename T>
 struct __repacked_tuple

From 929f628ffd492c504a40e70f552b5bf38b4137bd Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 1 Mar 2024 14:32:27 +0100
Subject: [PATCH 502/566] include/oneapi/dpl/pstl/omp/util.h - rename
 __buffer_impl to __buffer + _ExecutionPolicy ->
 oneapi::dpl::__internal::__omp_backend_tag

---
 include/oneapi/dpl/pstl/omp/util.h | 19 +++++++++++--------
 1 file changed, 11 insertions(+), 8 deletions(-)

diff --git a/include/oneapi/dpl/pstl/omp/util.h b/include/oneapi/dpl/pstl/omp/util.h
index 923d4504ea7..e3b75e3b516 100644
--- a/include/oneapi/dpl/pstl/omp/util.h
+++ b/include/oneapi/dpl/pstl/omp/util.h
@@ -57,8 +57,11 @@ __cancel_execution(oneapi::dpl::__internal::__omp_backend_tag)
 // raw buffer
 //------------------------------------------------------------------------
 
-template <typename _ExecutionPolicy, typename _Tp>
-class __buffer_impl
+template <typename _BackendTag, typename _Tp>
+class __buffer_impl;
+
+template <typename _Tp>
+class __buffer_impl<oneapi::dpl::__internal::__omp_backend_tag, _Tp>
 {
     std::allocator<_Tp> __allocator_;
     _Tp* __ptr_;
@@ -68,9 +71,10 @@ class __buffer_impl
     operator=(const __buffer_impl&) = delete;
 
   public:
-    static_assert(::std::is_same_v<_ExecutionPolicy, ::std::decay_t<_ExecutionPolicy>>);
-
-    __buffer_impl(std::size_t __n) : __allocator_(), __ptr_(__allocator_.allocate(__n)), __buf_size_(__n) {}
+    __buffer_impl(oneapi::dpl::__internal::__omp_backend_tag, std::size_t __n)
+        : __allocator_(), __ptr_(__allocator_.allocate(__n)), __buf_size_(__n)
+    {
+    }
 
     operator bool() const { return __ptr_ != nullptr; }
 
@@ -82,9 +86,8 @@ class __buffer_impl
     ~__buffer_impl() { __allocator_.deallocate(__ptr_, __buf_size_); }
 };
 
-template <class _Tag, typename _ExecutionPolicy, typename _Tp>
-using __buffer = ::std::enable_if_t<::std::is_same_v<_Tag, oneapi::dpl::__internal::__omp_backend_tag>,
-                                    __buffer_impl<::std::decay_t<_ExecutionPolicy>, _Tp>>;
+template <typename _Tp>
+using __buffer = __buffer_impl<oneapi::dpl::__internal::__omp_backend_tag, _Tp>;
 
 // Preliminary size of each chunk: requires further discussion
 constexpr std::size_t __default_chunk_size = 2048;

From 60b2a864bd73b57ef1c3819ecec4fa084c6c7607 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 1 Mar 2024 14:34:42 +0100
Subject: [PATCH 503/566] include/oneapi/dpl/pstl/parallel_backend_serial.h -
 rename __buffer_impl to __buffer + _ExecutionPolicy ->
 oneapi::dpl::__internal::__serial_backend_tag

---
 .../oneapi/dpl/pstl/parallel_backend_serial.h | 19 +++++++++++--------
 1 file changed, 11 insertions(+), 8 deletions(-)

diff --git a/include/oneapi/dpl/pstl/parallel_backend_serial.h b/include/oneapi/dpl/pstl/parallel_backend_serial.h
index a3f0ce2e38f..23cf5ca9805 100644
--- a/include/oneapi/dpl/pstl/parallel_backend_serial.h
+++ b/include/oneapi/dpl/pstl/parallel_backend_serial.h
@@ -32,8 +32,11 @@ namespace dpl
 namespace __serial_backend
 {
 
-template <typename _ExecutionPolicy, typename _Tp>
-class __buffer_impl
+template <typename _BackendTag, typename _Tp>
+class __buffer_impl;
+
+template <typename _Tp>
+class __buffer_impl<oneapi::dpl::__internal::__serial_backend_tag, _Tp>
 {
     ::std::allocator<_Tp> __allocator_;
     _Tp* __ptr_;
@@ -43,9 +46,10 @@ class __buffer_impl
     operator=(const __buffer_impl&) = delete;
 
   public:
-    static_assert(::std::is_same_v<_ExecutionPolicy, ::std::decay_t<_ExecutionPolicy>>);
-
-    __buffer_impl(::std::size_t __n) : __allocator_(), __ptr_(__allocator_.allocate(__n)), __buf_size_(__n) {}
+    __buffer_impl(oneapi::dpl::__internal::__serial_backend_tag, ::std::size_t __n)
+        : __allocator_(), __ptr_(__allocator_.allocate(__n)), __buf_size_(__n)
+    {
+    }
 
     operator bool() const { return __ptr_ != nullptr; }
     _Tp*
@@ -56,9 +60,8 @@ class __buffer_impl
     ~__buffer_impl() { __allocator_.deallocate(__ptr_, __buf_size_); }
 };
 
-template <class _Tag, typename _ExecutionPolicy, typename _Tp>
-using __buffer = ::std::enable_if_t<::std::is_same_v<_Tag, oneapi::dpl::__internal::__serial_backend_tag>,
-                                    __buffer_impl<::std::decay_t<_ExecutionPolicy>, _Tp>>;
+template <typename _Tp>
+using __buffer = __buffer_impl<oneapi::dpl::__internal::__serial_backend_tag, _Tp>;
 
 inline void
 __cancel_execution(oneapi::dpl::__internal::__serial_backend_tag)

From ed68f27d911eaf73723caa402711849f004a90ae Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 1 Mar 2024 14:35:32 +0100
Subject: [PATCH 504/566] include/oneapi/dpl/pstl/parallel_backend_tbb.h -
 rename __buffer_impl to __buffer + _ExecutionPolicy ->
 oneapi::dpl::__internal::__tbb_backend_tag

---
 .../oneapi/dpl/pstl/parallel_backend_tbb.h    | 23 +++++++++++--------
 1 file changed, 13 insertions(+), 10 deletions(-)

diff --git a/include/oneapi/dpl/pstl/parallel_backend_tbb.h b/include/oneapi/dpl/pstl/parallel_backend_tbb.h
index 4292f0ec405..80421a32109 100644
--- a/include/oneapi/dpl/pstl/parallel_backend_tbb.h
+++ b/include/oneapi/dpl/pstl/parallel_backend_tbb.h
@@ -49,13 +49,16 @@ namespace dpl
 namespace __tbb_backend
 {
 
+template <typename _BackendTag, typename _Tp>
+class __buffer_impl;
+
 //! Raw memory buffer with automatic freeing and no exceptions.
 /** Some of our algorithms need to start with raw memory buffer,
 not an initialize array, because initialization/destruction
 would make the span be at least O(N). */
 // tbb::allocator can improve performance in some cases.
-template <typename _ExecutionPolicy, typename _Tp>
-class __buffer_impl
+template <typename _Tp>
+class __buffer_impl<oneapi::dpl::__internal::__tbb_backend_tag, _Tp>
 {
     tbb::tbb_allocator<_Tp> _M_allocator;
     _Tp* _M_ptr;
@@ -65,10 +68,12 @@ class __buffer_impl
     operator=(const __buffer_impl&) = delete;
 
   public:
-    static_assert(::std::is_same_v<_ExecutionPolicy, ::std::decay_t<_ExecutionPolicy>>);
 
     //! Try to obtain buffer of given size to store objects of _Tp type
-    __buffer_impl(const ::std::size_t __n) : _M_allocator(), _M_ptr(_M_allocator.allocate(__n)), _M_buf_size(__n) {}
+    __buffer_impl(oneapi::dpl::__internal::__tbb_backend_tag, const ::std::size_t __n)
+        : _M_allocator(), _M_ptr(_M_allocator.allocate(__n)), _M_buf_size(__n)
+    {
+    }
     //! True if buffer was successfully obtained, zero otherwise.
     operator bool() const { return _M_ptr != nullptr; }
     //! Return pointer to buffer, or nullptr if buffer could not be obtained.
@@ -81,9 +86,8 @@ class __buffer_impl
     ~__buffer_impl() { _M_allocator.deallocate(_M_ptr, _M_buf_size); }
 };
 
-template <class _Tag, typename _ExecutionPolicy, typename _Tp>
-using __buffer = ::std::enable_if_t<::std::is_same_v<_Tag, oneapi::dpl::__internal::__tbb_backend_tag>,
-                                    __buffer_impl<::std::decay_t<_ExecutionPolicy>, _Tp>>;
+template <typename _Tp>
+using __buffer = __buffer_impl<oneapi::dpl::__internal::__tbb_backend_tag, _Tp>;
 
 // Wrapper for tbb::task
 inline void
@@ -391,7 +395,7 @@ __parallel_strict_scan(oneapi::dpl::__internal::__tbb_backend_tag, _ExecutionPol
             const _Index __slack = 4;
             _Index __tilesize = (__n - 1) / (__slack * __p) + 1;
             _Index __m = (__n - 1) / __tilesize;
-            __tbb_backend::__buffer<oneapi::dpl::__internal::__tbb_backend_tag, _ExecutionPolicy, _Tp> __buf(__m + 1);
+            __tbb_backend::__buffer<_Tp> __buf(oneapi::dpl::__internal::__tbb_backend_tag{}, __m + 1);
             _Tp* __r = __buf.get();
             __tbb_backend::__upsweep(_Index(0), _Index(__m + 1), __tilesize, __r, __n - __m * __tilesize, __reduce,
                                      __combine);
@@ -1196,8 +1200,7 @@ __parallel_stable_sort(oneapi::dpl::__internal::__tbb_backend_tag, _ExecutionPol
         const _DifferenceType __sort_cut_off = _ONEDPL_STABLE_SORT_CUT_OFF;
         if (__n > __sort_cut_off)
         {
-            __tbb_backend::__buffer<oneapi::dpl::__internal::__tbb_backend_tag, _ExecutionPolicy, _ValueType> __buf(
-                __n);
+            __tbb_backend::__buffer<_ValueType> __buf(oneapi::dpl::__internal::__tbb_backend_tag{}, __n);
             __root_task<__stable_sort_func<_RandomAccessIterator, _ValueType*, _Compare, _LeafSort>> __root{
                 __xs, __xe, __buf.get(), true, __comp, __leaf_sort, __nsort, __xs, __buf.get()};
             __task::spawn_root_and_wait(__root);

From 09611bbcaf4886d729b4ed2348f2e1317ae323a1 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 1 Mar 2024 14:27:21 +0100
Subject: [PATCH 505/566] __par_backend::__buffer usage

---
 .../internal/exclusive_scan_by_segment_impl.h |  8 +++----
 .../internal/inclusive_scan_by_segment_impl.h |  4 ++--
 .../dpl/internal/reduce_by_segment_impl.h     | 20 ++++++----------
 .../dpl/internal/scan_by_segment_impl.h       |  7 ++----
 include/oneapi/dpl/pstl/algorithm_impl.h      | 20 ++++++++--------
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 24 +++++++++----------
 .../hetero/algorithm_ranges_impl_hetero.h     | 18 +++++---------
 .../pstl/hetero/dpcpp/parallel_backend_sycl.h |  8 +++----
 .../dpcpp/parallel_backend_sycl_histogram.h   |  4 +---
 .../dpcpp/parallel_backend_sycl_radix_sort.h  |  5 ++--
 .../dpl/pstl/hetero/numeric_impl_hetero.h     |  2 +-
 include/oneapi/dpl/pstl/omp/parallel_scan.h   |  2 +-
 12 files changed, 50 insertions(+), 72 deletions(-)

diff --git a/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h b/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h
index a2c0eaebc4b..e1496d38e13 100644
--- a/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h
@@ -66,7 +66,7 @@ pattern_exclusive_scan_by_segment(_Tag, Policy&& policy, InputIterator1 first1,
     InputIterator2 last2 = first2 + n;
 
     // compute head flags
-    oneapi::dpl::__par_backend::__buffer<__backend_tag, Policy, FlagType> _flags(n);
+    oneapi::dpl::__par_backend::__buffer<FlagType> _flags(__backend_tag{}, n);
     auto flags = _flags.get();
     flags[0] = 1;
 
@@ -74,7 +74,7 @@ pattern_exclusive_scan_by_segment(_Tag, Policy&& policy, InputIterator1 first1,
               oneapi::dpl::__internal::__not_pred<BinaryPredicate>(binary_pred));
 
     // shift input one to the right and initialize segments with init
-    oneapi::dpl::__par_backend::__buffer<__backend_tag, Policy, OutputType> _temp(n);
+    oneapi::dpl::__par_backend::__buffer<OutputType> _temp(__backend_tag{}, n);
     auto temp = _temp.get();
 
     temp[0] = init;
@@ -125,7 +125,7 @@ exclusive_scan_by_segment_impl(__internal::__hetero_tag<_BackendTag>, Policy&& p
     InputIterator2 last2 = first2 + n;
 
     // compute head flags
-    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, Policy, FlagType> _flags(policy, n);
+    oneapi::dpl::__par_backend_hetero::__buffer<FlagType> _flags(_BackendTag{}, n);
     {
         auto flag_buf = _flags.get_buffer();
         auto flags = flag_buf.get_host_access(sycl::read_write);
@@ -136,7 +136,7 @@ exclusive_scan_by_segment_impl(__internal::__hetero_tag<_BackendTag>, Policy&& p
               oneapi::dpl::__internal::__not_pred<BinaryPredicate>(binary_pred));
 
     // shift input one to the right and initialize segments with init
-    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, Policy, OutputType> _temp(policy, n);
+    oneapi::dpl::__par_backend_hetero::__buffer<OutputType> _temp(_BackendTag{}, n);
     {
         auto temp_buf = _temp.get_buffer();
         auto temp = temp_buf.get_host_access(sycl::read_write);
diff --git a/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h b/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h
index f4d2b8d6a33..7dd345ba534 100644
--- a/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h
@@ -61,7 +61,7 @@ pattern_inclusive_scan_by_segment(_Tag, Policy&& policy, InputIterator1 first1,
     typedef unsigned int FlagType;
     typedef typename ::std::iterator_traits<InputIterator2>::value_type ValueType;
 
-    oneapi::dpl::__par_backend::__buffer<__backend_tag, Policy, FlagType> _mask(n);
+    oneapi::dpl::__par_backend::__buffer<FlagType> _mask(__backend_tag{}, n);
     auto mask = _mask.get();
 
     mask[0] = 1;
@@ -111,7 +111,7 @@ inclusive_scan_by_segment_impl(__internal::__hetero_tag<_BackendTag>, Policy&& p
 
     FlagType initial_mask = 1;
 
-    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, Policy, FlagType> _mask(policy, n);
+    oneapi::dpl::__par_backend_hetero::__buffer<FlagType> _mask(_BackendTag{}, n);
     {
         auto mask_buf = _mask.get_buffer();
         auto mask = mask_buf.get_host_access(sycl::read_write);
diff --git a/include/oneapi/dpl/internal/reduce_by_segment_impl.h b/include/oneapi/dpl/internal/reduce_by_segment_impl.h
index 37a7aad1379..0c879f71978 100644
--- a/include/oneapi/dpl/internal/reduce_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/reduce_by_segment_impl.h
@@ -116,7 +116,7 @@ reduce_by_segment_impl(_Tag, Policy&& policy, InputIterator1 first1, InputIterat
 
     // buffer that is used to store a flag indicating if the associated key is not equal to
     // the next key, and thus its associated sum should be part of the final result
-    oneapi::dpl::__par_backend::__buffer<__backend_tag, Policy, FlagType> _mask(n + 1);
+    oneapi::dpl::__par_backend::__buffer<FlagType> _mask(__backend_tag{}, n + 1);
     auto mask = _mask.get();
     mask[0] = 1;
 
@@ -132,11 +132,11 @@ reduce_by_segment_impl(_Tag, Policy&& policy, InputIterator1 first1, InputIterat
     // buffer stores the sums of values associated with a given key. Sums are copied with
     // a shift into result2, and the shift is computed at the same time as the sums, so the
     // sums can't be written to result2 directly.
-    oneapi::dpl::__par_backend::__buffer<__backend_tag, Policy, ValueType> _scanned_values(n);
+    oneapi::dpl::__par_backend::__buffer<ValueType> _scanned_values(__backend_tag{}, n);
 
     // Buffer is used to store results of the scan of the mask. Values indicate which position
     // in result2 needs to be written with the scanned_values element.
-    oneapi::dpl::__par_backend::__buffer<__backend_tag, Policy, FlagType> _scanned_tail_flags(n);
+    oneapi::dpl::__par_backend::__buffer<FlagType> _scanned_tail_flags(__backend_tag{}, n);
 
     // Compute the sum of the segments. scanned_tail_flags values are not used.
     inclusive_scan(policy, make_zip_iterator(first2, _mask.get()), make_zip_iterator(first2, _mask.get()) + n,
@@ -264,22 +264,16 @@ __sycl_reduce_by_segment(__internal::__hetero_tag<_BackendTag>, _ExecutionPolicy
     ::std::size_t __n_groups = oneapi::dpl::__internal::__dpl_ceiling_div(__n, __wgroup_size * __vals_per_item);
 
     // intermediate reductions within a workgroup
-    auto __partials =
-        oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, __val_type>(__exec, __n_groups)
-            .get_buffer();
+    auto __partials = oneapi::dpl::__par_backend_hetero::__buffer<__val_type>(_BackendTag{}, __n_groups).get_buffer();
 
-    auto __end_idx =
-        oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, __diff_type>(__exec, 1).get_buffer();
+    auto __end_idx = oneapi::dpl::__par_backend_hetero::__buffer<__diff_type>(_BackendTag{}, 1).get_buffer();
 
     // the number of segment ends found in each work group
-    auto __seg_ends =
-        oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, __diff_type>(__exec, __n_groups)
-            .get_buffer();
+    auto __seg_ends = oneapi::dpl::__par_backend_hetero::__buffer<__diff_type>(_BackendTag{}, __n_groups).get_buffer();
 
     // buffer that stores an exclusive scan of the results
     auto __seg_ends_scanned =
-        oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, __diff_type>(__exec, __n_groups)
-            .get_buffer();
+        oneapi::dpl::__par_backend_hetero::__buffer<__diff_type>(_BackendTag{}, __n_groups).get_buffer();
 
     // 1. Count the segment ends in each workgroup
     auto __seg_end_identification = __exec.queue().submit([&](sycl::handler& __cgh) {
diff --git a/include/oneapi/dpl/internal/scan_by_segment_impl.h b/include/oneapi/dpl/internal/scan_by_segment_impl.h
index 85903a93fdf..0df27861946 100644
--- a/include/oneapi/dpl/internal/scan_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/scan_by_segment_impl.h
@@ -145,13 +145,10 @@ struct __sycl_scan_by_segment_impl
         ::std::size_t __n_groups = __internal::__dpl_ceiling_div(__n, __wgroup_size * __vals_per_item);
 
         auto __partials =
-            oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, __val_type>(__exec, __n_groups)
-                .get_buffer();
+            oneapi::dpl::__par_backend_hetero::__buffer<__val_type>(_BackendTag{}, __n_groups).get_buffer();
 
         // the number of segment ends found in each work group
-        auto __seg_ends =
-            oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, bool>(__exec, __n_groups)
-                .get_buffer();
+        auto __seg_ends = oneapi::dpl::__par_backend_hetero::__buffer<bool>(_BackendTag{}, __n_groups).get_buffer();
 
         // 1. Work group reduction
         auto __wg_scan = __exec.queue().submit([&](sycl::handler& __cgh) {
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index ee475895a70..dff58a82e02 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -1346,7 +1346,7 @@ __pattern_copy_if(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomA
     const _DifferenceType __n = __last - __first;
     if (_DifferenceType(1) < __n)
     {
-        __par_backend::__buffer<__backend_tag, _ExecutionPolicy, bool> __mask_buf(__n);
+        __par_backend::__buffer<bool> __mask_buf(__backend_tag{}, __n);
         return __internal::__except_handler([&__exec, __n, __first, __result, __pred, &__mask_buf]() {
             bool* __mask = __mask_buf.get();
             _DifferenceType __m{};
@@ -1467,7 +1467,7 @@ __remove_elements(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Forward
     typedef typename ::std::iterator_traits<_ForwardIterator>::difference_type _DifferenceType;
     typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _Tp;
     _DifferenceType __n = __last - __first;
-    __par_backend::__buffer<__backend_tag, _ExecutionPolicy, bool> __mask_buf(__n);
+    __par_backend::__buffer<bool> __mask_buf(__backend_tag{}, __n);
     // 1. find a first iterator that should be removed
     return __internal::__except_handler([&]() {
         bool* __mask = __mask_buf.get();
@@ -1504,7 +1504,7 @@ __remove_elements(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Forward
         __n -= __min;
         __first += __min;
 
-        __par_backend::__buffer<__backend_tag, _ExecutionPolicy, _Tp> __buf(__n);
+        __par_backend::__buffer<_Tp> __buf(__backend_tag{}, __n);
         _Tp* __result = __buf.get();
         __mask += __min;
         _DifferenceType __m{};
@@ -1633,7 +1633,7 @@ __pattern_unique_copy(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Ran
     const _DifferenceType __n = __last - __first;
     if (_DifferenceType(2) < __n)
     {
-        __par_backend::__buffer<__backend_tag, _ExecutionPolicy, bool> __mask_buf(__n);
+        __par_backend::__buffer<bool> __mask_buf(__backend_tag{}, __n);
         if (_DifferenceType(2) < __n)
         {
             return __internal::__except_handler([&__exec, __n, __first, __result, __pred, &__mask_buf]() {
@@ -1877,7 +1877,7 @@ __pattern_rotate(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAc
     auto __m = __middle - __first;
     if (__m <= __n / 2)
     {
-        __par_backend::__buffer<__backend_tag, _ExecutionPolicy, _Tp> __buf(__n - __m);
+        __par_backend::__buffer<_Tp> __buf(__backend_tag{}, __n - __m);
         return __internal::__except_handler([&__exec, __n, __m, __first, __middle, __last, &__buf]() {
             _Tp* __result = __buf.get();
             __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __middle, __last,
@@ -1903,7 +1903,7 @@ __pattern_rotate(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAc
     }
     else
     {
-        __par_backend::__buffer<__backend_tag, _ExecutionPolicy, _Tp> __buf(__m);
+        __par_backend::__buffer<_Tp> __buf(__backend_tag{}, __m);
         return __internal::__except_handler([&__exec, __n, __m, __first, __middle, __last, &__buf]() {
             _Tp* __result = __buf.get();
             __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __middle,
@@ -2386,7 +2386,7 @@ __pattern_partition_copy(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _
     const _DifferenceType __n = __last - __first;
     if (_DifferenceType(1) < __n)
     {
-        __par_backend::__buffer<__backend_tag, _ExecutionPolicy, bool> __mask_buf(__n);
+        __par_backend::__buffer<bool> __mask_buf(__backend_tag{}, __n);
         return __internal::__except_handler([&__exec, __n, __first, __out_true, __out_false, __pred, &__mask_buf]() {
             bool* __mask = __mask_buf.get();
             _ReturnType __m{};
@@ -2617,7 +2617,7 @@ __pattern_partial_sort_copy(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec
         {
             typedef typename ::std::iterator_traits<_RandomAccessIterator1>::value_type _T1;
             typedef typename ::std::iterator_traits<_RandomAccessIterator2>::value_type _T2;
-            __par_backend::__buffer<__backend_tag, _ExecutionPolicy, _T1> __buf(__n1);
+            __par_backend::__buffer<_T1> __buf(__backend_tag{}, __n1);
             _T1* __r = __buf.get();
 
             __par_backend::__parallel_stable_sort(
@@ -3122,7 +3122,7 @@ __pattern_inplace_merge(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _R
 
     typedef typename ::std::iterator_traits<_RandomAccessIterator>::value_type _Tp;
     auto __n = __last - __first;
-    __par_backend::__buffer<__backend_tag, _ExecutionPolicy, _Tp> __buf(__n);
+    __par_backend::__buffer<_Tp> __buf(__backend_tag{}, __n);
     _Tp* __r = __buf.get();
     __internal::__except_handler([&]() {
         auto __move_values = [](_RandomAccessIterator __x, _Tp* __z) {
@@ -3257,7 +3257,7 @@ __parallel_set_op(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Forward
     const _DifferenceType __n1 = __last1 - __first1;
     const _DifferenceType __n2 = __last2 - __first2;
 
-    __par_backend::__buffer<__backend_tag, _ExecutionPolicy, _T> __buf(__size_func(__n1, __n2));
+    __par_backend::__buffer<_T> __buf(__backend_tag{}, __size_func(__n1, __n2));
 
     return __internal::__except_handler([&__exec, __n1, __first1, __last1, __first2, __last2, __result, __comp,
                                          __size_func, __set_op, &__buf]() {
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 6a42ad15792..64e8316d88a 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -1032,8 +1032,7 @@ __pattern_remove_if(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec,
 
     using _ValueType = typename ::std::iterator_traits<_Iterator>::value_type;
 
-    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _ValueType> __buf(__exec,
-                                                                                                 __last - __first);
+    oneapi::dpl::__par_backend_hetero::__buffer<_ValueType> __buf(_BackendTag{}, __last - __first);
     auto __copy_first = __buf.get();
 
     auto __copy_last = __pattern_copy_if(__tag, __exec, __first, __last, __copy_first, __not_pred<_Predicate>{__pred});
@@ -1054,8 +1053,7 @@ __pattern_unique(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _It
 
     using _ValueType = typename ::std::iterator_traits<_Iterator>::value_type;
 
-    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _ValueType> __buf(__exec,
-                                                                                                 __last - __first);
+    oneapi::dpl::__par_backend_hetero::__buffer<_ValueType> __buf(_BackendTag{}, __last - __first);
     auto __copy_first = __buf.get();
     auto __copy_last = __pattern_unique_copy(__tag, __exec, __first, __last, __copy_first, __pred);
 
@@ -1227,7 +1225,7 @@ __pattern_inplace_merge(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __ex
     assert(__first < __middle && __middle < __last);
 
     auto __n = __last - __first;
-    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _ValueType> __buf(__exec, __n);
+    oneapi::dpl::__par_backend_hetero::__buffer<_ValueType> __buf(_BackendTag{}, __n);
     auto __copy_first = __buf.get();
     auto __copy_last = __copy_first + __n;
 
@@ -1315,8 +1313,8 @@ __pattern_stable_partition(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& _
 
     auto __n = __last - __first;
 
-    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _ValueType> __true_buf(__exec, __n);
-    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _ValueType> __false_buf(__exec, __n);
+    oneapi::dpl::__par_backend_hetero::__buffer<_ValueType> __true_buf(_BackendTag{}, __n);
+    oneapi::dpl::__par_backend_hetero::__buffer<_ValueType> __false_buf(_BackendTag{}, __n);
     auto __true_result = __true_buf.get();
     auto __false_result = __false_buf.get();
 
@@ -1512,7 +1510,7 @@ __pattern_partial_sort_copy(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&&
         // - create a temporary buffer and copy all the elements from the input buffer there
         // - run partial sort on the temporary buffer
         // - copy k elements from the temporary buffer to the output buffer.
-        oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _ValueType> __buf(__exec, __in_size);
+        oneapi::dpl::__par_backend_hetero::__buffer<_ValueType> __buf(_BackendTag{}, __in_size);
 
         auto __buf_first = __buf.get();
 
@@ -1629,7 +1627,7 @@ __pattern_rotate(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Iterator
 
     auto __keep = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read_write, _Iterator>();
     auto __buf = __keep(__first, __last);
-    auto __temp_buf = oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _Tp>(__exec, __n);
+    auto __temp_buf = oneapi::dpl::__par_backend_hetero::__buffer<_Tp>(_BackendTag{}, __n);
 
     auto __temp_rng =
         oneapi::dpl::__ranges::all_view<_Tp, __par_backend_hetero::access_mode::write>(__temp_buf.get_buffer());
@@ -1712,7 +1710,7 @@ __pattern_hetero_set_op(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _F
         __comp, __n1, __n2};
 
     // temporary buffer to store boolean mask
-    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, int32_t> __mask_buf(__exec, __n1);
+    oneapi::dpl::__par_backend_hetero::__buffer<int32_t> __mask_buf(_BackendTag{}, __n1);
 
     auto __keep1 =
         oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _ForwardIterator1>();
@@ -1838,7 +1836,7 @@ __pattern_set_union(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec,
 
     // temporary buffer to store intermediate result
     const auto __n2 = __last2 - __first2;
-    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _ValueType> __diff(__exec, __n2);
+    oneapi::dpl::__par_backend_hetero::__buffer<_ValueType> __diff(_BackendTag{}, __n2);
     auto __buf = __diff.get();
 
     //1. Calc difference {2} \ {1}
@@ -1917,10 +1915,10 @@ __pattern_set_symmetric_difference(__hetero_tag<_BackendTag> __tag, _ExecutionPo
 
     // temporary buffers to store intermediate result
     const auto __n1 = __last1 - __first1;
-    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _ValueType> __diff_1(__exec, __n1);
+    oneapi::dpl::__par_backend_hetero::__buffer<_ValueType> __diff_1(_BackendTag{}, __n1);
     auto __buf_1 = __diff_1.get();
     const auto __n2 = __last2 - __first2;
-    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _ValueType> __diff_2(__exec, __n2);
+    oneapi::dpl::__par_backend_hetero::__buffer<_ValueType> __diff_2(_BackendTag{}, __n2);
     auto __buf_2 = __diff_2.get();
 
     //1. Calc difference {1} \ {2}
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
index 45863f75615..7071d49d066 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
@@ -360,8 +360,7 @@ __pattern_scan_copy(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Range
     _DataAcc __get_data_op;
     _MaskAssigner __add_mask_op;
 
-    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, int32_t> __mask_buf(__exec,
-                                                                                                   __rng1.size());
+    oneapi::dpl::__par_backend_hetero::__buffer<int32_t> __mask_buf(_BackendTag{}, __rng1.size());
 
     auto __res =
         __par_backend_hetero::__parallel_transform_scan_base(
@@ -414,7 +413,7 @@ __pattern_remove_if(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec,
 
     using _ValueType = oneapi::dpl::__internal::__value_t<_Range>;
 
-    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _ValueType> __buf(__exec, __rng.size());
+    oneapi::dpl::__par_backend_hetero::__buffer<_ValueType> __buf(_BackendTag{}, __rng.size());
     auto __copy_rng = oneapi::dpl::__ranges::views::all(__buf.get_buffer());
 
     auto __copy_last_id = __ranges::__pattern_copy_if(__tag, __exec, __rng, __copy_rng, __not_pred<_Predicate>{__pred},
@@ -462,7 +461,7 @@ __pattern_unique(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Ra
 
     using _ValueType = oneapi::dpl::__internal::__value_t<_Range>;
 
-    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _ValueType> __buf(__exec, __rng.size());
+    oneapi::dpl::__par_backend_hetero::__buffer<_ValueType> __buf(_BackendTag{}, __rng.size());
     auto res_rng = oneapi::dpl::__ranges::views::all(__buf.get_buffer());
     auto res = __ranges::__pattern_unique_copy(__tag, __exec, __rng, res_rng, __pred,
                                                oneapi::dpl::__internal::__pstl_assign());
@@ -715,14 +714,9 @@ __pattern_reduce_by_segment(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&&
     // Round 1: reduce with extra indices added to avoid long segments
     // TODO: At threshold points check if the key is equal to the key at the previous threshold point, indicating a long sequence.
     // Skip a round of copy_if and reduces if there are none.
-    auto __idx = oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, __diff_type>(__exec, __n)
-                     .get_buffer();
-    auto __tmp_out_keys =
-        oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, __key_type>(__exec, __n)
-            .get_buffer();
-    auto __tmp_out_values =
-        oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, __val_type>(__exec, __n)
-            .get_buffer();
+    auto __idx = oneapi::dpl::__par_backend_hetero::__buffer<__diff_type>(_BackendTag{}, __n).get_buffer();
+    auto __tmp_out_keys = oneapi::dpl::__par_backend_hetero::__buffer<__key_type>(_BackendTag{}, __n).get_buffer();
+    auto __tmp_out_values = oneapi::dpl::__par_backend_hetero::__buffer<__val_type>(_BackendTag{}, __n).get_buffer();
 
     // Replicating first element of keys view to be able to compare (i-1)-th and (i)-th key with aligned sequences,
     //  dropping the last key for the i-1 sequence.
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
index d2a935a8463..f4d938d1011 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
@@ -871,9 +871,7 @@ __parallel_scan_copy(oneapi::dpl::__internal::__device_backend_tag __backend_tag
     _MaskAssigner __add_mask_op;
 
     // temporary buffer to store boolean mask
-    oneapi::dpl::__par_backend_hetero::__buffer<oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy,
-                                                int32_t>
-        __mask_buf(__exec, __n);
+    oneapi::dpl::__par_backend_hetero::__buffer<int32_t> __mask_buf(oneapi::dpl::__internal::__device_backend_tag{}, __n);
 
     return __parallel_transform_scan_base(
         __backend_tag, ::std::forward<_ExecutionPolicy>(__exec),
@@ -1581,7 +1579,7 @@ struct __parallel_sort_submitter<_BackendTag, _IdType, __internal::__optional_ke
         });
 
         // 2. Merge sorting
-        oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _Tp> __temp_buf(__exec, __n);
+        oneapi::dpl::__par_backend_hetero::__buffer<_Tp> __temp_buf(_BackendTag{}, __n);
         auto __temp = __temp_buf.get_buffer();
         bool __data_in_temp = false;
         _IdType __n_sorted = __leaf;
@@ -1702,7 +1700,7 @@ struct __parallel_partial_sort_submitter<_BackendTag, __internal::__optional_ker
         _Size __n = __rng.size();
         assert(__n > 1);
 
-        oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _Tp> __temp_buf(__exec, __n);
+        oneapi::dpl::__par_backend_hetero::__buffer<_Tp> __temp_buf(_BackendTag{}, __n);
         auto __temp = __temp_buf.get_buffer();
         _PRINT_INFO_IN_DEBUG_MODE(__exec);
 
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h
index 083bffcefc0..ef15febc4b6 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h
@@ -427,9 +427,7 @@ struct __histogram_general_private_global_atomics_submitter<_BackendTag,
             oneapi::dpl::__internal::__dpl_ceiling_div(__n, __work_group_size * __iters_per_work_item);
 
         auto __private_histograms =
-            oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _bin_type>(
-                __exec, __segments * __num_bins)
-                .get_buffer();
+            oneapi::dpl::__par_backend_hetero::__buffer<_bin_type>(_BackendTag{}, __segments * __num_bins).get_buffer();
 
         return __exec.queue().submit([&](auto& __h) {
             __h.depends_on(__init_event);
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_radix_sort.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_radix_sort.h
index 221031e4ac8..d16610390f6 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_radix_sort.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_radix_sort.h
@@ -813,9 +813,8 @@ __parallel_radix_sort(oneapi::dpl::__internal::__device_backend_tag, _ExecutionP
         sycl::buffer<::std::uint32_t, 1> __tmp_buf{sycl::range<1>(__tmp_buf_size)};
 
         // memory for storing values sorted for an iteration
-        oneapi::dpl::__par_backend_hetero::__buffer<oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy,
-                                                    _ValueT>
-            __out_buffer_holder{__exec, __n};
+        oneapi::dpl::__par_backend_hetero::__buffer<_ValueT> __out_buffer_holder{
+            oneapi::dpl::__internal::__device_backend_tag{}, __n};
         auto __out_rng = oneapi::dpl::__ranges::all_view<_ValueT, __par_backend_hetero::access_mode::read_write>(
             __out_buffer_holder.get_buffer());
 
diff --git a/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
index 216c84820f2..90e477eb188 100644
--- a/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
@@ -164,7 +164,7 @@ __pattern_transform_scan_base(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&
         using _NewExecutionPolicy = decltype(__policy);
 
         // Create temporary buffer
-        oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _NewExecutionPolicy, _Type> __tmp_buf(__policy, __n);
+        oneapi::dpl::__par_backend_hetero::__buffer<_Type> __tmp_buf(_BackendTag{}, __n);
         auto __first_tmp = __tmp_buf.get();
         auto __last_tmp = __first_tmp + __n;
         auto __keep2 = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::write, _Iterator2>();
diff --git a/include/oneapi/dpl/pstl/omp/parallel_scan.h b/include/oneapi/dpl/pstl/omp/parallel_scan.h
index 90e0a150152..c316ed3a104 100644
--- a/include/oneapi/dpl/pstl/omp/parallel_scan.h
+++ b/include/oneapi/dpl/pstl/omp/parallel_scan.h
@@ -88,7 +88,7 @@ __parallel_strict_scan_body(_Index __n, _Tp __initial, _Rp __reduce, _Cp __combi
     const _Index __slack = 4;
     _Index __tilesize = (__n - 1) / (__slack * __p) + 1;
     _Index __m = (__n - 1) / __tilesize;
-    __buffer<oneapi::dpl::__internal::__omp_backend_tag, _ExecutionPolicy, _Tp> __buf(__m + 1);
+    __buffer<_Tp> __buf(oneapi::dpl::__internal::__omp_backend_tag{}, __m + 1);
     _Tp* __r = __buf.get();
 
     oneapi::dpl::__omp_backend::__upsweep(_Index(0), _Index(__m + 1), __tilesize, __r, __n - __m * __tilesize, __reduce,

From 70b983aa0773bbda32d9dc399a921858b224c745 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 1 Mar 2024 16:40:49 +0100
Subject: [PATCH 506/566] 
 include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h - fix review
 comment: Do we want static_assert instead of enable_if?

---
 include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h b/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h
index 01589e1ff20..80a19db8130 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h
@@ -297,9 +297,10 @@ struct __device_backend_tag
 };
 
 template <class... _IteratorTypes, typename _KernelName>
-::std::enable_if_t<__is_random_access_iterator_v<_IteratorTypes...>, __hetero_tag<__device_backend_tag>>
+__hetero_tag<__device_backend_tag>
 __select_backend(const execution::device_policy<_KernelName>&, _IteratorTypes&&...)
 {
+    static_assert(__is_random_access_iterator_v<_IteratorTypes...>);
     return {};
 }
 
@@ -309,9 +310,10 @@ struct __fpga_backend_tag : __device_backend_tag
 };
 
 template <class... _IteratorTypes, unsigned int _Factor, typename _KernelName>
-::std::enable_if_t<__is_random_access_iterator_v<_IteratorTypes...>, __hetero_tag<__fpga_backend_tag>>
+__hetero_tag<__fpga_backend_tag>
 __select_backend(const execution::fpga_policy<_Factor, _KernelName>&, _IteratorTypes&&...)
 {
+    static_assert(__is_random_access_iterator_v<_IteratorTypes...>);
     return {};
 }
 #endif

From b69c31081c09f9373c0254c05ea0bfdb9d15b125 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 5 Mar 2024 09:09:18 +0100
Subject: [PATCH 507/566] Fix review comment: What's the point of reading the
 __backend_tag, while we already have it deduced as __BackendTag.

---
 include/oneapi/dpl/internal/binary_search_impl.h     | 12 +++---------
 .../oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h   |  4 +---
 include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h |  6 ++----
 3 files changed, 6 insertions(+), 16 deletions(-)

diff --git a/include/oneapi/dpl/internal/binary_search_impl.h b/include/oneapi/dpl/internal/binary_search_impl.h
index 92ffd046be9..29c30e79e2c 100644
--- a/include/oneapi/dpl/internal/binary_search_impl.h
+++ b/include/oneapi/dpl/internal/binary_search_impl.h
@@ -117,8 +117,6 @@ OutputIterator
 lower_bound_impl(__internal::__hetero_tag<_BackendTag>, Policy&& policy, InputIterator1 start, InputIterator1 end,
                  InputIterator2 value_start, InputIterator2 value_end, OutputIterator result, StrictWeakOrdering comp)
 {
-    using __backend_tag = typename __internal::__hetero_tag<_BackendTag>::__backend_tag;
-
     namespace __bknd = __par_backend_hetero;
     const auto size = ::std::distance(start, end);
 
@@ -136,7 +134,7 @@ lower_bound_impl(__internal::__hetero_tag<_BackendTag>, Policy&& policy, InputIt
     auto keep_result = oneapi::dpl::__ranges::__get_sycl_range<__bknd::access_mode::read_write, OutputIterator>();
     auto result_buf = keep_result(result, result + value_size);
     auto zip_vw = make_zip_view(input_buf.all_view(), value_buf.all_view(), result_buf.all_view());
-    __bknd::__parallel_for(__backend_tag{}, ::std::forward<Policy>(policy),
+    __bknd::__parallel_for(_BackendTag{}, ::std::forward<Policy>(policy),
                            custom_brick<StrictWeakOrdering, decltype(size), lower_bound>{comp, size}, value_size,
                            zip_vw)
         .wait();
@@ -149,8 +147,6 @@ OutputIterator
 upper_bound_impl(__internal::__hetero_tag<_BackendTag>, Policy&& policy, InputIterator1 start, InputIterator1 end,
                  InputIterator2 value_start, InputIterator2 value_end, OutputIterator result, StrictWeakOrdering comp)
 {
-    using __backend_tag = typename __internal::__hetero_tag<_BackendTag>::__backend_tag;
-
     namespace __bknd = __par_backend_hetero;
     const auto size = ::std::distance(start, end);
 
@@ -168,7 +164,7 @@ upper_bound_impl(__internal::__hetero_tag<_BackendTag>, Policy&& policy, InputIt
     auto keep_result = oneapi::dpl::__ranges::__get_sycl_range<__bknd::access_mode::read_write, OutputIterator>();
     auto result_buf = keep_result(result, result + value_size);
     auto zip_vw = make_zip_view(input_buf.all_view(), value_buf.all_view(), result_buf.all_view());
-    __bknd::__parallel_for(__backend_tag{}, ::std::forward<Policy>(policy),
+    __bknd::__parallel_for(_BackendTag{}, ::std::forward<Policy>(policy),
                            custom_brick<StrictWeakOrdering, decltype(size), upper_bound>{comp, size}, value_size,
                            zip_vw)
         .wait();
@@ -181,8 +177,6 @@ OutputIterator
 binary_search_impl(__internal::__hetero_tag<_BackendTag>, Policy&& policy, InputIterator1 start, InputIterator1 end,
                    InputIterator2 value_start, InputIterator2 value_end, OutputIterator result, StrictWeakOrdering comp)
 {
-    using __backend_tag = typename __internal::__hetero_tag<_BackendTag>::__backend_tag;
-
     namespace __bknd = __par_backend_hetero;
     const auto size = ::std::distance(start, end);
 
@@ -200,7 +194,7 @@ binary_search_impl(__internal::__hetero_tag<_BackendTag>, Policy&& policy, Input
     auto keep_result = oneapi::dpl::__ranges::__get_sycl_range<__bknd::access_mode::read_write, OutputIterator>();
     auto result_buf = keep_result(result, result + value_size);
     auto zip_vw = make_zip_view(input_buf.all_view(), value_buf.all_view(), result_buf.all_view());
-    __bknd::__parallel_for(__backend_tag{}, ::std::forward<Policy>(policy),
+    __bknd::__parallel_for(_BackendTag{}, ::std::forward<Policy>(policy),
                            custom_brick<StrictWeakOrdering, decltype(size), binary_search>{comp, size}, value_size,
                            zip_vw)
         .wait();
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 64e8316d88a..f975af77358 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -1558,8 +1558,6 @@ template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator>
 void
 __pattern_reverse(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Iterator __first, _Iterator __last)
 {
-    using __backend_tag = typename __hetero_tag<_BackendTag>::__backend_tag;
-
     auto __n = __last - __first;
     if (__n <= 0)
         return;
@@ -1567,7 +1565,7 @@ __pattern_reverse(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Iterato
     auto __keep = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read_write, _Iterator>();
     auto __buf = __keep(__first, __last);
     oneapi::dpl::__par_backend_hetero::__parallel_for(
-        __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
+        _BackendTag{}, ::std::forward<_ExecutionPolicy>(__exec),
         unseq_backend::__reverse_functor<typename ::std::iterator_traits<_Iterator>::difference_type>{__n}, __n / 2,
         __buf.all_view())
         .wait();
diff --git a/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
index 90e477eb188..d8ea609c7e0 100644
--- a/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
@@ -129,8 +129,6 @@ __pattern_transform_scan_base(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&
                               _Iterator1 __last, _Iterator2 __result, _UnaryOperation __unary_op, _InitType __init,
                               _BinaryOperation __binary_op, _Inclusive)
 {
-    using __backend_tag = typename __hetero_tag<_BackendTag>::__backend_tag;
-
     if (__first == __last)
         return __result;
 
@@ -147,7 +145,7 @@ __pattern_transform_scan_base(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&
         auto __buf2 = __keep2(__result, __result + __n);
 
         oneapi::dpl::__par_backend_hetero::__parallel_transform_scan(
-            __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __buf1.all_view(), __buf2.all_view(), __n,
+            _BackendTag{}, ::std::forward<_ExecutionPolicy>(__exec), __buf1.all_view(), __buf2.all_view(), __n,
             __unary_op, __init, __binary_op, _Inclusive{})
             .wait();
     }
@@ -171,7 +169,7 @@ __pattern_transform_scan_base(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&
         auto __buf2 = __keep2(__first_tmp, __last_tmp);
 
         // Run main algorithm and save data into temporary buffer
-        oneapi::dpl::__par_backend_hetero::__parallel_transform_scan(__backend_tag{}, __policy, __buf1.all_view(),
+        oneapi::dpl::__par_backend_hetero::__parallel_transform_scan(_BackendTag{}, __policy, __buf1.all_view(),
                                                                      __buf2.all_view(), __n, __unary_op, __init,
                                                                      __binary_op, _Inclusive{})
             .wait();

From 23ef7d99223cb0b90e99f2d1b4baf67ffa4065cb Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 5 Mar 2024 09:14:27 +0100
Subject: [PATCH 508/566] Fix review comment: if if we imagine that we need
 that check, I don't feel we need std::decay

---
 .../oneapi/dpl/internal/binary_search_impl.h  |   6 +-
 .../internal/exclusive_scan_by_segment_impl.h |   2 +-
 .../internal/inclusive_scan_by_segment_impl.h |   2 +-
 .../dpl/internal/reduce_by_segment_impl.h     |   2 +-
 include/oneapi/dpl/pstl/algorithm_impl.h      | 236 +++++++++---------
 .../experimental/internal/for_loop_impl.h     |  16 +-
 include/oneapi/dpl/pstl/histogram_impl.h      |   4 +-
 include/oneapi/dpl/pstl/numeric_impl.h        |  18 +-
 8 files changed, 143 insertions(+), 143 deletions(-)

diff --git a/include/oneapi/dpl/internal/binary_search_impl.h b/include/oneapi/dpl/internal/binary_search_impl.h
index 29c30e79e2c..ee5240b92cf 100644
--- a/include/oneapi/dpl/internal/binary_search_impl.h
+++ b/include/oneapi/dpl/internal/binary_search_impl.h
@@ -74,7 +74,7 @@ OutputIterator
 lower_bound_impl(_Tag, Policy&& policy, InputIterator1 start, InputIterator1 end, InputIterator2 value_start,
                  InputIterator2 value_end, OutputIterator result, StrictWeakOrdering comp)
 {
-    static_assert(__internal::__is_host_backend_tag_v<::std::decay_t<_Tag>>);
+    static_assert(__internal::__is_host_backend_tag_v<_Tag>);
 
     return oneapi::dpl::transform(policy, value_start, value_end, result,
                                   [=](typename ::std::iterator_traits<InputIterator2>::reference val) {
@@ -88,7 +88,7 @@ OutputIterator
 upper_bound_impl(_Tag, Policy&& policy, InputIterator1 start, InputIterator1 end, InputIterator2 value_start,
                  InputIterator2 value_end, OutputIterator result, StrictWeakOrdering comp)
 {
-    static_assert(__internal::__is_host_backend_tag_v<::std::decay_t<_Tag>>);
+    static_assert(__internal::__is_host_backend_tag_v<_Tag>);
 
     return oneapi::dpl::transform(policy, value_start, value_end, result,
                                   [=](typename ::std::iterator_traits<InputIterator2>::reference val) {
@@ -102,7 +102,7 @@ OutputIterator
 binary_search_impl(_Tag, Policy&& policy, InputIterator1 start, InputIterator1 end, InputIterator2 value_start,
                    InputIterator2 value_end, OutputIterator result, StrictWeakOrdering comp)
 {
-    static_assert(__internal::__is_host_backend_tag_v<::std::decay_t<_Tag>>);
+    static_assert(__internal::__is_host_backend_tag_v<_Tag>);
 
     return oneapi::dpl::transform(policy, value_start, value_end, result,
                                   [=](typename ::std::iterator_traits<InputIterator2>::reference val) {
diff --git a/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h b/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h
index e1496d38e13..a39ff79f7e0 100644
--- a/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h
@@ -44,7 +44,7 @@ pattern_exclusive_scan_by_segment(_Tag, Policy&& policy, InputIterator1 first1,
                                   InputIterator2 first2, OutputIterator result, T init, BinaryPredicate binary_pred,
                                   Operator binary_op)
 {
-    static_assert(__internal::__is_host_backend_tag_v<::std::decay_t<_Tag>>);
+    static_assert(__internal::__is_host_backend_tag_v<_Tag>);
 
     using __backend_tag = typename _Tag::__backend_tag;
 
diff --git a/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h b/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h
index 7dd345ba534..eaca4b20b43 100644
--- a/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h
@@ -43,7 +43,7 @@ pattern_inclusive_scan_by_segment(_Tag, Policy&& policy, InputIterator1 first1,
                                   InputIterator2 first2, OutputIterator result, BinaryPredicate binary_pred,
                                   BinaryOperator binary_op)
 {
-    static_assert(__internal::__is_host_backend_tag_v<::std::decay_t<_Tag>>);
+    static_assert(__internal::__is_host_backend_tag_v<_Tag>);
 
     using __backend_tag = typename _Tag::__backend_tag;
 
diff --git a/include/oneapi/dpl/internal/reduce_by_segment_impl.h b/include/oneapi/dpl/internal/reduce_by_segment_impl.h
index 0c879f71978..8e634f1567c 100644
--- a/include/oneapi/dpl/internal/reduce_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/reduce_by_segment_impl.h
@@ -85,7 +85,7 @@ reduce_by_segment_impl(_Tag, Policy&& policy, InputIterator1 first1, InputIterat
                        OutputIterator1 result1, OutputIterator2 result2, BinaryPred binary_pred,
                        BinaryOperator binary_op)
 {
-    static_assert(__internal::__is_host_backend_tag_v<::std::decay_t<_Tag>>);
+    static_assert(__internal::__is_host_backend_tag_v<_Tag>);
 
     using __backend_tag = typename _Tag::__backend_tag;
 
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index dff58a82e02..607d3b44cf1 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -67,8 +67,8 @@ template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Pre
 bool
 __pattern_any_of(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _Pred __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
-                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<_Tag> ||
+                  __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_any_of(__first, __last, __pred, typename _Tag::__is_vector{});
 }
@@ -140,7 +140,7 @@ template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Fun
 void
 __pattern_walk1(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _Function __f) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<_Tag>);
 
     __internal::__brick_walk1(__first, __last, __f, typename _Tag::__is_vector{});
 }
@@ -180,8 +180,8 @@ void
 __pattern_walk_brick(_Tag, _ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last,
                      _Brick __brick) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
-                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<_Tag> ||
+                  __is_backend_tag_parallel_forward_v<_Tag>);
 
     __brick(__first, __last, typename _Tag::__is_vector{});
 }
@@ -223,8 +223,8 @@ template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Siz
 _ForwardIterator
 __pattern_walk1_n(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _Size __n, _Function __f) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
-                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<_Tag> ||
+                  __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_walk1_n(__first, __n, __f, typename _Tag::__is_vector{});
 }
@@ -243,8 +243,8 @@ template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Siz
 _ForwardIterator
 __pattern_walk_brick_n(_Tag, _ExecutionPolicy&& __exec, _ForwardIterator __first, _Size __n, _Brick __brick) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
-                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<_Tag> ||
+                  __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __brick(__first, __n, typename _Tag::__is_vector{});
 }
@@ -311,7 +311,7 @@ _ForwardIterator2
 __pattern_walk2(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
                 _ForwardIterator2 __first2, _Function __f) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<_Tag>);
 
     return __internal::__brick_walk2(__first1, __last1, __first2, __f, typename _Tag::__is_vector{});
 }
@@ -368,8 +368,8 @@ _ForwardIterator2
 __pattern_walk2_n(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _Size __n, _ForwardIterator2 __first2,
                   _Function __f) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
-                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<_Tag> ||
+                  __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_walk2_n(__first1, __n, __first2, __f, typename _Tag::__is_vector{});
 }
@@ -389,7 +389,7 @@ _ForwardIterator2
 __pattern_walk2_brick(_Tag, _ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
                       _ForwardIterator2 __first2, _Brick __brick) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<_Tag>);
 
     return __brick(__first1, __last1, __first2, typename _Tag::__is_vector{});
 }
@@ -466,8 +466,8 @@ _ForwardIterator2
 __pattern_walk2_brick_n(_Tag, _ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _Size __n,
                         _ForwardIterator2 __first2, _Brick __brick) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
-                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<_Tag> ||
+                  __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __brick(__first1, __n, __first2, typename _Tag::__is_vector{});
 }
@@ -501,7 +501,7 @@ _ForwardIterator3
 __pattern_walk3(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
                 _ForwardIterator2 __first2, _ForwardIterator3 __first3, _Function __f) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<_Tag>);
 
     return __internal::__brick_walk3(__first1, __last1, __first2, __first3, __f, typename _Tag::__is_vector{});
 }
@@ -567,7 +567,7 @@ _ForwardIterator2
 __pattern_walk2_transform_if(_Tag __tag, _ExecutionPolicy&& __exec, _ForwardIterator1 __first1,
                              _ForwardIterator1 __last1, _ForwardIterator2 __first2, _Function __func) noexcept
 {
-    static_assert(__is_host_backend_tag_v<::std::decay_t<_Tag>>);
+    static_assert(__is_host_backend_tag_v<_Tag>);
 
     return __pattern_walk2(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __func);
 }
@@ -579,7 +579,7 @@ __pattern_walk3_transform_if(_Tag __tag, _ExecutionPolicy&& __exec, _ForwardIter
                              _ForwardIterator1 __last1, _ForwardIterator2 __first2, _ForwardIterator3 __first3,
                              _Function __func) noexcept
 {
-    static_assert(__is_host_backend_tag_v<::std::decay_t<_Tag>>);
+    static_assert(__is_host_backend_tag_v<_Tag>);
 
     return __pattern_walk3(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __first3,
                            __func);
@@ -614,8 +614,8 @@ bool
 __pattern_equal(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
                 _ForwardIterator2 __first2, _ForwardIterator2 __last2, _BinaryPredicate __p) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
-                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<_Tag> ||
+                  __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_equal(__first1, __last1, __first2, __last2, __p, typename _Tag::__is_vector{});
 }
@@ -666,8 +666,8 @@ bool
 __pattern_equal(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
                 _ForwardIterator2 __first2, _BinaryPredicate __p) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
-                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<_Tag> ||
+                  __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_equal(__first1, __last1, __first2, __p, typename _Tag::__is_vector{});
 }
@@ -714,8 +714,8 @@ _ForwardIterator
 __pattern_find_if(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                   _Predicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
-                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<_Tag> ||
+                  __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_find_if(__first, __last, __pred, typename _Tag::__is_vector{});
 }
@@ -858,8 +858,8 @@ _ForwardIterator1
 __pattern_find_end(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first, _ForwardIterator1 __last,
                    _ForwardIterator2 __s_first, _ForwardIterator2 __s_last, _BinaryPredicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
-                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<_Tag> ||
+                  __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_find_end(__first, __last, __s_first, __s_last, __pred, typename _Tag::__is_vector{});
 }
@@ -915,8 +915,8 @@ _ForwardIterator1
 __pattern_find_first_of(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first, _ForwardIterator1 __last,
                         _ForwardIterator2 __s_first, _ForwardIterator2 __s_last, _BinaryPredicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
-                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<_Tag> ||
+                  __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_find_first_of(__first, __last, __s_first, __s_last, __pred,
                                              typename _Tag::__is_vector{});
@@ -965,8 +965,8 @@ _ForwardIterator1
 __pattern_search(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first, _ForwardIterator1 __last,
                  _ForwardIterator2 __s_first, _ForwardIterator2 __s_last, _BinaryPredicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
-                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<_Tag> ||
+                  __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_search(__first, __last, __s_first, __s_last, __pred, typename _Tag::__is_vector{});
 }
@@ -1022,8 +1022,8 @@ _ForwardIterator
 __pattern_search_n(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _Size __count,
                    const _Tp& __value, _BinaryPredicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
-                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<_Tag> ||
+                  __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_search_n(__first, __last, __count, __value, __pred, typename _Tag::__is_vector{});
 }
@@ -1064,7 +1064,7 @@ __pattern_search_n(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _
 template <class _Tag>
 struct __brick_copy_n
 {
-    static_assert(__is_host_backend_tag_v<::std::decay_t<_Tag>>);
+    static_assert(__is_host_backend_tag_v<_Tag>);
 
     template <typename _RandomAccessIterator1, typename _Size, typename _RandomAccessIterator2>
     _RandomAccessIterator2
@@ -1091,7 +1091,7 @@ struct __brick_copy_n
 template <class _Tag>
 struct __brick_copy
 {
-    static_assert(__is_host_backend_tag_v<::std::decay_t<_Tag>>);
+    static_assert(__is_host_backend_tag_v<_Tag>);
 
     template <typename _RandomAccessIterator1, typename _RandomAccessIterator2>
     _RandomAccessIterator2
@@ -1125,7 +1125,7 @@ struct __brick_copy
 template <class _Tag>
 struct __brick_move
 {
-    static_assert(__is_host_backend_tag_v<::std::decay_t<_Tag>>);
+    static_assert(__is_host_backend_tag_v<_Tag>);
 
     template <typename _RandomAccessIterator1, typename _RandomAccessIterator2>
     _RandomAccessIterator2
@@ -1155,7 +1155,7 @@ struct __brick_move
 template <class _Tag>
 struct __brick_move_destroy
 {
-    static_assert(__is_host_backend_tag_v<::std::decay_t<_Tag>>);
+    static_assert(__is_host_backend_tag_v<_Tag>);
 
     template <typename _RandomAccessIterator1, typename _RandomAccessIterator2>
     _RandomAccessIterator2
@@ -1328,8 +1328,8 @@ _OutputIterator
 __pattern_copy_if(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _OutputIterator __result,
                   _UnaryPredicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
-                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<_Tag> ||
+                  __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_copy_if(__first, __last, __result, __pred, typename _Tag::__is_vector{});
 }
@@ -1394,8 +1394,8 @@ template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Pre
 typename ::std::iterator_traits<_ForwardIterator>::difference_type
 __pattern_count(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _Predicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
-                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<_Tag> ||
+                  __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_count(__first, __last, __pred, typename _Tag::__is_vector{});
 }
@@ -1449,8 +1449,8 @@ _ForwardIterator
 __pattern_unique(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                  _BinaryPredicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
-                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<_Tag> ||
+                  __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_unique(__first, __last, __pred, typename _Tag::__is_vector{});
 }
@@ -1593,8 +1593,8 @@ _OutputIterator
 __pattern_unique_copy(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                       _OutputIterator __result, _BinaryPredicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
-                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<_Tag> ||
+                  __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_unique_copy(__first, __last, __result, __pred, typename _Tag::__is_vector{});
 }
@@ -1730,8 +1730,8 @@ template <class _Tag, class _ExecutionPolicy, class _BidirectionalIterator>
 void
 __pattern_reverse(_Tag, _ExecutionPolicy&&, _BidirectionalIterator __first, _BidirectionalIterator __last) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
-                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<_Tag> ||
+                  __is_backend_tag_parallel_forward_v<_Tag>);
 
     __internal::__brick_reverse(__first, __last, typename _Tag::__is_vector{});
 }
@@ -1779,8 +1779,8 @@ _OutputIterator
 __pattern_reverse_copy(_Tag, _ExecutionPolicy&&, _BidirectionalIterator __first, _BidirectionalIterator __last,
                        _OutputIterator __d_first) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
-                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<_Tag> ||
+                  __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_reverse_copy(__first, __last, __d_first, typename _Tag::__is_vector{});
 }
@@ -1859,8 +1859,8 @@ _ForwardIterator
 __pattern_rotate(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __middle,
                  _ForwardIterator __last) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
-                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<_Tag> ||
+                  __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_rotate(__first, __middle, __last, typename _Tag::__is_vector{});
 }
@@ -1938,8 +1938,8 @@ _OutputIterator
 __brick_rotate_copy(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __middle,
                     _ForwardIterator __last, _OutputIterator __result) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
-                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<_Tag> ||
+                  __is_backend_tag_parallel_forward_v<_Tag>);
 
     return ::std::rotate_copy(__first, __middle, __last, __result);
 }
@@ -1959,8 +1959,8 @@ _OutputIterator
 __pattern_rotate_copy(_Tag __tag, _ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __middle,
                       _ForwardIterator __last, _OutputIterator __result) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
-                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<_Tag> ||
+                  __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_rotate_copy(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __middle, __last,
                                            __result);
@@ -2042,8 +2042,8 @@ bool
 __pattern_is_partitioned(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                          _UnaryPredicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
-                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<_Tag> ||
+                  __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_is_partitioned(__first, __last, __pred, typename _Tag::__is_vector{});
 }
@@ -2173,8 +2173,8 @@ _ForwardIterator
 __pattern_partition(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                     _UnaryPredicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
-                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<_Tag> ||
+                  __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_partition(__first, __last, __pred, typename _Tag::__is_vector{});
 }
@@ -2274,8 +2274,8 @@ _BidirectionalIterator
 __pattern_stable_partition(_Tag, _ExecutionPolicy&&, _BidirectionalIterator __first, _BidirectionalIterator __last,
                            _UnaryPredicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
-                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<_Tag> ||
+                  __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_stable_partition(__first, __last, __pred, typename _Tag::__is_vector{});
 }
@@ -2365,8 +2365,8 @@ ::std::pair<_OutputIterator1, _OutputIterator2>
 __pattern_partition_copy(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                          _OutputIterator1 __out_true, _OutputIterator2 __out_false, _UnaryPredicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
-                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<_Tag> ||
+                  __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_partition_copy(__first, __last, __out_true, __out_false, __pred,
                                               typename _Tag::__is_vector{});
@@ -2422,8 +2422,8 @@ void
 __pattern_sort(_Tag, _ExecutionPolicy&&, _RandomAccessIterator __first, _RandomAccessIterator __last, _Compare __comp,
                _IsMoveConstructible) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
-                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<_Tag> ||
+                  __is_backend_tag_parallel_forward_v<_Tag>);
 
     ::std::sort(__first, __last, __comp);
 }
@@ -2455,8 +2455,8 @@ void
 __pattern_stable_sort(_Tag, _ExecutionPolicy&&, _RandomAccessIterator __first, _RandomAccessIterator __last,
                       _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
-                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<_Tag> ||
+                  __is_backend_tag_parallel_forward_v<_Tag>);
 
     ::std::stable_sort(__first, __last, __comp);
 }
@@ -2489,8 +2489,8 @@ __pattern_sort_by_key(_Tag, _ExecutionPolicy&& __exec, _RandomAccessIterator1 __
                       _RandomAccessIterator1 __keys_last, _RandomAccessIterator2 __values_first,
                       _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
-                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<_Tag> ||
+                  __is_backend_tag_parallel_forward_v<_Tag>);
 
     auto __beg = oneapi::dpl::make_zip_iterator(__keys_first, __values_first);
     auto __end = __beg + (__keys_last - __keys_first);
@@ -2536,8 +2536,8 @@ void
 __pattern_partial_sort(_Tag, _ExecutionPolicy&&, _RandomAccessIterator __first, _RandomAccessIterator __middle,
                        _RandomAccessIterator __last, _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
-                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<_Tag> ||
+                  __is_backend_tag_parallel_forward_v<_Tag>);
 
     ::std::partial_sort(__first, __middle, __last, __comp);
 }
@@ -2575,8 +2575,8 @@ _RandomAccessIterator
 __pattern_partial_sort_copy(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                             _RandomAccessIterator __d_first, _RandomAccessIterator __d_last, _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
-                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<_Tag> ||
+                  __is_backend_tag_parallel_forward_v<_Tag>);
 
     return ::std::partial_sort_copy(__first, __last, __d_first, __d_last, __comp);
 }
@@ -2680,8 +2680,8 @@ _ForwardIterator
 __pattern_adjacent_find(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                         _BinaryPredicate __pred, _Semantic) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
-                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<_Tag> ||
+                  __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_adjacent_find(__first, __last, __pred, typename _Tag::__is_vector{}, _Semantic::value);
 }
@@ -2741,8 +2741,8 @@ void
 __pattern_nth_element(_Tag, _ExecutionPolicy&&, _RandomAccessIterator __first, _RandomAccessIterator __nth,
                       _RandomAccessIterator __last, _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
-                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<_Tag> ||
+                  __is_backend_tag_parallel_forward_v<_Tag>);
 
     ::std::nth_element(__first, __nth, __last, __comp);
 }
@@ -2794,7 +2794,7 @@ __pattern_nth_element(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec
 template <class _Tag, typename _Tp>
 struct __brick_fill
 {
-    static_assert(__is_host_backend_tag_v<::std::decay_t<_Tag>>);
+    static_assert(__is_host_backend_tag_v<_Tag>);
 
     const _Tp& __value;
 
@@ -2819,8 +2819,8 @@ template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Tp>
 void
 __pattern_fill(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, const _Tp& __value) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
-                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<_Tag> ||
+                  __is_backend_tag_parallel_forward_v<_Tag>);
 
     __internal::__brick_fill<_Tag, _Tp>{__value}(__first, __last, typename _Tag::__is_vector{});
 }
@@ -2845,7 +2845,7 @@ __pattern_fill(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAcce
 template <class _Tag, typename _Tp>
 struct __brick_fill_n
 {
-    static_assert(__is_host_backend_tag_v<::std::decay_t<_Tag>>);
+    static_assert(__is_host_backend_tag_v<_Tag>);
 
     const _Tp& __value;
 
@@ -2870,8 +2870,8 @@ template <class _Tag, class _ExecutionPolicy, class _OutputIterator, class _Size
 _OutputIterator
 __pattern_fill_n(_Tag, _ExecutionPolicy&&, _OutputIterator __first, _Size __count, const _Tp& __value) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
-                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<_Tag> ||
+                  __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_fill_n<_Tag, _Tp>{__value}(__first, __count, typename _Tag::__is_vector{});
 }
@@ -2908,8 +2908,8 @@ template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Gen
 void
 __pattern_generate(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _Generator __g) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
-                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<_Tag> ||
+                  __is_backend_tag_parallel_forward_v<_Tag>);
 
     __internal::__brick_generate(__first, __last, __g, typename _Tag::__is_vector{});
 }
@@ -2949,8 +2949,8 @@ template <class _Tag, class _ExecutionPolicy, class _OutputIterator, class _Size
 _OutputIterator
 __pattern_generate_n(_Tag, _ExecutionPolicy&&, _OutputIterator __first, _Size __count, _Generator __g) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
-                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<_Tag> ||
+                  __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_generate_n(__first, __count, __g, typename _Tag::__is_vector{});
 }
@@ -2995,8 +2995,8 @@ _ForwardIterator
 __pattern_remove_if(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                     _UnaryPredicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
-                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<_Tag> ||
+                  __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_remove_if(__first, __last, __pred, typename _Tag::__is_vector{});
 }
@@ -3052,8 +3052,8 @@ __pattern_merge(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIt
                 _ForwardIterator2 __first2, _ForwardIterator2 __last2, _OutputIterator __d_first,
                 _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
-                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<_Tag> ||
+                  __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_merge(__first1, __last1, __first2, __last2, __d_first, __comp,
                                      typename _Tag::__is_vector{});
@@ -3102,8 +3102,8 @@ void
 __pattern_inplace_merge(_Tag, _ExecutionPolicy&&, _BidirectionalIterator __first, _BidirectionalIterator __middle,
                         _BidirectionalIterator __last, _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
-                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<_Tag> ||
+                  __is_backend_tag_parallel_forward_v<_Tag>);
 
     __internal::__brick_inplace_merge(__first, __middle, __last, __comp, typename _Tag::__is_vector{});
 }
@@ -3162,8 +3162,8 @@ bool
 __pattern_includes(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
                    _ForwardIterator2 __first2, _ForwardIterator2 __last2, _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
-                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<_Tag> ||
+                  __is_backend_tag_parallel_forward_v<_Tag>);
 
     return ::std::includes(__first1, __last1, __first2, __last2, __comp);
 }
@@ -3478,8 +3478,8 @@ __pattern_set_union(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _Forwa
                     _ForwardIterator2 __first2, _ForwardIterator2 __last2, _OutputIterator __result,
                     _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
-                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<_Tag> ||
+                  __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_set_union(__first1, __last1, __first2, __last2, __result, __comp,
                                          typename _Tag::__is_vector{});
@@ -3540,8 +3540,8 @@ __pattern_set_intersection(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1,
                            _ForwardIterator2 __first2, _ForwardIterator2 __last2, _OutputIterator __result,
                            _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
-                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<_Tag> ||
+                  __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_set_intersection(__first1, __last1, __first2, __last2, __result, __comp,
                                                 typename _Tag::__is_vector{});
@@ -3639,8 +3639,8 @@ __pattern_set_difference(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _
                          _ForwardIterator2 __first2, _ForwardIterator2 __last2, _OutputIterator __result,
                          _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
-                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<_Tag> ||
+                  __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_set_difference(__first1, __last1, __first2, __last2, __result, __comp,
                                               typename _Tag::__is_vector{});
@@ -3727,8 +3727,8 @@ __pattern_set_symmetric_difference(_Tag, _ExecutionPolicy&&, _ForwardIterator1 _
                                    _ForwardIterator2 __first2, _ForwardIterator2 __last2, _OutputIterator __result,
                                    _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
-                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<_Tag> ||
+                  __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_set_symmetric_difference(__first1, __last1, __first2, __last2, __result, __comp,
                                                         typename _Tag::__is_vector{});
@@ -3787,8 +3787,8 @@ _RandomAccessIterator
 __pattern_is_heap_until(_Tag, _ExecutionPolicy&&, _RandomAccessIterator __first, _RandomAccessIterator __last,
                         _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
-                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<_Tag> ||
+                  __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_is_heap_until(__first, __last, __comp, typename _Tag::__is_vector{});
 }
@@ -3876,8 +3876,8 @@ bool
 __pattern_is_heap(_Tag, _ExecutionPolicy&&, _RandomAccessIterator __first, _RandomAccessIterator __last,
                   _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
-                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<_Tag> ||
+                  __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_is_heap(__first, __last, __comp, typename _Tag::__is_vector{});
 }
@@ -3925,8 +3925,8 @@ _ForwardIterator
 __pattern_min_element(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                       _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
-                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<_Tag> ||
+                  __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_min_element(__first, __last, __comp, typename _Tag::__is_vector{});
 }
@@ -3993,8 +3993,8 @@ ::std::pair<_ForwardIterator, _ForwardIterator>
 __pattern_minmax_element(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                          _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
-                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<_Tag> ||
+                  __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_minmax_element(__first, __last, __comp, typename _Tag::__is_vector{});
 }
@@ -4075,8 +4075,8 @@ ::std::pair<_ForwardIterator1, _ForwardIterator2>
 __pattern_mismatch(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
                    _ForwardIterator2 __first2, _ForwardIterator2 __last2, _Predicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
-                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<_Tag> ||
+                  __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_mismatch(__first1, __last1, __first2, __last2, __pred, typename _Tag::__is_vector{});
 }
@@ -4157,8 +4157,8 @@ bool
 __pattern_lexicographical_compare(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
                                   _ForwardIterator2 __first2, _ForwardIterator2 __last2, _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
-                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<_Tag> ||
+                  __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_lexicographical_compare(__first1, __last1, __first2, __last2, __comp,
                                                        typename _Tag::__is_vector{});
@@ -4220,7 +4220,7 @@ _ForwardIterator2
 __pattern_swap(_Tag __tag, _ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
                _ForwardIterator2 __first2, _Function __f)
 {
-    static_assert(__is_host_backend_tag_v<::std::decay_t<_Tag>>);
+    static_assert(__is_host_backend_tag_v<_Tag>);
 
     return __pattern_walk2(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __f);
 }
@@ -4298,8 +4298,8 @@ _ForwardIterator
 __pattern_shift_left(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                      typename ::std::iterator_traits<_ForwardIterator>::difference_type __n) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
-                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<_Tag> ||
+                  __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __brick_shift_left(__first, __last, __n, typename _Tag::__is_vector{});
 }
@@ -4355,7 +4355,7 @@ __pattern_shift_right(_Tag __tag, _ExecutionPolicy&& __exec, _BidirectionalItera
                       _BidirectionalIterator __last,
                       typename ::std::iterator_traits<_BidirectionalIterator>::difference_type __n)
 {
-    static_assert(__is_host_backend_tag_v<::std::decay_t<_Tag>>);
+    static_assert(__is_host_backend_tag_v<_Tag>);
 
     using _ReverseIterator = typename ::std::reverse_iterator<_BidirectionalIterator>;
 
diff --git a/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h b/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
index 88d08abb3a1..b41ec62b84a 100644
--- a/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
+++ b/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
@@ -167,8 +167,8 @@ void
 __pattern_for_loop_n(_Tag, _ExecutionPolicy&&, _Ip __first, _Size __n, _Function __f, __single_stride_type,
                      _Rest&&... __rest) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
-                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<_Tag> ||
+                  __is_backend_tag_parallel_forward_v<_Tag>);
 
     __reduction_pack<_Rest...> __pack{__reduction_pack_tag(), ::std::forward<_Rest>(__rest)...};
 
@@ -193,8 +193,8 @@ void
 __pattern_for_loop_n(_Tag, _ExecutionPolicy&&, _Ip __first, _Size __n, _Function __f, _Sp __stride,
                      _Rest&&... __rest) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
-                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<_Tag> ||
+                  __is_backend_tag_parallel_forward_v<_Tag>);
 
     __reduction_pack<_Rest...> __pack{__reduction_pack_tag(), ::std::forward<_Rest>(__rest)...};
 
@@ -267,8 +267,8 @@ void
 __pattern_for_loop(_Tag __tag, _ExecutionPolicy&& __exec, _Ip __first, _Ip __last, _Function __f, _Sp __stride,
                    _Rest&&... __rest) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
-                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<_Tag> ||
+                  __is_backend_tag_parallel_forward_v<_Tag>);
 
     if constexpr (typename _Tag::__is_vector{} || __is_random_access_or_integral_v<_Ip>)
     {
@@ -369,8 +369,8 @@ void
 __pattern_for_loop(_Tag, _ExecutionPolicy&& __exec, _Ip __first, _Ip __last, _Function __f, __single_stride_type,
                    _Rest&&... __rest) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
-                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<_Tag> ||
+                  __is_backend_tag_parallel_forward_v<_Tag>);
 
     __reduction_pack<_Rest...> __pack{__reduction_pack_tag(), ::std::forward<_Rest>(__rest)...};
 
diff --git a/include/oneapi/dpl/pstl/histogram_impl.h b/include/oneapi/dpl/pstl/histogram_impl.h
index 5d27df53749..bdc34746bc5 100644
--- a/include/oneapi/dpl/pstl/histogram_impl.h
+++ b/include/oneapi/dpl/pstl/histogram_impl.h
@@ -38,8 +38,8 @@ void
 __pattern_histogram(_Tag, _ExecutionPolicy&& exec, _RandomAccessIterator1 __first, _RandomAccessIterator1 __last,
                     _Size __num_bins, _IdxHashFunc __func, _RandomAccessIterator2 __histogram_first)
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
-                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<_Tag> ||
+                  __is_backend_tag_parallel_forward_v<_Tag>);
 
     static_assert(sizeof(_Size) == 0 /*false*/,
                   "Histogram API is currently unsupported for policies other than device execution policies");
diff --git a/include/oneapi/dpl/pstl/numeric_impl.h b/include/oneapi/dpl/pstl/numeric_impl.h
index b03ee25340a..85d248ac348 100644
--- a/include/oneapi/dpl/pstl/numeric_impl.h
+++ b/include/oneapi/dpl/pstl/numeric_impl.h
@@ -68,8 +68,8 @@ __pattern_transform_reduce(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1,
                            _ForwardIterator2 __first2, _Tp __init, _BinaryOperation1 __binary_op1,
                            _BinaryOperation2 __binary_op2) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
-                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<_Tag> ||
+                  __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __brick_transform_reduce(__first1, __last1, __first2, __init, __binary_op1, __binary_op2,
                                     typename _Tag::__is_vector{});
@@ -134,8 +134,8 @@ _Tp
 __pattern_transform_reduce(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _Tp __init,
                            _BinaryOperation __binary_op, _UnaryOperation __unary_op) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
-                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<_Tag> ||
+                  __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_transform_reduce(__first, __last, __init, __binary_op, __unary_op,
                                                 typename _Tag::__is_vector{});
@@ -245,8 +245,8 @@ __pattern_transform_scan(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _Fo
                          _OutputIterator __result, _UnaryOperation __unary_op, _Tp __init, _BinaryOperation __binary_op,
                          _Inclusive) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
-                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<_Tag> ||
+                  __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_transform_scan(__first, __last, __result, __unary_op, __init, __binary_op, _Inclusive(),
                                               typename _Tag::__is_vector{})
@@ -330,7 +330,7 @@ _OutputIterator
 __pattern_transform_scan(_Tag __tag, _ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last,
                          _OutputIterator __result, _UnaryOperation __unary_op, _BinaryOperation __binary_op, _Inclusive)
 {
-    static_assert(__is_host_backend_tag_v<::std::decay_t<_Tag>>);
+    static_assert(__is_host_backend_tag_v<_Tag>);
 
     typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _ValueType;
     if (__first != __last)
@@ -382,8 +382,8 @@ _OutputIterator
 __pattern_adjacent_difference(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                               _OutputIterator __d_first, _BinaryOperation __op) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<::std::decay_t<_Tag>> ||
-                  __is_backend_tag_parallel_forward_v<::std::decay_t<_Tag>>);
+    static_assert(__is_backend_tag_serial_v<_Tag> ||
+                  __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_adjacent_difference(__first, __last, __d_first, __op, typename _Tag::__is_vector{});
 }

From 93036b23906274fa2c7db95f652d3f7706f90815 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 5 Mar 2024 09:16:20 +0100
Subject: [PATCH 509/566] 
 include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h - fix review
 comment: Why do we need using base then?

---
 .../oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h | 10 ++++------
 1 file changed, 4 insertions(+), 6 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h b/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h
index 80a19db8130..bdc2e8ac7bd 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h
@@ -69,13 +69,11 @@ struct DefaultKernelNameFPGA;
 template <unsigned int factor = 1, typename KernelName = DefaultKernelNameFPGA>
 class fpga_policy : public device_policy<KernelName>
 {
-    using base = device_policy<KernelName>;
-
   public:
     static constexpr unsigned int unroll_factor = factor;
 
     fpga_policy()
-        : base(sycl::queue(
+        : device_policy<KernelName>(sycl::queue(
 #    if _ONEDPL_FPGA_EMU
               __dpl_sycl::__fpga_emulator_selector()
 #    else
@@ -86,9 +84,9 @@ class fpga_policy : public device_policy<KernelName>
     }
 
     template <unsigned int other_factor, typename OtherName>
-    fpga_policy(const fpga_policy<other_factor, OtherName>& other) : base(other.queue()){};
-    explicit fpga_policy(sycl::queue q) : base(q) {}
-    explicit fpga_policy(sycl::device d) : base(d) {}
+    fpga_policy(const fpga_policy<other_factor, OtherName>& other) : device_policy<KernelName>(other.queue()){};
+    explicit fpga_policy(sycl::queue q) : device_policy<KernelName>(q) {}
+    explicit fpga_policy(sycl::device d) : device_policy<KernelName>(d) {}
 };
 
 #endif // _ONEDPL_FPGA_DEVICE

From bac418f5879490324cb6ccd45a61666071b44e27 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 5 Mar 2024 09:49:39 +0100
Subject: [PATCH 510/566] Fix review comment: Why is policy removed? The
 agreement was that we keep both a policy and a tag.

---
 .../internal/exclusive_scan_by_segment_impl.h |  8 +++----
 .../internal/inclusive_scan_by_segment_impl.h |  4 ++--
 .../dpl/internal/reduce_by_segment_impl.h     | 14 ++++++------
 .../dpl/internal/scan_by_segment_impl.h       |  4 ++--
 include/oneapi/dpl/pstl/algorithm_impl.h      | 20 ++++++++---------
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 22 +++++++++----------
 .../hetero/algorithm_ranges_impl_hetero.h     | 12 +++++-----
 .../pstl/hetero/dpcpp/parallel_backend_sycl.h |  6 ++---
 .../dpcpp/parallel_backend_sycl_histogram.h   |  2 +-
 .../dpcpp/parallel_backend_sycl_radix_sort.h  |  4 ++--
 .../dpcpp/parallel_backend_sycl_utils.h       | 16 +++++++++-----
 .../dpl/pstl/hetero/numeric_impl_hetero.h     |  2 +-
 include/oneapi/dpl/pstl/omp/parallel_scan.h   |  2 +-
 include/oneapi/dpl/pstl/omp/util.h            | 14 +++++++-----
 .../oneapi/dpl/pstl/parallel_backend_serial.h | 12 +++++-----
 .../oneapi/dpl/pstl/parallel_backend_tbb.h    | 19 +++++++++-------
 16 files changed, 86 insertions(+), 75 deletions(-)

diff --git a/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h b/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h
index a39ff79f7e0..f9beef59f2b 100644
--- a/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h
@@ -66,7 +66,7 @@ pattern_exclusive_scan_by_segment(_Tag, Policy&& policy, InputIterator1 first1,
     InputIterator2 last2 = first2 + n;
 
     // compute head flags
-    oneapi::dpl::__par_backend::__buffer<FlagType> _flags(__backend_tag{}, n);
+    oneapi::dpl::__par_backend::__buffer<Policy, FlagType> _flags(__backend_tag{}, policy, n);
     auto flags = _flags.get();
     flags[0] = 1;
 
@@ -74,7 +74,7 @@ pattern_exclusive_scan_by_segment(_Tag, Policy&& policy, InputIterator1 first1,
               oneapi::dpl::__internal::__not_pred<BinaryPredicate>(binary_pred));
 
     // shift input one to the right and initialize segments with init
-    oneapi::dpl::__par_backend::__buffer<OutputType> _temp(__backend_tag{}, n);
+    oneapi::dpl::__par_backend::__buffer<Policy, OutputType> _temp(__backend_tag{}, policy, n);
     auto temp = _temp.get();
 
     temp[0] = init;
@@ -125,7 +125,7 @@ exclusive_scan_by_segment_impl(__internal::__hetero_tag<_BackendTag>, Policy&& p
     InputIterator2 last2 = first2 + n;
 
     // compute head flags
-    oneapi::dpl::__par_backend_hetero::__buffer<FlagType> _flags(_BackendTag{}, n);
+    oneapi::dpl::__par_backend_hetero::__buffer<Policy, FlagType> _flags(_BackendTag{}, policy, n);
     {
         auto flag_buf = _flags.get_buffer();
         auto flags = flag_buf.get_host_access(sycl::read_write);
@@ -136,7 +136,7 @@ exclusive_scan_by_segment_impl(__internal::__hetero_tag<_BackendTag>, Policy&& p
               oneapi::dpl::__internal::__not_pred<BinaryPredicate>(binary_pred));
 
     // shift input one to the right and initialize segments with init
-    oneapi::dpl::__par_backend_hetero::__buffer<OutputType> _temp(_BackendTag{}, n);
+    oneapi::dpl::__par_backend_hetero::__buffer<Policy, OutputType> _temp(_BackendTag{}, policy, n);
     {
         auto temp_buf = _temp.get_buffer();
         auto temp = temp_buf.get_host_access(sycl::read_write);
diff --git a/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h b/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h
index eaca4b20b43..fbe094104fe 100644
--- a/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h
@@ -61,7 +61,7 @@ pattern_inclusive_scan_by_segment(_Tag, Policy&& policy, InputIterator1 first1,
     typedef unsigned int FlagType;
     typedef typename ::std::iterator_traits<InputIterator2>::value_type ValueType;
 
-    oneapi::dpl::__par_backend::__buffer<FlagType> _mask(__backend_tag{}, n);
+    oneapi::dpl::__par_backend::__buffer<Policy, FlagType> _mask(__backend_tag{}, policy, n);
     auto mask = _mask.get();
 
     mask[0] = 1;
@@ -111,7 +111,7 @@ inclusive_scan_by_segment_impl(__internal::__hetero_tag<_BackendTag>, Policy&& p
 
     FlagType initial_mask = 1;
 
-    oneapi::dpl::__par_backend_hetero::__buffer<FlagType> _mask(_BackendTag{}, n);
+    oneapi::dpl::__par_backend_hetero::__buffer<Policy, FlagType> _mask(_BackendTag{}, policy, n);
     {
         auto mask_buf = _mask.get_buffer();
         auto mask = mask_buf.get_host_access(sycl::read_write);
diff --git a/include/oneapi/dpl/internal/reduce_by_segment_impl.h b/include/oneapi/dpl/internal/reduce_by_segment_impl.h
index 8e634f1567c..919da41ae80 100644
--- a/include/oneapi/dpl/internal/reduce_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/reduce_by_segment_impl.h
@@ -116,7 +116,7 @@ reduce_by_segment_impl(_Tag, Policy&& policy, InputIterator1 first1, InputIterat
 
     // buffer that is used to store a flag indicating if the associated key is not equal to
     // the next key, and thus its associated sum should be part of the final result
-    oneapi::dpl::__par_backend::__buffer<FlagType> _mask(__backend_tag{}, n + 1);
+    oneapi::dpl::__par_backend::__buffer<Policy, FlagType> _mask(__backend_tag{}, n + 1);
     auto mask = _mask.get();
     mask[0] = 1;
 
@@ -132,11 +132,11 @@ reduce_by_segment_impl(_Tag, Policy&& policy, InputIterator1 first1, InputIterat
     // buffer stores the sums of values associated with a given key. Sums are copied with
     // a shift into result2, and the shift is computed at the same time as the sums, so the
     // sums can't be written to result2 directly.
-    oneapi::dpl::__par_backend::__buffer<ValueType> _scanned_values(__backend_tag{}, n);
+    oneapi::dpl::__par_backend::__buffer<Policy, ValueType> _scanned_values(__backend_tag{}, n);
 
     // Buffer is used to store results of the scan of the mask. Values indicate which position
     // in result2 needs to be written with the scanned_values element.
-    oneapi::dpl::__par_backend::__buffer<FlagType> _scanned_tail_flags(__backend_tag{}, n);
+    oneapi::dpl::__par_backend::__buffer<Policy, FlagType> _scanned_tail_flags(__backend_tag{}, n);
 
     // Compute the sum of the segments. scanned_tail_flags values are not used.
     inclusive_scan(policy, make_zip_iterator(first2, _mask.get()), make_zip_iterator(first2, _mask.get()) + n,
@@ -264,16 +264,16 @@ __sycl_reduce_by_segment(__internal::__hetero_tag<_BackendTag>, _ExecutionPolicy
     ::std::size_t __n_groups = oneapi::dpl::__internal::__dpl_ceiling_div(__n, __wgroup_size * __vals_per_item);
 
     // intermediate reductions within a workgroup
-    auto __partials = oneapi::dpl::__par_backend_hetero::__buffer<__val_type>(_BackendTag{}, __n_groups).get_buffer();
+    auto __partials = oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, __val_type>(_BackendTag{}, __exec, __n_groups).get_buffer();
 
-    auto __end_idx = oneapi::dpl::__par_backend_hetero::__buffer<__diff_type>(_BackendTag{}, 1).get_buffer();
+    auto __end_idx = oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, __diff_type>(_BackendTag{}, __exec, 1).get_buffer();
 
     // the number of segment ends found in each work group
-    auto __seg_ends = oneapi::dpl::__par_backend_hetero::__buffer<__diff_type>(_BackendTag{}, __n_groups).get_buffer();
+    auto __seg_ends = oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, __diff_type>(_BackendTag{}, __exec, __n_groups).get_buffer();
 
     // buffer that stores an exclusive scan of the results
     auto __seg_ends_scanned =
-        oneapi::dpl::__par_backend_hetero::__buffer<__diff_type>(_BackendTag{}, __n_groups).get_buffer();
+        oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, __diff_type>(_BackendTag{}, __exec, __n_groups).get_buffer();
 
     // 1. Count the segment ends in each workgroup
     auto __seg_end_identification = __exec.queue().submit([&](sycl::handler& __cgh) {
diff --git a/include/oneapi/dpl/internal/scan_by_segment_impl.h b/include/oneapi/dpl/internal/scan_by_segment_impl.h
index 0df27861946..bbbc1db0d52 100644
--- a/include/oneapi/dpl/internal/scan_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/scan_by_segment_impl.h
@@ -145,10 +145,10 @@ struct __sycl_scan_by_segment_impl
         ::std::size_t __n_groups = __internal::__dpl_ceiling_div(__n, __wgroup_size * __vals_per_item);
 
         auto __partials =
-            oneapi::dpl::__par_backend_hetero::__buffer<__val_type>(_BackendTag{}, __n_groups).get_buffer();
+            oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, __val_type>(_BackendTag{}, __exec, __n_groups).get_buffer();
 
         // the number of segment ends found in each work group
-        auto __seg_ends = oneapi::dpl::__par_backend_hetero::__buffer<bool>(_BackendTag{}, __n_groups).get_buffer();
+        auto __seg_ends = oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, bool>(_BackendTag{}, __exec, __n_groups).get_buffer();
 
         // 1. Work group reduction
         auto __wg_scan = __exec.queue().submit([&](sycl::handler& __cgh) {
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 607d3b44cf1..1fe1f07ec6f 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -1346,7 +1346,7 @@ __pattern_copy_if(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomA
     const _DifferenceType __n = __last - __first;
     if (_DifferenceType(1) < __n)
     {
-        __par_backend::__buffer<bool> __mask_buf(__backend_tag{}, __n);
+        __par_backend::__buffer<_ExecutionPolicy, bool> __mask_buf(__backend_tag{}, __n);
         return __internal::__except_handler([&__exec, __n, __first, __result, __pred, &__mask_buf]() {
             bool* __mask = __mask_buf.get();
             _DifferenceType __m{};
@@ -1467,7 +1467,7 @@ __remove_elements(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Forward
     typedef typename ::std::iterator_traits<_ForwardIterator>::difference_type _DifferenceType;
     typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _Tp;
     _DifferenceType __n = __last - __first;
-    __par_backend::__buffer<bool> __mask_buf(__backend_tag{}, __n);
+    __par_backend::__buffer<_ExecutionPolicy, bool> __mask_buf(__backend_tag{}, __n);
     // 1. find a first iterator that should be removed
     return __internal::__except_handler([&]() {
         bool* __mask = __mask_buf.get();
@@ -1504,7 +1504,7 @@ __remove_elements(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Forward
         __n -= __min;
         __first += __min;
 
-        __par_backend::__buffer<_Tp> __buf(__backend_tag{}, __n);
+        __par_backend::__buffer<_ExecutionPolicy, _Tp> __buf(__backend_tag{}, __n);
         _Tp* __result = __buf.get();
         __mask += __min;
         _DifferenceType __m{};
@@ -1633,7 +1633,7 @@ __pattern_unique_copy(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Ran
     const _DifferenceType __n = __last - __first;
     if (_DifferenceType(2) < __n)
     {
-        __par_backend::__buffer<bool> __mask_buf(__backend_tag{}, __n);
+        __par_backend::__buffer<_ExecutionPolicy, bool> __mask_buf(__backend_tag{}, __n);
         if (_DifferenceType(2) < __n)
         {
             return __internal::__except_handler([&__exec, __n, __first, __result, __pred, &__mask_buf]() {
@@ -1877,7 +1877,7 @@ __pattern_rotate(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAc
     auto __m = __middle - __first;
     if (__m <= __n / 2)
     {
-        __par_backend::__buffer<_Tp> __buf(__backend_tag{}, __n - __m);
+        __par_backend::__buffer<_ExecutionPolicy, _Tp> __buf(__backend_tag{}, __n - __m);
         return __internal::__except_handler([&__exec, __n, __m, __first, __middle, __last, &__buf]() {
             _Tp* __result = __buf.get();
             __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __middle, __last,
@@ -1903,7 +1903,7 @@ __pattern_rotate(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAc
     }
     else
     {
-        __par_backend::__buffer<_Tp> __buf(__backend_tag{}, __m);
+        __par_backend::__buffer<_ExecutionPolicy, _Tp> __buf(__backend_tag{}, __m);
         return __internal::__except_handler([&__exec, __n, __m, __first, __middle, __last, &__buf]() {
             _Tp* __result = __buf.get();
             __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __middle,
@@ -2386,7 +2386,7 @@ __pattern_partition_copy(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _
     const _DifferenceType __n = __last - __first;
     if (_DifferenceType(1) < __n)
     {
-        __par_backend::__buffer<bool> __mask_buf(__backend_tag{}, __n);
+        __par_backend::__buffer<_ExecutionPolicy, bool> __mask_buf(__backend_tag{}, __n);
         return __internal::__except_handler([&__exec, __n, __first, __out_true, __out_false, __pred, &__mask_buf]() {
             bool* __mask = __mask_buf.get();
             _ReturnType __m{};
@@ -2617,7 +2617,7 @@ __pattern_partial_sort_copy(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec
         {
             typedef typename ::std::iterator_traits<_RandomAccessIterator1>::value_type _T1;
             typedef typename ::std::iterator_traits<_RandomAccessIterator2>::value_type _T2;
-            __par_backend::__buffer<_T1> __buf(__backend_tag{}, __n1);
+            __par_backend::__buffer<_ExecutionPolicy, _T1> __buf(__backend_tag{}, __n1);
             _T1* __r = __buf.get();
 
             __par_backend::__parallel_stable_sort(
@@ -3122,7 +3122,7 @@ __pattern_inplace_merge(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _R
 
     typedef typename ::std::iterator_traits<_RandomAccessIterator>::value_type _Tp;
     auto __n = __last - __first;
-    __par_backend::__buffer<_Tp> __buf(__backend_tag{}, __n);
+    __par_backend::__buffer<_ExecutionPolicy, _Tp> __buf(__backend_tag{}, __n);
     _Tp* __r = __buf.get();
     __internal::__except_handler([&]() {
         auto __move_values = [](_RandomAccessIterator __x, _Tp* __z) {
@@ -3257,7 +3257,7 @@ __parallel_set_op(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Forward
     const _DifferenceType __n1 = __last1 - __first1;
     const _DifferenceType __n2 = __last2 - __first2;
 
-    __par_backend::__buffer<_T> __buf(__backend_tag{}, __size_func(__n1, __n2));
+    __par_backend::__buffer<_ExecutionPolicy, _T> __buf(__backend_tag{}, __size_func(__n1, __n2));
 
     return __internal::__except_handler([&__exec, __n1, __first1, __last1, __first2, __last2, __result, __comp,
                                          __size_func, __set_op, &__buf]() {
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index f975af77358..fd958bf8b1b 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -1032,7 +1032,7 @@ __pattern_remove_if(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec,
 
     using _ValueType = typename ::std::iterator_traits<_Iterator>::value_type;
 
-    oneapi::dpl::__par_backend_hetero::__buffer<_ValueType> __buf(_BackendTag{}, __last - __first);
+    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __buf(__backend_tag{}, __exec, __last - __first);
     auto __copy_first = __buf.get();
 
     auto __copy_last = __pattern_copy_if(__tag, __exec, __first, __last, __copy_first, __not_pred<_Predicate>{__pred});
@@ -1053,7 +1053,7 @@ __pattern_unique(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _It
 
     using _ValueType = typename ::std::iterator_traits<_Iterator>::value_type;
 
-    oneapi::dpl::__par_backend_hetero::__buffer<_ValueType> __buf(_BackendTag{}, __last - __first);
+    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __buf(__backend_tag{}, __exec, __last - __first);
     auto __copy_first = __buf.get();
     auto __copy_last = __pattern_unique_copy(__tag, __exec, __first, __last, __copy_first, __pred);
 
@@ -1225,7 +1225,7 @@ __pattern_inplace_merge(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __ex
     assert(__first < __middle && __middle < __last);
 
     auto __n = __last - __first;
-    oneapi::dpl::__par_backend_hetero::__buffer<_ValueType> __buf(_BackendTag{}, __n);
+    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __buf(__backend_tag{}, __exec, __n);
     auto __copy_first = __buf.get();
     auto __copy_last = __copy_first + __n;
 
@@ -1313,8 +1313,8 @@ __pattern_stable_partition(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& _
 
     auto __n = __last - __first;
 
-    oneapi::dpl::__par_backend_hetero::__buffer<_ValueType> __true_buf(_BackendTag{}, __n);
-    oneapi::dpl::__par_backend_hetero::__buffer<_ValueType> __false_buf(_BackendTag{}, __n);
+    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __true_buf(__backend_tag{}, __exec, __n);
+    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __false_buf(__backend_tag{}, __exec, __n);
     auto __true_result = __true_buf.get();
     auto __false_result = __false_buf.get();
 
@@ -1510,7 +1510,7 @@ __pattern_partial_sort_copy(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&&
         // - create a temporary buffer and copy all the elements from the input buffer there
         // - run partial sort on the temporary buffer
         // - copy k elements from the temporary buffer to the output buffer.
-        oneapi::dpl::__par_backend_hetero::__buffer<_ValueType> __buf(_BackendTag{}, __in_size);
+        oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __buf(__backend_tag{}, __exec, __in_size);
 
         auto __buf_first = __buf.get();
 
@@ -1625,7 +1625,7 @@ __pattern_rotate(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Iterator
 
     auto __keep = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read_write, _Iterator>();
     auto __buf = __keep(__first, __last);
-    auto __temp_buf = oneapi::dpl::__par_backend_hetero::__buffer<_Tp>(_BackendTag{}, __n);
+    auto __temp_buf = oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _Tp>(__backend_tag{}, __exec, __n);
 
     auto __temp_rng =
         oneapi::dpl::__ranges::all_view<_Tp, __par_backend_hetero::access_mode::write>(__temp_buf.get_buffer());
@@ -1708,7 +1708,7 @@ __pattern_hetero_set_op(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _F
         __comp, __n1, __n2};
 
     // temporary buffer to store boolean mask
-    oneapi::dpl::__par_backend_hetero::__buffer<int32_t> __mask_buf(_BackendTag{}, __n1);
+    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, int32_t> __mask_buf(__backend_tag{}, __exec, __n1);
 
     auto __keep1 =
         oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _ForwardIterator1>();
@@ -1834,7 +1834,7 @@ __pattern_set_union(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec,
 
     // temporary buffer to store intermediate result
     const auto __n2 = __last2 - __first2;
-    oneapi::dpl::__par_backend_hetero::__buffer<_ValueType> __diff(_BackendTag{}, __n2);
+    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __diff(__backend_tag{}, __exec, __n2);
     auto __buf = __diff.get();
 
     //1. Calc difference {2} \ {1}
@@ -1913,10 +1913,10 @@ __pattern_set_symmetric_difference(__hetero_tag<_BackendTag> __tag, _ExecutionPo
 
     // temporary buffers to store intermediate result
     const auto __n1 = __last1 - __first1;
-    oneapi::dpl::__par_backend_hetero::__buffer<_ValueType> __diff_1(_BackendTag{}, __n1);
+    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __diff_1(__backend_tag{}, __exec, __n1);
     auto __buf_1 = __diff_1.get();
     const auto __n2 = __last2 - __first2;
-    oneapi::dpl::__par_backend_hetero::__buffer<_ValueType> __diff_2(_BackendTag{}, __n2);
+    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __diff_2(__backend_tag{}, __exec, __n2);
     auto __buf_2 = __diff_2.get();
 
     //1. Calc difference {1} \ {2}
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
index 7071d49d066..a8e788f60e8 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
@@ -360,7 +360,7 @@ __pattern_scan_copy(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Range
     _DataAcc __get_data_op;
     _MaskAssigner __add_mask_op;
 
-    oneapi::dpl::__par_backend_hetero::__buffer<int32_t> __mask_buf(_BackendTag{}, __rng1.size());
+    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, int32_t> __mask_buf(_BackendTag{}, __exec, __rng1.size());
 
     auto __res =
         __par_backend_hetero::__parallel_transform_scan_base(
@@ -413,7 +413,7 @@ __pattern_remove_if(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec,
 
     using _ValueType = oneapi::dpl::__internal::__value_t<_Range>;
 
-    oneapi::dpl::__par_backend_hetero::__buffer<_ValueType> __buf(_BackendTag{}, __rng.size());
+    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __buf(_BackendTag{}, __exec, __rng.size());
     auto __copy_rng = oneapi::dpl::__ranges::views::all(__buf.get_buffer());
 
     auto __copy_last_id = __ranges::__pattern_copy_if(__tag, __exec, __rng, __copy_rng, __not_pred<_Predicate>{__pred},
@@ -461,7 +461,7 @@ __pattern_unique(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Ra
 
     using _ValueType = oneapi::dpl::__internal::__value_t<_Range>;
 
-    oneapi::dpl::__par_backend_hetero::__buffer<_ValueType> __buf(_BackendTag{}, __rng.size());
+    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __buf(_BackendTag{}, __exec, __rng.size());
     auto res_rng = oneapi::dpl::__ranges::views::all(__buf.get_buffer());
     auto res = __ranges::__pattern_unique_copy(__tag, __exec, __rng, res_rng, __pred,
                                                oneapi::dpl::__internal::__pstl_assign());
@@ -714,9 +714,9 @@ __pattern_reduce_by_segment(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&&
     // Round 1: reduce with extra indices added to avoid long segments
     // TODO: At threshold points check if the key is equal to the key at the previous threshold point, indicating a long sequence.
     // Skip a round of copy_if and reduces if there are none.
-    auto __idx = oneapi::dpl::__par_backend_hetero::__buffer<__diff_type>(_BackendTag{}, __n).get_buffer();
-    auto __tmp_out_keys = oneapi::dpl::__par_backend_hetero::__buffer<__key_type>(_BackendTag{}, __n).get_buffer();
-    auto __tmp_out_values = oneapi::dpl::__par_backend_hetero::__buffer<__val_type>(_BackendTag{}, __n).get_buffer();
+    auto __idx = oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, __diff_type>(_BackendTag{}, __exec, __n).get_buffer();
+    auto __tmp_out_keys = oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, __key_type>(_BackendTag{}, __exec, __n) .get_buffer();
+    auto __tmp_out_values = oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, __val_type>(_BackendTag{}, __exec, __n) .get_buffer();
 
     // Replicating first element of keys view to be able to compare (i-1)-th and (i)-th key with aligned sequences,
     //  dropping the last key for the i-1 sequence.
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
index f4d938d1011..2a43411bdca 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
@@ -871,7 +871,7 @@ __parallel_scan_copy(oneapi::dpl::__internal::__device_backend_tag __backend_tag
     _MaskAssigner __add_mask_op;
 
     // temporary buffer to store boolean mask
-    oneapi::dpl::__par_backend_hetero::__buffer<int32_t> __mask_buf(oneapi::dpl::__internal::__device_backend_tag{}, __n);
+    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, int32_t> __mask_buf(oneapi::dpl::__internal::__device_backend_tag{}, __exec, __n);
 
     return __parallel_transform_scan_base(
         __backend_tag, ::std::forward<_ExecutionPolicy>(__exec),
@@ -1579,7 +1579,7 @@ struct __parallel_sort_submitter<_BackendTag, _IdType, __internal::__optional_ke
         });
 
         // 2. Merge sorting
-        oneapi::dpl::__par_backend_hetero::__buffer<_Tp> __temp_buf(_BackendTag{}, __n);
+        oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _Tp> __temp_buf(_BackendTag{}, __exec, __n);
         auto __temp = __temp_buf.get_buffer();
         bool __data_in_temp = false;
         _IdType __n_sorted = __leaf;
@@ -1700,7 +1700,7 @@ struct __parallel_partial_sort_submitter<_BackendTag, __internal::__optional_ker
         _Size __n = __rng.size();
         assert(__n > 1);
 
-        oneapi::dpl::__par_backend_hetero::__buffer<_Tp> __temp_buf(_BackendTag{}, __n);
+        oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _Tp> __temp_buf(_BackendTag{}, __exec, __n);
         auto __temp = __temp_buf.get_buffer();
         _PRINT_INFO_IN_DEBUG_MODE(__exec);
 
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h
index ef15febc4b6..4ea22c31b2a 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h
@@ -427,7 +427,7 @@ struct __histogram_general_private_global_atomics_submitter<_BackendTag,
             oneapi::dpl::__internal::__dpl_ceiling_div(__n, __work_group_size * __iters_per_work_item);
 
         auto __private_histograms =
-            oneapi::dpl::__par_backend_hetero::__buffer<_bin_type>(_BackendTag{}, __segments * __num_bins).get_buffer();
+            oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _bin_type>(_BackendTag{}, __exec, __segments * __num_bins).get_buffer();
 
         return __exec.queue().submit([&](auto& __h) {
             __h.depends_on(__init_event);
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_radix_sort.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_radix_sort.h
index d16610390f6..bb9c08bbe4c 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_radix_sort.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_radix_sort.h
@@ -813,8 +813,8 @@ __parallel_radix_sort(oneapi::dpl::__internal::__device_backend_tag, _ExecutionP
         sycl::buffer<::std::uint32_t, 1> __tmp_buf{sycl::range<1>(__tmp_buf_size)};
 
         // memory for storing values sorted for an iteration
-        oneapi::dpl::__par_backend_hetero::__buffer<_ValueT> __out_buffer_holder{
-            oneapi::dpl::__internal::__device_backend_tag{}, __n};
+        oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy,_ValueT> __out_buffer_holder{
+            oneapi::dpl::__internal::__device_backend_tag{}, __exec, __n};
         auto __out_rng = oneapi::dpl::__ranges::all_view<_ValueT, __par_backend_hetero::access_mode::read_write>(
             __out_buffer_holder.get_buffer());
 
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_utils.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_utils.h
index fcf9a915211..3d4166507a0 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_utils.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_utils.h
@@ -376,12 +376,12 @@ struct __local_buffer<sycl::buffer<::std::tuple<_T...>, __dim, _AllocT>>
     using type = sycl::buffer<oneapi::dpl::__internal::tuple<_T...>, __dim, _AllocT>;
 };
 
-template <typename _BackendTag, typename _Tp>
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Tp>
 class __buffer_impl;
 
 // impl for sycl::buffer<...>
-template <typename _T>
-class __buffer_impl<oneapi::dpl::__internal::__device_backend_tag, _T>
+template <typename _ExecutionPolicy, typename _T>
+class __buffer_impl<oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy, _T>
 {
   private:
     using __container_t = typename __local_buffer<sycl::buffer<_T>>::type;
@@ -390,7 +390,10 @@ class __buffer_impl<oneapi::dpl::__internal::__device_backend_tag, _T>
 
   public:
 
-    __buffer_impl(oneapi::dpl::__internal::__device_backend_tag, ::std::size_t __n_elements)
+    static_assert(::std::is_same_v<_ExecutionPolicy, ::std::decay_t<_ExecutionPolicy>>);
+
+    __buffer_impl(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy /*__exec*/,
+                  ::std::size_t __n_elements)
         : __container{sycl::range<1>(__n_elements)}
     {
     }
@@ -451,8 +454,9 @@ struct __memobj_traits<_T*>
 
 } // namespace __internal
 
-template <typename _T>
-using __buffer = __internal::__buffer_impl<oneapi::dpl::__internal::__device_backend_tag, _T>;
+template <typename _ExecutionPolicy, typename _T>
+using __buffer =
+    __internal::__buffer_impl<oneapi::dpl::__internal::__device_backend_tag, ::std::decay_t<_ExecutionPolicy>, _T>;
 
 template <typename T>
 struct __repacked_tuple
diff --git a/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
index d8ea609c7e0..ccba7e7dafc 100644
--- a/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
@@ -162,7 +162,7 @@ __pattern_transform_scan_base(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&
         using _NewExecutionPolicy = decltype(__policy);
 
         // Create temporary buffer
-        oneapi::dpl::__par_backend_hetero::__buffer<_Type> __tmp_buf(_BackendTag{}, __n);
+        oneapi::dpl::__par_backend_hetero::__buffer<_NewExecutionPolicy, _Type> __tmp_buf(_BackendTag{}, __policy, __n);
         auto __first_tmp = __tmp_buf.get();
         auto __last_tmp = __first_tmp + __n;
         auto __keep2 = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::write, _Iterator2>();
diff --git a/include/oneapi/dpl/pstl/omp/parallel_scan.h b/include/oneapi/dpl/pstl/omp/parallel_scan.h
index c316ed3a104..e9afe34c6a1 100644
--- a/include/oneapi/dpl/pstl/omp/parallel_scan.h
+++ b/include/oneapi/dpl/pstl/omp/parallel_scan.h
@@ -88,7 +88,7 @@ __parallel_strict_scan_body(_Index __n, _Tp __initial, _Rp __reduce, _Cp __combi
     const _Index __slack = 4;
     _Index __tilesize = (__n - 1) / (__slack * __p) + 1;
     _Index __m = (__n - 1) / __tilesize;
-    __buffer<_Tp> __buf(oneapi::dpl::__internal::__omp_backend_tag{}, __m + 1);
+    __buffer<_ExecutionPolicy, _Tp> __buf(oneapi::dpl::__internal::__omp_backend_tag{}, __m + 1);
     _Tp* __r = __buf.get();
 
     oneapi::dpl::__omp_backend::__upsweep(_Index(0), _Index(__m + 1), __tilesize, __r, __n - __m * __tilesize, __reduce,
diff --git a/include/oneapi/dpl/pstl/omp/util.h b/include/oneapi/dpl/pstl/omp/util.h
index e3b75e3b516..6bbc3d3b94d 100644
--- a/include/oneapi/dpl/pstl/omp/util.h
+++ b/include/oneapi/dpl/pstl/omp/util.h
@@ -57,11 +57,11 @@ __cancel_execution(oneapi::dpl::__internal::__omp_backend_tag)
 // raw buffer
 //------------------------------------------------------------------------
 
-template <typename _BackendTag, typename _Tp>
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Tp>
 class __buffer_impl;
 
-template <typename _Tp>
-class __buffer_impl<oneapi::dpl::__internal::__omp_backend_tag, _Tp>
+template <typename _ExecutionPolicy, typename _Tp>
+class __buffer_impl<oneapi::dpl::__internal::__omp_backend_tag, _ExecutionPolicy, _Tp>
 {
     std::allocator<_Tp> __allocator_;
     _Tp* __ptr_;
@@ -71,7 +71,9 @@ class __buffer_impl<oneapi::dpl::__internal::__omp_backend_tag, _Tp>
     operator=(const __buffer_impl&) = delete;
 
   public:
-    __buffer_impl(oneapi::dpl::__internal::__omp_backend_tag, std::size_t __n)
+    static_assert(::std::is_same_v<_ExecutionPolicy, ::std::decay_t<_ExecutionPolicy>>);
+
+    __buffer_impl(oneapi::dpl::__internal::__omp_backend_tag, _ExecutionPolicy /*__exec*/, std::size_t __n)
         : __allocator_(), __ptr_(__allocator_.allocate(__n)), __buf_size_(__n)
     {
     }
@@ -86,8 +88,8 @@ class __buffer_impl<oneapi::dpl::__internal::__omp_backend_tag, _Tp>
     ~__buffer_impl() { __allocator_.deallocate(__ptr_, __buf_size_); }
 };
 
-template <typename _Tp>
-using __buffer = __buffer_impl<oneapi::dpl::__internal::__omp_backend_tag, _Tp>;
+template <typename _ExecutionPolicy, typename _Tp>
+using __buffer = __buffer_impl<oneapi::dpl::__internal::__omp_backend_tag, ::std::decay_t<_ExecutionPolicy>, _Tp>;
 
 // Preliminary size of each chunk: requires further discussion
 constexpr std::size_t __default_chunk_size = 2048;
diff --git a/include/oneapi/dpl/pstl/parallel_backend_serial.h b/include/oneapi/dpl/pstl/parallel_backend_serial.h
index 23cf5ca9805..832bc0b8a92 100644
--- a/include/oneapi/dpl/pstl/parallel_backend_serial.h
+++ b/include/oneapi/dpl/pstl/parallel_backend_serial.h
@@ -32,11 +32,11 @@ namespace dpl
 namespace __serial_backend
 {
 
-template <typename _BackendTag, typename _Tp>
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Tp>
 class __buffer_impl;
 
-template <typename _Tp>
-class __buffer_impl<oneapi::dpl::__internal::__serial_backend_tag, _Tp>
+template <typename _ExecutionPolicy, typename _Tp>
+class __buffer_impl<oneapi::dpl::__internal::__serial_backend_tag, _ExecutionPolicy, _Tp>
 {
     ::std::allocator<_Tp> __allocator_;
     _Tp* __ptr_;
@@ -46,6 +46,8 @@ class __buffer_impl<oneapi::dpl::__internal::__serial_backend_tag, _Tp>
     operator=(const __buffer_impl&) = delete;
 
   public:
+    static_assert(::std::is_same_v<_ExecutionPolicy, ::std::decay_t<_ExecutionPolicy>>);
+
     __buffer_impl(oneapi::dpl::__internal::__serial_backend_tag, ::std::size_t __n)
         : __allocator_(), __ptr_(__allocator_.allocate(__n)), __buf_size_(__n)
     {
@@ -60,8 +62,8 @@ class __buffer_impl<oneapi::dpl::__internal::__serial_backend_tag, _Tp>
     ~__buffer_impl() { __allocator_.deallocate(__ptr_, __buf_size_); }
 };
 
-template <typename _Tp>
-using __buffer = __buffer_impl<oneapi::dpl::__internal::__serial_backend_tag, _Tp>;
+template <typename _ExecutionPolicy, typename _Tp>
+using __buffer = __buffer_impl<oneapi::dpl::__internal::__serial_backend_tag, ::std::decay_t<_ExecutionPolicy>, _Tp>;
 
 inline void
 __cancel_execution(oneapi::dpl::__internal::__serial_backend_tag)
diff --git a/include/oneapi/dpl/pstl/parallel_backend_tbb.h b/include/oneapi/dpl/pstl/parallel_backend_tbb.h
index 80421a32109..fac8eb5c0b4 100644
--- a/include/oneapi/dpl/pstl/parallel_backend_tbb.h
+++ b/include/oneapi/dpl/pstl/parallel_backend_tbb.h
@@ -49,7 +49,7 @@ namespace dpl
 namespace __tbb_backend
 {
 
-template <typename _BackendTag, typename _Tp>
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Tp>
 class __buffer_impl;
 
 //! Raw memory buffer with automatic freeing and no exceptions.
@@ -57,8 +57,8 @@ class __buffer_impl;
 not an initialize array, because initialization/destruction
 would make the span be at least O(N). */
 // tbb::allocator can improve performance in some cases.
-template <typename _Tp>
-class __buffer_impl<oneapi::dpl::__internal::__tbb_backend_tag, _Tp>
+template <typename _ExecutionPolicy, typename _Tp>
+class __buffer_impl<oneapi::dpl::__internal::__tbb_backend_tag, _ExecutionPolicy, _Tp>
 {
     tbb::tbb_allocator<_Tp> _M_allocator;
     _Tp* _M_ptr;
@@ -68,9 +68,10 @@ class __buffer_impl<oneapi::dpl::__internal::__tbb_backend_tag, _Tp>
     operator=(const __buffer_impl&) = delete;
 
   public:
+    static_assert(::std::is_same_v<_ExecutionPolicy, ::std::decay_t<_ExecutionPolicy>>);
 
     //! Try to obtain buffer of given size to store objects of _Tp type
-    __buffer_impl(oneapi::dpl::__internal::__tbb_backend_tag, const ::std::size_t __n)
+    __buffer_impl(oneapi::dpl::__internal::__tbb_backend_tag, _ExecutionPolicy /*__exec*/, const ::std::size_t __n)
         : _M_allocator(), _M_ptr(_M_allocator.allocate(__n)), _M_buf_size(__n)
     {
     }
@@ -86,8 +87,8 @@ class __buffer_impl<oneapi::dpl::__internal::__tbb_backend_tag, _Tp>
     ~__buffer_impl() { _M_allocator.deallocate(_M_ptr, _M_buf_size); }
 };
 
-template <typename _Tp>
-using __buffer = __buffer_impl<oneapi::dpl::__internal::__tbb_backend_tag, _Tp>;
+template <typename _ExecutionPolicy, typename _Tp>
+using __buffer = __buffer_impl<oneapi::dpl::__internal::__tbb_backend_tag, ::std::decay_t<_ExecutionPolicy>, _Tp>;
 
 // Wrapper for tbb::task
 inline void
@@ -395,7 +396,8 @@ __parallel_strict_scan(oneapi::dpl::__internal::__tbb_backend_tag, _ExecutionPol
             const _Index __slack = 4;
             _Index __tilesize = (__n - 1) / (__slack * __p) + 1;
             _Index __m = (__n - 1) / __tilesize;
-            __tbb_backend::__buffer<_Tp> __buf(oneapi::dpl::__internal::__tbb_backend_tag{}, __m + 1);
+            __tbb_backend::__buffer<_ExecutionPolicy, _Tp> __buf(oneapi::dpl::__internal::__tbb_backend_tag{}, __exec,
+                                                                 __m + 1);
             _Tp* __r = __buf.get();
             __tbb_backend::__upsweep(_Index(0), _Index(__m + 1), __tilesize, __r, __n - __m * __tilesize, __reduce,
                                      __combine);
@@ -1200,7 +1202,8 @@ __parallel_stable_sort(oneapi::dpl::__internal::__tbb_backend_tag, _ExecutionPol
         const _DifferenceType __sort_cut_off = _ONEDPL_STABLE_SORT_CUT_OFF;
         if (__n > __sort_cut_off)
         {
-            __tbb_backend::__buffer<_ValueType> __buf(oneapi::dpl::__internal::__tbb_backend_tag{}, __n);
+            __tbb_backend::__buffer<_ExecutionPolicy, _ValueType> __buf(oneapi::dpl::__internal::__tbb_backend_tag{},
+                                                                        __exec, __n);
             __root_task<__stable_sort_func<_RandomAccessIterator, _ValueType*, _Compare, _LeafSort>> __root{
                 __xs, __xe, __buf.get(), true, __comp, __leaf_sort, __nsort, __xs, __buf.get()};
             __task::spawn_root_and_wait(__root);

From c044137bbda413b2518c51c475b2e0b2a05d2bbc Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 5 Mar 2024 09:59:24 +0100
Subject: [PATCH 511/566] Fix review comment: Why is policy removed? The
 agreement was that we keep both a policy and a tag - align all __buffer_impl
 API: pass ExecutionPolicy into constructor

---
 .../dpl/internal/reduce_by_segment_impl.h     |  6 +++---
 include/oneapi/dpl/pstl/algorithm_impl.h      | 20 +++++++++----------
 include/oneapi/dpl/pstl/omp/parallel_scan.h   | 15 +++++++-------
 .../oneapi/dpl/pstl/parallel_backend_serial.h |  2 +-
 .../oneapi/dpl/pstl/parallel_backend_tbb.h    |  4 ++--
 5 files changed, 24 insertions(+), 23 deletions(-)

diff --git a/include/oneapi/dpl/internal/reduce_by_segment_impl.h b/include/oneapi/dpl/internal/reduce_by_segment_impl.h
index 919da41ae80..900a5c16c5b 100644
--- a/include/oneapi/dpl/internal/reduce_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/reduce_by_segment_impl.h
@@ -116,7 +116,7 @@ reduce_by_segment_impl(_Tag, Policy&& policy, InputIterator1 first1, InputIterat
 
     // buffer that is used to store a flag indicating if the associated key is not equal to
     // the next key, and thus its associated sum should be part of the final result
-    oneapi::dpl::__par_backend::__buffer<Policy, FlagType> _mask(__backend_tag{}, n + 1);
+    oneapi::dpl::__par_backend::__buffer<Policy, FlagType> _mask(__backend_tag{}, policy, n + 1);
     auto mask = _mask.get();
     mask[0] = 1;
 
@@ -132,11 +132,11 @@ reduce_by_segment_impl(_Tag, Policy&& policy, InputIterator1 first1, InputIterat
     // buffer stores the sums of values associated with a given key. Sums are copied with
     // a shift into result2, and the shift is computed at the same time as the sums, so the
     // sums can't be written to result2 directly.
-    oneapi::dpl::__par_backend::__buffer<Policy, ValueType> _scanned_values(__backend_tag{}, n);
+    oneapi::dpl::__par_backend::__buffer<Policy, ValueType> _scanned_values(__backend_tag{}, policy, n);
 
     // Buffer is used to store results of the scan of the mask. Values indicate which position
     // in result2 needs to be written with the scanned_values element.
-    oneapi::dpl::__par_backend::__buffer<Policy, FlagType> _scanned_tail_flags(__backend_tag{}, n);
+    oneapi::dpl::__par_backend::__buffer<Policy, FlagType> _scanned_tail_flags(__backend_tag{}, policy, n);
 
     // Compute the sum of the segments. scanned_tail_flags values are not used.
     inclusive_scan(policy, make_zip_iterator(first2, _mask.get()), make_zip_iterator(first2, _mask.get()) + n,
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 1fe1f07ec6f..09fdb63bea5 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -1346,7 +1346,7 @@ __pattern_copy_if(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomA
     const _DifferenceType __n = __last - __first;
     if (_DifferenceType(1) < __n)
     {
-        __par_backend::__buffer<_ExecutionPolicy, bool> __mask_buf(__backend_tag{}, __n);
+        __par_backend::__buffer<_ExecutionPolicy, bool> __mask_buf(__backend_tag{}, __exec, __n);
         return __internal::__except_handler([&__exec, __n, __first, __result, __pred, &__mask_buf]() {
             bool* __mask = __mask_buf.get();
             _DifferenceType __m{};
@@ -1467,7 +1467,7 @@ __remove_elements(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Forward
     typedef typename ::std::iterator_traits<_ForwardIterator>::difference_type _DifferenceType;
     typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _Tp;
     _DifferenceType __n = __last - __first;
-    __par_backend::__buffer<_ExecutionPolicy, bool> __mask_buf(__backend_tag{}, __n);
+    __par_backend::__buffer<_ExecutionPolicy, bool> __mask_buf(__backend_tag{}, __exec, __n);
     // 1. find a first iterator that should be removed
     return __internal::__except_handler([&]() {
         bool* __mask = __mask_buf.get();
@@ -1504,7 +1504,7 @@ __remove_elements(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Forward
         __n -= __min;
         __first += __min;
 
-        __par_backend::__buffer<_ExecutionPolicy, _Tp> __buf(__backend_tag{}, __n);
+        __par_backend::__buffer<_ExecutionPolicy, _Tp> __buf(__backend_tag{}, __exec, __n);
         _Tp* __result = __buf.get();
         __mask += __min;
         _DifferenceType __m{};
@@ -1633,7 +1633,7 @@ __pattern_unique_copy(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Ran
     const _DifferenceType __n = __last - __first;
     if (_DifferenceType(2) < __n)
     {
-        __par_backend::__buffer<_ExecutionPolicy, bool> __mask_buf(__backend_tag{}, __n);
+        __par_backend::__buffer<_ExecutionPolicy, bool> __mask_buf(__backend_tag{}, __exec, __n);
         if (_DifferenceType(2) < __n)
         {
             return __internal::__except_handler([&__exec, __n, __first, __result, __pred, &__mask_buf]() {
@@ -1877,7 +1877,7 @@ __pattern_rotate(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAc
     auto __m = __middle - __first;
     if (__m <= __n / 2)
     {
-        __par_backend::__buffer<_ExecutionPolicy, _Tp> __buf(__backend_tag{}, __n - __m);
+        __par_backend::__buffer<_ExecutionPolicy, _Tp> __buf(__backend_tag{}, __exec, __n - __m);
         return __internal::__except_handler([&__exec, __n, __m, __first, __middle, __last, &__buf]() {
             _Tp* __result = __buf.get();
             __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __middle, __last,
@@ -1903,7 +1903,7 @@ __pattern_rotate(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAc
     }
     else
     {
-        __par_backend::__buffer<_ExecutionPolicy, _Tp> __buf(__backend_tag{}, __m);
+        __par_backend::__buffer<_ExecutionPolicy, _Tp> __buf(__backend_tag{}, __exec, __m);
         return __internal::__except_handler([&__exec, __n, __m, __first, __middle, __last, &__buf]() {
             _Tp* __result = __buf.get();
             __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __middle,
@@ -2386,7 +2386,7 @@ __pattern_partition_copy(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _
     const _DifferenceType __n = __last - __first;
     if (_DifferenceType(1) < __n)
     {
-        __par_backend::__buffer<_ExecutionPolicy, bool> __mask_buf(__backend_tag{}, __n);
+        __par_backend::__buffer<_ExecutionPolicy, bool> __mask_buf(__backend_tag{}, __exec, __n);
         return __internal::__except_handler([&__exec, __n, __first, __out_true, __out_false, __pred, &__mask_buf]() {
             bool* __mask = __mask_buf.get();
             _ReturnType __m{};
@@ -2617,7 +2617,7 @@ __pattern_partial_sort_copy(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec
         {
             typedef typename ::std::iterator_traits<_RandomAccessIterator1>::value_type _T1;
             typedef typename ::std::iterator_traits<_RandomAccessIterator2>::value_type _T2;
-            __par_backend::__buffer<_ExecutionPolicy, _T1> __buf(__backend_tag{}, __n1);
+            __par_backend::__buffer<_ExecutionPolicy, _T1> __buf(__backend_tag{}, __exec, __n1);
             _T1* __r = __buf.get();
 
             __par_backend::__parallel_stable_sort(
@@ -3122,7 +3122,7 @@ __pattern_inplace_merge(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _R
 
     typedef typename ::std::iterator_traits<_RandomAccessIterator>::value_type _Tp;
     auto __n = __last - __first;
-    __par_backend::__buffer<_ExecutionPolicy, _Tp> __buf(__backend_tag{}, __n);
+    __par_backend::__buffer<_ExecutionPolicy, _Tp> __buf(__backend_tag{}, __exec, __n);
     _Tp* __r = __buf.get();
     __internal::__except_handler([&]() {
         auto __move_values = [](_RandomAccessIterator __x, _Tp* __z) {
@@ -3257,7 +3257,7 @@ __parallel_set_op(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Forward
     const _DifferenceType __n1 = __last1 - __first1;
     const _DifferenceType __n2 = __last2 - __first2;
 
-    __par_backend::__buffer<_ExecutionPolicy, _T> __buf(__backend_tag{}, __size_func(__n1, __n2));
+    __par_backend::__buffer<_ExecutionPolicy, _T> __buf(__backend_tag{}, __exec, __size_func(__n1, __n2));
 
     return __internal::__except_handler([&__exec, __n1, __first1, __last1, __first2, __last2, __result, __comp,
                                          __size_func, __set_op, &__buf]() {
diff --git a/include/oneapi/dpl/pstl/omp/parallel_scan.h b/include/oneapi/dpl/pstl/omp/parallel_scan.h
index e9afe34c6a1..0bcbc85ac66 100644
--- a/include/oneapi/dpl/pstl/omp/parallel_scan.h
+++ b/include/oneapi/dpl/pstl/omp/parallel_scan.h
@@ -82,13 +82,14 @@ __downsweep(_Index __i, _Index __m, _Index __tilesize, _Tp* __r, _Index __lastsi
 template <typename _ExecutionPolicy, typename _Index, typename _Tp, typename _Rp, typename _Cp, typename _Sp,
           typename _Ap>
 void
-__parallel_strict_scan_body(_Index __n, _Tp __initial, _Rp __reduce, _Cp __combine, _Sp __scan, _Ap __apex)
+__parallel_strict_scan_body(_ExecutionPolicy&& __exec, _Index __n, _Tp __initial, _Rp __reduce, _Cp __combine, _Sp __scan,
+                            _Ap __apex)
 {
     _Index __p = omp_get_num_threads();
     const _Index __slack = 4;
     _Index __tilesize = (__n - 1) / (__slack * __p) + 1;
     _Index __m = (__n - 1) / __tilesize;
-    __buffer<_ExecutionPolicy, _Tp> __buf(oneapi::dpl::__internal::__omp_backend_tag{}, __m + 1);
+    __buffer<_ExecutionPolicy, _Tp> __buf(oneapi::dpl::__internal::__omp_backend_tag{}, __exec, __m + 1);
     _Tp* __r = __buf.get();
 
     oneapi::dpl::__omp_backend::__upsweep(_Index(0), _Index(__m + 1), __tilesize, __r, __n - __m * __tilesize, __reduce,
@@ -108,7 +109,7 @@ __parallel_strict_scan_body(_Index __n, _Tp __initial, _Rp __reduce, _Cp __combi
 
 template <class _ExecutionPolicy, typename _Index, typename _Tp, typename _Rp, typename _Cp, typename _Sp, typename _Ap>
 void
-__parallel_strict_scan(oneapi::dpl::__internal::__omp_backend_tag, _ExecutionPolicy&&, _Index __n, _Tp __initial,
+__parallel_strict_scan(oneapi::dpl::__internal::__omp_backend_tag, _ExecutionPolicy&& __exec, _Index __n, _Tp __initial,
                        _Rp __reduce, _Cp __combine, _Sp __scan, _Ap __apex)
 {
     if (__n <= __default_chunk_size)
@@ -128,16 +129,16 @@ __parallel_strict_scan(oneapi::dpl::__internal::__omp_backend_tag, _ExecutionPol
 
     if (omp_in_parallel())
     {
-        oneapi::dpl::__omp_backend::__parallel_strict_scan_body<_ExecutionPolicy>(__n, __initial, __reduce, __combine,
-                                                                                  __scan, __apex);
+        oneapi::dpl::__omp_backend::__parallel_strict_scan_body(::std::forward<_ExecutionPolicy>(__exec), __n,
+                                                                __initial, __reduce, __combine, __scan, __apex);
     }
     else
     {
         _PSTL_PRAGMA(omp parallel)
         _PSTL_PRAGMA(omp single nowait)
         {
-            oneapi::dpl::__omp_backend::__parallel_strict_scan_body<_ExecutionPolicy>(__n, __initial, __reduce,
-                                                                                      __combine, __scan, __apex);
+            oneapi::dpl::__omp_backend::__parallel_strict_scan_body(::std::forward<_ExecutionPolicy>(__exec), __n,
+                                                                    __initial, __reduce, __combine, __scan, __apex);
         }
     }
 }
diff --git a/include/oneapi/dpl/pstl/parallel_backend_serial.h b/include/oneapi/dpl/pstl/parallel_backend_serial.h
index 832bc0b8a92..c26626ecbbd 100644
--- a/include/oneapi/dpl/pstl/parallel_backend_serial.h
+++ b/include/oneapi/dpl/pstl/parallel_backend_serial.h
@@ -48,7 +48,7 @@ class __buffer_impl<oneapi::dpl::__internal::__serial_backend_tag, _ExecutionPol
   public:
     static_assert(::std::is_same_v<_ExecutionPolicy, ::std::decay_t<_ExecutionPolicy>>);
 
-    __buffer_impl(oneapi::dpl::__internal::__serial_backend_tag, ::std::size_t __n)
+    __buffer_impl(oneapi::dpl::__internal::__serial_backend_tag, _ExecutionPolicy /*__exec*/, ::std::size_t __n)
         : __allocator_(), __ptr_(__allocator_.allocate(__n)), __buf_size_(__n)
     {
     }
diff --git a/include/oneapi/dpl/pstl/parallel_backend_tbb.h b/include/oneapi/dpl/pstl/parallel_backend_tbb.h
index fac8eb5c0b4..79a9fb44f52 100644
--- a/include/oneapi/dpl/pstl/parallel_backend_tbb.h
+++ b/include/oneapi/dpl/pstl/parallel_backend_tbb.h
@@ -386,7 +386,7 @@ __downsweep(_Index __i, _Index __m, _Index __tilesize, _Tp* __r, _Index __lastsi
 // T must have a trivial constructor and destructor.
 template <class _ExecutionPolicy, typename _Index, typename _Tp, typename _Rp, typename _Cp, typename _Sp, typename _Ap>
 void
-__parallel_strict_scan(oneapi::dpl::__internal::__tbb_backend_tag, _ExecutionPolicy&&, _Index __n, _Tp __initial,
+__parallel_strict_scan(oneapi::dpl::__internal::__tbb_backend_tag, _ExecutionPolicy&& __exec, _Index __n, _Tp __initial,
                        _Rp __reduce, _Cp __combine, _Sp __scan, _Ap __apex)
 {
     tbb::this_task_arena::isolate([=, &__combine]() {
@@ -1190,7 +1190,7 @@ __stable_sort_func<_RandomAccessIterator1, _RandomAccessIterator2, _Compare, _Le
 
 template <class _ExecutionPolicy, typename _RandomAccessIterator, typename _Compare, typename _LeafSort>
 void
-__parallel_stable_sort(oneapi::dpl::__internal::__tbb_backend_tag, _ExecutionPolicy&&, _RandomAccessIterator __xs,
+__parallel_stable_sort(oneapi::dpl::__internal::__tbb_backend_tag, _ExecutionPolicy&& __exec, _RandomAccessIterator __xs,
                        _RandomAccessIterator __xe, _Compare __comp, _LeafSort __leaf_sort, ::std::size_t __nsort)
 {
     tbb::this_task_arena::isolate([=, &__nsort]() {

From 557f61c0c6640319506dd5a493d578093c99e7f9 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 5 Mar 2024 10:39:27 +0100
Subject: [PATCH 512/566] Fix review comment: Why do we remove execution policy
 from bricks?

---
 .../dpl/internal/async_impl/glue_async_impl.h |  2 +-
 include/oneapi/dpl/pstl/algorithm_fwd.h       | 10 +--
 include/oneapi/dpl/pstl/algorithm_impl.h      | 64 ++++++++++---------
 include/oneapi/dpl/pstl/glue_algorithm_impl.h |  6 +-
 .../dpl/pstl/glue_algorithm_ranges_impl.h     |  4 +-
 include/oneapi/dpl/pstl/glue_memory_impl.h    | 25 ++++----
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 62 ++++++++++--------
 .../hetero/algorithm_ranges_impl_hetero.h     | 17 ++---
 .../dpl/pstl/hetero/numeric_impl_hetero.h     |  4 +-
 9 files changed, 104 insertions(+), 90 deletions(-)

diff --git a/include/oneapi/dpl/internal/async_impl/glue_async_impl.h b/include/oneapi/dpl/internal/async_impl/glue_async_impl.h
index 4dd32e3a012..194f6a84f08 100644
--- a/include/oneapi/dpl/internal/async_impl/glue_async_impl.h
+++ b/include/oneapi/dpl/internal/async_impl/glue_async_impl.h
@@ -82,7 +82,7 @@ copy_async(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _ForwardIterato
     wait_for_all(::std::forward<_Events>(__dependencies)...);
     auto ret_val = oneapi::dpl::__internal::__pattern_walk2_brick_async(
         __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result,
-        oneapi::dpl::__internal::__brick_copy<::std::decay_t<decltype(__dispatch_tag)>>{});
+        oneapi::dpl::__internal::__brick_copy<::std::decay_t<decltype(__dispatch_tag)>, _ExecutionPolicy>{});
     return ret_val;
 }
 
diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index 56dc6b54d6a..ce504e6fa33 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -406,13 +406,13 @@ __pattern_search_n(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessI
 // copy_n
 //------------------------------------------------------------------------
 
-template <class _Tag>
+template <class _Tag, typename _ExecutionPolicy>
 struct __brick_copy_n;
 
-template <class _Tag>
+template <class _Tag, typename _ExecutionPolicy>
 struct __brick_copy;
 
-template <class _Tag>
+template <class _Tag, typename _ExecutionPolicy>
 struct __brick_move;
 
 //------------------------------------------------------------------------
@@ -853,7 +853,7 @@ __pattern_nth_element(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAcce
 //------------------------------------------------------------------------
 // fill, fill_n
 //------------------------------------------------------------------------
-template <class _Tag, typename _Tp>
+template <class _Tag, typename _ExecutionPolicy, typename _Tp>
 struct __brick_fill;
 
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Tp>
@@ -864,7 +864,7 @@ template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator,
 _RandomAccessIterator
 __pattern_fill(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, const _Tp&);
 
-template <class _Tag, typename _Tp>
+template <class _Tag, typename _ExecutionPolicy, typename _Tp>
 struct __brick_fill_n;
 
 template <class _Tag, class _ExecutionPolicy, class _OutputIterator, class _Size, class _Tp>
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 09fdb63bea5..5f7efcc7036 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -1061,7 +1061,7 @@ __pattern_search_n(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _
 // clear that doing so is worth the trouble and extra layers of call chain.
 // Sometimes a little duplication for sake of regularity is better than the alternative.
 
-template <class _Tag>
+template <class _Tag, typename _ExecutionPolicy>
 struct __brick_copy_n
 {
     static_assert(__is_host_backend_tag_v<_Tag>);
@@ -1088,7 +1088,7 @@ struct __brick_copy_n
 // copy
 //------------------------------------------------------------------------
 
-template <class _Tag>
+template <class _Tag, typename _ExecutionPolicy>
 struct __brick_copy
 {
     static_assert(__is_host_backend_tag_v<_Tag>);
@@ -1122,7 +1122,7 @@ struct __brick_copy
 // move
 //------------------------------------------------------------------------
 
-template <class _Tag>
+template <class _Tag, typename _ExecutionPolicy>
 struct __brick_move
 {
     static_assert(__is_host_backend_tag_v<_Tag>);
@@ -1152,7 +1152,7 @@ struct __brick_move
     }
 };
 
-template <class _Tag>
+template <class _Tag, typename _ExecutionPolicy>
 struct __brick_move_destroy
 {
     static_assert(__is_host_backend_tag_v<_Tag>);
@@ -1532,7 +1532,7 @@ __remove_elements(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Forward
         // 3. Elements from result are moved to [first, last)
         __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __result,
                                       __result + __m, [__result, __first](_Tp* __i, _Tp* __j) {
-                                          __brick_move_destroy<__parallel_tag<_IsVector>>{}(
+                                          __brick_move_destroy<__parallel_tag<_IsVector>, _ExecutionPolicy>{}(
                                               __i, __j, __first + (__i - __result), _IsVector{});
                                       });
         return __first + __m;
@@ -1888,13 +1888,13 @@ __pattern_rotate(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAc
 
             __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __middle,
                                           [__last, __middle](_RandomAccessIterator __b, _RandomAccessIterator __e) {
-                                              __internal::__brick_move<__parallel_tag<_IsVector>>{}(
+                                              __internal::__brick_move<__parallel_tag<_IsVector>, _ExecutionPolicy>{}(
                                                   __b, __e, __b + (__last - __middle), _IsVector{});
                                           });
 
             __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __result,
                                           __result + (__n - __m), [__first, __result](_Tp* __b, _Tp* __e) {
-                                              __brick_move_destroy<__parallel_tag<_IsVector>>{}(
+                                              __brick_move_destroy<__parallel_tag<_IsVector>, _ExecutionPolicy>{}(
                                                   __b, __e, __first + (__b - __result), _IsVector{});
                                           });
 
@@ -1914,13 +1914,13 @@ __pattern_rotate(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAc
 
             __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __middle, __last,
                                           [__first, __middle](_RandomAccessIterator __b, _RandomAccessIterator __e) {
-                                              __internal::__brick_move<__parallel_tag<_IsVector>>{}(
+                                              __internal::__brick_move<__parallel_tag<_IsVector>, _ExecutionPolicy>{}(
                                                   __b, __e, __first + (__b - __middle), _IsVector{});
                                           });
 
             __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __result,
                                           __result + __m, [__n, __m, __first, __result](_Tp* __b, _Tp* __e) {
-                                              __brick_move_destroy<__parallel_tag<_IsVector>>{}(
+                                              __brick_move_destroy<__parallel_tag<_IsVector>, _ExecutionPolicy>{}(
                                                   __b, __e, __first + ((__n - __m) + (__b - __result)), _IsVector{});
                                           });
 
@@ -1950,8 +1950,9 @@ __brick_rotate_copy(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccess
                     _RandomAccessIterator1 __middle, _RandomAccessIterator1 __last,
                     _RandomAccessIterator2 __result) noexcept
 {
-    _RandomAccessIterator2 __res = __brick_copy<__parallel_tag<_IsVector>>{}(__middle, __last, __result);
-    return __internal::__brick_copy<__parallel_tag<_IsVector>>{}(__first, __middle, __res);
+    _RandomAccessIterator2 __res =
+        __brick_copy<__parallel_tag<_IsVector>, _ExecutionPolicy>{}(__middle, __last, __result);
+    return __internal::__brick_copy<__parallel_tag<_IsVector>, _ExecutionPolicy>{}(__first, __middle, __res);
 }
 
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _OutputIterator>
@@ -1976,7 +1977,7 @@ __pattern_rotate_copy(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Ran
     __par_backend::__parallel_for(
         __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
         [__first, __last, __middle, __result](_RandomAccessIterator1 __b, _RandomAccessIterator1 __e) {
-            __internal::__brick_copy<__parallel_tag<_IsVector>> __copy{};
+            __internal::__brick_copy<__parallel_tag<_IsVector>, _ExecutionPolicy> __copy{};
             if (__b > __middle)
             {
                 __copy(__b, __e, __result + (__b - __middle), _IsVector{});
@@ -2606,7 +2607,7 @@ __pattern_partial_sort_copy(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec
                     _RandomAccessIterator1 __j1 = __first + (__j - __d_first);
 
                     // 1. Copy elements from input to output
-                    __brick_copy<__parallel_tag<_IsVector>>{}(__i1, __j1, __i, _IsVector{});
+                    __brick_copy<__parallel_tag<_IsVector>, _ExecutionPolicy>{}(__i1, __j1, __i, _IsVector{});
                     // 2. Sort elements in output sequence
                     ::std::sort(__i, __j, __comp);
                 },
@@ -2642,7 +2643,7 @@ __pattern_partial_sort_copy(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec
             // 3. Move elements from temporary buffer to output
             __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __r, __r + __n2,
                                           [__r, __d_first](_T1* __i, _T1* __j) {
-                                              __brick_move_destroy<__parallel_tag<_IsVector>>{}(
+                                              __brick_move_destroy<__parallel_tag<_IsVector>, _ExecutionPolicy>{}(
                                                   __i, __j, __d_first + (__i - __r), _IsVector{});
                                           });
 
@@ -2791,7 +2792,7 @@ __pattern_nth_element(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec
 //------------------------------------------------------------------------
 // fill, fill_n
 //------------------------------------------------------------------------
-template <class _Tag, typename _Tp>
+template <class _Tag, typename _ExecutionPolicy, typename _Tp>
 struct __brick_fill
 {
     static_assert(__is_host_backend_tag_v<_Tag>);
@@ -2822,7 +2823,7 @@ __pattern_fill(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardItera
     static_assert(__is_backend_tag_serial_v<_Tag> ||
                   __is_backend_tag_parallel_forward_v<_Tag>);
 
-    __internal::__brick_fill<_Tag, _Tp>{__value}(__first, __last, typename _Tag::__is_vector{});
+    __internal::__brick_fill<_Tag, _ExecutionPolicy, _Tp>{__value}(__first, __last, typename _Tag::__is_vector{});
 }
 
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Tp>
@@ -2835,14 +2836,14 @@ __pattern_fill(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAcce
     return __internal::__except_handler([&__exec, __first, __last, &__value]() {
         __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
                                       [&__value](_RandomAccessIterator __begin, _RandomAccessIterator __end) {
-                                          __internal::__brick_fill<__parallel_tag<_IsVector>, _Tp>{__value}(
-                                              __begin, __end, _IsVector{});
+                                          __internal::__brick_fill<__parallel_tag<_IsVector>, _ExecutionPolicy, _Tp>{
+                                              __value}(__begin, __end, _IsVector{});
                                       });
         return __last;
     });
 }
 
-template <class _Tag, typename _Tp>
+template <class _Tag, typename _ExecutionPolicy, typename _Tp>
 struct __brick_fill_n
 {
     static_assert(__is_host_backend_tag_v<_Tag>);
@@ -2873,7 +2874,8 @@ __pattern_fill_n(_Tag, _ExecutionPolicy&&, _OutputIterator __first, _Size __coun
     static_assert(__is_backend_tag_serial_v<_Tag> ||
                   __is_backend_tag_parallel_forward_v<_Tag>);
 
-    return __internal::__brick_fill_n<_Tag, _Tp>{__value}(__first, __count, typename _Tag::__is_vector{});
+    return __internal::__brick_fill_n<_Tag, _ExecutionPolicy, _Tp>{__value}(__first, __count,
+                                                                            typename _Tag::__is_vector{});
 }
 
 template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Size, class _Tp>
@@ -3147,7 +3149,7 @@ __pattern_inplace_merge(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _R
             });
         __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __r, __r + __n,
                                       [__r, __first](_Tp* __i, _Tp* __j) {
-                                          __brick_move_destroy<__parallel_tag<_IsVector>>{}(
+                                          __brick_move_destroy<__parallel_tag<_IsVector>, _ExecutionPolicy>{}(
                                               __i, __j, __first + (__i - __r), _IsVector{});
                                       });
     });
@@ -3265,9 +3267,9 @@ __parallel_set_op(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Forward
         _DifferenceType __m{};
         auto __scan = [=](_DifferenceType, _DifferenceType, const _SetRange& __s) { // Scan
             if (!__s.empty())
-                __brick_move_destroy<__parallel_tag<_IsVector>>{}(__tmp_memory + __s.__buf_pos,
-                                                                  __tmp_memory + (__s.__buf_pos + __s.__len),
-                                                                  __result + __s.__pos, _IsVector{});
+                __brick_move_destroy<__parallel_tag<_IsVector>, _ExecutionPolicy>{}(
+                    __tmp_memory + __s.__buf_pos, __tmp_memory + (__s.__buf_pos + __s.__len), __result + __s.__pos,
+                    _IsVector{});
         };
         __par_backend::__parallel_strict_scan(
             __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __n1, _SetRange{0, 0, 0}, //-1, 0},
@@ -3339,7 +3341,7 @@ __parallel_set_union_op(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __ex
     const auto __n1 = __last1 - __first1;
     const auto __n2 = __last2 - __first2;
 
-    __brick_copy<__parallel_tag<_IsVector>> __copy_range{};
+    __brick_copy<__parallel_tag<_IsVector>, _ExecutionPolicy> __copy_range{};
 
     // {1} {}: parallel copying just first sequence
     if (__n2 == 0)
@@ -3666,21 +3668,21 @@ __pattern_set_difference(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __e
     // {1} \ {}: parallel copying just first sequence
     if (__n2 == 0)
         return __pattern_walk2_brick(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __result,
-                                     __internal::__brick_copy<__parallel_tag<_IsVector>>{});
+                                     __internal::__brick_copy<__parallel_tag<_IsVector>, _ExecutionPolicy>{});
 
     // testing  whether the sequences are intersected
     _RandomAccessIterator1 __left_bound_seq_1 = ::std::lower_bound(__first1, __last1, *__first2, __comp);
     //{1} < {2}: seq 2 is wholly greater than seq 1, so, parallel copying just first sequence
     if (__left_bound_seq_1 == __last1)
         return __pattern_walk2_brick(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __result,
-                                     __internal::__brick_copy<__parallel_tag<_IsVector>>{});
+                                     __internal::__brick_copy<__parallel_tag<_IsVector>, _ExecutionPolicy>{});
 
     // testing  whether the sequences are intersected
     _RandomAccessIterator2 __left_bound_seq_2 = ::std::lower_bound(__first2, __last2, *__first1, __comp);
     //{2} < {1}: seq 1 is wholly greater than seq 2, so, parallel copying just first sequence
     if (__left_bound_seq_2 == __last2)
         return __internal::__pattern_walk2_brick(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1,
-                                                 __result, __brick_copy<__parallel_tag<_IsVector>>{});
+                                                 __result, __brick_copy<__parallel_tag<_IsVector>, _ExecutionPolicy>{});
 
     if (__n1 + __n2 > __set_algo_cut_off)
         return __parallel_set_op(
@@ -4328,8 +4330,8 @@ __pattern_shift_left(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Forw
     {
         __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __n, __size,
                                       [__first, __n](_DiffType __i, _DiffType __j) {
-                                          __brick_move<__parallel_tag<_IsVector>>{}(__first + __i, __first + __j,
-                                                                                    __first + __i - __n, _IsVector{});
+                                          __brick_move<__parallel_tag<_IsVector>, _ExecutionPolicy>{}(
+                                              __first + __i, __first + __j, __first + __i - __n, _IsVector{});
                                       });
     }
     else //2. n < size/2; there is not enough memory to parallel copying; doing parallel copying by n elements
@@ -4340,7 +4342,7 @@ __pattern_shift_left(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Forw
             auto __end = ::std::min(__k + __n, __size);
             __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __k, __end,
                                           [__first, __n](_DiffType __i, _DiffType __j) {
-                                              __brick_move<__parallel_tag<_IsVector>>{}(
+                                              __brick_move<__parallel_tag<_IsVector>, _ExecutionPolicy>{}(
                                                   __first + __i, __first + __j, __first + __i - __n, _IsVector{});
                                           });
         }
diff --git a/include/oneapi/dpl/pstl/glue_algorithm_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
index 35bdfac332b..fa9bbb43051 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
@@ -271,7 +271,7 @@ copy(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _ForwardIterator1 __l
 
     return oneapi::dpl::__internal::__pattern_walk2_brick(
         __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result,
-        oneapi::dpl::__internal::__brick_copy<::std::decay_t<decltype(__dispatch_tag)>>{});
+        oneapi::dpl::__internal::__brick_copy<::std::decay_t<decltype(__dispatch_tag)>, _ExecutionPolicy>{});
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator1, class _Size, class _ForwardIterator2>
@@ -284,7 +284,7 @@ copy_n(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _Size __n, _Forward
 
     return oneapi::dpl::__internal::__pattern_walk2_brick_n(
         __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __n, __result,
-        oneapi::dpl::__internal::__brick_copy_n<::std::decay_t<decltype(__dispatch_tag)>>{});
+        oneapi::dpl::__internal::__brick_copy_n<::std::decay_t<decltype(__dispatch_tag)>, _DecayedExecutionPolicy>{});
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Predicate>
@@ -818,7 +818,7 @@ move(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _ForwardIterator1 __l
 
     return oneapi::dpl::__internal::__pattern_walk2_brick(
         __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __d_first,
-        oneapi::dpl::__internal::__brick_move<::std::decay_t<decltype(__dispatch_tag)>>{});
+        oneapi::dpl::__internal::__brick_move<::std::decay_t<decltype(__dispatch_tag)>, _DecayedExecutionPolicy>{});
 }
 
 // [partial.sort]
diff --git a/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
index 32f178b5153..3796812f620 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
@@ -253,7 +253,7 @@ copy(_ExecutionPolicy&& __exec, _Range1&& __rng, _Range2&& __result)
 
     oneapi::dpl::__internal::__ranges::__pattern_walk_n(
         __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
-        oneapi::dpl::__internal::__brick_copy<::std::decay_t<decltype(__dispatch_tag)>>{},
+        oneapi::dpl::__internal::__brick_copy<::std::decay_t<decltype(__dispatch_tag)>, _ExecutionPolicy>{},
         views::all_read(::std::forward<_Range1>(__rng)), views::all_write(::std::forward<_Range2>(__result)));
 }
 
@@ -612,7 +612,7 @@ move(_ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&& __rng2)
 
     oneapi::dpl::__internal::__ranges::__pattern_walk_n(
         __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
-        oneapi::dpl::__internal::__brick_move<::std::decay_t<decltype(__dispatch_tag)>>{},
+        oneapi::dpl::__internal::__brick_move<::std::decay_t<decltype(__dispatch_tag)>, _DecayedExecutionPolicy>{},
         views::all_read(::std::forward<_Range1>(__rng1)), views::all_write(::std::forward<_Range2>(__rng2)));
 }
 
diff --git a/include/oneapi/dpl/pstl/glue_memory_impl.h b/include/oneapi/dpl/pstl/glue_memory_impl.h
index edd93953626..f0026f9842e 100644
--- a/include/oneapi/dpl/pstl/glue_memory_impl.h
+++ b/include/oneapi/dpl/pstl/glue_memory_impl.h
@@ -51,7 +51,7 @@ uninitialized_copy(_ExecutionPolicy&& __exec, _InputIterator __first, _InputIter
     {
         return oneapi::dpl::__internal::__pattern_walk2_brick(
             __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result,
-            oneapi::dpl::__internal::__brick_copy<::std::decay_t<decltype(__dispatch_tag)>>{});
+            oneapi::dpl::__internal::__brick_copy<::std::decay_t<decltype(__dispatch_tag)>, _DecayedExecutionPolicy>{});
     }
     else
     {
@@ -75,7 +75,8 @@ uninitialized_copy_n(_ExecutionPolicy&& __exec, _InputIterator __first, _Size __
     {
         return oneapi::dpl::__internal::__pattern_walk2_brick_n(
             __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __n, __result,
-            oneapi::dpl::__internal::__brick_copy_n<::std::decay_t<decltype(__dispatch_tag)>>{});
+            oneapi::dpl::__internal::__brick_copy_n<::std::decay_t<decltype(__dispatch_tag)>,
+                                                    _DecayedExecutionPolicy>{});
     }
     else
     {
@@ -101,7 +102,7 @@ uninitialized_move(_ExecutionPolicy&& __exec, _InputIterator __first, _InputIter
     {
         return oneapi::dpl::__internal::__pattern_walk2_brick(
             __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result,
-            oneapi::dpl::__internal::__brick_copy<::std::decay_t<decltype(__dispatch_tag)>>{});
+            oneapi::dpl::__internal::__brick_copy<::std::decay_t<decltype(__dispatch_tag)>, _DecayedExecutionPolicy>{});
     }
     else
     {
@@ -125,7 +126,8 @@ uninitialized_move_n(_ExecutionPolicy&& __exec, _InputIterator __first, _Size __
     {
         return oneapi::dpl::__internal::__pattern_walk2_brick_n(
             __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __n, __result,
-            oneapi::dpl::__internal::__brick_copy_n<::std::decay_t<decltype(__dispatch_tag)>>{});
+            oneapi::dpl::__internal::__brick_copy_n<::std::decay_t<decltype(__dispatch_tag)>,
+                                                    _DecayedExecutionPolicy>{});
     }
     else
     {
@@ -150,8 +152,8 @@ uninitialized_fill(_ExecutionPolicy&& __exec, _ForwardIterator __first, _Forward
     {
         oneapi::dpl::__internal::__pattern_walk_brick(
             __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
-            oneapi::dpl::__internal::__brick_fill<::std::decay_t<decltype(__dispatch_tag)>, _ValueType>{
-                _ValueType(__value)});
+            oneapi::dpl::__internal::__brick_fill<::std::decay_t<decltype(__dispatch_tag)>, _DecayedExecutionPolicy,
+                                                  _ValueType>{_ValueType(__value)});
     }
     else
     {
@@ -174,8 +176,8 @@ uninitialized_fill_n(_ExecutionPolicy&& __exec, _ForwardIterator __first, _Size
     {
         return oneapi::dpl::__internal::__pattern_walk_brick_n(
             __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __n,
-            oneapi::dpl::__internal::__brick_fill_n<::std::decay_t<decltype(__dispatch_tag)>, _ValueType>{
-                _ValueType(__value)});
+            oneapi::dpl::__internal::__brick_fill_n<::std::decay_t<decltype(__dispatch_tag)>, _DecayedExecutionPolicy,
+                                                    _ValueType>{_ValueType(__value)});
     }
     else
     {
@@ -312,7 +314,8 @@ uninitialized_value_construct(_ExecutionPolicy&& __exec, _ForwardIterator __firs
     {
         oneapi::dpl::__internal::__pattern_walk_brick(
             __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
-            oneapi::dpl::__internal::__brick_fill<::std::decay_t<decltype(__dispatch_tag)>, _ValueType>{_ValueType()});
+            oneapi::dpl::__internal::__brick_fill<::std::decay_t<decltype(__dispatch_tag)>, _DecayedExecutionPolicy,
+                                                  _ValueType>{_ValueType()});
     }
     else
     {
@@ -335,8 +338,8 @@ uninitialized_value_construct_n(_ExecutionPolicy&& __exec, _ForwardIterator __fi
     {
         return oneapi::dpl::__internal::__pattern_walk_brick_n(
             __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __n,
-            oneapi::dpl::__internal::__brick_fill_n<::std::decay_t<decltype(__dispatch_tag)>, _ValueType>{
-                _ValueType()});
+            oneapi::dpl::__internal::__brick_fill_n<::std::decay_t<decltype(__dispatch_tag)>, _DecayedExecutionPolicy,
+                                                    _ValueType>{_ValueType()});
     }
     else
     {
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index fd958bf8b1b..0b6b9a517d9 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -375,8 +375,8 @@ __pattern_generate_n(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec,
 // brick_copy, brick_move
 //------------------------------------------------------------------------
 
-template <typename _BackendTag>
-struct __brick_copy_n<__hetero_tag<_BackendTag>>
+template <typename _BackendTag, typename _ExecutionPolicy>
+struct __brick_copy_n<__hetero_tag<_BackendTag>, _ExecutionPolicy>
 {
     template <typename _SourceT, typename _TargetT>
     void
@@ -386,8 +386,8 @@ struct __brick_copy_n<__hetero_tag<_BackendTag>>
     }
 };
 
-template <typename _BackendTag>
-struct __brick_copy<__hetero_tag<_BackendTag>>
+template <typename _BackendTag, typename _ExecutionPolicy>
+struct __brick_copy<__hetero_tag<_BackendTag>, _ExecutionPolicy>
 {
     template <typename _SourceT, typename _TargetT>
     void
@@ -397,8 +397,8 @@ struct __brick_copy<__hetero_tag<_BackendTag>>
     }
 };
 
-template <typename _BackendTag>
-struct __brick_move<__hetero_tag<_BackendTag>>
+template <typename _BackendTag, typename _ExecutionPolicy>
+struct __brick_move<__hetero_tag<_BackendTag>, _ExecutionPolicy>
 {
     template <typename _SourceT, typename _TargetT>
     void
@@ -408,8 +408,8 @@ struct __brick_move<__hetero_tag<_BackendTag>>
     }
 };
 
-template <typename _BackendTag, typename _SourceT>
-struct __brick_fill<__hetero_tag<_BackendTag>, _SourceT>
+template <typename _BackendTag, typename _ExecutionPolicy, typename _SourceT>
+struct __brick_fill<__hetero_tag<_BackendTag>, _ExecutionPolicy, _SourceT>
 {
     _SourceT __value;
     template <typename _TargetT>
@@ -420,8 +420,8 @@ struct __brick_fill<__hetero_tag<_BackendTag>, _SourceT>
     }
 };
 
-template <typename _BackendTag, typename _SourceT>
-struct __brick_fill_n<__hetero_tag<_BackendTag>, _SourceT>
+template <typename _BackendTag, typename _ExecutionPolicy, typename _SourceT>
+struct __brick_fill_n<__hetero_tag<_BackendTag>, _ExecutionPolicy, _SourceT>
 {
     _SourceT __value;
     template <typename _TargetT>
@@ -1040,7 +1040,7 @@ __pattern_remove_if(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec,
     //TODO: optimize copy back depending on Iterator, i.e. set_final_data for host iterator/pointer
     return __pattern_walk2(
         __tag, __par_backend_hetero::make_wrapped_policy<copy_back_wrapper>(::std::forward<_ExecutionPolicy>(__exec)),
-        __copy_first, __copy_last, __first, __brick_copy<__hetero_tag<_BackendTag>>{});
+        __copy_first, __copy_last, __first, __brick_copy<__hetero_tag<_BackendTag>, _ExecutionPolicy>{});
 }
 
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Iterator, typename _BinaryPredicate>
@@ -1061,7 +1061,7 @@ __pattern_unique(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _It
     return __pattern_walk2<_BackendTag, /*_IsSync=*/::std::true_type, __par_backend_hetero::access_mode::read_write,
                            __par_backend_hetero::access_mode::read_write>(
         __tag, __par_backend_hetero::make_wrapped_policy<copy_back_wrapper>(::std::forward<_ExecutionPolicy>(__exec)),
-        __copy_first, __copy_last, __first, __brick_copy<__hetero_tag<_BackendTag>>{});
+        __copy_first, __copy_last, __first, __brick_copy<__hetero_tag<_BackendTag>, _ExecutionPolicy>{});
 }
 
 //------------------------------------------------------------------------
@@ -1184,13 +1184,15 @@ __pattern_merge(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Ite
             __tag,
             oneapi::dpl::__par_backend_hetero::make_wrapped_policy<copy_back_wrapper>(
                 ::std::forward<_ExecutionPolicy>(__exec)),
-            __first2, __last2, __d_first, oneapi::dpl::__internal::__brick_copy<__hetero_tag<_BackendTag>>{});
+            __first2, __last2, __d_first,
+            oneapi::dpl::__internal::__brick_copy<__hetero_tag<_BackendTag>, _ExecutionPolicy>{});
     else if (__n2 == 0)
         oneapi::dpl::__internal::__pattern_walk2_brick(
             __tag,
             oneapi::dpl::__par_backend_hetero::make_wrapped_policy<copy_back_wrapper2>(
                 ::std::forward<_ExecutionPolicy>(__exec)),
-            __first1, __last1, __d_first, oneapi::dpl::__internal::__brick_copy<__hetero_tag<_BackendTag>>{});
+            __first1, __last1, __d_first,
+            oneapi::dpl::__internal::__brick_copy<__hetero_tag<_BackendTag>, _ExecutionPolicy>{});
     else
     {
         auto __keep1 = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _Iterator1>();
@@ -1239,7 +1241,7 @@ __pattern_inplace_merge(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __ex
     //TODO: optimize copy back depending on Iterator, i.e. set_final_data for host iterator/pointer
     __pattern_walk2(
         __tag, __par_backend_hetero::make_wrapped_policy<copy_back_wrapper>(::std::forward<_ExecutionPolicy>(__exec)),
-        __copy_first, __copy_last, __first, __brick_move<__hetero_tag<_BackendTag>>{});
+        __copy_first, __copy_last, __first, __brick_move<__hetero_tag<_BackendTag>, _ExecutionPolicy>{});
 }
 
 //------------------------------------------------------------------------
@@ -1324,11 +1326,12 @@ __pattern_stable_partition(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& _
     //TODO: optimize copy back if possible (inplace, decrease number of submits)
     __pattern_walk2<_BackendTag, /*_IsSync=*/::std::false_type>(
         __tag, __par_backend_hetero::make_wrapped_policy<copy_back_wrapper>(__exec), __true_result, copy_result.first,
-        __first, __brick_move<__hetero_tag<_BackendTag>>{});
+        __first, __brick_move<__hetero_tag<_BackendTag>, _ExecutionPolicy>{});
 
     __pattern_walk2(
         __tag, __par_backend_hetero::make_wrapped_policy<copy_back_wrapper2>(::std::forward<_ExecutionPolicy>(__exec)),
-        __false_result, copy_result.second, __first + true_count, __brick_move<__hetero_tag<_BackendTag>>{});
+        __false_result, copy_result.second, __first + true_count,
+        __brick_move<__hetero_tag<_BackendTag>, _ExecutionPolicy>{});
 
     return __first + true_count;
 }
@@ -1494,7 +1497,7 @@ __pattern_partial_sort_copy(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&&
         // full sort on them.
         auto __out_end = __pattern_walk2<_BackendTag, /*_IsSync=*/::std::false_type>(
             __tag, __par_backend_hetero::make_wrapped_policy<__initial_copy_1>(__exec), __first, __last, __out_first,
-            __brick_copy<__hetero_tag<_BackendTag>>{});
+            __brick_copy<__hetero_tag<_BackendTag>, _ExecutionPolicy>{});
 
         // Use regular sort as partial_sort isn't required to be stable
         __pattern_sort(
@@ -1516,7 +1519,7 @@ __pattern_partial_sort_copy(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&&
 
         auto __buf_last = __pattern_walk2<_BackendTag, /*_IsSync=*/::std::false_type>(
             __tag, __par_backend_hetero::make_wrapped_policy<__initial_copy_2>(__exec), __first, __last, __buf_first,
-            __brick_copy<__hetero_tag<_BackendTag>>{});
+            __brick_copy<__hetero_tag<_BackendTag>, _ExecutionPolicy>{});
 
         auto __buf_mid = __buf_first + __out_size;
 
@@ -1528,7 +1531,7 @@ __pattern_partial_sort_copy(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&&
 
         return __pattern_walk2(
             __tag, __par_backend_hetero::make_wrapped_policy<__copy_back>(::std::forward<_ExecutionPolicy>(__exec)),
-            __buf_first, __buf_mid, __out_first, __brick_copy<__hetero_tag<_BackendTag>>{});
+            __buf_first, __buf_mid, __out_first, __brick_copy<__hetero_tag<_BackendTag>, _ExecutionPolicy>{});
     }
 }
 
@@ -1636,7 +1639,7 @@ __pattern_rotate(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Iterator
         unseq_backend::__rotate_copy<typename ::std::iterator_traits<_Iterator>::difference_type>{__n, __shift}, __n,
         __buf.all_view(), __temp_rng);
 
-    using _Function = __brick_move<__hetero_tag<_BackendTag>>;
+    using _Function = __brick_move<__hetero_tag<_BackendTag>, _ExecutionPolicy>;
     auto __brick = unseq_backend::walk_n<_ExecutionPolicy, _Function>{_Function{}};
 
     oneapi::dpl::__par_backend_hetero::__parallel_for(_BackendTag{}, ::std::forward<_ExecutionPolicy>(__exec), __brick,
@@ -1782,7 +1785,8 @@ __pattern_set_difference(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __e
             __tag,
             oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__set_difference_copy_case_1>(
                 ::std::forward<_ExecutionPolicy>(__exec)),
-            __first1, __last1, __result, oneapi::dpl::__internal::__brick_copy<__hetero_tag<_BackendTag>>{});
+            __first1, __last1, __result,
+            oneapi::dpl::__internal::__brick_copy<__hetero_tag<_BackendTag>, _ExecutionPolicy>{});
     }
 
     return __pattern_hetero_set_op(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2,
@@ -1817,7 +1821,8 @@ __pattern_set_union(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec,
             __tag,
             oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__set_union_copy_case_1>(
                 ::std::forward<_ExecutionPolicy>(__exec)),
-            __first2, __last2, __result, oneapi::dpl::__internal::__brick_copy<__hetero_tag<_BackendTag>>{});
+            __first2, __last2, __result,
+            oneapi::dpl::__internal::__brick_copy<__hetero_tag<_BackendTag>, _ExecutionPolicy>{});
     }
 
     //{2} is empty
@@ -1827,7 +1832,8 @@ __pattern_set_union(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec,
             __tag,
             oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__set_union_copy_case_2>(
                 ::std::forward<_ExecutionPolicy>(__exec)),
-            __first1, __last1, __result, oneapi::dpl::__internal::__brick_copy<__hetero_tag<_BackendTag>>{});
+            __first1, __last1, __result,
+            oneapi::dpl::__internal::__brick_copy<__hetero_tag<_BackendTag>, _ExecutionPolicy>{});
     }
 
     typedef typename ::std::iterator_traits<_OutputIterator>::value_type _ValueType;
@@ -1896,7 +1902,8 @@ __pattern_set_symmetric_difference(__hetero_tag<_BackendTag> __tag, _ExecutionPo
             __tag,
             oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__set_symmetric_difference_copy_case_1>(
                 ::std::forward<_ExecutionPolicy>(__exec)),
-            __first2, __last2, __result, oneapi::dpl::__internal::__brick_copy<__hetero_tag<_BackendTag>>{});
+            __first2, __last2, __result,
+            oneapi::dpl::__internal::__brick_copy<__hetero_tag<_BackendTag>, _ExecutionPolicy>{});
     }
 
     //{2} is empty
@@ -1906,7 +1913,8 @@ __pattern_set_symmetric_difference(__hetero_tag<_BackendTag> __tag, _ExecutionPo
             __tag,
             oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__set_symmetric_difference_copy_case_2>(
                 ::std::forward<_ExecutionPolicy>(__exec)),
-            __first1, __last1, __result, oneapi::dpl::__internal::__brick_copy<__hetero_tag<_BackendTag>>{});
+            __first1, __last1, __result,
+            oneapi::dpl::__internal::__brick_copy<__hetero_tag<_BackendTag>, _ExecutionPolicy>{});
     }
 
     typedef typename ::std::iterator_traits<_OutputIterator>::value_type _ValueType;
@@ -1961,7 +1969,7 @@ __pattern_shift_left(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Rang
     //1. n >= size/2; 'size - _n' parallel copying
     if (__n >= __mid)
     {
-        using _Function = __brick_move<__hetero_tag<_BackendTag>>;
+        using _Function = __brick_move<__hetero_tag<_BackendTag>, _ExecutionPolicy>;
         auto __brick = oneapi::dpl::unseq_backend::walk_n<_ExecutionPolicy, _Function>{_Function{}};
 
         //TODO: to consider use just "read" access mode for a source range and just "write" - for a destination range.
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
index a8e788f60e8..d8fc18d4a64 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
@@ -422,7 +422,7 @@ __pattern_remove_if(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec,
 
     oneapi::dpl::__internal::__ranges::__pattern_walk_n(
         __tag, ::std::forward<_ExecutionPolicy>(__exec),
-        oneapi::dpl::__internal::__brick_copy<__hetero_tag<_BackendTag>>{}, __copy_rng_truncated,
+        oneapi::dpl::__internal::__brick_copy<__hetero_tag<_BackendTag>, _ExecutionPolicy>{}, __copy_rng_truncated,
         ::std::forward<_Range>(__rng));
 
     return __copy_last_id;
@@ -466,8 +466,9 @@ __pattern_unique(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Ra
     auto res = __ranges::__pattern_unique_copy(__tag, __exec, __rng, res_rng, __pred,
                                                oneapi::dpl::__internal::__pstl_assign());
 
-    __pattern_walk_n(__tag, ::std::forward<_ExecutionPolicy>(__exec), __brick_copy<__hetero_tag<_BackendTag>>{},
-                     res_rng, ::std::forward<_Range>(__rng));
+    __pattern_walk_n(__tag, ::std::forward<_ExecutionPolicy>(__exec),
+                     __brick_copy<__hetero_tag<_BackendTag>, _ExecutionPolicy>{}, res_rng,
+                     ::std::forward<_Range>(__rng));
     return res;
 }
 
@@ -504,8 +505,8 @@ __pattern_merge(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Ran
             __tag,
             oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__copy1_wrapper>(
                 ::std::forward<_ExecutionPolicy>(__exec)),
-            oneapi::dpl::__internal::__brick_copy<__hetero_tag<_BackendTag>>{}, ::std::forward<_Range2>(__rng2),
-            ::std::forward<_Range3>(__rng3));
+            oneapi::dpl::__internal::__brick_copy<__hetero_tag<_BackendTag>, _ExecutionPolicy>{},
+            ::std::forward<_Range2>(__rng2), ::std::forward<_Range3>(__rng3));
     }
     else if (__n2 == 0)
     {
@@ -513,8 +514,8 @@ __pattern_merge(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Ran
             __tag,
             oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__copy2_wrapper>(
                 ::std::forward<_ExecutionPolicy>(__exec)),
-            oneapi::dpl::__internal::__brick_copy<__hetero_tag<_BackendTag>>{}, ::std::forward<_Range1>(__rng1),
-            ::std::forward<_Range3>(__rng3));
+            oneapi::dpl::__internal::__brick_copy<__hetero_tag<_BackendTag>, _ExecutionPolicy>{},
+            ::std::forward<_Range1>(__rng1), ::std::forward<_Range3>(__rng3));
     }
     else
     {
@@ -692,7 +693,7 @@ __pattern_reduce_by_segment(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&&
 
     if (__n == 1)
     {
-        __brick_copy<__hetero_tag<_BackendTag>> __copy_range{};
+        __brick_copy<__hetero_tag<_BackendTag>, _ExecutionPolicy> __copy_range{};
 
         oneapi::dpl::__internal::__ranges::__pattern_walk_n(
             __tag, oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__copy_keys_wrapper>(__exec), __copy_range,
diff --git a/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
index ccba7e7dafc..a24a2de0ec1 100644
--- a/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
@@ -177,7 +177,7 @@ __pattern_transform_scan_base(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&
         // Move data from temporary buffer into results
         oneapi::dpl::__internal::__pattern_walk2_brick(
             __tag, ::std::move(__policy), __first_tmp, __last_tmp, __result,
-            oneapi::dpl::__internal::__brick_move<__hetero_tag<_BackendTag>>{});
+            oneapi::dpl::__internal::__brick_move<__hetero_tag<_BackendTag>, _ExecutionPolicy>{});
 
         //TODO: optimize copy back depending on Iterator, i.e. set_final_data for host iterator/pointer
     }
@@ -248,7 +248,7 @@ __pattern_adjacent_difference(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&
                 ::std::forward<_ExecutionPolicy>(__exec));
 
             __internal::__pattern_walk2_brick(__tag, __wrapped_policy, __first, __last, __d_first,
-                                              __internal::__brick_copy<__hetero_tag<_BackendTag>>{});
+                                              __internal::__brick_copy<__hetero_tag<_BackendTag>, _ExecutionPolicy>{});
 
             return __d_last;
         });

From 3f3c9e5f76eb0b79149a9b35bd18a16a5bd0fb17 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 5 Mar 2024 11:07:52 +0100
Subject: [PATCH 513/566] 
 include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h - fix compile errors

---
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 25 +++++++++++--------
 1 file changed, 14 insertions(+), 11 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 0b6b9a517d9..1ef0900744f 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -1032,7 +1032,8 @@ __pattern_remove_if(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec,
 
     using _ValueType = typename ::std::iterator_traits<_Iterator>::value_type;
 
-    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __buf(__backend_tag{}, __exec, __last - __first);
+    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __buf(_BackendTag{}, __exec,
+                                                                                    __last - __first);
     auto __copy_first = __buf.get();
 
     auto __copy_last = __pattern_copy_if(__tag, __exec, __first, __last, __copy_first, __not_pred<_Predicate>{__pred});
@@ -1053,7 +1054,8 @@ __pattern_unique(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _It
 
     using _ValueType = typename ::std::iterator_traits<_Iterator>::value_type;
 
-    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __buf(__backend_tag{}, __exec, __last - __first);
+    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __buf(_BackendTag{}, __exec,
+                                                                                    __last - __first);
     auto __copy_first = __buf.get();
     auto __copy_last = __pattern_unique_copy(__tag, __exec, __first, __last, __copy_first, __pred);
 
@@ -1227,7 +1229,7 @@ __pattern_inplace_merge(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __ex
     assert(__first < __middle && __middle < __last);
 
     auto __n = __last - __first;
-    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __buf(__backend_tag{}, __exec, __n);
+    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __buf(_BackendTag{}, __exec, __n);
     auto __copy_first = __buf.get();
     auto __copy_last = __copy_first + __n;
 
@@ -1315,8 +1317,8 @@ __pattern_stable_partition(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& _
 
     auto __n = __last - __first;
 
-    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __true_buf(__backend_tag{}, __exec, __n);
-    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __false_buf(__backend_tag{}, __exec, __n);
+    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __true_buf(_BackendTag{}, __exec, __n);
+    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __false_buf(_BackendTag{}, __exec, __n);
     auto __true_result = __true_buf.get();
     auto __false_result = __false_buf.get();
 
@@ -1513,7 +1515,8 @@ __pattern_partial_sort_copy(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&&
         // - create a temporary buffer and copy all the elements from the input buffer there
         // - run partial sort on the temporary buffer
         // - copy k elements from the temporary buffer to the output buffer.
-        oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __buf(__backend_tag{}, __exec, __in_size);
+        oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __buf(_BackendTag{}, __exec,
+                                                                                        __in_size);
 
         auto __buf_first = __buf.get();
 
@@ -1628,7 +1631,7 @@ __pattern_rotate(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Iterator
 
     auto __keep = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read_write, _Iterator>();
     auto __buf = __keep(__first, __last);
-    auto __temp_buf = oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _Tp>(__backend_tag{}, __exec, __n);
+    auto __temp_buf = oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _Tp>(_BackendTag{}, __exec, __n);
 
     auto __temp_rng =
         oneapi::dpl::__ranges::all_view<_Tp, __par_backend_hetero::access_mode::write>(__temp_buf.get_buffer());
@@ -1711,7 +1714,7 @@ __pattern_hetero_set_op(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _F
         __comp, __n1, __n2};
 
     // temporary buffer to store boolean mask
-    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, int32_t> __mask_buf(__backend_tag{}, __exec, __n1);
+    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, int32_t> __mask_buf(_BackendTag{}, __exec, __n1);
 
     auto __keep1 =
         oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _ForwardIterator1>();
@@ -1840,7 +1843,7 @@ __pattern_set_union(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec,
 
     // temporary buffer to store intermediate result
     const auto __n2 = __last2 - __first2;
-    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __diff(__backend_tag{}, __exec, __n2);
+    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __diff(_BackendTag{}, __exec, __n2);
     auto __buf = __diff.get();
 
     //1. Calc difference {2} \ {1}
@@ -1921,10 +1924,10 @@ __pattern_set_symmetric_difference(__hetero_tag<_BackendTag> __tag, _ExecutionPo
 
     // temporary buffers to store intermediate result
     const auto __n1 = __last1 - __first1;
-    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __diff_1(__backend_tag{}, __exec, __n1);
+    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __diff_1(_BackendTag{}, __exec, __n1);
     auto __buf_1 = __diff_1.get();
     const auto __n2 = __last2 - __first2;
-    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __diff_2(__backend_tag{}, __exec, __n2);
+    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __diff_2(_BackendTag{}, __exec, __n2);
     auto __buf_2 = __diff_2.get();
 
     //1. Calc difference {1} \ {2}

From cf7e231d190a261a8da12de97f4f0bd37025ecc7 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 5 Mar 2024 13:08:46 +0100
Subject: [PATCH 514/566] Apply GitHUB clang format

---
 .../dpl/internal/reduce_by_segment_impl.h     |  15 +-
 .../dpl/internal/scan_by_segment_impl.h       |   7 +-
 include/oneapi/dpl/pstl/algorithm_impl.h      | 156 ++++++------------
 .../experimental/internal/for_loop_impl.h     |  12 +-
 .../hetero/algorithm_ranges_impl_hetero.h     |  20 ++-
 .../pstl/hetero/dpcpp/execution_sycl_defs.h   |   2 +-
 .../pstl/hetero/dpcpp/parallel_backend_sycl.h |   3 +-
 .../dpcpp/parallel_backend_sycl_histogram.h   |   5 +-
 .../dpcpp/parallel_backend_sycl_radix_sort.h  |   2 +-
 .../dpcpp/parallel_backend_sycl_utils.h       |   1 -
 include/oneapi/dpl/pstl/histogram_impl.h      |   3 +-
 include/oneapi/dpl/pstl/numeric_impl.h        |  12 +-
 include/oneapi/dpl/pstl/omp/parallel_scan.h   |   4 +-
 .../oneapi/dpl/pstl/parallel_backend_tbb.h    |   5 +-
 14 files changed, 103 insertions(+), 144 deletions(-)

diff --git a/include/oneapi/dpl/internal/reduce_by_segment_impl.h b/include/oneapi/dpl/internal/reduce_by_segment_impl.h
index 900a5c16c5b..9ee25b41913 100644
--- a/include/oneapi/dpl/internal/reduce_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/reduce_by_segment_impl.h
@@ -264,16 +264,23 @@ __sycl_reduce_by_segment(__internal::__hetero_tag<_BackendTag>, _ExecutionPolicy
     ::std::size_t __n_groups = oneapi::dpl::__internal::__dpl_ceiling_div(__n, __wgroup_size * __vals_per_item);
 
     // intermediate reductions within a workgroup
-    auto __partials = oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, __val_type>(_BackendTag{}, __exec, __n_groups).get_buffer();
+    auto __partials =
+        oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, __val_type>(_BackendTag{}, __exec, __n_groups)
+            .get_buffer();
 
-    auto __end_idx = oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, __diff_type>(_BackendTag{}, __exec, 1).get_buffer();
+    auto __end_idx =
+        oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, __diff_type>(_BackendTag{}, __exec, 1)
+            .get_buffer();
 
     // the number of segment ends found in each work group
-    auto __seg_ends = oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, __diff_type>(_BackendTag{}, __exec, __n_groups).get_buffer();
+    auto __seg_ends =
+        oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, __diff_type>(_BackendTag{}, __exec, __n_groups)
+            .get_buffer();
 
     // buffer that stores an exclusive scan of the results
     auto __seg_ends_scanned =
-        oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, __diff_type>(_BackendTag{}, __exec, __n_groups).get_buffer();
+        oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, __diff_type>(_BackendTag{}, __exec, __n_groups)
+            .get_buffer();
 
     // 1. Count the segment ends in each workgroup
     auto __seg_end_identification = __exec.queue().submit([&](sycl::handler& __cgh) {
diff --git a/include/oneapi/dpl/internal/scan_by_segment_impl.h b/include/oneapi/dpl/internal/scan_by_segment_impl.h
index bbbc1db0d52..65b236ba852 100644
--- a/include/oneapi/dpl/internal/scan_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/scan_by_segment_impl.h
@@ -145,10 +145,13 @@ struct __sycl_scan_by_segment_impl
         ::std::size_t __n_groups = __internal::__dpl_ceiling_div(__n, __wgroup_size * __vals_per_item);
 
         auto __partials =
-            oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, __val_type>(_BackendTag{}, __exec, __n_groups).get_buffer();
+            oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, __val_type>(_BackendTag{}, __exec, __n_groups)
+                .get_buffer();
 
         // the number of segment ends found in each work group
-        auto __seg_ends = oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, bool>(_BackendTag{}, __exec, __n_groups).get_buffer();
+        auto __seg_ends =
+            oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, bool>(_BackendTag{}, __exec, __n_groups)
+                .get_buffer();
 
         // 1. Work group reduction
         auto __wg_scan = __exec.queue().submit([&](sycl::handler& __cgh) {
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 5f7efcc7036..8ef28edae67 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -67,8 +67,7 @@ template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Pre
 bool
 __pattern_any_of(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _Pred __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> ||
-                  __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_any_of(__first, __last, __pred, typename _Tag::__is_vector{});
 }
@@ -180,8 +179,7 @@ void
 __pattern_walk_brick(_Tag, _ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last,
                      _Brick __brick) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> ||
-                  __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     __brick(__first, __last, typename _Tag::__is_vector{});
 }
@@ -223,8 +221,7 @@ template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Siz
 _ForwardIterator
 __pattern_walk1_n(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _Size __n, _Function __f) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> ||
-                  __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_walk1_n(__first, __n, __f, typename _Tag::__is_vector{});
 }
@@ -243,8 +240,7 @@ template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Siz
 _ForwardIterator
 __pattern_walk_brick_n(_Tag, _ExecutionPolicy&& __exec, _ForwardIterator __first, _Size __n, _Brick __brick) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> ||
-                  __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __brick(__first, __n, typename _Tag::__is_vector{});
 }
@@ -368,8 +364,7 @@ _ForwardIterator2
 __pattern_walk2_n(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _Size __n, _ForwardIterator2 __first2,
                   _Function __f) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> ||
-                  __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_walk2_n(__first1, __n, __first2, __f, typename _Tag::__is_vector{});
 }
@@ -466,8 +461,7 @@ _ForwardIterator2
 __pattern_walk2_brick_n(_Tag, _ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _Size __n,
                         _ForwardIterator2 __first2, _Brick __brick) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> ||
-                  __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __brick(__first1, __n, __first2, typename _Tag::__is_vector{});
 }
@@ -614,8 +608,7 @@ bool
 __pattern_equal(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
                 _ForwardIterator2 __first2, _ForwardIterator2 __last2, _BinaryPredicate __p) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> ||
-                  __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_equal(__first1, __last1, __first2, __last2, __p, typename _Tag::__is_vector{});
 }
@@ -666,8 +659,7 @@ bool
 __pattern_equal(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
                 _ForwardIterator2 __first2, _BinaryPredicate __p) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> ||
-                  __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_equal(__first1, __last1, __first2, __p, typename _Tag::__is_vector{});
 }
@@ -714,8 +706,7 @@ _ForwardIterator
 __pattern_find_if(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                   _Predicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> ||
-                  __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_find_if(__first, __last, __pred, typename _Tag::__is_vector{});
 }
@@ -858,8 +849,7 @@ _ForwardIterator1
 __pattern_find_end(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first, _ForwardIterator1 __last,
                    _ForwardIterator2 __s_first, _ForwardIterator2 __s_last, _BinaryPredicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> ||
-                  __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_find_end(__first, __last, __s_first, __s_last, __pred, typename _Tag::__is_vector{});
 }
@@ -915,8 +905,7 @@ _ForwardIterator1
 __pattern_find_first_of(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first, _ForwardIterator1 __last,
                         _ForwardIterator2 __s_first, _ForwardIterator2 __s_last, _BinaryPredicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> ||
-                  __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_find_first_of(__first, __last, __s_first, __s_last, __pred,
                                              typename _Tag::__is_vector{});
@@ -965,8 +954,7 @@ _ForwardIterator1
 __pattern_search(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first, _ForwardIterator1 __last,
                  _ForwardIterator2 __s_first, _ForwardIterator2 __s_last, _BinaryPredicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> ||
-                  __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_search(__first, __last, __s_first, __s_last, __pred, typename _Tag::__is_vector{});
 }
@@ -1022,8 +1010,7 @@ _ForwardIterator
 __pattern_search_n(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _Size __count,
                    const _Tp& __value, _BinaryPredicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> ||
-                  __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_search_n(__first, __last, __count, __value, __pred, typename _Tag::__is_vector{});
 }
@@ -1328,8 +1315,7 @@ _OutputIterator
 __pattern_copy_if(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _OutputIterator __result,
                   _UnaryPredicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> ||
-                  __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_copy_if(__first, __last, __result, __pred, typename _Tag::__is_vector{});
 }
@@ -1394,8 +1380,7 @@ template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Pre
 typename ::std::iterator_traits<_ForwardIterator>::difference_type
 __pattern_count(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _Predicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> ||
-                  __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_count(__first, __last, __pred, typename _Tag::__is_vector{});
 }
@@ -1449,8 +1434,7 @@ _ForwardIterator
 __pattern_unique(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                  _BinaryPredicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> ||
-                  __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_unique(__first, __last, __pred, typename _Tag::__is_vector{});
 }
@@ -1593,8 +1577,7 @@ _OutputIterator
 __pattern_unique_copy(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                       _OutputIterator __result, _BinaryPredicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> ||
-                  __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_unique_copy(__first, __last, __result, __pred, typename _Tag::__is_vector{});
 }
@@ -1730,8 +1713,7 @@ template <class _Tag, class _ExecutionPolicy, class _BidirectionalIterator>
 void
 __pattern_reverse(_Tag, _ExecutionPolicy&&, _BidirectionalIterator __first, _BidirectionalIterator __last) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> ||
-                  __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     __internal::__brick_reverse(__first, __last, typename _Tag::__is_vector{});
 }
@@ -1779,8 +1761,7 @@ _OutputIterator
 __pattern_reverse_copy(_Tag, _ExecutionPolicy&&, _BidirectionalIterator __first, _BidirectionalIterator __last,
                        _OutputIterator __d_first) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> ||
-                  __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_reverse_copy(__first, __last, __d_first, typename _Tag::__is_vector{});
 }
@@ -1859,8 +1840,7 @@ _ForwardIterator
 __pattern_rotate(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __middle,
                  _ForwardIterator __last) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> ||
-                  __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_rotate(__first, __middle, __last, typename _Tag::__is_vector{});
 }
@@ -1938,8 +1918,7 @@ _OutputIterator
 __brick_rotate_copy(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __middle,
                     _ForwardIterator __last, _OutputIterator __result) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> ||
-                  __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return ::std::rotate_copy(__first, __middle, __last, __result);
 }
@@ -1960,8 +1939,7 @@ _OutputIterator
 __pattern_rotate_copy(_Tag __tag, _ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __middle,
                       _ForwardIterator __last, _OutputIterator __result) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> ||
-                  __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_rotate_copy(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __middle, __last,
                                            __result);
@@ -2043,8 +2021,7 @@ bool
 __pattern_is_partitioned(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                          _UnaryPredicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> ||
-                  __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_is_partitioned(__first, __last, __pred, typename _Tag::__is_vector{});
 }
@@ -2174,8 +2151,7 @@ _ForwardIterator
 __pattern_partition(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                     _UnaryPredicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> ||
-                  __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_partition(__first, __last, __pred, typename _Tag::__is_vector{});
 }
@@ -2275,8 +2251,7 @@ _BidirectionalIterator
 __pattern_stable_partition(_Tag, _ExecutionPolicy&&, _BidirectionalIterator __first, _BidirectionalIterator __last,
                            _UnaryPredicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> ||
-                  __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_stable_partition(__first, __last, __pred, typename _Tag::__is_vector{});
 }
@@ -2366,8 +2341,7 @@ ::std::pair<_OutputIterator1, _OutputIterator2>
 __pattern_partition_copy(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                          _OutputIterator1 __out_true, _OutputIterator2 __out_false, _UnaryPredicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> ||
-                  __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_partition_copy(__first, __last, __out_true, __out_false, __pred,
                                               typename _Tag::__is_vector{});
@@ -2423,8 +2397,7 @@ void
 __pattern_sort(_Tag, _ExecutionPolicy&&, _RandomAccessIterator __first, _RandomAccessIterator __last, _Compare __comp,
                _IsMoveConstructible) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> ||
-                  __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     ::std::sort(__first, __last, __comp);
 }
@@ -2456,8 +2429,7 @@ void
 __pattern_stable_sort(_Tag, _ExecutionPolicy&&, _RandomAccessIterator __first, _RandomAccessIterator __last,
                       _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> ||
-                  __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     ::std::stable_sort(__first, __last, __comp);
 }
@@ -2490,8 +2462,7 @@ __pattern_sort_by_key(_Tag, _ExecutionPolicy&& __exec, _RandomAccessIterator1 __
                       _RandomAccessIterator1 __keys_last, _RandomAccessIterator2 __values_first,
                       _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> ||
-                  __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     auto __beg = oneapi::dpl::make_zip_iterator(__keys_first, __values_first);
     auto __end = __beg + (__keys_last - __keys_first);
@@ -2537,8 +2508,7 @@ void
 __pattern_partial_sort(_Tag, _ExecutionPolicy&&, _RandomAccessIterator __first, _RandomAccessIterator __middle,
                        _RandomAccessIterator __last, _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> ||
-                  __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     ::std::partial_sort(__first, __middle, __last, __comp);
 }
@@ -2576,8 +2546,7 @@ _RandomAccessIterator
 __pattern_partial_sort_copy(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                             _RandomAccessIterator __d_first, _RandomAccessIterator __d_last, _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> ||
-                  __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return ::std::partial_sort_copy(__first, __last, __d_first, __d_last, __comp);
 }
@@ -2681,8 +2650,7 @@ _ForwardIterator
 __pattern_adjacent_find(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                         _BinaryPredicate __pred, _Semantic) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> ||
-                  __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_adjacent_find(__first, __last, __pred, typename _Tag::__is_vector{}, _Semantic::value);
 }
@@ -2742,8 +2710,7 @@ void
 __pattern_nth_element(_Tag, _ExecutionPolicy&&, _RandomAccessIterator __first, _RandomAccessIterator __nth,
                       _RandomAccessIterator __last, _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> ||
-                  __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     ::std::nth_element(__first, __nth, __last, __comp);
 }
@@ -2820,8 +2787,7 @@ template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Tp>
 void
 __pattern_fill(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, const _Tp& __value) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> ||
-                  __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     __internal::__brick_fill<_Tag, _ExecutionPolicy, _Tp>{__value}(__first, __last, typename _Tag::__is_vector{});
 }
@@ -2871,8 +2837,7 @@ template <class _Tag, class _ExecutionPolicy, class _OutputIterator, class _Size
 _OutputIterator
 __pattern_fill_n(_Tag, _ExecutionPolicy&&, _OutputIterator __first, _Size __count, const _Tp& __value) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> ||
-                  __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_fill_n<_Tag, _ExecutionPolicy, _Tp>{__value}(__first, __count,
                                                                             typename _Tag::__is_vector{});
@@ -2910,8 +2875,7 @@ template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Gen
 void
 __pattern_generate(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _Generator __g) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> ||
-                  __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     __internal::__brick_generate(__first, __last, __g, typename _Tag::__is_vector{});
 }
@@ -2951,8 +2915,7 @@ template <class _Tag, class _ExecutionPolicy, class _OutputIterator, class _Size
 _OutputIterator
 __pattern_generate_n(_Tag, _ExecutionPolicy&&, _OutputIterator __first, _Size __count, _Generator __g) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> ||
-                  __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_generate_n(__first, __count, __g, typename _Tag::__is_vector{});
 }
@@ -2997,8 +2960,7 @@ _ForwardIterator
 __pattern_remove_if(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                     _UnaryPredicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> ||
-                  __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_remove_if(__first, __last, __pred, typename _Tag::__is_vector{});
 }
@@ -3054,8 +3016,7 @@ __pattern_merge(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIt
                 _ForwardIterator2 __first2, _ForwardIterator2 __last2, _OutputIterator __d_first,
                 _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> ||
-                  __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_merge(__first1, __last1, __first2, __last2, __d_first, __comp,
                                      typename _Tag::__is_vector{});
@@ -3104,8 +3065,7 @@ void
 __pattern_inplace_merge(_Tag, _ExecutionPolicy&&, _BidirectionalIterator __first, _BidirectionalIterator __middle,
                         _BidirectionalIterator __last, _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> ||
-                  __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     __internal::__brick_inplace_merge(__first, __middle, __last, __comp, typename _Tag::__is_vector{});
 }
@@ -3164,8 +3124,7 @@ bool
 __pattern_includes(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
                    _ForwardIterator2 __first2, _ForwardIterator2 __last2, _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> ||
-                  __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return ::std::includes(__first1, __last1, __first2, __last2, __comp);
 }
@@ -3480,8 +3439,7 @@ __pattern_set_union(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _Forwa
                     _ForwardIterator2 __first2, _ForwardIterator2 __last2, _OutputIterator __result,
                     _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> ||
-                  __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_set_union(__first1, __last1, __first2, __last2, __result, __comp,
                                          typename _Tag::__is_vector{});
@@ -3542,8 +3500,7 @@ __pattern_set_intersection(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1,
                            _ForwardIterator2 __first2, _ForwardIterator2 __last2, _OutputIterator __result,
                            _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> ||
-                  __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_set_intersection(__first1, __last1, __first2, __last2, __result, __comp,
                                                 typename _Tag::__is_vector{});
@@ -3641,8 +3598,7 @@ __pattern_set_difference(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _
                          _ForwardIterator2 __first2, _ForwardIterator2 __last2, _OutputIterator __result,
                          _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> ||
-                  __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_set_difference(__first1, __last1, __first2, __last2, __result, __comp,
                                               typename _Tag::__is_vector{});
@@ -3729,8 +3685,7 @@ __pattern_set_symmetric_difference(_Tag, _ExecutionPolicy&&, _ForwardIterator1 _
                                    _ForwardIterator2 __first2, _ForwardIterator2 __last2, _OutputIterator __result,
                                    _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> ||
-                  __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_set_symmetric_difference(__first1, __last1, __first2, __last2, __result, __comp,
                                                         typename _Tag::__is_vector{});
@@ -3789,8 +3744,7 @@ _RandomAccessIterator
 __pattern_is_heap_until(_Tag, _ExecutionPolicy&&, _RandomAccessIterator __first, _RandomAccessIterator __last,
                         _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> ||
-                  __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_is_heap_until(__first, __last, __comp, typename _Tag::__is_vector{});
 }
@@ -3878,8 +3832,7 @@ bool
 __pattern_is_heap(_Tag, _ExecutionPolicy&&, _RandomAccessIterator __first, _RandomAccessIterator __last,
                   _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> ||
-                  __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_is_heap(__first, __last, __comp, typename _Tag::__is_vector{});
 }
@@ -3927,8 +3880,7 @@ _ForwardIterator
 __pattern_min_element(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                       _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> ||
-                  __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_min_element(__first, __last, __comp, typename _Tag::__is_vector{});
 }
@@ -3995,8 +3947,7 @@ ::std::pair<_ForwardIterator, _ForwardIterator>
 __pattern_minmax_element(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                          _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> ||
-                  __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_minmax_element(__first, __last, __comp, typename _Tag::__is_vector{});
 }
@@ -4077,8 +4028,7 @@ ::std::pair<_ForwardIterator1, _ForwardIterator2>
 __pattern_mismatch(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
                    _ForwardIterator2 __first2, _ForwardIterator2 __last2, _Predicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> ||
-                  __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_mismatch(__first1, __last1, __first2, __last2, __pred, typename _Tag::__is_vector{});
 }
@@ -4159,8 +4109,7 @@ bool
 __pattern_lexicographical_compare(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
                                   _ForwardIterator2 __first2, _ForwardIterator2 __last2, _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> ||
-                  __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_lexicographical_compare(__first1, __last1, __first2, __last2, __comp,
                                                        typename _Tag::__is_vector{});
@@ -4300,8 +4249,7 @@ _ForwardIterator
 __pattern_shift_left(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                      typename ::std::iterator_traits<_ForwardIterator>::difference_type __n) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> ||
-                  __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __brick_shift_left(__first, __last, __n, typename _Tag::__is_vector{});
 }
diff --git a/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h b/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
index b41ec62b84a..a6a6952738a 100644
--- a/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
+++ b/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
@@ -167,8 +167,7 @@ void
 __pattern_for_loop_n(_Tag, _ExecutionPolicy&&, _Ip __first, _Size __n, _Function __f, __single_stride_type,
                      _Rest&&... __rest) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> ||
-                  __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     __reduction_pack<_Rest...> __pack{__reduction_pack_tag(), ::std::forward<_Rest>(__rest)...};
 
@@ -193,8 +192,7 @@ void
 __pattern_for_loop_n(_Tag, _ExecutionPolicy&&, _Ip __first, _Size __n, _Function __f, _Sp __stride,
                      _Rest&&... __rest) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> ||
-                  __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     __reduction_pack<_Rest...> __pack{__reduction_pack_tag(), ::std::forward<_Rest>(__rest)...};
 
@@ -267,8 +265,7 @@ void
 __pattern_for_loop(_Tag __tag, _ExecutionPolicy&& __exec, _Ip __first, _Ip __last, _Function __f, _Sp __stride,
                    _Rest&&... __rest) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> ||
-                  __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     if constexpr (typename _Tag::__is_vector{} || __is_random_access_or_integral_v<_Ip>)
     {
@@ -369,8 +366,7 @@ void
 __pattern_for_loop(_Tag, _ExecutionPolicy&& __exec, _Ip __first, _Ip __last, _Function __f, __single_stride_type,
                    _Rest&&... __rest) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> ||
-                  __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     __reduction_pack<_Rest...> __pack{__reduction_pack_tag(), ::std::forward<_Rest>(__rest)...};
 
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
index d8fc18d4a64..013133eebbf 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
@@ -360,7 +360,8 @@ __pattern_scan_copy(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Range
     _DataAcc __get_data_op;
     _MaskAssigner __add_mask_op;
 
-    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, int32_t> __mask_buf(_BackendTag{}, __exec, __rng1.size());
+    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, int32_t> __mask_buf(_BackendTag{}, __exec,
+                                                                                      __rng1.size());
 
     auto __res =
         __par_backend_hetero::__parallel_transform_scan_base(
@@ -413,7 +414,8 @@ __pattern_remove_if(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec,
 
     using _ValueType = oneapi::dpl::__internal::__value_t<_Range>;
 
-    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __buf(_BackendTag{}, __exec, __rng.size());
+    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __buf(_BackendTag{}, __exec,
+                                                                                    __rng.size());
     auto __copy_rng = oneapi::dpl::__ranges::views::all(__buf.get_buffer());
 
     auto __copy_last_id = __ranges::__pattern_copy_if(__tag, __exec, __rng, __copy_rng, __not_pred<_Predicate>{__pred},
@@ -461,7 +463,8 @@ __pattern_unique(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Ra
 
     using _ValueType = oneapi::dpl::__internal::__value_t<_Range>;
 
-    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __buf(_BackendTag{}, __exec, __rng.size());
+    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __buf(_BackendTag{}, __exec,
+                                                                                    __rng.size());
     auto res_rng = oneapi::dpl::__ranges::views::all(__buf.get_buffer());
     auto res = __ranges::__pattern_unique_copy(__tag, __exec, __rng, res_rng, __pred,
                                                oneapi::dpl::__internal::__pstl_assign());
@@ -715,9 +718,14 @@ __pattern_reduce_by_segment(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&&
     // Round 1: reduce with extra indices added to avoid long segments
     // TODO: At threshold points check if the key is equal to the key at the previous threshold point, indicating a long sequence.
     // Skip a round of copy_if and reduces if there are none.
-    auto __idx = oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, __diff_type>(_BackendTag{}, __exec, __n).get_buffer();
-    auto __tmp_out_keys = oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, __key_type>(_BackendTag{}, __exec, __n) .get_buffer();
-    auto __tmp_out_values = oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, __val_type>(_BackendTag{}, __exec, __n) .get_buffer();
+    auto __idx = oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, __diff_type>(_BackendTag{}, __exec, __n)
+                     .get_buffer();
+    auto __tmp_out_keys =
+        oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, __key_type>(_BackendTag{}, __exec, __n)
+            .get_buffer();
+    auto __tmp_out_values =
+        oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, __val_type>(_BackendTag{}, __exec, __n)
+            .get_buffer();
 
     // Replicating first element of keys view to be able to compare (i-1)-th and (i)-th key with aligned sequences,
     //  dropping the last key for the i-1 sequence.
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h b/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h
index bdc2e8ac7bd..7640f6fbf86 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h
@@ -79,7 +79,7 @@ class fpga_policy : public device_policy<KernelName>
 #    else
               __dpl_sycl::__fpga_selector()
 #    endif // _ONEDPL_FPGA_EMU
-              ))
+                  ))
     {
     }
 
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
index 2a43411bdca..f7553d2d6ea 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
@@ -871,7 +871,8 @@ __parallel_scan_copy(oneapi::dpl::__internal::__device_backend_tag __backend_tag
     _MaskAssigner __add_mask_op;
 
     // temporary buffer to store boolean mask
-    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, int32_t> __mask_buf(oneapi::dpl::__internal::__device_backend_tag{}, __exec, __n);
+    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, int32_t> __mask_buf(
+        oneapi::dpl::__internal::__device_backend_tag{}, __exec, __n);
 
     return __parallel_transform_scan_base(
         __backend_tag, ::std::forward<_ExecutionPolicy>(__exec),
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h
index 4ea22c31b2a..ecaabd2ae48 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h
@@ -426,8 +426,9 @@ struct __histogram_general_private_global_atomics_submitter<_BackendTag,
         ::std::size_t __segments =
             oneapi::dpl::__internal::__dpl_ceiling_div(__n, __work_group_size * __iters_per_work_item);
 
-        auto __private_histograms =
-            oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _bin_type>(_BackendTag{}, __exec, __segments * __num_bins).get_buffer();
+        auto __private_histograms = oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _bin_type>(
+                                        _BackendTag{}, __exec, __segments * __num_bins)
+                                        .get_buffer();
 
         return __exec.queue().submit([&](auto& __h) {
             __h.depends_on(__init_event);
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_radix_sort.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_radix_sort.h
index bb9c08bbe4c..d754b901117 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_radix_sort.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_radix_sort.h
@@ -813,7 +813,7 @@ __parallel_radix_sort(oneapi::dpl::__internal::__device_backend_tag, _ExecutionP
         sycl::buffer<::std::uint32_t, 1> __tmp_buf{sycl::range<1>(__tmp_buf_size)};
 
         // memory for storing values sorted for an iteration
-        oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy,_ValueT> __out_buffer_holder{
+        oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueT> __out_buffer_holder{
             oneapi::dpl::__internal::__device_backend_tag{}, __exec, __n};
         auto __out_rng = oneapi::dpl::__ranges::all_view<_ValueT, __par_backend_hetero::access_mode::read_write>(
             __out_buffer_holder.get_buffer());
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_utils.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_utils.h
index 3d4166507a0..6a744fa2c25 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_utils.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_utils.h
@@ -389,7 +389,6 @@ class __buffer_impl<oneapi::dpl::__internal::__device_backend_tag, _ExecutionPol
     __container_t __container;
 
   public:
-
     static_assert(::std::is_same_v<_ExecutionPolicy, ::std::decay_t<_ExecutionPolicy>>);
 
     __buffer_impl(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy /*__exec*/,
diff --git a/include/oneapi/dpl/pstl/histogram_impl.h b/include/oneapi/dpl/pstl/histogram_impl.h
index bdc34746bc5..885600f0288 100644
--- a/include/oneapi/dpl/pstl/histogram_impl.h
+++ b/include/oneapi/dpl/pstl/histogram_impl.h
@@ -38,8 +38,7 @@ void
 __pattern_histogram(_Tag, _ExecutionPolicy&& exec, _RandomAccessIterator1 __first, _RandomAccessIterator1 __last,
                     _Size __num_bins, _IdxHashFunc __func, _RandomAccessIterator2 __histogram_first)
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> ||
-                  __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     static_assert(sizeof(_Size) == 0 /*false*/,
                   "Histogram API is currently unsupported for policies other than device execution policies");
diff --git a/include/oneapi/dpl/pstl/numeric_impl.h b/include/oneapi/dpl/pstl/numeric_impl.h
index 85d248ac348..2b8102dfe7a 100644
--- a/include/oneapi/dpl/pstl/numeric_impl.h
+++ b/include/oneapi/dpl/pstl/numeric_impl.h
@@ -68,8 +68,7 @@ __pattern_transform_reduce(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1,
                            _ForwardIterator2 __first2, _Tp __init, _BinaryOperation1 __binary_op1,
                            _BinaryOperation2 __binary_op2) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> ||
-                  __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __brick_transform_reduce(__first1, __last1, __first2, __init, __binary_op1, __binary_op2,
                                     typename _Tag::__is_vector{});
@@ -134,8 +133,7 @@ _Tp
 __pattern_transform_reduce(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _Tp __init,
                            _BinaryOperation __binary_op, _UnaryOperation __unary_op) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> ||
-                  __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_transform_reduce(__first, __last, __init, __binary_op, __unary_op,
                                                 typename _Tag::__is_vector{});
@@ -245,8 +243,7 @@ __pattern_transform_scan(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _Fo
                          _OutputIterator __result, _UnaryOperation __unary_op, _Tp __init, _BinaryOperation __binary_op,
                          _Inclusive) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> ||
-                  __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_transform_scan(__first, __last, __result, __unary_op, __init, __binary_op, _Inclusive(),
                                               typename _Tag::__is_vector{})
@@ -382,8 +379,7 @@ _OutputIterator
 __pattern_adjacent_difference(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                               _OutputIterator __d_first, _BinaryOperation __op) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> ||
-                  __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
 
     return __internal::__brick_adjacent_difference(__first, __last, __d_first, __op, typename _Tag::__is_vector{});
 }
diff --git a/include/oneapi/dpl/pstl/omp/parallel_scan.h b/include/oneapi/dpl/pstl/omp/parallel_scan.h
index 0bcbc85ac66..958161a2325 100644
--- a/include/oneapi/dpl/pstl/omp/parallel_scan.h
+++ b/include/oneapi/dpl/pstl/omp/parallel_scan.h
@@ -82,8 +82,8 @@ __downsweep(_Index __i, _Index __m, _Index __tilesize, _Tp* __r, _Index __lastsi
 template <typename _ExecutionPolicy, typename _Index, typename _Tp, typename _Rp, typename _Cp, typename _Sp,
           typename _Ap>
 void
-__parallel_strict_scan_body(_ExecutionPolicy&& __exec, _Index __n, _Tp __initial, _Rp __reduce, _Cp __combine, _Sp __scan,
-                            _Ap __apex)
+__parallel_strict_scan_body(_ExecutionPolicy&& __exec, _Index __n, _Tp __initial, _Rp __reduce, _Cp __combine,
+                            _Sp __scan, _Ap __apex)
 {
     _Index __p = omp_get_num_threads();
     const _Index __slack = 4;
diff --git a/include/oneapi/dpl/pstl/parallel_backend_tbb.h b/include/oneapi/dpl/pstl/parallel_backend_tbb.h
index 79a9fb44f52..c2bf8bda04a 100644
--- a/include/oneapi/dpl/pstl/parallel_backend_tbb.h
+++ b/include/oneapi/dpl/pstl/parallel_backend_tbb.h
@@ -1190,8 +1190,9 @@ __stable_sort_func<_RandomAccessIterator1, _RandomAccessIterator2, _Compare, _Le
 
 template <class _ExecutionPolicy, typename _RandomAccessIterator, typename _Compare, typename _LeafSort>
 void
-__parallel_stable_sort(oneapi::dpl::__internal::__tbb_backend_tag, _ExecutionPolicy&& __exec, _RandomAccessIterator __xs,
-                       _RandomAccessIterator __xe, _Compare __comp, _LeafSort __leaf_sort, ::std::size_t __nsort)
+__parallel_stable_sort(oneapi::dpl::__internal::__tbb_backend_tag, _ExecutionPolicy&& __exec,
+                       _RandomAccessIterator __xs, _RandomAccessIterator __xe, _Compare __comp, _LeafSort __leaf_sort,
+                       ::std::size_t __nsort)
 {
     tbb::this_task_arena::isolate([=, &__nsort]() {
         //sorting based on task tree and parallel merge

From 33b3c94b43900eb580c6dcca25a4a10a1d5d51c0 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 5 Mar 2024 13:25:10 +0100
Subject: [PATCH 515/566] include/oneapi/dpl/pstl/algorithm_impl.h - remove
 using __backend_tag = ... as not used

---
 include/oneapi/dpl/pstl/algorithm_impl.h | 2 --
 1 file changed, 2 deletions(-)

diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 8ef28edae67..937a74a9b94 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -918,8 +918,6 @@ __pattern_find_first_of(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __ex
                         _ForwardIterator1 __last, _ForwardIterator2 __s_first, _ForwardIterator2 __s_last,
                         _BinaryPredicate __pred)
 {
-    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
-
     return __internal::__except_handler([&]() {
         return __internal::__parallel_find(
             __tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,

From 759566ec45e80daaee59b749c9dc010345a54e7a Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 6 Mar 2024 09:23:42 +0100
Subject: [PATCH 516/566] Revert
 "include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h - fix review
 comment: Why do we need using base then?"

This reverts commit 93036b23906274fa2c7db95f652d3f7706f90815.
---
 .../oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h | 10 ++++++----
 1 file changed, 6 insertions(+), 4 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h b/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h
index 7640f6fbf86..93d291b661b 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h
@@ -69,11 +69,13 @@ struct DefaultKernelNameFPGA;
 template <unsigned int factor = 1, typename KernelName = DefaultKernelNameFPGA>
 class fpga_policy : public device_policy<KernelName>
 {
+    using base = device_policy<KernelName>;
+
   public:
     static constexpr unsigned int unroll_factor = factor;
 
     fpga_policy()
-        : device_policy<KernelName>(sycl::queue(
+        : base(sycl::queue(
 #    if _ONEDPL_FPGA_EMU
               __dpl_sycl::__fpga_emulator_selector()
 #    else
@@ -84,9 +86,9 @@ class fpga_policy : public device_policy<KernelName>
     }
 
     template <unsigned int other_factor, typename OtherName>
-    fpga_policy(const fpga_policy<other_factor, OtherName>& other) : device_policy<KernelName>(other.queue()){};
-    explicit fpga_policy(sycl::queue q) : device_policy<KernelName>(q) {}
-    explicit fpga_policy(sycl::device d) : device_policy<KernelName>(d) {}
+    fpga_policy(const fpga_policy<other_factor, OtherName>& other) : base(other.queue()){};
+    explicit fpga_policy(sycl::queue q) : base(q) {}
+    explicit fpga_policy(sycl::device d) : base(d) {}
 };
 
 #endif // _ONEDPL_FPGA_DEVICE

From 6bfeb8702ea01b6d85ce46f336ed46608d099764 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 6 Mar 2024 09:53:26 +0100
Subject: [PATCH 517/566] Fix review comment: Please leave the old semantics
 with a partial specialization. That means keeping enable_if but change the
 condition to check a tag type

---
 include/oneapi/dpl/pstl/algorithm_fwd.h       | 10 +++----
 include/oneapi/dpl/pstl/algorithm_impl.h      | 27 +++++++------------
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   |  8 +++---
 3 files changed, 18 insertions(+), 27 deletions(-)

diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index ce504e6fa33..cada84d0f22 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -406,13 +406,13 @@ __pattern_search_n(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessI
 // copy_n
 //------------------------------------------------------------------------
 
-template <class _Tag, typename _ExecutionPolicy>
+template <class _Tag, typename _ExecutionPolicy, typename = void>
 struct __brick_copy_n;
 
-template <class _Tag, typename _ExecutionPolicy>
+template <class _Tag, typename _ExecutionPolicy, typename = void>
 struct __brick_copy;
 
-template <class _Tag, typename _ExecutionPolicy>
+template <class _Tag, typename _ExecutionPolicy, typename = void>
 struct __brick_move;
 
 //------------------------------------------------------------------------
@@ -853,7 +853,7 @@ __pattern_nth_element(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAcce
 //------------------------------------------------------------------------
 // fill, fill_n
 //------------------------------------------------------------------------
-template <class _Tag, typename _ExecutionPolicy, typename _Tp>
+template <class _Tag, typename _ExecutionPolicy, typename _Tp, typename = void>
 struct __brick_fill;
 
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Tp>
@@ -864,7 +864,7 @@ template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator,
 _RandomAccessIterator
 __pattern_fill(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, const _Tp&);
 
-template <class _Tag, typename _ExecutionPolicy, typename _Tp>
+template <class _Tag, typename _ExecutionPolicy, typename _Tp, typename = void>
 struct __brick_fill_n;
 
 template <class _Tag, class _ExecutionPolicy, class _OutputIterator, class _Size, class _Tp>
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 937a74a9b94..7287c2b544d 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -1046,11 +1046,12 @@ __pattern_search_n(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _
 // clear that doing so is worth the trouble and extra layers of call chain.
 // Sometimes a little duplication for sake of regularity is better than the alternative.
 
+//template <class _Tag, typename _ExecutionPolicy, ::std::enable_if_t<oneapi::dpl::__internal::__is_host_backend_tag_v<_Tag>>>
+//template <class _Tag, typename _ExecutionPolicy, typename _D> // +
+//template <class _Tag, typename _ExecutionPolicy, typename = void>
 template <class _Tag, typename _ExecutionPolicy>
-struct __brick_copy_n
+struct __brick_copy_n<_Tag, _ExecutionPolicy, ::std::enable_if_t<oneapi::dpl::__internal::__is_host_backend_tag_v<_Tag>>>
 {
-    static_assert(__is_host_backend_tag_v<_Tag>);
-
     template <typename _RandomAccessIterator1, typename _Size, typename _RandomAccessIterator2>
     _RandomAccessIterator2
     operator()(_RandomAccessIterator1 __first, _Size __n, _RandomAccessIterator2 __result,
@@ -1074,10 +1075,8 @@ struct __brick_copy_n
 //------------------------------------------------------------------------
 
 template <class _Tag, typename _ExecutionPolicy>
-struct __brick_copy
+struct __brick_copy<_Tag, _ExecutionPolicy, ::std::enable_if_t<__is_host_backend_tag_v<_Tag>>>
 {
-    static_assert(__is_host_backend_tag_v<_Tag>);
-
     template <typename _RandomAccessIterator1, typename _RandomAccessIterator2>
     _RandomAccessIterator2
     operator()(_RandomAccessIterator1 __first, _RandomAccessIterator1 __last, _RandomAccessIterator2 __result,
@@ -1108,10 +1107,8 @@ struct __brick_copy
 //------------------------------------------------------------------------
 
 template <class _Tag, typename _ExecutionPolicy>
-struct __brick_move
+struct __brick_move<_Tag, _ExecutionPolicy, ::std::enable_if_t<__is_host_backend_tag_v<_Tag>>>
 {
-    static_assert(__is_host_backend_tag_v<_Tag>);
-
     template <typename _RandomAccessIterator1, typename _RandomAccessIterator2>
     _RandomAccessIterator2
     operator()(_RandomAccessIterator1 __first, _RandomAccessIterator1 __last, _RandomAccessIterator2 __result,
@@ -1137,11 +1134,9 @@ struct __brick_move
     }
 };
 
-template <class _Tag, typename _ExecutionPolicy>
+template <class _Tag, typename _ExecutionPolicy, typename = ::std::enable_if_t<__is_host_backend_tag_v<_Tag>>>
 struct __brick_move_destroy
 {
-    static_assert(__is_host_backend_tag_v<_Tag>);
-
     template <typename _RandomAccessIterator1, typename _RandomAccessIterator2>
     _RandomAccessIterator2
     operator()(_RandomAccessIterator1 __first, _RandomAccessIterator1 __last, _RandomAccessIterator2 __result,
@@ -2758,10 +2753,8 @@ __pattern_nth_element(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec
 // fill, fill_n
 //------------------------------------------------------------------------
 template <class _Tag, typename _ExecutionPolicy, typename _Tp>
-struct __brick_fill
+struct __brick_fill<_Tag, _ExecutionPolicy, _Tp, ::std::enable_if_t<__is_host_backend_tag_v<_Tag>>>
 {
-    static_assert(__is_host_backend_tag_v<_Tag>);
-
     const _Tp& __value;
 
     template <typename _RandomAccessIterator>
@@ -2808,10 +2801,8 @@ __pattern_fill(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAcce
 }
 
 template <class _Tag, typename _ExecutionPolicy, typename _Tp>
-struct __brick_fill_n
+struct __brick_fill_n<_Tag, _ExecutionPolicy, _Tp, ::std::enable_if_t<__is_host_backend_tag_v<_Tag>>>
 {
-    static_assert(__is_host_backend_tag_v<_Tag>);
-
     const _Tp& __value;
 
     template <typename _RandomAccessIterator, typename _Size>
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 1ef0900744f..57524d0ea05 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -387,7 +387,7 @@ struct __brick_copy_n<__hetero_tag<_BackendTag>, _ExecutionPolicy>
 };
 
 template <typename _BackendTag, typename _ExecutionPolicy>
-struct __brick_copy<__hetero_tag<_BackendTag>, _ExecutionPolicy>
+struct __brick_copy<__hetero_tag<_BackendTag>, _ExecutionPolicy, void>
 {
     template <typename _SourceT, typename _TargetT>
     void
@@ -398,7 +398,7 @@ struct __brick_copy<__hetero_tag<_BackendTag>, _ExecutionPolicy>
 };
 
 template <typename _BackendTag, typename _ExecutionPolicy>
-struct __brick_move<__hetero_tag<_BackendTag>, _ExecutionPolicy>
+struct __brick_move<__hetero_tag<_BackendTag>, _ExecutionPolicy, void>
 {
     template <typename _SourceT, typename _TargetT>
     void
@@ -409,7 +409,7 @@ struct __brick_move<__hetero_tag<_BackendTag>, _ExecutionPolicy>
 };
 
 template <typename _BackendTag, typename _ExecutionPolicy, typename _SourceT>
-struct __brick_fill<__hetero_tag<_BackendTag>, _ExecutionPolicy, _SourceT>
+struct __brick_fill<__hetero_tag<_BackendTag>, _ExecutionPolicy, _SourceT, void>
 {
     _SourceT __value;
     template <typename _TargetT>
@@ -421,7 +421,7 @@ struct __brick_fill<__hetero_tag<_BackendTag>, _ExecutionPolicy, _SourceT>
 };
 
 template <typename _BackendTag, typename _ExecutionPolicy, typename _SourceT>
-struct __brick_fill_n<__hetero_tag<_BackendTag>, _ExecutionPolicy, _SourceT>
+struct __brick_fill_n<__hetero_tag<_BackendTag>, _ExecutionPolicy, _SourceT, void>
 {
     _SourceT __value;
     template <typename _TargetT>

From b2216a004f0172bb44649d7cb3af650459fe03bd Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 6 Mar 2024 10:00:02 +0100
Subject: [PATCH 518/566] Fix review comment: Please leave the old semantics
 with a partial specialization. That means keeping enable_if but change the
 condition to check a tag type - replace void to enable_if_t

---
 .../dpl/pstl/hetero/algorithm_impl_hetero.h    | 10 +++++-----
 .../pstl/hetero/dpcpp/execution_sycl_defs.h    | 18 ++++++++++++++++++
 2 files changed, 23 insertions(+), 5 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 57524d0ea05..9b917ff658e 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -376,7 +376,7 @@ __pattern_generate_n(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec,
 //------------------------------------------------------------------------
 
 template <typename _BackendTag, typename _ExecutionPolicy>
-struct __brick_copy_n<__hetero_tag<_BackendTag>, _ExecutionPolicy>
+struct __brick_copy_n<__hetero_tag<_BackendTag>, _ExecutionPolicy, ::std::enable_if_t<__is_backend_tag_hetero_v<__hetero_tag<_BackendTag>>>>
 {
     template <typename _SourceT, typename _TargetT>
     void
@@ -387,7 +387,7 @@ struct __brick_copy_n<__hetero_tag<_BackendTag>, _ExecutionPolicy>
 };
 
 template <typename _BackendTag, typename _ExecutionPolicy>
-struct __brick_copy<__hetero_tag<_BackendTag>, _ExecutionPolicy, void>
+struct __brick_copy<__hetero_tag<_BackendTag>, _ExecutionPolicy, ::std::enable_if_t<__is_backend_tag_hetero_v<__hetero_tag<_BackendTag>>>>
 {
     template <typename _SourceT, typename _TargetT>
     void
@@ -398,7 +398,7 @@ struct __brick_copy<__hetero_tag<_BackendTag>, _ExecutionPolicy, void>
 };
 
 template <typename _BackendTag, typename _ExecutionPolicy>
-struct __brick_move<__hetero_tag<_BackendTag>, _ExecutionPolicy, void>
+struct __brick_move<__hetero_tag<_BackendTag>, _ExecutionPolicy, ::std::enable_if_t<__is_backend_tag_hetero_v<__hetero_tag<_BackendTag>>>>
 {
     template <typename _SourceT, typename _TargetT>
     void
@@ -409,7 +409,7 @@ struct __brick_move<__hetero_tag<_BackendTag>, _ExecutionPolicy, void>
 };
 
 template <typename _BackendTag, typename _ExecutionPolicy, typename _SourceT>
-struct __brick_fill<__hetero_tag<_BackendTag>, _ExecutionPolicy, _SourceT, void>
+struct __brick_fill<__hetero_tag<_BackendTag>, _ExecutionPolicy, _SourceT, ::std::enable_if_t<__is_backend_tag_hetero_v<__hetero_tag<_BackendTag>>>>
 {
     _SourceT __value;
     template <typename _TargetT>
@@ -421,7 +421,7 @@ struct __brick_fill<__hetero_tag<_BackendTag>, _ExecutionPolicy, _SourceT, void>
 };
 
 template <typename _BackendTag, typename _ExecutionPolicy, typename _SourceT>
-struct __brick_fill_n<__hetero_tag<_BackendTag>, _ExecutionPolicy, _SourceT, void>
+struct __brick_fill_n<__hetero_tag<_BackendTag>, _ExecutionPolicy, _SourceT, ::std::enable_if_t<__is_backend_tag_hetero_v<__hetero_tag<_BackendTag>>>>
 {
     _SourceT __value;
     template <typename _TargetT>
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h b/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h
index 93d291b661b..fbc4350eca2 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h
@@ -316,6 +316,24 @@ __select_backend(const execution::fpga_policy<_Factor, _KernelName>&, _IteratorT
     static_assert(__is_random_access_iterator_v<_IteratorTypes...>);
     return {};
 }
+
+//----------------------------------------------------------
+// __is_backend_tag_hetero, __is_backend_tag_hetero_v
+//----------------------------------------------------------
+
+template <class _Tag>
+struct __is_backend_tag_hetero : ::std::false_type
+{
+};
+
+template <typename _BackendTag>
+struct __is_backend_tag_hetero<__hetero_tag<_BackendTag>> : ::std::true_type
+{
+};
+
+template <class _Tag>
+inline constexpr bool __is_backend_tag_hetero_v = __is_backend_tag_hetero<_Tag>::value;
+
 #endif
 
 } // namespace __internal

From 69ee18ba8524abe23e56d23773ba453b1e22c36f Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 6 Mar 2024 10:29:15 +0100
Subject: [PATCH 519/566] Revert "Fix review comment: Please leave the old
 semantics with a partial specialization. That means keeping enable_if but
 change the condition to check a tag type - replace void to enable_if_t"

This reverts commit b2216a004f0172bb44649d7cb3af650459fe03bd.
---
 .../dpl/pstl/hetero/algorithm_impl_hetero.h    | 10 +++++-----
 .../pstl/hetero/dpcpp/execution_sycl_defs.h    | 18 ------------------
 2 files changed, 5 insertions(+), 23 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 9b917ff658e..57524d0ea05 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -376,7 +376,7 @@ __pattern_generate_n(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec,
 //------------------------------------------------------------------------
 
 template <typename _BackendTag, typename _ExecutionPolicy>
-struct __brick_copy_n<__hetero_tag<_BackendTag>, _ExecutionPolicy, ::std::enable_if_t<__is_backend_tag_hetero_v<__hetero_tag<_BackendTag>>>>
+struct __brick_copy_n<__hetero_tag<_BackendTag>, _ExecutionPolicy>
 {
     template <typename _SourceT, typename _TargetT>
     void
@@ -387,7 +387,7 @@ struct __brick_copy_n<__hetero_tag<_BackendTag>, _ExecutionPolicy, ::std::enable
 };
 
 template <typename _BackendTag, typename _ExecutionPolicy>
-struct __brick_copy<__hetero_tag<_BackendTag>, _ExecutionPolicy, ::std::enable_if_t<__is_backend_tag_hetero_v<__hetero_tag<_BackendTag>>>>
+struct __brick_copy<__hetero_tag<_BackendTag>, _ExecutionPolicy, void>
 {
     template <typename _SourceT, typename _TargetT>
     void
@@ -398,7 +398,7 @@ struct __brick_copy<__hetero_tag<_BackendTag>, _ExecutionPolicy, ::std::enable_i
 };
 
 template <typename _BackendTag, typename _ExecutionPolicy>
-struct __brick_move<__hetero_tag<_BackendTag>, _ExecutionPolicy, ::std::enable_if_t<__is_backend_tag_hetero_v<__hetero_tag<_BackendTag>>>>
+struct __brick_move<__hetero_tag<_BackendTag>, _ExecutionPolicy, void>
 {
     template <typename _SourceT, typename _TargetT>
     void
@@ -409,7 +409,7 @@ struct __brick_move<__hetero_tag<_BackendTag>, _ExecutionPolicy, ::std::enable_i
 };
 
 template <typename _BackendTag, typename _ExecutionPolicy, typename _SourceT>
-struct __brick_fill<__hetero_tag<_BackendTag>, _ExecutionPolicy, _SourceT, ::std::enable_if_t<__is_backend_tag_hetero_v<__hetero_tag<_BackendTag>>>>
+struct __brick_fill<__hetero_tag<_BackendTag>, _ExecutionPolicy, _SourceT, void>
 {
     _SourceT __value;
     template <typename _TargetT>
@@ -421,7 +421,7 @@ struct __brick_fill<__hetero_tag<_BackendTag>, _ExecutionPolicy, _SourceT, ::std
 };
 
 template <typename _BackendTag, typename _ExecutionPolicy, typename _SourceT>
-struct __brick_fill_n<__hetero_tag<_BackendTag>, _ExecutionPolicy, _SourceT, ::std::enable_if_t<__is_backend_tag_hetero_v<__hetero_tag<_BackendTag>>>>
+struct __brick_fill_n<__hetero_tag<_BackendTag>, _ExecutionPolicy, _SourceT, void>
 {
     _SourceT __value;
     template <typename _TargetT>
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h b/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h
index fbc4350eca2..93d291b661b 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h
@@ -316,24 +316,6 @@ __select_backend(const execution::fpga_policy<_Factor, _KernelName>&, _IteratorT
     static_assert(__is_random_access_iterator_v<_IteratorTypes...>);
     return {};
 }
-
-//----------------------------------------------------------
-// __is_backend_tag_hetero, __is_backend_tag_hetero_v
-//----------------------------------------------------------
-
-template <class _Tag>
-struct __is_backend_tag_hetero : ::std::false_type
-{
-};
-
-template <typename _BackendTag>
-struct __is_backend_tag_hetero<__hetero_tag<_BackendTag>> : ::std::true_type
-{
-};
-
-template <class _Tag>
-inline constexpr bool __is_backend_tag_hetero_v = __is_backend_tag_hetero<_Tag>::value;
-
 #endif
 
 } // namespace __internal

From 8e4f5ea2e5a0b843985b21ae1970db4dd4fbe8d0 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 6 Mar 2024 10:54:14 +0100
Subject: [PATCH 520/566] Add _BackendTag as template param into __buffer
 implementations

---
 .../internal/exclusive_scan_by_segment_impl.h |  8 ++---
 .../internal/inclusive_scan_by_segment_impl.h |  4 +--
 .../dpl/internal/reduce_by_segment_impl.h     | 30 ++++++++--------
 .../dpl/internal/scan_by_segment_impl.h       | 12 +++----
 include/oneapi/dpl/pstl/algorithm_impl.h      | 21 ++++++------
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 34 +++++++++++--------
 .../hetero/algorithm_ranges_impl_hetero.h     | 27 ++++++++-------
 .../pstl/hetero/dpcpp/parallel_backend_sycl.h | 11 +++---
 .../dpcpp/parallel_backend_sycl_histogram.h   |  7 ++--
 .../dpcpp/parallel_backend_sycl_radix_sort.h  |  5 +--
 .../dpcpp/parallel_backend_sycl_utils.h       | 14 ++++----
 .../dpl/pstl/hetero/numeric_impl_hetero.h     |  3 +-
 include/oneapi/dpl/pstl/omp/parallel_scan.h   |  3 +-
 include/oneapi/dpl/pstl/omp/util.h            | 11 +++---
 .../oneapi/dpl/pstl/parallel_backend_serial.h | 11 +++---
 .../oneapi/dpl/pstl/parallel_backend_tbb.h    | 19 ++++++-----
 16 files changed, 119 insertions(+), 101 deletions(-)

diff --git a/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h b/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h
index f9beef59f2b..50211ab30ec 100644
--- a/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h
@@ -66,7 +66,7 @@ pattern_exclusive_scan_by_segment(_Tag, Policy&& policy, InputIterator1 first1,
     InputIterator2 last2 = first2 + n;
 
     // compute head flags
-    oneapi::dpl::__par_backend::__buffer<Policy, FlagType> _flags(__backend_tag{}, policy, n);
+    oneapi::dpl::__par_backend::__buffer<__backend_tag, Policy, FlagType> _flags(__backend_tag{}, policy, n);
     auto flags = _flags.get();
     flags[0] = 1;
 
@@ -74,7 +74,7 @@ pattern_exclusive_scan_by_segment(_Tag, Policy&& policy, InputIterator1 first1,
               oneapi::dpl::__internal::__not_pred<BinaryPredicate>(binary_pred));
 
     // shift input one to the right and initialize segments with init
-    oneapi::dpl::__par_backend::__buffer<Policy, OutputType> _temp(__backend_tag{}, policy, n);
+    oneapi::dpl::__par_backend::__buffer<__backend_tag, Policy, OutputType> _temp(__backend_tag{}, policy, n);
     auto temp = _temp.get();
 
     temp[0] = init;
@@ -125,7 +125,7 @@ exclusive_scan_by_segment_impl(__internal::__hetero_tag<_BackendTag>, Policy&& p
     InputIterator2 last2 = first2 + n;
 
     // compute head flags
-    oneapi::dpl::__par_backend_hetero::__buffer<Policy, FlagType> _flags(_BackendTag{}, policy, n);
+    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, Policy, FlagType> _flags(_BackendTag{}, policy, n);
     {
         auto flag_buf = _flags.get_buffer();
         auto flags = flag_buf.get_host_access(sycl::read_write);
@@ -136,7 +136,7 @@ exclusive_scan_by_segment_impl(__internal::__hetero_tag<_BackendTag>, Policy&& p
               oneapi::dpl::__internal::__not_pred<BinaryPredicate>(binary_pred));
 
     // shift input one to the right and initialize segments with init
-    oneapi::dpl::__par_backend_hetero::__buffer<Policy, OutputType> _temp(_BackendTag{}, policy, n);
+    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, Policy, OutputType> _temp(_BackendTag{}, policy, n);
     {
         auto temp_buf = _temp.get_buffer();
         auto temp = temp_buf.get_host_access(sycl::read_write);
diff --git a/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h b/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h
index fbe094104fe..23d197adc9a 100644
--- a/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h
@@ -61,7 +61,7 @@ pattern_inclusive_scan_by_segment(_Tag, Policy&& policy, InputIterator1 first1,
     typedef unsigned int FlagType;
     typedef typename ::std::iterator_traits<InputIterator2>::value_type ValueType;
 
-    oneapi::dpl::__par_backend::__buffer<Policy, FlagType> _mask(__backend_tag{}, policy, n);
+    oneapi::dpl::__par_backend::__buffer<__backend_tag, Policy, FlagType> _mask(__backend_tag{}, policy, n);
     auto mask = _mask.get();
 
     mask[0] = 1;
@@ -111,7 +111,7 @@ inclusive_scan_by_segment_impl(__internal::__hetero_tag<_BackendTag>, Policy&& p
 
     FlagType initial_mask = 1;
 
-    oneapi::dpl::__par_backend_hetero::__buffer<Policy, FlagType> _mask(_BackendTag{}, policy, n);
+    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, Policy, FlagType> _mask(_BackendTag{}, policy, n);
     {
         auto mask_buf = _mask.get_buffer();
         auto mask = mask_buf.get_host_access(sycl::read_write);
diff --git a/include/oneapi/dpl/internal/reduce_by_segment_impl.h b/include/oneapi/dpl/internal/reduce_by_segment_impl.h
index 9ee25b41913..834cd9ed767 100644
--- a/include/oneapi/dpl/internal/reduce_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/reduce_by_segment_impl.h
@@ -116,7 +116,7 @@ reduce_by_segment_impl(_Tag, Policy&& policy, InputIterator1 first1, InputIterat
 
     // buffer that is used to store a flag indicating if the associated key is not equal to
     // the next key, and thus its associated sum should be part of the final result
-    oneapi::dpl::__par_backend::__buffer<Policy, FlagType> _mask(__backend_tag{}, policy, n + 1);
+    oneapi::dpl::__par_backend::__buffer<__backend_tag, Policy, FlagType> _mask(__backend_tag{}, policy, n + 1);
     auto mask = _mask.get();
     mask[0] = 1;
 
@@ -132,11 +132,11 @@ reduce_by_segment_impl(_Tag, Policy&& policy, InputIterator1 first1, InputIterat
     // buffer stores the sums of values associated with a given key. Sums are copied with
     // a shift into result2, and the shift is computed at the same time as the sums, so the
     // sums can't be written to result2 directly.
-    oneapi::dpl::__par_backend::__buffer<Policy, ValueType> _scanned_values(__backend_tag{}, policy, n);
+    oneapi::dpl::__par_backend::__buffer<__backend_tag, Policy, ValueType> _scanned_values(__backend_tag{}, policy, n);
 
     // Buffer is used to store results of the scan of the mask. Values indicate which position
     // in result2 needs to be written with the scanned_values element.
-    oneapi::dpl::__par_backend::__buffer<Policy, FlagType> _scanned_tail_flags(__backend_tag{}, policy, n);
+    oneapi::dpl::__par_backend::__buffer<__backend_tag, Policy, FlagType> _scanned_tail_flags(__backend_tag{}, policy, n);
 
     // Compute the sum of the segments. scanned_tail_flags values are not used.
     inclusive_scan(policy, make_zip_iterator(first2, _mask.get()), make_zip_iterator(first2, _mask.get()) + n,
@@ -264,23 +264,23 @@ __sycl_reduce_by_segment(__internal::__hetero_tag<_BackendTag>, _ExecutionPolicy
     ::std::size_t __n_groups = oneapi::dpl::__internal::__dpl_ceiling_div(__n, __wgroup_size * __vals_per_item);
 
     // intermediate reductions within a workgroup
-    auto __partials =
-        oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, __val_type>(_BackendTag{}, __exec, __n_groups)
-            .get_buffer();
+    auto __partials = oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, __val_type>(
+                          _BackendTag{}, __exec, __n_groups)
+                          .get_buffer();
 
-    auto __end_idx =
-        oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, __diff_type>(_BackendTag{}, __exec, 1)
-            .get_buffer();
+    auto __end_idx = oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, __diff_type>(
+                         _BackendTag{}, __exec, 1)
+                         .get_buffer();
 
     // the number of segment ends found in each work group
-    auto __seg_ends =
-        oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, __diff_type>(_BackendTag{}, __exec, __n_groups)
-            .get_buffer();
+    auto __seg_ends = oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, __diff_type>(
+                          _BackendTag{}, __exec, __n_groups)
+                          .get_buffer();
 
     // buffer that stores an exclusive scan of the results
-    auto __seg_ends_scanned =
-        oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, __diff_type>(_BackendTag{}, __exec, __n_groups)
-            .get_buffer();
+    auto __seg_ends_scanned = oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, __diff_type>(
+                                  _BackendTag{}, __exec, __n_groups)
+                                  .get_buffer();
 
     // 1. Count the segment ends in each workgroup
     auto __seg_end_identification = __exec.queue().submit([&](sycl::handler& __cgh) {
diff --git a/include/oneapi/dpl/internal/scan_by_segment_impl.h b/include/oneapi/dpl/internal/scan_by_segment_impl.h
index 65b236ba852..9104ac18e42 100644
--- a/include/oneapi/dpl/internal/scan_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/scan_by_segment_impl.h
@@ -144,14 +144,14 @@ struct __sycl_scan_by_segment_impl
 
         ::std::size_t __n_groups = __internal::__dpl_ceiling_div(__n, __wgroup_size * __vals_per_item);
 
-        auto __partials =
-            oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, __val_type>(_BackendTag{}, __exec, __n_groups)
-                .get_buffer();
+        auto __partials = oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, __val_type>(
+                              _BackendTag{}, __exec, __n_groups)
+                              .get_buffer();
 
         // the number of segment ends found in each work group
-        auto __seg_ends =
-            oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, bool>(_BackendTag{}, __exec, __n_groups)
-                .get_buffer();
+        auto __seg_ends = oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, bool>(
+                              _BackendTag{}, __exec, __n_groups)
+                              .get_buffer();
 
         // 1. Work group reduction
         auto __wg_scan = __exec.queue().submit([&](sycl::handler& __cgh) {
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 7287c2b544d..8620b81b353 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -1325,7 +1325,7 @@ __pattern_copy_if(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomA
     const _DifferenceType __n = __last - __first;
     if (_DifferenceType(1) < __n)
     {
-        __par_backend::__buffer<_ExecutionPolicy, bool> __mask_buf(__backend_tag{}, __exec, __n);
+        __par_backend::__buffer<__backend_tag, _ExecutionPolicy, bool> __mask_buf(__backend_tag{}, __exec, __n);
         return __internal::__except_handler([&__exec, __n, __first, __result, __pred, &__mask_buf]() {
             bool* __mask = __mask_buf.get();
             _DifferenceType __m{};
@@ -1444,7 +1444,7 @@ __remove_elements(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Forward
     typedef typename ::std::iterator_traits<_ForwardIterator>::difference_type _DifferenceType;
     typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _Tp;
     _DifferenceType __n = __last - __first;
-    __par_backend::__buffer<_ExecutionPolicy, bool> __mask_buf(__backend_tag{}, __exec, __n);
+    __par_backend::__buffer<__backend_tag, _ExecutionPolicy, bool> __mask_buf(__backend_tag{}, __exec, __n);
     // 1. find a first iterator that should be removed
     return __internal::__except_handler([&]() {
         bool* __mask = __mask_buf.get();
@@ -1481,7 +1481,7 @@ __remove_elements(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Forward
         __n -= __min;
         __first += __min;
 
-        __par_backend::__buffer<_ExecutionPolicy, _Tp> __buf(__backend_tag{}, __exec, __n);
+        __par_backend::__buffer<__backend_tag, _ExecutionPolicy, _Tp> __buf(__backend_tag{}, __exec, __n);
         _Tp* __result = __buf.get();
         __mask += __min;
         _DifferenceType __m{};
@@ -1609,7 +1609,7 @@ __pattern_unique_copy(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Ran
     const _DifferenceType __n = __last - __first;
     if (_DifferenceType(2) < __n)
     {
-        __par_backend::__buffer<_ExecutionPolicy, bool> __mask_buf(__backend_tag{}, __exec, __n);
+        __par_backend::__buffer<__backend_tag, _ExecutionPolicy, bool> __mask_buf(__backend_tag{}, __exec, __n);
         if (_DifferenceType(2) < __n)
         {
             return __internal::__except_handler([&__exec, __n, __first, __result, __pred, &__mask_buf]() {
@@ -1850,7 +1850,7 @@ __pattern_rotate(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAc
     auto __m = __middle - __first;
     if (__m <= __n / 2)
     {
-        __par_backend::__buffer<_ExecutionPolicy, _Tp> __buf(__backend_tag{}, __exec, __n - __m);
+        __par_backend::__buffer<__backend_tag, _ExecutionPolicy, _Tp> __buf(__backend_tag{}, __exec, __n - __m);
         return __internal::__except_handler([&__exec, __n, __m, __first, __middle, __last, &__buf]() {
             _Tp* __result = __buf.get();
             __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __middle, __last,
@@ -1876,7 +1876,7 @@ __pattern_rotate(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAc
     }
     else
     {
-        __par_backend::__buffer<_ExecutionPolicy, _Tp> __buf(__backend_tag{}, __exec, __m);
+        __par_backend::__buffer<__backend_tag, _ExecutionPolicy, _Tp> __buf(__backend_tag{}, __exec, __m);
         return __internal::__except_handler([&__exec, __n, __m, __first, __middle, __last, &__buf]() {
             _Tp* __result = __buf.get();
             __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __middle,
@@ -2354,7 +2354,7 @@ __pattern_partition_copy(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _
     const _DifferenceType __n = __last - __first;
     if (_DifferenceType(1) < __n)
     {
-        __par_backend::__buffer<_ExecutionPolicy, bool> __mask_buf(__backend_tag{}, __exec, __n);
+        __par_backend::__buffer<__backend_tag, _ExecutionPolicy, bool> __mask_buf(__backend_tag{}, __exec, __n);
         return __internal::__except_handler([&__exec, __n, __first, __out_true, __out_false, __pred, &__mask_buf]() {
             bool* __mask = __mask_buf.get();
             _ReturnType __m{};
@@ -2580,7 +2580,7 @@ __pattern_partial_sort_copy(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec
         {
             typedef typename ::std::iterator_traits<_RandomAccessIterator1>::value_type _T1;
             typedef typename ::std::iterator_traits<_RandomAccessIterator2>::value_type _T2;
-            __par_backend::__buffer<_ExecutionPolicy, _T1> __buf(__backend_tag{}, __exec, __n1);
+            __par_backend::__buffer<__backend_tag, _ExecutionPolicy, _T1> __buf(__backend_tag{}, __exec, __n1);
             _T1* __r = __buf.get();
 
             __par_backend::__parallel_stable_sort(
@@ -3073,7 +3073,7 @@ __pattern_inplace_merge(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _R
 
     typedef typename ::std::iterator_traits<_RandomAccessIterator>::value_type _Tp;
     auto __n = __last - __first;
-    __par_backend::__buffer<_ExecutionPolicy, _Tp> __buf(__backend_tag{}, __exec, __n);
+    __par_backend::__buffer<__backend_tag, _ExecutionPolicy, _Tp> __buf(__backend_tag{}, __exec, __n);
     _Tp* __r = __buf.get();
     __internal::__except_handler([&]() {
         auto __move_values = [](_RandomAccessIterator __x, _Tp* __z) {
@@ -3207,7 +3207,8 @@ __parallel_set_op(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Forward
     const _DifferenceType __n1 = __last1 - __first1;
     const _DifferenceType __n2 = __last2 - __first2;
 
-    __par_backend::__buffer<_ExecutionPolicy, _T> __buf(__backend_tag{}, __exec, __size_func(__n1, __n2));
+    __par_backend::__buffer<__backend_tag, _ExecutionPolicy, _T> __buf(__backend_tag{}, __exec,
+                                                                       __size_func(__n1, __n2));
 
     return __internal::__except_handler([&__exec, __n1, __first1, __last1, __first2, __last2, __result, __comp,
                                          __size_func, __set_op, &__buf]() {
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 57524d0ea05..64ab0123d88 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -1032,8 +1032,8 @@ __pattern_remove_if(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec,
 
     using _ValueType = typename ::std::iterator_traits<_Iterator>::value_type;
 
-    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __buf(_BackendTag{}, __exec,
-                                                                                    __last - __first);
+    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _ValueType> __buf(_BackendTag{}, __exec,
+                                                                                                 __last - __first);
     auto __copy_first = __buf.get();
 
     auto __copy_last = __pattern_copy_if(__tag, __exec, __first, __last, __copy_first, __not_pred<_Predicate>{__pred});
@@ -1054,8 +1054,8 @@ __pattern_unique(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _It
 
     using _ValueType = typename ::std::iterator_traits<_Iterator>::value_type;
 
-    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __buf(_BackendTag{}, __exec,
-                                                                                    __last - __first);
+    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _ValueType> __buf(_BackendTag{}, __exec,
+                                                                                                 __last - __first);
     auto __copy_first = __buf.get();
     auto __copy_last = __pattern_unique_copy(__tag, __exec, __first, __last, __copy_first, __pred);
 
@@ -1229,7 +1229,8 @@ __pattern_inplace_merge(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __ex
     assert(__first < __middle && __middle < __last);
 
     auto __n = __last - __first;
-    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __buf(_BackendTag{}, __exec, __n);
+    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _ValueType> __buf(_BackendTag{}, __exec,
+                                                                                                 __n);
     auto __copy_first = __buf.get();
     auto __copy_last = __copy_first + __n;
 
@@ -1317,8 +1318,8 @@ __pattern_stable_partition(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& _
 
     auto __n = __last - __first;
 
-    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __true_buf(_BackendTag{}, __exec, __n);
-    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __false_buf(_BackendTag{}, __exec, __n);
+    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _ValueType> __true_buf(_BackendTag{}, __exec, __n);
+    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _ValueType> __false_buf(_BackendTag{}, __exec, __n);
     auto __true_result = __true_buf.get();
     auto __false_result = __false_buf.get();
 
@@ -1515,8 +1516,8 @@ __pattern_partial_sort_copy(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&&
         // - create a temporary buffer and copy all the elements from the input buffer there
         // - run partial sort on the temporary buffer
         // - copy k elements from the temporary buffer to the output buffer.
-        oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __buf(_BackendTag{}, __exec,
-                                                                                        __in_size);
+        oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _ValueType> __buf(_BackendTag{},
+                                                                                                     __exec, __in_size);
 
         auto __buf_first = __buf.get();
 
@@ -1631,7 +1632,8 @@ __pattern_rotate(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Iterator
 
     auto __keep = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read_write, _Iterator>();
     auto __buf = __keep(__first, __last);
-    auto __temp_buf = oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _Tp>(_BackendTag{}, __exec, __n);
+    auto __temp_buf =
+        oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _Tp>(_BackendTag{}, __exec, __n);
 
     auto __temp_rng =
         oneapi::dpl::__ranges::all_view<_Tp, __par_backend_hetero::access_mode::write>(__temp_buf.get_buffer());
@@ -1714,7 +1716,8 @@ __pattern_hetero_set_op(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _F
         __comp, __n1, __n2};
 
     // temporary buffer to store boolean mask
-    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, int32_t> __mask_buf(_BackendTag{}, __exec, __n1);
+    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, int32_t> __mask_buf(_BackendTag{}, __exec,
+                                                                                                 __n1);
 
     auto __keep1 =
         oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _ForwardIterator1>();
@@ -1843,7 +1846,8 @@ __pattern_set_union(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec,
 
     // temporary buffer to store intermediate result
     const auto __n2 = __last2 - __first2;
-    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __diff(_BackendTag{}, __exec, __n2);
+    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _ValueType> __diff(_BackendTag{}, __exec,
+                                                                                                  __n2);
     auto __buf = __diff.get();
 
     //1. Calc difference {2} \ {1}
@@ -1924,10 +1928,12 @@ __pattern_set_symmetric_difference(__hetero_tag<_BackendTag> __tag, _ExecutionPo
 
     // temporary buffers to store intermediate result
     const auto __n1 = __last1 - __first1;
-    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __diff_1(_BackendTag{}, __exec, __n1);
+    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _ValueType> __diff_1(_BackendTag{},
+                                                                                                    __exec, __n1);
     auto __buf_1 = __diff_1.get();
     const auto __n2 = __last2 - __first2;
-    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __diff_2(_BackendTag{}, __exec, __n2);
+    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _ValueType> __diff_2(_BackendTag{},
+                                                                                                    __exec, __n2);
     auto __buf_2 = __diff_2.get();
 
     //1. Calc difference {1} \ {2}
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
index 013133eebbf..715c0d720ef 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
@@ -360,8 +360,8 @@ __pattern_scan_copy(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Range
     _DataAcc __get_data_op;
     _MaskAssigner __add_mask_op;
 
-    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, int32_t> __mask_buf(_BackendTag{}, __exec,
-                                                                                      __rng1.size());
+    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, int32_t> __mask_buf(
+        _BackendTag{}, __exec, __rng1.size());
 
     auto __res =
         __par_backend_hetero::__parallel_transform_scan_base(
@@ -414,8 +414,8 @@ __pattern_remove_if(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec,
 
     using _ValueType = oneapi::dpl::__internal::__value_t<_Range>;
 
-    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __buf(_BackendTag{}, __exec,
-                                                                                    __rng.size());
+    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _ValueType> __buf(_BackendTag{}, __exec,
+                                                                                                 __rng.size());
     auto __copy_rng = oneapi::dpl::__ranges::views::all(__buf.get_buffer());
 
     auto __copy_last_id = __ranges::__pattern_copy_if(__tag, __exec, __rng, __copy_rng, __not_pred<_Predicate>{__pred},
@@ -463,8 +463,8 @@ __pattern_unique(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Ra
 
     using _ValueType = oneapi::dpl::__internal::__value_t<_Range>;
 
-    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __buf(_BackendTag{}, __exec,
-                                                                                    __rng.size());
+    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _ValueType> __buf(_BackendTag{}, __exec,
+                                                                                                 __rng.size());
     auto res_rng = oneapi::dpl::__ranges::views::all(__buf.get_buffer());
     auto res = __ranges::__pattern_unique_copy(__tag, __exec, __rng, res_rng, __pred,
                                                oneapi::dpl::__internal::__pstl_assign());
@@ -718,14 +718,15 @@ __pattern_reduce_by_segment(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&&
     // Round 1: reduce with extra indices added to avoid long segments
     // TODO: At threshold points check if the key is equal to the key at the previous threshold point, indicating a long sequence.
     // Skip a round of copy_if and reduces if there are none.
-    auto __idx = oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, __diff_type>(_BackendTag{}, __exec, __n)
+    auto __idx = oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, __diff_type>(_BackendTag{},
+                                                                                                         __exec, __n)
                      .get_buffer();
-    auto __tmp_out_keys =
-        oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, __key_type>(_BackendTag{}, __exec, __n)
-            .get_buffer();
-    auto __tmp_out_values =
-        oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, __val_type>(_BackendTag{}, __exec, __n)
-            .get_buffer();
+    auto __tmp_out_keys = oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, __key_type>(
+                              _BackendTag{}, __exec, __n)
+                              .get_buffer();
+    auto __tmp_out_values = oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, __val_type>(
+                                _BackendTag{}, __exec, __n)
+                                .get_buffer();
 
     // Replicating first element of keys view to be able to compare (i-1)-th and (i)-th key with aligned sequences,
     //  dropping the last key for the i-1 sequence.
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
index 8c40edefbd2..46c02feb3a7 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
@@ -871,8 +871,9 @@ __parallel_scan_copy(oneapi::dpl::__internal::__device_backend_tag __backend_tag
     _MaskAssigner __add_mask_op;
 
     // temporary buffer to store boolean mask
-    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, int32_t> __mask_buf(
-        oneapi::dpl::__internal::__device_backend_tag{}, __exec, __n);
+    oneapi::dpl::__par_backend_hetero::__buffer<oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy,
+                                                int32_t>
+        __mask_buf(oneapi::dpl::__internal::__device_backend_tag{}, __exec, __n);
 
     return __parallel_transform_scan_base(
         __backend_tag, ::std::forward<_ExecutionPolicy>(__exec),
@@ -1580,7 +1581,8 @@ struct __parallel_sort_submitter<_BackendTag, _IdType, __internal::__optional_ke
         });
 
         // 2. Merge sorting
-        oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _Tp> __temp_buf(_BackendTag{}, __exec, __n);
+        oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _Tp> __temp_buf(_BackendTag{},
+                                                                                                   __exec, __n);
         auto __temp = __temp_buf.get_buffer();
         bool __data_in_temp = false;
         _IdType __n_sorted = __leaf;
@@ -1701,7 +1703,8 @@ struct __parallel_partial_sort_submitter<_BackendTag, __internal::__optional_ker
         _Size __n = __rng.size();
         assert(__n > 1);
 
-        oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _Tp> __temp_buf(_BackendTag{}, __exec, __n);
+        oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _Tp> __temp_buf(_BackendTag{}, __exec,
+                                                                                                 __n);
         auto __temp = __temp_buf.get_buffer();
         _PRINT_INFO_IN_DEBUG_MODE(__exec);
 
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h
index ecaabd2ae48..d31b1e6d346 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h
@@ -426,9 +426,10 @@ struct __histogram_general_private_global_atomics_submitter<_BackendTag,
         ::std::size_t __segments =
             oneapi::dpl::__internal::__dpl_ceiling_div(__n, __work_group_size * __iters_per_work_item);
 
-        auto __private_histograms = oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _bin_type>(
-                                        _BackendTag{}, __exec, __segments * __num_bins)
-                                        .get_buffer();
+        auto __private_histograms =
+            oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _bin_type>(
+                _BackendTag{}, __exec, __segments * __num_bins)
+                .get_buffer();
 
         return __exec.queue().submit([&](auto& __h) {
             __h.depends_on(__init_event);
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_radix_sort.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_radix_sort.h
index 0c495f8d498..df1649f99cb 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_radix_sort.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_radix_sort.h
@@ -827,8 +827,9 @@ __parallel_radix_sort(oneapi::dpl::__internal::__device_backend_tag, _ExecutionP
         sycl::buffer<::std::uint32_t, 1> __tmp_buf{sycl::range<1>(__tmp_buf_size)};
 
         // memory for storing values sorted for an iteration
-        oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueT> __out_buffer_holder{
-            oneapi::dpl::__internal::__device_backend_tag{}, __exec, __n};
+        oneapi::dpl::__par_backend_hetero::__buffer<oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy,
+                                                    _ValueT>
+            __out_buffer_holder{oneapi::dpl::__internal::__device_backend_tag{}, __exec, __n};
         auto __out_rng = oneapi::dpl::__ranges::all_view<_ValueT, __par_backend_hetero::access_mode::read_write>(
             __out_buffer_holder.get_buffer());
 
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_utils.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_utils.h
index 6a744fa2c25..853cd3deb93 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_utils.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_utils.h
@@ -376,15 +376,16 @@ struct __local_buffer<sycl::buffer<::std::tuple<_T...>, __dim, _AllocT>>
     using type = sycl::buffer<oneapi::dpl::__internal::tuple<_T...>, __dim, _AllocT>;
 };
 
-template <typename _BackendTag, typename _ExecutionPolicy, typename _Tp>
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Tp, typename = void>
 class __buffer_impl;
 
 // impl for sycl::buffer<...>
-template <typename _ExecutionPolicy, typename _T>
-class __buffer_impl<oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy, _T>
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Tp>
+class __buffer_impl<_BackendTag, _ExecutionPolicy, _Tp,
+                    ::std::enable_if_t<::std::is_same_v<_BackendTag, oneapi::dpl::__internal::__device_backend_tag>>>
 {
   private:
-    using __container_t = typename __local_buffer<sycl::buffer<_T>>::type;
+    using __container_t = typename __local_buffer<sycl::buffer<_Tp>>::type;
 
     __container_t __container;
 
@@ -453,9 +454,8 @@ struct __memobj_traits<_T*>
 
 } // namespace __internal
 
-template <typename _ExecutionPolicy, typename _T>
-using __buffer =
-    __internal::__buffer_impl<oneapi::dpl::__internal::__device_backend_tag, ::std::decay_t<_ExecutionPolicy>, _T>;
+template <typename _BackendTag, typename _ExecutionPolicy, typename _T>
+using __buffer = __internal::__buffer_impl<_BackendTag, ::std::decay_t<_ExecutionPolicy>, _T>;
 
 template <typename T>
 struct __repacked_tuple
diff --git a/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
index a24a2de0ec1..a5df37dd935 100644
--- a/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
@@ -162,7 +162,8 @@ __pattern_transform_scan_base(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&
         using _NewExecutionPolicy = decltype(__policy);
 
         // Create temporary buffer
-        oneapi::dpl::__par_backend_hetero::__buffer<_NewExecutionPolicy, _Type> __tmp_buf(_BackendTag{}, __policy, __n);
+        oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _NewExecutionPolicy, _Type> __tmp_buf(_BackendTag{},
+                                                                                                     __policy, __n);
         auto __first_tmp = __tmp_buf.get();
         auto __last_tmp = __first_tmp + __n;
         auto __keep2 = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::write, _Iterator2>();
diff --git a/include/oneapi/dpl/pstl/omp/parallel_scan.h b/include/oneapi/dpl/pstl/omp/parallel_scan.h
index 958161a2325..da272bbb557 100644
--- a/include/oneapi/dpl/pstl/omp/parallel_scan.h
+++ b/include/oneapi/dpl/pstl/omp/parallel_scan.h
@@ -89,7 +89,8 @@ __parallel_strict_scan_body(_ExecutionPolicy&& __exec, _Index __n, _Tp __initial
     const _Index __slack = 4;
     _Index __tilesize = (__n - 1) / (__slack * __p) + 1;
     _Index __m = (__n - 1) / __tilesize;
-    __buffer<_ExecutionPolicy, _Tp> __buf(oneapi::dpl::__internal::__omp_backend_tag{}, __exec, __m + 1);
+    __buffer<oneapi::dpl::__internal::__omp_backend_tag, _ExecutionPolicy, _Tp> __buf(
+        oneapi::dpl::__internal::__omp_backend_tag{}, __exec, __m + 1);
     _Tp* __r = __buf.get();
 
     oneapi::dpl::__omp_backend::__upsweep(_Index(0), _Index(__m + 1), __tilesize, __r, __n - __m * __tilesize, __reduce,
diff --git a/include/oneapi/dpl/pstl/omp/util.h b/include/oneapi/dpl/pstl/omp/util.h
index 6bbc3d3b94d..92a3f8b6258 100644
--- a/include/oneapi/dpl/pstl/omp/util.h
+++ b/include/oneapi/dpl/pstl/omp/util.h
@@ -57,11 +57,12 @@ __cancel_execution(oneapi::dpl::__internal::__omp_backend_tag)
 // raw buffer
 //------------------------------------------------------------------------
 
-template <typename _BackendTag, typename _ExecutionPolicy, typename _Tp>
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Tp, typename = void>
 class __buffer_impl;
 
-template <typename _ExecutionPolicy, typename _Tp>
-class __buffer_impl<oneapi::dpl::__internal::__omp_backend_tag, _ExecutionPolicy, _Tp>
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Tp>
+class __buffer_impl<_BackendTag, _ExecutionPolicy, _Tp,
+                    ::std::enable_if_t<::std::is_same_v<_BackendTag, oneapi::dpl::__internal::__omp_backend_tag>>>
 {
     std::allocator<_Tp> __allocator_;
     _Tp* __ptr_;
@@ -88,8 +89,8 @@ class __buffer_impl<oneapi::dpl::__internal::__omp_backend_tag, _ExecutionPolicy
     ~__buffer_impl() { __allocator_.deallocate(__ptr_, __buf_size_); }
 };
 
-template <typename _ExecutionPolicy, typename _Tp>
-using __buffer = __buffer_impl<oneapi::dpl::__internal::__omp_backend_tag, ::std::decay_t<_ExecutionPolicy>, _Tp>;
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Tp>
+using __buffer = __buffer_impl<_BackendTag, ::std::decay_t<_ExecutionPolicy>, _Tp>;
 
 // Preliminary size of each chunk: requires further discussion
 constexpr std::size_t __default_chunk_size = 2048;
diff --git a/include/oneapi/dpl/pstl/parallel_backend_serial.h b/include/oneapi/dpl/pstl/parallel_backend_serial.h
index c26626ecbbd..ed8e94a0704 100644
--- a/include/oneapi/dpl/pstl/parallel_backend_serial.h
+++ b/include/oneapi/dpl/pstl/parallel_backend_serial.h
@@ -32,11 +32,12 @@ namespace dpl
 namespace __serial_backend
 {
 
-template <typename _BackendTag, typename _ExecutionPolicy, typename _Tp>
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Tp, typename = void>
 class __buffer_impl;
 
-template <typename _ExecutionPolicy, typename _Tp>
-class __buffer_impl<oneapi::dpl::__internal::__serial_backend_tag, _ExecutionPolicy, _Tp>
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Tp>
+class __buffer_impl<_BackendTag, _ExecutionPolicy, _Tp,
+                    ::std::enable_if_t<::std::is_same_v<_BackendTag, oneapi::dpl::__internal::__serial_backend_tag>>>
 {
     ::std::allocator<_Tp> __allocator_;
     _Tp* __ptr_;
@@ -62,8 +63,8 @@ class __buffer_impl<oneapi::dpl::__internal::__serial_backend_tag, _ExecutionPol
     ~__buffer_impl() { __allocator_.deallocate(__ptr_, __buf_size_); }
 };
 
-template <typename _ExecutionPolicy, typename _Tp>
-using __buffer = __buffer_impl<oneapi::dpl::__internal::__serial_backend_tag, ::std::decay_t<_ExecutionPolicy>, _Tp>;
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Tp>
+using __buffer = __buffer_impl<_BackendTag, ::std::decay_t<_ExecutionPolicy>, _Tp>;
 
 inline void
 __cancel_execution(oneapi::dpl::__internal::__serial_backend_tag)
diff --git a/include/oneapi/dpl/pstl/parallel_backend_tbb.h b/include/oneapi/dpl/pstl/parallel_backend_tbb.h
index c2bf8bda04a..0426030e92a 100644
--- a/include/oneapi/dpl/pstl/parallel_backend_tbb.h
+++ b/include/oneapi/dpl/pstl/parallel_backend_tbb.h
@@ -49,7 +49,7 @@ namespace dpl
 namespace __tbb_backend
 {
 
-template <typename _BackendTag, typename _ExecutionPolicy, typename _Tp>
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Tp, typename = void>
 class __buffer_impl;
 
 //! Raw memory buffer with automatic freeing and no exceptions.
@@ -57,8 +57,9 @@ class __buffer_impl;
 not an initialize array, because initialization/destruction
 would make the span be at least O(N). */
 // tbb::allocator can improve performance in some cases.
-template <typename _ExecutionPolicy, typename _Tp>
-class __buffer_impl<oneapi::dpl::__internal::__tbb_backend_tag, _ExecutionPolicy, _Tp>
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Tp>
+class __buffer_impl<_BackendTag, _ExecutionPolicy, _Tp,
+                    ::std::enable_if_t<::std::is_same_v<_BackendTag, oneapi::dpl::__internal::__tbb_backend_tag>>>
 {
     tbb::tbb_allocator<_Tp> _M_allocator;
     _Tp* _M_ptr;
@@ -87,8 +88,8 @@ class __buffer_impl<oneapi::dpl::__internal::__tbb_backend_tag, _ExecutionPolicy
     ~__buffer_impl() { _M_allocator.deallocate(_M_ptr, _M_buf_size); }
 };
 
-template <typename _ExecutionPolicy, typename _Tp>
-using __buffer = __buffer_impl<oneapi::dpl::__internal::__tbb_backend_tag, ::std::decay_t<_ExecutionPolicy>, _Tp>;
+template <typename _BackendTag, typename _ExecutionPolicy, typename _Tp>
+using __buffer = __buffer_impl<_BackendTag, ::std::decay_t<_ExecutionPolicy>, _Tp>;
 
 // Wrapper for tbb::task
 inline void
@@ -396,8 +397,8 @@ __parallel_strict_scan(oneapi::dpl::__internal::__tbb_backend_tag, _ExecutionPol
             const _Index __slack = 4;
             _Index __tilesize = (__n - 1) / (__slack * __p) + 1;
             _Index __m = (__n - 1) / __tilesize;
-            __tbb_backend::__buffer<_ExecutionPolicy, _Tp> __buf(oneapi::dpl::__internal::__tbb_backend_tag{}, __exec,
-                                                                 __m + 1);
+            __tbb_backend::__buffer<oneapi::dpl::__internal::__tbb_backend_tag, _ExecutionPolicy, _Tp> __buf(
+                oneapi::dpl::__internal::__tbb_backend_tag{}, __exec, __m + 1);
             _Tp* __r = __buf.get();
             __tbb_backend::__upsweep(_Index(0), _Index(__m + 1), __tilesize, __r, __n - __m * __tilesize, __reduce,
                                      __combine);
@@ -1203,8 +1204,8 @@ __parallel_stable_sort(oneapi::dpl::__internal::__tbb_backend_tag, _ExecutionPol
         const _DifferenceType __sort_cut_off = _ONEDPL_STABLE_SORT_CUT_OFF;
         if (__n > __sort_cut_off)
         {
-            __tbb_backend::__buffer<_ExecutionPolicy, _ValueType> __buf(oneapi::dpl::__internal::__tbb_backend_tag{},
-                                                                        __exec, __n);
+            __tbb_backend::__buffer<oneapi::dpl::__internal::__tbb_backend_tag, _ExecutionPolicy, _ValueType> __buf(
+                oneapi::dpl::__internal::__tbb_backend_tag{}, __exec, __n);
             __root_task<__stable_sort_func<_RandomAccessIterator, _ValueType*, _Compare, _LeafSort>> __root{
                 __xs, __xe, __buf.get(), true, __comp, __leaf_sort, __nsort, __xs, __buf.get()};
             __task::spawn_root_and_wait(__root);

From 73963234d3e2e2ec106706cb8cb99b0f73aa2e32 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 6 Mar 2024 11:07:38 +0100
Subject: [PATCH 521/566] Remove _BackendTag from params of __buffer contructor

---
 .../internal/exclusive_scan_by_segment_impl.h |  8 ++---
 .../internal/inclusive_scan_by_segment_impl.h |  4 +--
 .../dpl/internal/reduce_by_segment_impl.h     | 29 +++++++++----------
 .../dpl/internal/scan_by_segment_impl.h       | 12 ++++----
 include/oneapi/dpl/pstl/algorithm_impl.h      | 21 +++++++-------
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 28 +++++++-----------
 .../hetero/algorithm_ranges_impl_hetero.h     | 25 +++++++---------
 .../pstl/hetero/dpcpp/parallel_backend_sycl.h |  8 ++---
 .../dpcpp/parallel_backend_sycl_histogram.h   |  2 +-
 .../dpcpp/parallel_backend_sycl_radix_sort.h  |  2 +-
 .../dpcpp/parallel_backend_sycl_utils.h       |  4 +--
 .../dpl/pstl/hetero/numeric_impl_hetero.h     |  3 +-
 include/oneapi/dpl/pstl/omp/parallel_scan.h   |  3 +-
 include/oneapi/dpl/pstl/omp/util.h            |  2 +-
 .../oneapi/dpl/pstl/parallel_backend_serial.h |  2 +-
 .../oneapi/dpl/pstl/parallel_backend_tbb.h    |  8 ++---
 16 files changed, 72 insertions(+), 89 deletions(-)

diff --git a/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h b/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h
index 50211ab30ec..47cb14dd67c 100644
--- a/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h
@@ -66,7 +66,7 @@ pattern_exclusive_scan_by_segment(_Tag, Policy&& policy, InputIterator1 first1,
     InputIterator2 last2 = first2 + n;
 
     // compute head flags
-    oneapi::dpl::__par_backend::__buffer<__backend_tag, Policy, FlagType> _flags(__backend_tag{}, policy, n);
+    oneapi::dpl::__par_backend::__buffer<__backend_tag, Policy, FlagType> _flags(policy, n);
     auto flags = _flags.get();
     flags[0] = 1;
 
@@ -74,7 +74,7 @@ pattern_exclusive_scan_by_segment(_Tag, Policy&& policy, InputIterator1 first1,
               oneapi::dpl::__internal::__not_pred<BinaryPredicate>(binary_pred));
 
     // shift input one to the right and initialize segments with init
-    oneapi::dpl::__par_backend::__buffer<__backend_tag, Policy, OutputType> _temp(__backend_tag{}, policy, n);
+    oneapi::dpl::__par_backend::__buffer<__backend_tag, Policy, OutputType> _temp(policy, n);
     auto temp = _temp.get();
 
     temp[0] = init;
@@ -125,7 +125,7 @@ exclusive_scan_by_segment_impl(__internal::__hetero_tag<_BackendTag>, Policy&& p
     InputIterator2 last2 = first2 + n;
 
     // compute head flags
-    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, Policy, FlagType> _flags(_BackendTag{}, policy, n);
+    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, Policy, FlagType> _flags(policy, n);
     {
         auto flag_buf = _flags.get_buffer();
         auto flags = flag_buf.get_host_access(sycl::read_write);
@@ -136,7 +136,7 @@ exclusive_scan_by_segment_impl(__internal::__hetero_tag<_BackendTag>, Policy&& p
               oneapi::dpl::__internal::__not_pred<BinaryPredicate>(binary_pred));
 
     // shift input one to the right and initialize segments with init
-    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, Policy, OutputType> _temp(_BackendTag{}, policy, n);
+    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, Policy, OutputType> _temp(policy, n);
     {
         auto temp_buf = _temp.get_buffer();
         auto temp = temp_buf.get_host_access(sycl::read_write);
diff --git a/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h b/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h
index 23d197adc9a..d61b08ed552 100644
--- a/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h
@@ -61,7 +61,7 @@ pattern_inclusive_scan_by_segment(_Tag, Policy&& policy, InputIterator1 first1,
     typedef unsigned int FlagType;
     typedef typename ::std::iterator_traits<InputIterator2>::value_type ValueType;
 
-    oneapi::dpl::__par_backend::__buffer<__backend_tag, Policy, FlagType> _mask(__backend_tag{}, policy, n);
+    oneapi::dpl::__par_backend::__buffer<__backend_tag, Policy, FlagType> _mask(policy, n);
     auto mask = _mask.get();
 
     mask[0] = 1;
@@ -111,7 +111,7 @@ inclusive_scan_by_segment_impl(__internal::__hetero_tag<_BackendTag>, Policy&& p
 
     FlagType initial_mask = 1;
 
-    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, Policy, FlagType> _mask(_BackendTag{}, policy, n);
+    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, Policy, FlagType> _mask(policy, n);
     {
         auto mask_buf = _mask.get_buffer();
         auto mask = mask_buf.get_host_access(sycl::read_write);
diff --git a/include/oneapi/dpl/internal/reduce_by_segment_impl.h b/include/oneapi/dpl/internal/reduce_by_segment_impl.h
index 834cd9ed767..c3e0dbc45a3 100644
--- a/include/oneapi/dpl/internal/reduce_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/reduce_by_segment_impl.h
@@ -116,7 +116,7 @@ reduce_by_segment_impl(_Tag, Policy&& policy, InputIterator1 first1, InputIterat
 
     // buffer that is used to store a flag indicating if the associated key is not equal to
     // the next key, and thus its associated sum should be part of the final result
-    oneapi::dpl::__par_backend::__buffer<__backend_tag, Policy, FlagType> _mask(__backend_tag{}, policy, n + 1);
+    oneapi::dpl::__par_backend::__buffer<__backend_tag, Policy, FlagType> _mask(policy, n + 1);
     auto mask = _mask.get();
     mask[0] = 1;
 
@@ -132,11 +132,11 @@ reduce_by_segment_impl(_Tag, Policy&& policy, InputIterator1 first1, InputIterat
     // buffer stores the sums of values associated with a given key. Sums are copied with
     // a shift into result2, and the shift is computed at the same time as the sums, so the
     // sums can't be written to result2 directly.
-    oneapi::dpl::__par_backend::__buffer<__backend_tag, Policy, ValueType> _scanned_values(__backend_tag{}, policy, n);
+    oneapi::dpl::__par_backend::__buffer<__backend_tag, Policy, ValueType> _scanned_values(policy, n);
 
     // Buffer is used to store results of the scan of the mask. Values indicate which position
     // in result2 needs to be written with the scanned_values element.
-    oneapi::dpl::__par_backend::__buffer<__backend_tag, Policy, FlagType> _scanned_tail_flags(__backend_tag{}, policy, n);
+    oneapi::dpl::__par_backend::__buffer<__backend_tag, Policy, FlagType> _scanned_tail_flags(policy, n);
 
     // Compute the sum of the segments. scanned_tail_flags values are not used.
     inclusive_scan(policy, make_zip_iterator(first2, _mask.get()), make_zip_iterator(first2, _mask.get()) + n,
@@ -264,23 +264,22 @@ __sycl_reduce_by_segment(__internal::__hetero_tag<_BackendTag>, _ExecutionPolicy
     ::std::size_t __n_groups = oneapi::dpl::__internal::__dpl_ceiling_div(__n, __wgroup_size * __vals_per_item);
 
     // intermediate reductions within a workgroup
-    auto __partials = oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, __val_type>(
-                          _BackendTag{}, __exec, __n_groups)
-                          .get_buffer();
+    auto __partials =
+        oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, __val_type>(__exec, __n_groups)
+            .get_buffer();
 
-    auto __end_idx = oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, __diff_type>(
-                         _BackendTag{}, __exec, 1)
-                         .get_buffer();
+    auto __end_idx =
+        oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, __diff_type>(__exec, 1).get_buffer();
 
     // the number of segment ends found in each work group
-    auto __seg_ends = oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, __diff_type>(
-                          _BackendTag{}, __exec, __n_groups)
-                          .get_buffer();
+    auto __seg_ends =
+        oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, __diff_type>(__exec, __n_groups)
+            .get_buffer();
 
     // buffer that stores an exclusive scan of the results
-    auto __seg_ends_scanned = oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, __diff_type>(
-                                  _BackendTag{}, __exec, __n_groups)
-                                  .get_buffer();
+    auto __seg_ends_scanned =
+        oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, __diff_type>(__exec, __n_groups)
+            .get_buffer();
 
     // 1. Count the segment ends in each workgroup
     auto __seg_end_identification = __exec.queue().submit([&](sycl::handler& __cgh) {
diff --git a/include/oneapi/dpl/internal/scan_by_segment_impl.h b/include/oneapi/dpl/internal/scan_by_segment_impl.h
index 9104ac18e42..85903a93fdf 100644
--- a/include/oneapi/dpl/internal/scan_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/scan_by_segment_impl.h
@@ -144,14 +144,14 @@ struct __sycl_scan_by_segment_impl
 
         ::std::size_t __n_groups = __internal::__dpl_ceiling_div(__n, __wgroup_size * __vals_per_item);
 
-        auto __partials = oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, __val_type>(
-                              _BackendTag{}, __exec, __n_groups)
-                              .get_buffer();
+        auto __partials =
+            oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, __val_type>(__exec, __n_groups)
+                .get_buffer();
 
         // the number of segment ends found in each work group
-        auto __seg_ends = oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, bool>(
-                              _BackendTag{}, __exec, __n_groups)
-                              .get_buffer();
+        auto __seg_ends =
+            oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, bool>(__exec, __n_groups)
+                .get_buffer();
 
         // 1. Work group reduction
         auto __wg_scan = __exec.queue().submit([&](sycl::handler& __cgh) {
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 8620b81b353..744a8064fcd 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -1325,7 +1325,7 @@ __pattern_copy_if(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomA
     const _DifferenceType __n = __last - __first;
     if (_DifferenceType(1) < __n)
     {
-        __par_backend::__buffer<__backend_tag, _ExecutionPolicy, bool> __mask_buf(__backend_tag{}, __exec, __n);
+        __par_backend::__buffer<__backend_tag, _ExecutionPolicy, bool> __mask_buf(__exec, __n);
         return __internal::__except_handler([&__exec, __n, __first, __result, __pred, &__mask_buf]() {
             bool* __mask = __mask_buf.get();
             _DifferenceType __m{};
@@ -1444,7 +1444,7 @@ __remove_elements(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Forward
     typedef typename ::std::iterator_traits<_ForwardIterator>::difference_type _DifferenceType;
     typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _Tp;
     _DifferenceType __n = __last - __first;
-    __par_backend::__buffer<__backend_tag, _ExecutionPolicy, bool> __mask_buf(__backend_tag{}, __exec, __n);
+    __par_backend::__buffer<__backend_tag, _ExecutionPolicy, bool> __mask_buf(__exec, __n);
     // 1. find a first iterator that should be removed
     return __internal::__except_handler([&]() {
         bool* __mask = __mask_buf.get();
@@ -1481,7 +1481,7 @@ __remove_elements(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Forward
         __n -= __min;
         __first += __min;
 
-        __par_backend::__buffer<__backend_tag, _ExecutionPolicy, _Tp> __buf(__backend_tag{}, __exec, __n);
+        __par_backend::__buffer<__backend_tag, _ExecutionPolicy, _Tp> __buf(__exec, __n);
         _Tp* __result = __buf.get();
         __mask += __min;
         _DifferenceType __m{};
@@ -1609,7 +1609,7 @@ __pattern_unique_copy(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Ran
     const _DifferenceType __n = __last - __first;
     if (_DifferenceType(2) < __n)
     {
-        __par_backend::__buffer<__backend_tag, _ExecutionPolicy, bool> __mask_buf(__backend_tag{}, __exec, __n);
+        __par_backend::__buffer<__backend_tag, _ExecutionPolicy, bool> __mask_buf(__exec, __n);
         if (_DifferenceType(2) < __n)
         {
             return __internal::__except_handler([&__exec, __n, __first, __result, __pred, &__mask_buf]() {
@@ -1850,7 +1850,7 @@ __pattern_rotate(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAc
     auto __m = __middle - __first;
     if (__m <= __n / 2)
     {
-        __par_backend::__buffer<__backend_tag, _ExecutionPolicy, _Tp> __buf(__backend_tag{}, __exec, __n - __m);
+        __par_backend::__buffer<__backend_tag, _ExecutionPolicy, _Tp> __buf(__exec, __n - __m);
         return __internal::__except_handler([&__exec, __n, __m, __first, __middle, __last, &__buf]() {
             _Tp* __result = __buf.get();
             __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __middle, __last,
@@ -1876,7 +1876,7 @@ __pattern_rotate(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAc
     }
     else
     {
-        __par_backend::__buffer<__backend_tag, _ExecutionPolicy, _Tp> __buf(__backend_tag{}, __exec, __m);
+        __par_backend::__buffer<__backend_tag, _ExecutionPolicy, _Tp> __buf(__exec, __m);
         return __internal::__except_handler([&__exec, __n, __m, __first, __middle, __last, &__buf]() {
             _Tp* __result = __buf.get();
             __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __middle,
@@ -2354,7 +2354,7 @@ __pattern_partition_copy(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _
     const _DifferenceType __n = __last - __first;
     if (_DifferenceType(1) < __n)
     {
-        __par_backend::__buffer<__backend_tag, _ExecutionPolicy, bool> __mask_buf(__backend_tag{}, __exec, __n);
+        __par_backend::__buffer<__backend_tag, _ExecutionPolicy, bool> __mask_buf(__exec, __n);
         return __internal::__except_handler([&__exec, __n, __first, __out_true, __out_false, __pred, &__mask_buf]() {
             bool* __mask = __mask_buf.get();
             _ReturnType __m{};
@@ -2580,7 +2580,7 @@ __pattern_partial_sort_copy(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec
         {
             typedef typename ::std::iterator_traits<_RandomAccessIterator1>::value_type _T1;
             typedef typename ::std::iterator_traits<_RandomAccessIterator2>::value_type _T2;
-            __par_backend::__buffer<__backend_tag, _ExecutionPolicy, _T1> __buf(__backend_tag{}, __exec, __n1);
+            __par_backend::__buffer<__backend_tag, _ExecutionPolicy, _T1> __buf(__exec, __n1);
             _T1* __r = __buf.get();
 
             __par_backend::__parallel_stable_sort(
@@ -3073,7 +3073,7 @@ __pattern_inplace_merge(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _R
 
     typedef typename ::std::iterator_traits<_RandomAccessIterator>::value_type _Tp;
     auto __n = __last - __first;
-    __par_backend::__buffer<__backend_tag, _ExecutionPolicy, _Tp> __buf(__backend_tag{}, __exec, __n);
+    __par_backend::__buffer<__backend_tag, _ExecutionPolicy, _Tp> __buf(__exec, __n);
     _Tp* __r = __buf.get();
     __internal::__except_handler([&]() {
         auto __move_values = [](_RandomAccessIterator __x, _Tp* __z) {
@@ -3207,8 +3207,7 @@ __parallel_set_op(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Forward
     const _DifferenceType __n1 = __last1 - __first1;
     const _DifferenceType __n2 = __last2 - __first2;
 
-    __par_backend::__buffer<__backend_tag, _ExecutionPolicy, _T> __buf(__backend_tag{}, __exec,
-                                                                       __size_func(__n1, __n2));
+    __par_backend::__buffer<__backend_tag, _ExecutionPolicy, _T> __buf(__exec, __size_func(__n1, __n2));
 
     return __internal::__except_handler([&__exec, __n1, __first1, __last1, __first2, __last2, __result, __comp,
                                          __size_func, __set_op, &__buf]() {
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 64ab0123d88..1de2a8f224d 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -1032,7 +1032,7 @@ __pattern_remove_if(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec,
 
     using _ValueType = typename ::std::iterator_traits<_Iterator>::value_type;
 
-    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _ValueType> __buf(_BackendTag{}, __exec,
+    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _ValueType> __buf(__exec,
                                                                                                  __last - __first);
     auto __copy_first = __buf.get();
 
@@ -1054,7 +1054,7 @@ __pattern_unique(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _It
 
     using _ValueType = typename ::std::iterator_traits<_Iterator>::value_type;
 
-    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _ValueType> __buf(_BackendTag{}, __exec,
+    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _ValueType> __buf(__exec,
                                                                                                  __last - __first);
     auto __copy_first = __buf.get();
     auto __copy_last = __pattern_unique_copy(__tag, __exec, __first, __last, __copy_first, __pred);
@@ -1229,8 +1229,7 @@ __pattern_inplace_merge(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __ex
     assert(__first < __middle && __middle < __last);
 
     auto __n = __last - __first;
-    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _ValueType> __buf(_BackendTag{}, __exec,
-                                                                                                 __n);
+    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _ValueType> __buf(__exec, __n);
     auto __copy_first = __buf.get();
     auto __copy_last = __copy_first + __n;
 
@@ -1318,8 +1317,8 @@ __pattern_stable_partition(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& _
 
     auto __n = __last - __first;
 
-    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _ValueType> __true_buf(_BackendTag{}, __exec, __n);
-    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _ValueType> __false_buf(_BackendTag{}, __exec, __n);
+    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _ValueType> __true_buf(__exec, __n);
+    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _ValueType> __false_buf(__exec, __n);
     auto __true_result = __true_buf.get();
     auto __false_result = __false_buf.get();
 
@@ -1516,8 +1515,7 @@ __pattern_partial_sort_copy(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&&
         // - create a temporary buffer and copy all the elements from the input buffer there
         // - run partial sort on the temporary buffer
         // - copy k elements from the temporary buffer to the output buffer.
-        oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _ValueType> __buf(_BackendTag{},
-                                                                                                     __exec, __in_size);
+        oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _ValueType> __buf(__exec, __in_size);
 
         auto __buf_first = __buf.get();
 
@@ -1633,7 +1631,7 @@ __pattern_rotate(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Iterator
     auto __keep = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read_write, _Iterator>();
     auto __buf = __keep(__first, __last);
     auto __temp_buf =
-        oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _Tp>(_BackendTag{}, __exec, __n);
+        oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _Tp>(__exec, __n);
 
     auto __temp_rng =
         oneapi::dpl::__ranges::all_view<_Tp, __par_backend_hetero::access_mode::write>(__temp_buf.get_buffer());
@@ -1716,8 +1714,7 @@ __pattern_hetero_set_op(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _F
         __comp, __n1, __n2};
 
     // temporary buffer to store boolean mask
-    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, int32_t> __mask_buf(_BackendTag{}, __exec,
-                                                                                                 __n1);
+    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, int32_t> __mask_buf(__exec, __n1);
 
     auto __keep1 =
         oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _ForwardIterator1>();
@@ -1846,8 +1843,7 @@ __pattern_set_union(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec,
 
     // temporary buffer to store intermediate result
     const auto __n2 = __last2 - __first2;
-    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _ValueType> __diff(_BackendTag{}, __exec,
-                                                                                                  __n2);
+    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _ValueType> __diff(__exec, __n2);
     auto __buf = __diff.get();
 
     //1. Calc difference {2} \ {1}
@@ -1928,12 +1924,10 @@ __pattern_set_symmetric_difference(__hetero_tag<_BackendTag> __tag, _ExecutionPo
 
     // temporary buffers to store intermediate result
     const auto __n1 = __last1 - __first1;
-    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _ValueType> __diff_1(_BackendTag{},
-                                                                                                    __exec, __n1);
+    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _ValueType> __diff_1(__exec, __n1);
     auto __buf_1 = __diff_1.get();
     const auto __n2 = __last2 - __first2;
-    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _ValueType> __diff_2(_BackendTag{},
-                                                                                                    __exec, __n2);
+    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _ValueType> __diff_2(__exec, __n2);
     auto __buf_2 = __diff_2.get();
 
     //1. Calc difference {1} \ {2}
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
index 715c0d720ef..227ff9761b5 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
@@ -360,8 +360,8 @@ __pattern_scan_copy(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Range
     _DataAcc __get_data_op;
     _MaskAssigner __add_mask_op;
 
-    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, int32_t> __mask_buf(
-        _BackendTag{}, __exec, __rng1.size());
+    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, int32_t> __mask_buf(__exec,
+                                                                                                   __rng1.size());
 
     auto __res =
         __par_backend_hetero::__parallel_transform_scan_base(
@@ -414,8 +414,7 @@ __pattern_remove_if(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec,
 
     using _ValueType = oneapi::dpl::__internal::__value_t<_Range>;
 
-    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _ValueType> __buf(_BackendTag{}, __exec,
-                                                                                                 __rng.size());
+    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _ValueType> __buf(__exec, __rng.size());
     auto __copy_rng = oneapi::dpl::__ranges::views::all(__buf.get_buffer());
 
     auto __copy_last_id = __ranges::__pattern_copy_if(__tag, __exec, __rng, __copy_rng, __not_pred<_Predicate>{__pred},
@@ -463,8 +462,7 @@ __pattern_unique(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Ra
 
     using _ValueType = oneapi::dpl::__internal::__value_t<_Range>;
 
-    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _ValueType> __buf(_BackendTag{}, __exec,
-                                                                                                 __rng.size());
+    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _ValueType> __buf(__exec, __rng.size());
     auto res_rng = oneapi::dpl::__ranges::views::all(__buf.get_buffer());
     auto res = __ranges::__pattern_unique_copy(__tag, __exec, __rng, res_rng, __pred,
                                                oneapi::dpl::__internal::__pstl_assign());
@@ -718,15 +716,14 @@ __pattern_reduce_by_segment(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&&
     // Round 1: reduce with extra indices added to avoid long segments
     // TODO: At threshold points check if the key is equal to the key at the previous threshold point, indicating a long sequence.
     // Skip a round of copy_if and reduces if there are none.
-    auto __idx = oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, __diff_type>(_BackendTag{},
-                                                                                                         __exec, __n)
+    auto __idx = oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, __diff_type>(__exec, __n)
                      .get_buffer();
-    auto __tmp_out_keys = oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, __key_type>(
-                              _BackendTag{}, __exec, __n)
-                              .get_buffer();
-    auto __tmp_out_values = oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, __val_type>(
-                                _BackendTag{}, __exec, __n)
-                                .get_buffer();
+    auto __tmp_out_keys =
+        oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, __key_type>(__exec, __n)
+            .get_buffer();
+    auto __tmp_out_values =
+        oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, __val_type>(__exec, __n)
+            .get_buffer();
 
     // Replicating first element of keys view to be able to compare (i-1)-th and (i)-th key with aligned sequences,
     //  dropping the last key for the i-1 sequence.
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
index 46c02feb3a7..e3c3f62d85a 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
@@ -873,7 +873,7 @@ __parallel_scan_copy(oneapi::dpl::__internal::__device_backend_tag __backend_tag
     // temporary buffer to store boolean mask
     oneapi::dpl::__par_backend_hetero::__buffer<oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy,
                                                 int32_t>
-        __mask_buf(oneapi::dpl::__internal::__device_backend_tag{}, __exec, __n);
+        __mask_buf(__exec, __n);
 
     return __parallel_transform_scan_base(
         __backend_tag, ::std::forward<_ExecutionPolicy>(__exec),
@@ -1581,8 +1581,7 @@ struct __parallel_sort_submitter<_BackendTag, _IdType, __internal::__optional_ke
         });
 
         // 2. Merge sorting
-        oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _Tp> __temp_buf(_BackendTag{},
-                                                                                                   __exec, __n);
+        oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _Tp> __temp_buf(__exec, __n);
         auto __temp = __temp_buf.get_buffer();
         bool __data_in_temp = false;
         _IdType __n_sorted = __leaf;
@@ -1703,8 +1702,7 @@ struct __parallel_partial_sort_submitter<_BackendTag, __internal::__optional_ker
         _Size __n = __rng.size();
         assert(__n > 1);
 
-        oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _Tp> __temp_buf(_BackendTag{}, __exec,
-                                                                                                 __n);
+        oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _Tp> __temp_buf(__exec, __n);
         auto __temp = __temp_buf.get_buffer();
         _PRINT_INFO_IN_DEBUG_MODE(__exec);
 
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h
index d31b1e6d346..083bffcefc0 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h
@@ -428,7 +428,7 @@ struct __histogram_general_private_global_atomics_submitter<_BackendTag,
 
         auto __private_histograms =
             oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _bin_type>(
-                _BackendTag{}, __exec, __segments * __num_bins)
+                __exec, __segments * __num_bins)
                 .get_buffer();
 
         return __exec.queue().submit([&](auto& __h) {
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_radix_sort.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_radix_sort.h
index df1649f99cb..05a1e8ccfc8 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_radix_sort.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_radix_sort.h
@@ -829,7 +829,7 @@ __parallel_radix_sort(oneapi::dpl::__internal::__device_backend_tag, _ExecutionP
         // memory for storing values sorted for an iteration
         oneapi::dpl::__par_backend_hetero::__buffer<oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy,
                                                     _ValueT>
-            __out_buffer_holder{oneapi::dpl::__internal::__device_backend_tag{}, __exec, __n};
+            __out_buffer_holder{__exec, __n};
         auto __out_rng = oneapi::dpl::__ranges::all_view<_ValueT, __par_backend_hetero::access_mode::read_write>(
             __out_buffer_holder.get_buffer());
 
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_utils.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_utils.h
index 853cd3deb93..f8bf122ad12 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_utils.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_utils.h
@@ -392,9 +392,7 @@ class __buffer_impl<_BackendTag, _ExecutionPolicy, _Tp,
   public:
     static_assert(::std::is_same_v<_ExecutionPolicy, ::std::decay_t<_ExecutionPolicy>>);
 
-    __buffer_impl(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy /*__exec*/,
-                  ::std::size_t __n_elements)
-        : __container{sycl::range<1>(__n_elements)}
+    __buffer_impl(_ExecutionPolicy /*__exec*/, ::std::size_t __n_elements) : __container{sycl::range<1>(__n_elements)}
     {
     }
 
diff --git a/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
index a5df37dd935..04f90cbdc04 100644
--- a/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
@@ -162,8 +162,7 @@ __pattern_transform_scan_base(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&
         using _NewExecutionPolicy = decltype(__policy);
 
         // Create temporary buffer
-        oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _NewExecutionPolicy, _Type> __tmp_buf(_BackendTag{},
-                                                                                                     __policy, __n);
+        oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _NewExecutionPolicy, _Type> __tmp_buf(__policy, __n);
         auto __first_tmp = __tmp_buf.get();
         auto __last_tmp = __first_tmp + __n;
         auto __keep2 = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::write, _Iterator2>();
diff --git a/include/oneapi/dpl/pstl/omp/parallel_scan.h b/include/oneapi/dpl/pstl/omp/parallel_scan.h
index da272bbb557..4c2ff2bc9e3 100644
--- a/include/oneapi/dpl/pstl/omp/parallel_scan.h
+++ b/include/oneapi/dpl/pstl/omp/parallel_scan.h
@@ -89,8 +89,7 @@ __parallel_strict_scan_body(_ExecutionPolicy&& __exec, _Index __n, _Tp __initial
     const _Index __slack = 4;
     _Index __tilesize = (__n - 1) / (__slack * __p) + 1;
     _Index __m = (__n - 1) / __tilesize;
-    __buffer<oneapi::dpl::__internal::__omp_backend_tag, _ExecutionPolicy, _Tp> __buf(
-        oneapi::dpl::__internal::__omp_backend_tag{}, __exec, __m + 1);
+    __buffer<oneapi::dpl::__internal::__omp_backend_tag, _ExecutionPolicy, _Tp> __buf(__exec, __m + 1);
     _Tp* __r = __buf.get();
 
     oneapi::dpl::__omp_backend::__upsweep(_Index(0), _Index(__m + 1), __tilesize, __r, __n - __m * __tilesize, __reduce,
diff --git a/include/oneapi/dpl/pstl/omp/util.h b/include/oneapi/dpl/pstl/omp/util.h
index 92a3f8b6258..d3fa736c20f 100644
--- a/include/oneapi/dpl/pstl/omp/util.h
+++ b/include/oneapi/dpl/pstl/omp/util.h
@@ -74,7 +74,7 @@ class __buffer_impl<_BackendTag, _ExecutionPolicy, _Tp,
   public:
     static_assert(::std::is_same_v<_ExecutionPolicy, ::std::decay_t<_ExecutionPolicy>>);
 
-    __buffer_impl(oneapi::dpl::__internal::__omp_backend_tag, _ExecutionPolicy /*__exec*/, std::size_t __n)
+    __buffer_impl(_ExecutionPolicy /*__exec*/, std::size_t __n)
         : __allocator_(), __ptr_(__allocator_.allocate(__n)), __buf_size_(__n)
     {
     }
diff --git a/include/oneapi/dpl/pstl/parallel_backend_serial.h b/include/oneapi/dpl/pstl/parallel_backend_serial.h
index ed8e94a0704..306921f61c1 100644
--- a/include/oneapi/dpl/pstl/parallel_backend_serial.h
+++ b/include/oneapi/dpl/pstl/parallel_backend_serial.h
@@ -49,7 +49,7 @@ class __buffer_impl<_BackendTag, _ExecutionPolicy, _Tp,
   public:
     static_assert(::std::is_same_v<_ExecutionPolicy, ::std::decay_t<_ExecutionPolicy>>);
 
-    __buffer_impl(oneapi::dpl::__internal::__serial_backend_tag, _ExecutionPolicy /*__exec*/, ::std::size_t __n)
+    __buffer_impl(_ExecutionPolicy /*__exec*/, ::std::size_t __n)
         : __allocator_(), __ptr_(__allocator_.allocate(__n)), __buf_size_(__n)
     {
     }
diff --git a/include/oneapi/dpl/pstl/parallel_backend_tbb.h b/include/oneapi/dpl/pstl/parallel_backend_tbb.h
index 0426030e92a..d0394b9369f 100644
--- a/include/oneapi/dpl/pstl/parallel_backend_tbb.h
+++ b/include/oneapi/dpl/pstl/parallel_backend_tbb.h
@@ -72,7 +72,7 @@ class __buffer_impl<_BackendTag, _ExecutionPolicy, _Tp,
     static_assert(::std::is_same_v<_ExecutionPolicy, ::std::decay_t<_ExecutionPolicy>>);
 
     //! Try to obtain buffer of given size to store objects of _Tp type
-    __buffer_impl(oneapi::dpl::__internal::__tbb_backend_tag, _ExecutionPolicy /*__exec*/, const ::std::size_t __n)
+    __buffer_impl(_ExecutionPolicy /*__exec*/, const ::std::size_t __n)
         : _M_allocator(), _M_ptr(_M_allocator.allocate(__n)), _M_buf_size(__n)
     {
     }
@@ -397,8 +397,8 @@ __parallel_strict_scan(oneapi::dpl::__internal::__tbb_backend_tag, _ExecutionPol
             const _Index __slack = 4;
             _Index __tilesize = (__n - 1) / (__slack * __p) + 1;
             _Index __m = (__n - 1) / __tilesize;
-            __tbb_backend::__buffer<oneapi::dpl::__internal::__tbb_backend_tag, _ExecutionPolicy, _Tp> __buf(
-                oneapi::dpl::__internal::__tbb_backend_tag{}, __exec, __m + 1);
+            __tbb_backend::__buffer<oneapi::dpl::__internal::__tbb_backend_tag, _ExecutionPolicy, _Tp> __buf(__exec,
+                                                                                                             __m + 1);
             _Tp* __r = __buf.get();
             __tbb_backend::__upsweep(_Index(0), _Index(__m + 1), __tilesize, __r, __n - __m * __tilesize, __reduce,
                                      __combine);
@@ -1205,7 +1205,7 @@ __parallel_stable_sort(oneapi::dpl::__internal::__tbb_backend_tag, _ExecutionPol
         if (__n > __sort_cut_off)
         {
             __tbb_backend::__buffer<oneapi::dpl::__internal::__tbb_backend_tag, _ExecutionPolicy, _ValueType> __buf(
-                oneapi::dpl::__internal::__tbb_backend_tag{}, __exec, __n);
+                __exec, __n);
             __root_task<__stable_sort_func<_RandomAccessIterator, _ValueType*, _Compare, _LeafSort>> __root{
                 __xs, __xe, __buf.get(), true, __comp, __leaf_sort, __nsort, __xs, __buf.get()};
             __task::spawn_root_and_wait(__root);

From 9f9f0178e1017efe707a15f5b9075bb7285b4da5 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 6 Mar 2024 11:08:57 +0100
Subject: [PATCH 522/566] Remove
 static_assert(::std::is_same_v<_ExecutionPolicy,
 ::std::decay_t<_ExecutionPolicy>>); from __buffer implementatios as already
 not required

---
 .../oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_utils.h  | 2 --
 include/oneapi/dpl/pstl/omp/util.h                              | 2 --
 include/oneapi/dpl/pstl/parallel_backend_serial.h               | 2 --
 include/oneapi/dpl/pstl/parallel_backend_tbb.h                  | 2 --
 4 files changed, 8 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_utils.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_utils.h
index f8bf122ad12..9319486f150 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_utils.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_utils.h
@@ -390,8 +390,6 @@ class __buffer_impl<_BackendTag, _ExecutionPolicy, _Tp,
     __container_t __container;
 
   public:
-    static_assert(::std::is_same_v<_ExecutionPolicy, ::std::decay_t<_ExecutionPolicy>>);
-
     __buffer_impl(_ExecutionPolicy /*__exec*/, ::std::size_t __n_elements) : __container{sycl::range<1>(__n_elements)}
     {
     }
diff --git a/include/oneapi/dpl/pstl/omp/util.h b/include/oneapi/dpl/pstl/omp/util.h
index d3fa736c20f..c616aedc206 100644
--- a/include/oneapi/dpl/pstl/omp/util.h
+++ b/include/oneapi/dpl/pstl/omp/util.h
@@ -72,8 +72,6 @@ class __buffer_impl<_BackendTag, _ExecutionPolicy, _Tp,
     operator=(const __buffer_impl&) = delete;
 
   public:
-    static_assert(::std::is_same_v<_ExecutionPolicy, ::std::decay_t<_ExecutionPolicy>>);
-
     __buffer_impl(_ExecutionPolicy /*__exec*/, std::size_t __n)
         : __allocator_(), __ptr_(__allocator_.allocate(__n)), __buf_size_(__n)
     {
diff --git a/include/oneapi/dpl/pstl/parallel_backend_serial.h b/include/oneapi/dpl/pstl/parallel_backend_serial.h
index 306921f61c1..867b925ff91 100644
--- a/include/oneapi/dpl/pstl/parallel_backend_serial.h
+++ b/include/oneapi/dpl/pstl/parallel_backend_serial.h
@@ -47,8 +47,6 @@ class __buffer_impl<_BackendTag, _ExecutionPolicy, _Tp,
     operator=(const __buffer_impl&) = delete;
 
   public:
-    static_assert(::std::is_same_v<_ExecutionPolicy, ::std::decay_t<_ExecutionPolicy>>);
-
     __buffer_impl(_ExecutionPolicy /*__exec*/, ::std::size_t __n)
         : __allocator_(), __ptr_(__allocator_.allocate(__n)), __buf_size_(__n)
     {
diff --git a/include/oneapi/dpl/pstl/parallel_backend_tbb.h b/include/oneapi/dpl/pstl/parallel_backend_tbb.h
index d0394b9369f..c9cc0a8ff52 100644
--- a/include/oneapi/dpl/pstl/parallel_backend_tbb.h
+++ b/include/oneapi/dpl/pstl/parallel_backend_tbb.h
@@ -69,8 +69,6 @@ class __buffer_impl<_BackendTag, _ExecutionPolicy, _Tp,
     operator=(const __buffer_impl&) = delete;
 
   public:
-    static_assert(::std::is_same_v<_ExecutionPolicy, ::std::decay_t<_ExecutionPolicy>>);
-
     //! Try to obtain buffer of given size to store objects of _Tp type
     __buffer_impl(_ExecutionPolicy /*__exec*/, const ::std::size_t __n)
         : _M_allocator(), _M_ptr(_M_allocator.allocate(__n)), _M_buf_size(__n)

From 25abaa9cd681d71dc3e0c760723dc8bcd261692f Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 6 Mar 2024 11:42:40 +0100
Subject: [PATCH 523/566] Rename dispatch tag check names

---
 .../oneapi/dpl/internal/binary_search_impl.h  |   6 +-
 .../internal/exclusive_scan_by_segment_impl.h |   2 +-
 .../internal/inclusive_scan_by_segment_impl.h |   2 +-
 .../dpl/internal/reduce_by_segment_impl.h     |   2 +-
 include/oneapi/dpl/pstl/algorithm_impl.h      | 134 +++++++++---------
 include/oneapi/dpl/pstl/execution_impl.h      |  30 ++--
 .../experimental/internal/for_loop_impl.h     |   8 +-
 include/oneapi/dpl/pstl/histogram_impl.h      |   2 +-
 include/oneapi/dpl/pstl/numeric_impl.h        |  10 +-
 9 files changed, 98 insertions(+), 98 deletions(-)

diff --git a/include/oneapi/dpl/internal/binary_search_impl.h b/include/oneapi/dpl/internal/binary_search_impl.h
index ee5240b92cf..59c2f74ea82 100644
--- a/include/oneapi/dpl/internal/binary_search_impl.h
+++ b/include/oneapi/dpl/internal/binary_search_impl.h
@@ -74,7 +74,7 @@ OutputIterator
 lower_bound_impl(_Tag, Policy&& policy, InputIterator1 start, InputIterator1 end, InputIterator2 value_start,
                  InputIterator2 value_end, OutputIterator result, StrictWeakOrdering comp)
 {
-    static_assert(__internal::__is_host_backend_tag_v<_Tag>);
+    static_assert(__internal::__is_host_dispatch_tag_v<_Tag>);
 
     return oneapi::dpl::transform(policy, value_start, value_end, result,
                                   [=](typename ::std::iterator_traits<InputIterator2>::reference val) {
@@ -88,7 +88,7 @@ OutputIterator
 upper_bound_impl(_Tag, Policy&& policy, InputIterator1 start, InputIterator1 end, InputIterator2 value_start,
                  InputIterator2 value_end, OutputIterator result, StrictWeakOrdering comp)
 {
-    static_assert(__internal::__is_host_backend_tag_v<_Tag>);
+    static_assert(__internal::__is_host_dispatch_tag_v<_Tag>);
 
     return oneapi::dpl::transform(policy, value_start, value_end, result,
                                   [=](typename ::std::iterator_traits<InputIterator2>::reference val) {
@@ -102,7 +102,7 @@ OutputIterator
 binary_search_impl(_Tag, Policy&& policy, InputIterator1 start, InputIterator1 end, InputIterator2 value_start,
                    InputIterator2 value_end, OutputIterator result, StrictWeakOrdering comp)
 {
-    static_assert(__internal::__is_host_backend_tag_v<_Tag>);
+    static_assert(__internal::__is_host_dispatch_tag_v<_Tag>);
 
     return oneapi::dpl::transform(policy, value_start, value_end, result,
                                   [=](typename ::std::iterator_traits<InputIterator2>::reference val) {
diff --git a/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h b/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h
index 47cb14dd67c..39f6d5ebf70 100644
--- a/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h
@@ -44,7 +44,7 @@ pattern_exclusive_scan_by_segment(_Tag, Policy&& policy, InputIterator1 first1,
                                   InputIterator2 first2, OutputIterator result, T init, BinaryPredicate binary_pred,
                                   Operator binary_op)
 {
-    static_assert(__internal::__is_host_backend_tag_v<_Tag>);
+    static_assert(__internal::__is_host_dispatch_tag_v<_Tag>);
 
     using __backend_tag = typename _Tag::__backend_tag;
 
diff --git a/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h b/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h
index d61b08ed552..36ef00a64f0 100644
--- a/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h
@@ -43,7 +43,7 @@ pattern_inclusive_scan_by_segment(_Tag, Policy&& policy, InputIterator1 first1,
                                   InputIterator2 first2, OutputIterator result, BinaryPredicate binary_pred,
                                   BinaryOperator binary_op)
 {
-    static_assert(__internal::__is_host_backend_tag_v<_Tag>);
+    static_assert(__internal::__is_host_dispatch_tag_v<_Tag>);
 
     using __backend_tag = typename _Tag::__backend_tag;
 
diff --git a/include/oneapi/dpl/internal/reduce_by_segment_impl.h b/include/oneapi/dpl/internal/reduce_by_segment_impl.h
index c3e0dbc45a3..84ec72c7ee9 100644
--- a/include/oneapi/dpl/internal/reduce_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/reduce_by_segment_impl.h
@@ -85,7 +85,7 @@ reduce_by_segment_impl(_Tag, Policy&& policy, InputIterator1 first1, InputIterat
                        OutputIterator1 result1, OutputIterator2 result2, BinaryPred binary_pred,
                        BinaryOperator binary_op)
 {
-    static_assert(__internal::__is_host_backend_tag_v<_Tag>);
+    static_assert(__internal::__is_host_dispatch_tag_v<_Tag>);
 
     using __backend_tag = typename _Tag::__backend_tag;
 
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 744a8064fcd..5fda420c564 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -67,7 +67,7 @@ template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Pre
 bool
 __pattern_any_of(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _Pred __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_serial_tag_v<_Tag> || __is_parallel_forward_tag_v<_Tag>);
 
     return __internal::__brick_any_of(__first, __last, __pred, typename _Tag::__is_vector{});
 }
@@ -139,7 +139,7 @@ template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Fun
 void
 __pattern_walk1(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _Function __f) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag>);
+    static_assert(__is_serial_tag_v<_Tag>);
 
     __internal::__brick_walk1(__first, __last, __f, typename _Tag::__is_vector{});
 }
@@ -179,7 +179,7 @@ void
 __pattern_walk_brick(_Tag, _ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last,
                      _Brick __brick) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_serial_tag_v<_Tag> || __is_parallel_forward_tag_v<_Tag>);
 
     __brick(__first, __last, typename _Tag::__is_vector{});
 }
@@ -221,7 +221,7 @@ template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Siz
 _ForwardIterator
 __pattern_walk1_n(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _Size __n, _Function __f) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_serial_tag_v<_Tag> || __is_parallel_forward_tag_v<_Tag>);
 
     return __internal::__brick_walk1_n(__first, __n, __f, typename _Tag::__is_vector{});
 }
@@ -240,7 +240,7 @@ template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Siz
 _ForwardIterator
 __pattern_walk_brick_n(_Tag, _ExecutionPolicy&& __exec, _ForwardIterator __first, _Size __n, _Brick __brick) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_serial_tag_v<_Tag> || __is_parallel_forward_tag_v<_Tag>);
 
     return __brick(__first, __n, typename _Tag::__is_vector{});
 }
@@ -307,7 +307,7 @@ _ForwardIterator2
 __pattern_walk2(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
                 _ForwardIterator2 __first2, _Function __f) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag>);
+    static_assert(__is_serial_tag_v<_Tag>);
 
     return __internal::__brick_walk2(__first1, __last1, __first2, __f, typename _Tag::__is_vector{});
 }
@@ -364,7 +364,7 @@ _ForwardIterator2
 __pattern_walk2_n(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _Size __n, _ForwardIterator2 __first2,
                   _Function __f) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_serial_tag_v<_Tag> || __is_parallel_forward_tag_v<_Tag>);
 
     return __internal::__brick_walk2_n(__first1, __n, __first2, __f, typename _Tag::__is_vector{});
 }
@@ -384,7 +384,7 @@ _ForwardIterator2
 __pattern_walk2_brick(_Tag, _ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
                       _ForwardIterator2 __first2, _Brick __brick) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag>);
+    static_assert(__is_serial_tag_v<_Tag>);
 
     return __brick(__first1, __last1, __first2, typename _Tag::__is_vector{});
 }
@@ -461,7 +461,7 @@ _ForwardIterator2
 __pattern_walk2_brick_n(_Tag, _ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _Size __n,
                         _ForwardIterator2 __first2, _Brick __brick) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_serial_tag_v<_Tag> || __is_parallel_forward_tag_v<_Tag>);
 
     return __brick(__first1, __n, __first2, typename _Tag::__is_vector{});
 }
@@ -495,7 +495,7 @@ _ForwardIterator3
 __pattern_walk3(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
                 _ForwardIterator2 __first2, _ForwardIterator3 __first3, _Function __f) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag>);
+    static_assert(__is_serial_tag_v<_Tag>);
 
     return __internal::__brick_walk3(__first1, __last1, __first2, __first3, __f, typename _Tag::__is_vector{});
 }
@@ -561,7 +561,7 @@ _ForwardIterator2
 __pattern_walk2_transform_if(_Tag __tag, _ExecutionPolicy&& __exec, _ForwardIterator1 __first1,
                              _ForwardIterator1 __last1, _ForwardIterator2 __first2, _Function __func) noexcept
 {
-    static_assert(__is_host_backend_tag_v<_Tag>);
+    static_assert(__is_host_dispatch_tag_v<_Tag>);
 
     return __pattern_walk2(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __func);
 }
@@ -573,7 +573,7 @@ __pattern_walk3_transform_if(_Tag __tag, _ExecutionPolicy&& __exec, _ForwardIter
                              _ForwardIterator1 __last1, _ForwardIterator2 __first2, _ForwardIterator3 __first3,
                              _Function __func) noexcept
 {
-    static_assert(__is_host_backend_tag_v<_Tag>);
+    static_assert(__is_host_dispatch_tag_v<_Tag>);
 
     return __pattern_walk3(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __first3,
                            __func);
@@ -608,7 +608,7 @@ bool
 __pattern_equal(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
                 _ForwardIterator2 __first2, _ForwardIterator2 __last2, _BinaryPredicate __p) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_serial_tag_v<_Tag> || __is_parallel_forward_tag_v<_Tag>);
 
     return __internal::__brick_equal(__first1, __last1, __first2, __last2, __p, typename _Tag::__is_vector{});
 }
@@ -659,7 +659,7 @@ bool
 __pattern_equal(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
                 _ForwardIterator2 __first2, _BinaryPredicate __p) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_serial_tag_v<_Tag> || __is_parallel_forward_tag_v<_Tag>);
 
     return __internal::__brick_equal(__first1, __last1, __first2, __p, typename _Tag::__is_vector{});
 }
@@ -706,7 +706,7 @@ _ForwardIterator
 __pattern_find_if(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                   _Predicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_serial_tag_v<_Tag> || __is_parallel_forward_tag_v<_Tag>);
 
     return __internal::__brick_find_if(__first, __last, __pred, typename _Tag::__is_vector{});
 }
@@ -849,7 +849,7 @@ _ForwardIterator1
 __pattern_find_end(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first, _ForwardIterator1 __last,
                    _ForwardIterator2 __s_first, _ForwardIterator2 __s_last, _BinaryPredicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_serial_tag_v<_Tag> || __is_parallel_forward_tag_v<_Tag>);
 
     return __internal::__brick_find_end(__first, __last, __s_first, __s_last, __pred, typename _Tag::__is_vector{});
 }
@@ -905,7 +905,7 @@ _ForwardIterator1
 __pattern_find_first_of(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first, _ForwardIterator1 __last,
                         _ForwardIterator2 __s_first, _ForwardIterator2 __s_last, _BinaryPredicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_serial_tag_v<_Tag> || __is_parallel_forward_tag_v<_Tag>);
 
     return __internal::__brick_find_first_of(__first, __last, __s_first, __s_last, __pred,
                                              typename _Tag::__is_vector{});
@@ -952,7 +952,7 @@ _ForwardIterator1
 __pattern_search(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first, _ForwardIterator1 __last,
                  _ForwardIterator2 __s_first, _ForwardIterator2 __s_last, _BinaryPredicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_serial_tag_v<_Tag> || __is_parallel_forward_tag_v<_Tag>);
 
     return __internal::__brick_search(__first, __last, __s_first, __s_last, __pred, typename _Tag::__is_vector{});
 }
@@ -1008,7 +1008,7 @@ _ForwardIterator
 __pattern_search_n(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _Size __count,
                    const _Tp& __value, _BinaryPredicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_serial_tag_v<_Tag> || __is_parallel_forward_tag_v<_Tag>);
 
     return __internal::__brick_search_n(__first, __last, __count, __value, __pred, typename _Tag::__is_vector{});
 }
@@ -1046,11 +1046,11 @@ __pattern_search_n(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _
 // clear that doing so is worth the trouble and extra layers of call chain.
 // Sometimes a little duplication for sake of regularity is better than the alternative.
 
-//template <class _Tag, typename _ExecutionPolicy, ::std::enable_if_t<oneapi::dpl::__internal::__is_host_backend_tag_v<_Tag>>>
+//template <class _Tag, typename _ExecutionPolicy, ::std::enable_if_t<oneapi::dpl::__internal::__is_host_dispatch_tag_v<_Tag>>>
 //template <class _Tag, typename _ExecutionPolicy, typename _D> // +
 //template <class _Tag, typename _ExecutionPolicy, typename = void>
 template <class _Tag, typename _ExecutionPolicy>
-struct __brick_copy_n<_Tag, _ExecutionPolicy, ::std::enable_if_t<oneapi::dpl::__internal::__is_host_backend_tag_v<_Tag>>>
+struct __brick_copy_n<_Tag, _ExecutionPolicy, ::std::enable_if_t<oneapi::dpl::__internal::__is_host_dispatch_tag_v<_Tag>>>
 {
     template <typename _RandomAccessIterator1, typename _Size, typename _RandomAccessIterator2>
     _RandomAccessIterator2
@@ -1075,7 +1075,7 @@ struct __brick_copy_n<_Tag, _ExecutionPolicy, ::std::enable_if_t<oneapi::dpl::__
 //------------------------------------------------------------------------
 
 template <class _Tag, typename _ExecutionPolicy>
-struct __brick_copy<_Tag, _ExecutionPolicy, ::std::enable_if_t<__is_host_backend_tag_v<_Tag>>>
+struct __brick_copy<_Tag, _ExecutionPolicy, ::std::enable_if_t<__is_host_dispatch_tag_v<_Tag>>>
 {
     template <typename _RandomAccessIterator1, typename _RandomAccessIterator2>
     _RandomAccessIterator2
@@ -1107,7 +1107,7 @@ struct __brick_copy<_Tag, _ExecutionPolicy, ::std::enable_if_t<__is_host_backend
 //------------------------------------------------------------------------
 
 template <class _Tag, typename _ExecutionPolicy>
-struct __brick_move<_Tag, _ExecutionPolicy, ::std::enable_if_t<__is_host_backend_tag_v<_Tag>>>
+struct __brick_move<_Tag, _ExecutionPolicy, ::std::enable_if_t<__is_host_dispatch_tag_v<_Tag>>>
 {
     template <typename _RandomAccessIterator1, typename _RandomAccessIterator2>
     _RandomAccessIterator2
@@ -1134,7 +1134,7 @@ struct __brick_move<_Tag, _ExecutionPolicy, ::std::enable_if_t<__is_host_backend
     }
 };
 
-template <class _Tag, typename _ExecutionPolicy, typename = ::std::enable_if_t<__is_host_backend_tag_v<_Tag>>>
+template <class _Tag, typename _ExecutionPolicy, typename = ::std::enable_if_t<__is_host_dispatch_tag_v<_Tag>>>
 struct __brick_move_destroy
 {
     template <typename _RandomAccessIterator1, typename _RandomAccessIterator2>
@@ -1308,7 +1308,7 @@ _OutputIterator
 __pattern_copy_if(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _OutputIterator __result,
                   _UnaryPredicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_serial_tag_v<_Tag> || __is_parallel_forward_tag_v<_Tag>);
 
     return __internal::__brick_copy_if(__first, __last, __result, __pred, typename _Tag::__is_vector{});
 }
@@ -1373,7 +1373,7 @@ template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Pre
 typename ::std::iterator_traits<_ForwardIterator>::difference_type
 __pattern_count(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _Predicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_serial_tag_v<_Tag> || __is_parallel_forward_tag_v<_Tag>);
 
     return __internal::__brick_count(__first, __last, __pred, typename _Tag::__is_vector{});
 }
@@ -1427,7 +1427,7 @@ _ForwardIterator
 __pattern_unique(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                  _BinaryPredicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_serial_tag_v<_Tag> || __is_parallel_forward_tag_v<_Tag>);
 
     return __internal::__brick_unique(__first, __last, __pred, typename _Tag::__is_vector{});
 }
@@ -1570,7 +1570,7 @@ _OutputIterator
 __pattern_unique_copy(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                       _OutputIterator __result, _BinaryPredicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_serial_tag_v<_Tag> || __is_parallel_forward_tag_v<_Tag>);
 
     return __internal::__brick_unique_copy(__first, __last, __result, __pred, typename _Tag::__is_vector{});
 }
@@ -1706,7 +1706,7 @@ template <class _Tag, class _ExecutionPolicy, class _BidirectionalIterator>
 void
 __pattern_reverse(_Tag, _ExecutionPolicy&&, _BidirectionalIterator __first, _BidirectionalIterator __last) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_serial_tag_v<_Tag> || __is_parallel_forward_tag_v<_Tag>);
 
     __internal::__brick_reverse(__first, __last, typename _Tag::__is_vector{});
 }
@@ -1754,7 +1754,7 @@ _OutputIterator
 __pattern_reverse_copy(_Tag, _ExecutionPolicy&&, _BidirectionalIterator __first, _BidirectionalIterator __last,
                        _OutputIterator __d_first) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_serial_tag_v<_Tag> || __is_parallel_forward_tag_v<_Tag>);
 
     return __internal::__brick_reverse_copy(__first, __last, __d_first, typename _Tag::__is_vector{});
 }
@@ -1833,7 +1833,7 @@ _ForwardIterator
 __pattern_rotate(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __middle,
                  _ForwardIterator __last) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_serial_tag_v<_Tag> || __is_parallel_forward_tag_v<_Tag>);
 
     return __internal::__brick_rotate(__first, __middle, __last, typename _Tag::__is_vector{});
 }
@@ -1911,7 +1911,7 @@ _OutputIterator
 __brick_rotate_copy(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __middle,
                     _ForwardIterator __last, _OutputIterator __result) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_serial_tag_v<_Tag> || __is_parallel_forward_tag_v<_Tag>);
 
     return ::std::rotate_copy(__first, __middle, __last, __result);
 }
@@ -1932,7 +1932,7 @@ _OutputIterator
 __pattern_rotate_copy(_Tag __tag, _ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __middle,
                       _ForwardIterator __last, _OutputIterator __result) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_serial_tag_v<_Tag> || __is_parallel_forward_tag_v<_Tag>);
 
     return __internal::__brick_rotate_copy(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __middle, __last,
                                            __result);
@@ -2014,7 +2014,7 @@ bool
 __pattern_is_partitioned(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                          _UnaryPredicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_serial_tag_v<_Tag> || __is_parallel_forward_tag_v<_Tag>);
 
     return __internal::__brick_is_partitioned(__first, __last, __pred, typename _Tag::__is_vector{});
 }
@@ -2144,7 +2144,7 @@ _ForwardIterator
 __pattern_partition(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                     _UnaryPredicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_serial_tag_v<_Tag> || __is_parallel_forward_tag_v<_Tag>);
 
     return __internal::__brick_partition(__first, __last, __pred, typename _Tag::__is_vector{});
 }
@@ -2244,7 +2244,7 @@ _BidirectionalIterator
 __pattern_stable_partition(_Tag, _ExecutionPolicy&&, _BidirectionalIterator __first, _BidirectionalIterator __last,
                            _UnaryPredicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_serial_tag_v<_Tag> || __is_parallel_forward_tag_v<_Tag>);
 
     return __internal::__brick_stable_partition(__first, __last, __pred, typename _Tag::__is_vector{});
 }
@@ -2334,7 +2334,7 @@ ::std::pair<_OutputIterator1, _OutputIterator2>
 __pattern_partition_copy(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                          _OutputIterator1 __out_true, _OutputIterator2 __out_false, _UnaryPredicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_serial_tag_v<_Tag> || __is_parallel_forward_tag_v<_Tag>);
 
     return __internal::__brick_partition_copy(__first, __last, __out_true, __out_false, __pred,
                                               typename _Tag::__is_vector{});
@@ -2390,7 +2390,7 @@ void
 __pattern_sort(_Tag, _ExecutionPolicy&&, _RandomAccessIterator __first, _RandomAccessIterator __last, _Compare __comp,
                _IsMoveConstructible) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_serial_tag_v<_Tag> || __is_parallel_forward_tag_v<_Tag>);
 
     ::std::sort(__first, __last, __comp);
 }
@@ -2422,7 +2422,7 @@ void
 __pattern_stable_sort(_Tag, _ExecutionPolicy&&, _RandomAccessIterator __first, _RandomAccessIterator __last,
                       _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_serial_tag_v<_Tag> || __is_parallel_forward_tag_v<_Tag>);
 
     ::std::stable_sort(__first, __last, __comp);
 }
@@ -2455,7 +2455,7 @@ __pattern_sort_by_key(_Tag, _ExecutionPolicy&& __exec, _RandomAccessIterator1 __
                       _RandomAccessIterator1 __keys_last, _RandomAccessIterator2 __values_first,
                       _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_serial_tag_v<_Tag> || __is_parallel_forward_tag_v<_Tag>);
 
     auto __beg = oneapi::dpl::make_zip_iterator(__keys_first, __values_first);
     auto __end = __beg + (__keys_last - __keys_first);
@@ -2501,7 +2501,7 @@ void
 __pattern_partial_sort(_Tag, _ExecutionPolicy&&, _RandomAccessIterator __first, _RandomAccessIterator __middle,
                        _RandomAccessIterator __last, _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_serial_tag_v<_Tag> || __is_parallel_forward_tag_v<_Tag>);
 
     ::std::partial_sort(__first, __middle, __last, __comp);
 }
@@ -2539,7 +2539,7 @@ _RandomAccessIterator
 __pattern_partial_sort_copy(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                             _RandomAccessIterator __d_first, _RandomAccessIterator __d_last, _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_serial_tag_v<_Tag> || __is_parallel_forward_tag_v<_Tag>);
 
     return ::std::partial_sort_copy(__first, __last, __d_first, __d_last, __comp);
 }
@@ -2643,7 +2643,7 @@ _ForwardIterator
 __pattern_adjacent_find(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                         _BinaryPredicate __pred, _Semantic) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_serial_tag_v<_Tag> || __is_parallel_forward_tag_v<_Tag>);
 
     return __internal::__brick_adjacent_find(__first, __last, __pred, typename _Tag::__is_vector{}, _Semantic::value);
 }
@@ -2703,7 +2703,7 @@ void
 __pattern_nth_element(_Tag, _ExecutionPolicy&&, _RandomAccessIterator __first, _RandomAccessIterator __nth,
                       _RandomAccessIterator __last, _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_serial_tag_v<_Tag> || __is_parallel_forward_tag_v<_Tag>);
 
     ::std::nth_element(__first, __nth, __last, __comp);
 }
@@ -2753,7 +2753,7 @@ __pattern_nth_element(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec
 // fill, fill_n
 //------------------------------------------------------------------------
 template <class _Tag, typename _ExecutionPolicy, typename _Tp>
-struct __brick_fill<_Tag, _ExecutionPolicy, _Tp, ::std::enable_if_t<__is_host_backend_tag_v<_Tag>>>
+struct __brick_fill<_Tag, _ExecutionPolicy, _Tp, ::std::enable_if_t<__is_host_dispatch_tag_v<_Tag>>>
 {
     const _Tp& __value;
 
@@ -2778,7 +2778,7 @@ template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Tp>
 void
 __pattern_fill(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, const _Tp& __value) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_serial_tag_v<_Tag> || __is_parallel_forward_tag_v<_Tag>);
 
     __internal::__brick_fill<_Tag, _ExecutionPolicy, _Tp>{__value}(__first, __last, typename _Tag::__is_vector{});
 }
@@ -2801,7 +2801,7 @@ __pattern_fill(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAcce
 }
 
 template <class _Tag, typename _ExecutionPolicy, typename _Tp>
-struct __brick_fill_n<_Tag, _ExecutionPolicy, _Tp, ::std::enable_if_t<__is_host_backend_tag_v<_Tag>>>
+struct __brick_fill_n<_Tag, _ExecutionPolicy, _Tp, ::std::enable_if_t<__is_host_dispatch_tag_v<_Tag>>>
 {
     const _Tp& __value;
 
@@ -2826,7 +2826,7 @@ template <class _Tag, class _ExecutionPolicy, class _OutputIterator, class _Size
 _OutputIterator
 __pattern_fill_n(_Tag, _ExecutionPolicy&&, _OutputIterator __first, _Size __count, const _Tp& __value) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_serial_tag_v<_Tag> || __is_parallel_forward_tag_v<_Tag>);
 
     return __internal::__brick_fill_n<_Tag, _ExecutionPolicy, _Tp>{__value}(__first, __count,
                                                                             typename _Tag::__is_vector{});
@@ -2864,7 +2864,7 @@ template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Gen
 void
 __pattern_generate(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _Generator __g) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_serial_tag_v<_Tag> || __is_parallel_forward_tag_v<_Tag>);
 
     __internal::__brick_generate(__first, __last, __g, typename _Tag::__is_vector{});
 }
@@ -2904,7 +2904,7 @@ template <class _Tag, class _ExecutionPolicy, class _OutputIterator, class _Size
 _OutputIterator
 __pattern_generate_n(_Tag, _ExecutionPolicy&&, _OutputIterator __first, _Size __count, _Generator __g) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_serial_tag_v<_Tag> || __is_parallel_forward_tag_v<_Tag>);
 
     return __internal::__brick_generate_n(__first, __count, __g, typename _Tag::__is_vector{});
 }
@@ -2949,7 +2949,7 @@ _ForwardIterator
 __pattern_remove_if(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                     _UnaryPredicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_serial_tag_v<_Tag> || __is_parallel_forward_tag_v<_Tag>);
 
     return __internal::__brick_remove_if(__first, __last, __pred, typename _Tag::__is_vector{});
 }
@@ -3005,7 +3005,7 @@ __pattern_merge(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIt
                 _ForwardIterator2 __first2, _ForwardIterator2 __last2, _OutputIterator __d_first,
                 _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_serial_tag_v<_Tag> || __is_parallel_forward_tag_v<_Tag>);
 
     return __internal::__brick_merge(__first1, __last1, __first2, __last2, __d_first, __comp,
                                      typename _Tag::__is_vector{});
@@ -3054,7 +3054,7 @@ void
 __pattern_inplace_merge(_Tag, _ExecutionPolicy&&, _BidirectionalIterator __first, _BidirectionalIterator __middle,
                         _BidirectionalIterator __last, _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_serial_tag_v<_Tag> || __is_parallel_forward_tag_v<_Tag>);
 
     __internal::__brick_inplace_merge(__first, __middle, __last, __comp, typename _Tag::__is_vector{});
 }
@@ -3113,7 +3113,7 @@ bool
 __pattern_includes(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
                    _ForwardIterator2 __first2, _ForwardIterator2 __last2, _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_serial_tag_v<_Tag> || __is_parallel_forward_tag_v<_Tag>);
 
     return ::std::includes(__first1, __last1, __first2, __last2, __comp);
 }
@@ -3428,7 +3428,7 @@ __pattern_set_union(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _Forwa
                     _ForwardIterator2 __first2, _ForwardIterator2 __last2, _OutputIterator __result,
                     _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_serial_tag_v<_Tag> || __is_parallel_forward_tag_v<_Tag>);
 
     return __internal::__brick_set_union(__first1, __last1, __first2, __last2, __result, __comp,
                                          typename _Tag::__is_vector{});
@@ -3489,7 +3489,7 @@ __pattern_set_intersection(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1,
                            _ForwardIterator2 __first2, _ForwardIterator2 __last2, _OutputIterator __result,
                            _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_serial_tag_v<_Tag> || __is_parallel_forward_tag_v<_Tag>);
 
     return __internal::__brick_set_intersection(__first1, __last1, __first2, __last2, __result, __comp,
                                                 typename _Tag::__is_vector{});
@@ -3587,7 +3587,7 @@ __pattern_set_difference(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _
                          _ForwardIterator2 __first2, _ForwardIterator2 __last2, _OutputIterator __result,
                          _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_serial_tag_v<_Tag> || __is_parallel_forward_tag_v<_Tag>);
 
     return __internal::__brick_set_difference(__first1, __last1, __first2, __last2, __result, __comp,
                                               typename _Tag::__is_vector{});
@@ -3674,7 +3674,7 @@ __pattern_set_symmetric_difference(_Tag, _ExecutionPolicy&&, _ForwardIterator1 _
                                    _ForwardIterator2 __first2, _ForwardIterator2 __last2, _OutputIterator __result,
                                    _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_serial_tag_v<_Tag> || __is_parallel_forward_tag_v<_Tag>);
 
     return __internal::__brick_set_symmetric_difference(__first1, __last1, __first2, __last2, __result, __comp,
                                                         typename _Tag::__is_vector{});
@@ -3733,7 +3733,7 @@ _RandomAccessIterator
 __pattern_is_heap_until(_Tag, _ExecutionPolicy&&, _RandomAccessIterator __first, _RandomAccessIterator __last,
                         _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_serial_tag_v<_Tag> || __is_parallel_forward_tag_v<_Tag>);
 
     return __internal::__brick_is_heap_until(__first, __last, __comp, typename _Tag::__is_vector{});
 }
@@ -3821,7 +3821,7 @@ bool
 __pattern_is_heap(_Tag, _ExecutionPolicy&&, _RandomAccessIterator __first, _RandomAccessIterator __last,
                   _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_serial_tag_v<_Tag> || __is_parallel_forward_tag_v<_Tag>);
 
     return __internal::__brick_is_heap(__first, __last, __comp, typename _Tag::__is_vector{});
 }
@@ -3869,7 +3869,7 @@ _ForwardIterator
 __pattern_min_element(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                       _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_serial_tag_v<_Tag> || __is_parallel_forward_tag_v<_Tag>);
 
     return __internal::__brick_min_element(__first, __last, __comp, typename _Tag::__is_vector{});
 }
@@ -3936,7 +3936,7 @@ ::std::pair<_ForwardIterator, _ForwardIterator>
 __pattern_minmax_element(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                          _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_serial_tag_v<_Tag> || __is_parallel_forward_tag_v<_Tag>);
 
     return __internal::__brick_minmax_element(__first, __last, __comp, typename _Tag::__is_vector{});
 }
@@ -4017,7 +4017,7 @@ ::std::pair<_ForwardIterator1, _ForwardIterator2>
 __pattern_mismatch(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
                    _ForwardIterator2 __first2, _ForwardIterator2 __last2, _Predicate __pred) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_serial_tag_v<_Tag> || __is_parallel_forward_tag_v<_Tag>);
 
     return __internal::__brick_mismatch(__first1, __last1, __first2, __last2, __pred, typename _Tag::__is_vector{});
 }
@@ -4098,7 +4098,7 @@ bool
 __pattern_lexicographical_compare(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
                                   _ForwardIterator2 __first2, _ForwardIterator2 __last2, _Compare __comp) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_serial_tag_v<_Tag> || __is_parallel_forward_tag_v<_Tag>);
 
     return __internal::__brick_lexicographical_compare(__first1, __last1, __first2, __last2, __comp,
                                                        typename _Tag::__is_vector{});
@@ -4160,7 +4160,7 @@ _ForwardIterator2
 __pattern_swap(_Tag __tag, _ExecutionPolicy&& __exec, _ForwardIterator1 __first1, _ForwardIterator1 __last1,
                _ForwardIterator2 __first2, _Function __f)
 {
-    static_assert(__is_host_backend_tag_v<_Tag>);
+    static_assert(__is_host_dispatch_tag_v<_Tag>);
 
     return __pattern_walk2(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __f);
 }
@@ -4238,7 +4238,7 @@ _ForwardIterator
 __pattern_shift_left(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                      typename ::std::iterator_traits<_ForwardIterator>::difference_type __n) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_serial_tag_v<_Tag> || __is_parallel_forward_tag_v<_Tag>);
 
     return __brick_shift_left(__first, __last, __n, typename _Tag::__is_vector{});
 }
@@ -4294,7 +4294,7 @@ __pattern_shift_right(_Tag __tag, _ExecutionPolicy&& __exec, _BidirectionalItera
                       _BidirectionalIterator __last,
                       typename ::std::iterator_traits<_BidirectionalIterator>::difference_type __n)
 {
-    static_assert(__is_host_backend_tag_v<_Tag>);
+    static_assert(__is_host_dispatch_tag_v<_Tag>);
 
     using _ReverseIterator = typename ::std::reverse_iterator<_BidirectionalIterator>;
 
diff --git a/include/oneapi/dpl/pstl/execution_impl.h b/include/oneapi/dpl/pstl/execution_impl.h
index b99114ebfc1..d2f277437fa 100644
--- a/include/oneapi/dpl/pstl/execution_impl.h
+++ b/include/oneapi/dpl/pstl/execution_impl.h
@@ -98,63 +98,63 @@ __select_backend(oneapi::dpl::execution::parallel_unsequenced_policy, _IteratorT
 }
 
 //----------------------------------------------------------
-// __is_backend_tag_serial, __is_backend_tag_serial_v
+// __is_serial_tag, __is_serial_tag_v
 //----------------------------------------------------------
 
 template <class _Tag>
-struct __is_backend_tag_serial : ::std::false_type
+struct __is_serial_tag : ::std::false_type
 {
 };
 
 template <class _IsVector>
-struct __is_backend_tag_serial<__serial_tag<_IsVector>> : ::std::true_type
+struct __is_serial_tag<__serial_tag<_IsVector>> : ::std::true_type
 {
 };
 
 template <class _Tag>
-inline constexpr bool __is_backend_tag_serial_v = __is_backend_tag_serial<_Tag>::value;
+inline constexpr bool __is_serial_tag_v = __is_serial_tag<_Tag>::value;
 
 //----------------------------------------------------------
-// __is_backend_tag_parallel_forward, __is_backend_tag_parallel_forward_v
+// __is_parallel_forward_tag, __is_parallel_forward_tag_v
 //----------------------------------------------------------
 
 template <class _Tag>
-struct __is_backend_tag_parallel_forward : ::std::false_type
+struct __is_parallel_forward_tag : ::std::false_type
 {
 };
 
 template <>
-struct __is_backend_tag_parallel_forward<__parallel_forward_tag> : ::std::true_type
+struct __is_parallel_forward_tag<__parallel_forward_tag> : ::std::true_type
 {
 };
 
 template <class _Tag>
-inline constexpr bool __is_backend_tag_parallel_forward_v = __is_backend_tag_parallel_forward<_Tag>::value;
+inline constexpr bool __is_parallel_forward_tag_v = __is_parallel_forward_tag<_Tag>::value;
 
 //----------------------------------------------------------
-// __is_backend_tag_parallel, __is_backend_tag_parallel_v
+// __is_parallel_tag, __is_parallel_tag_v
 //----------------------------------------------------------
 
 template <class _Tag>
-struct __is_backend_tag_parallel : ::std::false_type
+struct __is_parallel_tag : ::std::false_type
 {
 };
 
 template <class _IsVector>
-struct __is_backend_tag_parallel<__parallel_tag<_IsVector>> : ::std::true_type
+struct __is_parallel_tag<__parallel_tag<_IsVector>> : ::std::true_type
 {
 };
 
 template <class _Tag>
-inline constexpr bool __is_backend_tag_parallel_v = __is_backend_tag_parallel<_Tag>::value;
+inline constexpr bool __is_parallel_tag_v = __is_parallel_tag<_Tag>::value;
 
 //----------------------------------------------------------
-// __is_host_backend_tag_v
+// __is_host_dispatch_tag_v
 //----------------------------------------------------------
 
 template <class _Tag>
-inline constexpr bool __is_host_backend_tag_v =
-    __is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag> || __is_backend_tag_parallel_v<_Tag>;
+inline constexpr bool __is_host_dispatch_tag_v =
+    __is_serial_tag_v<_Tag> || __is_parallel_forward_tag_v<_Tag> || __is_parallel_tag_v<_Tag>;
 
 } // namespace __internal
 } // namespace dpl
diff --git a/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h b/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
index a6a6952738a..84888832193 100644
--- a/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
+++ b/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
@@ -167,7 +167,7 @@ void
 __pattern_for_loop_n(_Tag, _ExecutionPolicy&&, _Ip __first, _Size __n, _Function __f, __single_stride_type,
                      _Rest&&... __rest) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_serial_tag_v<_Tag> || __is_parallel_forward_tag_v<_Tag>);
 
     __reduction_pack<_Rest...> __pack{__reduction_pack_tag(), ::std::forward<_Rest>(__rest)...};
 
@@ -192,7 +192,7 @@ void
 __pattern_for_loop_n(_Tag, _ExecutionPolicy&&, _Ip __first, _Size __n, _Function __f, _Sp __stride,
                      _Rest&&... __rest) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_serial_tag_v<_Tag> || __is_parallel_forward_tag_v<_Tag>);
 
     __reduction_pack<_Rest...> __pack{__reduction_pack_tag(), ::std::forward<_Rest>(__rest)...};
 
@@ -265,7 +265,7 @@ void
 __pattern_for_loop(_Tag __tag, _ExecutionPolicy&& __exec, _Ip __first, _Ip __last, _Function __f, _Sp __stride,
                    _Rest&&... __rest) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_serial_tag_v<_Tag> || __is_parallel_forward_tag_v<_Tag>);
 
     if constexpr (typename _Tag::__is_vector{} || __is_random_access_or_integral_v<_Ip>)
     {
@@ -366,7 +366,7 @@ void
 __pattern_for_loop(_Tag, _ExecutionPolicy&& __exec, _Ip __first, _Ip __last, _Function __f, __single_stride_type,
                    _Rest&&... __rest) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_serial_tag_v<_Tag> || __is_parallel_forward_tag_v<_Tag>);
 
     __reduction_pack<_Rest...> __pack{__reduction_pack_tag(), ::std::forward<_Rest>(__rest)...};
 
diff --git a/include/oneapi/dpl/pstl/histogram_impl.h b/include/oneapi/dpl/pstl/histogram_impl.h
index 885600f0288..362685bd19c 100644
--- a/include/oneapi/dpl/pstl/histogram_impl.h
+++ b/include/oneapi/dpl/pstl/histogram_impl.h
@@ -38,7 +38,7 @@ void
 __pattern_histogram(_Tag, _ExecutionPolicy&& exec, _RandomAccessIterator1 __first, _RandomAccessIterator1 __last,
                     _Size __num_bins, _IdxHashFunc __func, _RandomAccessIterator2 __histogram_first)
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_serial_tag_v<_Tag> || __is_parallel_forward_tag_v<_Tag>);
 
     static_assert(sizeof(_Size) == 0 /*false*/,
                   "Histogram API is currently unsupported for policies other than device execution policies");
diff --git a/include/oneapi/dpl/pstl/numeric_impl.h b/include/oneapi/dpl/pstl/numeric_impl.h
index 2b8102dfe7a..b87a02ba428 100644
--- a/include/oneapi/dpl/pstl/numeric_impl.h
+++ b/include/oneapi/dpl/pstl/numeric_impl.h
@@ -68,7 +68,7 @@ __pattern_transform_reduce(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first1,
                            _ForwardIterator2 __first2, _Tp __init, _BinaryOperation1 __binary_op1,
                            _BinaryOperation2 __binary_op2) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_serial_tag_v<_Tag> || __is_parallel_forward_tag_v<_Tag>);
 
     return __brick_transform_reduce(__first1, __last1, __first2, __init, __binary_op1, __binary_op2,
                                     typename _Tag::__is_vector{});
@@ -133,7 +133,7 @@ _Tp
 __pattern_transform_reduce(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last, _Tp __init,
                            _BinaryOperation __binary_op, _UnaryOperation __unary_op) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_serial_tag_v<_Tag> || __is_parallel_forward_tag_v<_Tag>);
 
     return __internal::__brick_transform_reduce(__first, __last, __init, __binary_op, __unary_op,
                                                 typename _Tag::__is_vector{});
@@ -243,7 +243,7 @@ __pattern_transform_scan(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _Fo
                          _OutputIterator __result, _UnaryOperation __unary_op, _Tp __init, _BinaryOperation __binary_op,
                          _Inclusive) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_serial_tag_v<_Tag> || __is_parallel_forward_tag_v<_Tag>);
 
     return __internal::__brick_transform_scan(__first, __last, __result, __unary_op, __init, __binary_op, _Inclusive(),
                                               typename _Tag::__is_vector{})
@@ -327,7 +327,7 @@ _OutputIterator
 __pattern_transform_scan(_Tag __tag, _ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last,
                          _OutputIterator __result, _UnaryOperation __unary_op, _BinaryOperation __binary_op, _Inclusive)
 {
-    static_assert(__is_host_backend_tag_v<_Tag>);
+    static_assert(__is_host_dispatch_tag_v<_Tag>);
 
     typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _ValueType;
     if (__first != __last)
@@ -379,7 +379,7 @@ _OutputIterator
 __pattern_adjacent_difference(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIterator __last,
                               _OutputIterator __d_first, _BinaryOperation __op) noexcept
 {
-    static_assert(__is_backend_tag_serial_v<_Tag> || __is_backend_tag_parallel_forward_v<_Tag>);
+    static_assert(__is_serial_tag_v<_Tag> || __is_parallel_forward_tag_v<_Tag>);
 
     return __internal::__brick_adjacent_difference(__first, __last, __d_first, __op, typename _Tag::__is_vector{});
 }

From bd568cc8cc8272519a7f6d7e82898a95910d02f6 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 6 Mar 2024 12:38:07 +0100
Subject: [PATCH 524/566] Fix compile error: implicit instantiation of
 undefined template
 'oneapi::dpl::__par_backend_hetero::__internal::__buffer_impl<oneapi::dpl::__internal::__fpga_backend_tag,
 oneapi::dpl::execution::fpga_policy<>, int>'

/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h:1232:92: error: implicit instantiation of undefined template 'oneapi::dpl::__par_backend_hetero::__internal::__buffer_impl<oneapi::dpl::__internal::__fpga_backend_tag, oneapi::dpl::execution::fpga_policy<>, int>'
 1232 |     oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _ValueType> __buf(__exec, __n);
---
 .../pstl/hetero/dpcpp/execution_sycl_defs.h   | 24 +++++++++++++++++++
 .../dpcpp/parallel_backend_sycl_utils.h       |  2 +-
 2 files changed, 25 insertions(+), 1 deletion(-)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h b/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h
index 93d291b661b..65a0bdec9a6 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h
@@ -318,6 +318,30 @@ __select_backend(const execution::fpga_policy<_Factor, _KernelName>&, _IteratorT
 }
 #endif
 
+//----------------------------------------------------------
+// __is_hetero_backend_tag, __is_hetero_backend_tag_v
+//----------------------------------------------------------
+
+template <typename _BackendTag>
+struct __is_hetero_backend_tag : ::std::false_type
+{
+};
+
+template <>
+struct __is_hetero_backend_tag<__device_backend_tag> : ::std::true_type
+{
+};
+
+#if _ONEDPL_FPGA_DEVICE
+template <>
+struct __is_hetero_backend_tag<__fpga_backend_tag> : ::std::true_type
+{
+};
+#endif
+
+template <typename _BackendTag>
+inline constexpr bool __is_hetero_backend_tag_v = __is_hetero_backend_tag<_BackendTag>::value;
+
 } // namespace __internal
 
 } // namespace dpl
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_utils.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_utils.h
index 9319486f150..40b4b584c04 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_utils.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_utils.h
@@ -382,7 +382,7 @@ class __buffer_impl;
 // impl for sycl::buffer<...>
 template <typename _BackendTag, typename _ExecutionPolicy, typename _Tp>
 class __buffer_impl<_BackendTag, _ExecutionPolicy, _Tp,
-                    ::std::enable_if_t<::std::is_same_v<_BackendTag, oneapi::dpl::__internal::__device_backend_tag>>>
+                    ::std::enable_if_t<oneapi::dpl::__internal::__is_hetero_backend_tag_v<_BackendTag>>>
 {
   private:
     using __container_t = typename __local_buffer<sycl::buffer<_Tp>>::type;

From e4ad81f59250c3fb02ce1ca8a73bdc9a873aa770 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 6 Mar 2024 12:44:25 +0100
Subject: [PATCH 525/566] Apply GitHUB clang format

---
 include/oneapi/dpl/pstl/algorithm_impl.h               | 3 ++-
 include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h | 3 +--
 2 files changed, 3 insertions(+), 3 deletions(-)

diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 5fda420c564..5e4490301d2 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -1050,7 +1050,8 @@ __pattern_search_n(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _
 //template <class _Tag, typename _ExecutionPolicy, typename _D> // +
 //template <class _Tag, typename _ExecutionPolicy, typename = void>
 template <class _Tag, typename _ExecutionPolicy>
-struct __brick_copy_n<_Tag, _ExecutionPolicy, ::std::enable_if_t<oneapi::dpl::__internal::__is_host_dispatch_tag_v<_Tag>>>
+struct __brick_copy_n<_Tag, _ExecutionPolicy,
+                      ::std::enable_if_t<oneapi::dpl::__internal::__is_host_dispatch_tag_v<_Tag>>>
 {
     template <typename _RandomAccessIterator1, typename _Size, typename _RandomAccessIterator2>
     _RandomAccessIterator2
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 1de2a8f224d..66166a01f8b 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -1630,8 +1630,7 @@ __pattern_rotate(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Iterator
 
     auto __keep = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read_write, _Iterator>();
     auto __buf = __keep(__first, __last);
-    auto __temp_buf =
-        oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _Tp>(__exec, __n);
+    auto __temp_buf = oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _Tp>(__exec, __n);
 
     auto __temp_rng =
         oneapi::dpl::__ranges::all_view<_Tp, __par_backend_hetero::access_mode::write>(__temp_buf.get_buffer());

From aeefbcc16bd7b78fd57cbb843b4f567e880ea241 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 6 Mar 2024 15:38:06 +0100
Subject: [PATCH 526/566] Fix review comment: Why do we need __backend_tag type
 for the __serial_tag tag?

---
 .../internal/exclusive_scan_by_segment_impl.h |  6 +++---
 .../internal/inclusive_scan_by_segment_impl.h |  4 ++--
 .../dpl/internal/reduce_by_segment_impl.h     |  8 ++++----
 include/oneapi/dpl/pstl/execution_impl.h      | 19 +++++++++++++++++--
 4 files changed, 26 insertions(+), 11 deletions(-)

diff --git a/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h b/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h
index 39f6d5ebf70..30a9e9fd3a3 100644
--- a/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h
@@ -46,7 +46,7 @@ pattern_exclusive_scan_by_segment(_Tag, Policy&& policy, InputIterator1 first1,
 {
     static_assert(__internal::__is_host_dispatch_tag_v<_Tag>);
 
-    using __backend_tag = typename _Tag::__backend_tag;
+    using __buffer_backend_tag = typename __internal::__buffer_backend_tag_type<_Tag>::type;
 
     const auto n = ::std::distance(first1, last1);
 
@@ -66,7 +66,7 @@ pattern_exclusive_scan_by_segment(_Tag, Policy&& policy, InputIterator1 first1,
     InputIterator2 last2 = first2 + n;
 
     // compute head flags
-    oneapi::dpl::__par_backend::__buffer<__backend_tag, Policy, FlagType> _flags(policy, n);
+    oneapi::dpl::__par_backend::__buffer<__buffer_backend_tag, Policy, FlagType> _flags(policy, n);
     auto flags = _flags.get();
     flags[0] = 1;
 
@@ -74,7 +74,7 @@ pattern_exclusive_scan_by_segment(_Tag, Policy&& policy, InputIterator1 first1,
               oneapi::dpl::__internal::__not_pred<BinaryPredicate>(binary_pred));
 
     // shift input one to the right and initialize segments with init
-    oneapi::dpl::__par_backend::__buffer<__backend_tag, Policy, OutputType> _temp(policy, n);
+    oneapi::dpl::__par_backend::__buffer<__buffer_backend_tag, Policy, OutputType> _temp(policy, n);
     auto temp = _temp.get();
 
     temp[0] = init;
diff --git a/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h b/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h
index 36ef00a64f0..4c7405d4c5d 100644
--- a/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h
@@ -45,7 +45,7 @@ pattern_inclusive_scan_by_segment(_Tag, Policy&& policy, InputIterator1 first1,
 {
     static_assert(__internal::__is_host_dispatch_tag_v<_Tag>);
 
-    using __backend_tag = typename _Tag::__backend_tag;
+    using __buffer_backend_tag = typename __internal::__buffer_backend_tag_type<_Tag>::type;
 
     const auto n = ::std::distance(first1, last1);
 
@@ -61,7 +61,7 @@ pattern_inclusive_scan_by_segment(_Tag, Policy&& policy, InputIterator1 first1,
     typedef unsigned int FlagType;
     typedef typename ::std::iterator_traits<InputIterator2>::value_type ValueType;
 
-    oneapi::dpl::__par_backend::__buffer<__backend_tag, Policy, FlagType> _mask(policy, n);
+    oneapi::dpl::__par_backend::__buffer<__buffer_backend_tag, Policy, FlagType> _mask(policy, n);
     auto mask = _mask.get();
 
     mask[0] = 1;
diff --git a/include/oneapi/dpl/internal/reduce_by_segment_impl.h b/include/oneapi/dpl/internal/reduce_by_segment_impl.h
index 84ec72c7ee9..86252f6fe54 100644
--- a/include/oneapi/dpl/internal/reduce_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/reduce_by_segment_impl.h
@@ -87,7 +87,7 @@ reduce_by_segment_impl(_Tag, Policy&& policy, InputIterator1 first1, InputIterat
 {
     static_assert(__internal::__is_host_dispatch_tag_v<_Tag>);
 
-    using __backend_tag = typename _Tag::__backend_tag;
+    using __buffer_backend_tag = typename __internal::__buffer_backend_tag_type<_Tag>::type;
 
     // The algorithm reduces values in [first2, first2 + (last1-first1)) where the associated
     // keys for the values are equal to the adjacent key. This function's implementation is a derivative work
@@ -116,7 +116,7 @@ reduce_by_segment_impl(_Tag, Policy&& policy, InputIterator1 first1, InputIterat
 
     // buffer that is used to store a flag indicating if the associated key is not equal to
     // the next key, and thus its associated sum should be part of the final result
-    oneapi::dpl::__par_backend::__buffer<__backend_tag, Policy, FlagType> _mask(policy, n + 1);
+    oneapi::dpl::__par_backend::__buffer<__buffer_backend_tag, Policy, FlagType> _mask(policy, n + 1);
     auto mask = _mask.get();
     mask[0] = 1;
 
@@ -132,11 +132,11 @@ reduce_by_segment_impl(_Tag, Policy&& policy, InputIterator1 first1, InputIterat
     // buffer stores the sums of values associated with a given key. Sums are copied with
     // a shift into result2, and the shift is computed at the same time as the sums, so the
     // sums can't be written to result2 directly.
-    oneapi::dpl::__par_backend::__buffer<__backend_tag, Policy, ValueType> _scanned_values(policy, n);
+    oneapi::dpl::__par_backend::__buffer<__buffer_backend_tag, Policy, ValueType> _scanned_values(policy, n);
 
     // Buffer is used to store results of the scan of the mask. Values indicate which position
     // in result2 needs to be written with the scanned_values element.
-    oneapi::dpl::__par_backend::__buffer<__backend_tag, Policy, FlagType> _scanned_tail_flags(policy, n);
+    oneapi::dpl::__par_backend::__buffer<__buffer_backend_tag, Policy, FlagType> _scanned_tail_flags(policy, n);
 
     // Compute the sum of the segments. scanned_tail_flags values are not used.
     inclusive_scan(policy, make_zip_iterator(first2, _mask.get()), make_zip_iterator(first2, _mask.get()) + n,
diff --git a/include/oneapi/dpl/pstl/execution_impl.h b/include/oneapi/dpl/pstl/execution_impl.h
index d2f277437fa..485290209d6 100644
--- a/include/oneapi/dpl/pstl/execution_impl.h
+++ b/include/oneapi/dpl/pstl/execution_impl.h
@@ -48,7 +48,6 @@ template <class _IsVector>
 struct __serial_tag
 {
     using __is_vector = _IsVector;
-    using __backend_tag = __par_backend_tag;
 };
 
 template <class _IsVector>
@@ -152,10 +151,26 @@ inline constexpr bool __is_parallel_tag_v = __is_parallel_tag<_Tag>::value;
 // __is_host_dispatch_tag_v
 //----------------------------------------------------------
 
-template <class _Tag>
+template <class _Tag, typename = void>
 inline constexpr bool __is_host_dispatch_tag_v =
     __is_serial_tag_v<_Tag> || __is_parallel_forward_tag_v<_Tag> || __is_parallel_tag_v<_Tag>;
 
+//----------------------------------------------------------
+// __buffer_backend_tag_type
+//----------------------------------------------------------
+
+template <class _Tag, typename = void>
+struct __buffer_backend_tag_type
+{
+    using type = typename _Tag::__backend_tag;
+};
+
+template <class _Tag>
+struct __buffer_backend_tag_type<_Tag, ::std::enable_if_t<__is_serial_tag_v<_Tag>>>
+{
+    using type = __par_backend_tag;
+};
+
 } // namespace __internal
 } // namespace dpl
 } // namespace oneapi

From 80850dd4b7b395cbe4a2062c9feab8a7ceb74145 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 6 Mar 2024 15:48:37 +0100
Subject: [PATCH 527/566] Fix review comment: Why do we need __backend_tag type
 for the __serial_tag tag?

---
 include/oneapi/dpl/pstl/execution_impl.h | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/include/oneapi/dpl/pstl/execution_impl.h b/include/oneapi/dpl/pstl/execution_impl.h
index 485290209d6..468968292b8 100644
--- a/include/oneapi/dpl/pstl/execution_impl.h
+++ b/include/oneapi/dpl/pstl/execution_impl.h
@@ -159,14 +159,14 @@ inline constexpr bool __is_host_dispatch_tag_v =
 // __buffer_backend_tag_type
 //----------------------------------------------------------
 
-template <class _Tag, typename = void>
+template <class _Tag>
 struct __buffer_backend_tag_type
 {
     using type = typename _Tag::__backend_tag;
 };
 
-template <class _Tag>
-struct __buffer_backend_tag_type<_Tag, ::std::enable_if_t<__is_serial_tag_v<_Tag>>>
+template <class _IsVector>
+struct __buffer_backend_tag_type<__serial_tag<_IsVector>>
 {
     using type = __par_backend_tag;
 };

From 9b33ea7d9bcf535e2108739a6ea5496066a17118 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 6 Mar 2024 16:44:38 +0100
Subject: [PATCH 528/566] Some additional comments for __select_backend

---
 include/oneapi/dpl/pstl/execution_impl.h                   | 4 ++++
 include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h | 4 ++++
 include/oneapi/dpl/pstl/hetero/dpcpp/utils_ranges_sycl.h   | 4 ++++
 3 files changed, 12 insertions(+)

diff --git a/include/oneapi/dpl/pstl/execution_impl.h b/include/oneapi/dpl/pstl/execution_impl.h
index 468968292b8..dc653a3ec88 100644
--- a/include/oneapi/dpl/pstl/execution_impl.h
+++ b/include/oneapi/dpl/pstl/execution_impl.h
@@ -63,6 +63,10 @@ struct __parallel_forward_tag
     using __backend_tag = __par_backend_tag;
 };
 
+//----------------------------------------------------------
+// __select_backend (for the host policies)
+//----------------------------------------------------------
+
 template <class _IsVector, class... _IteratorTypes>
 using __tag_type = ::std::conditional_t<
     __internal::__is_random_access_iterator_v<_IteratorTypes...>, __parallel_tag<_IsVector>,
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h b/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h
index 65a0bdec9a6..e1e06e17e96 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/execution_sycl_defs.h
@@ -296,6 +296,10 @@ struct __device_backend_tag
 {
 };
 
+//----------------------------------------------------------
+// __select_backend (for the hetero policies)
+//----------------------------------------------------------
+
 template <class... _IteratorTypes, typename _KernelName>
 __hetero_tag<__device_backend_tag>
 __select_backend(const execution::device_policy<_KernelName>&, _IteratorTypes&&...)
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/utils_ranges_sycl.h b/include/oneapi/dpl/pstl/hetero/dpcpp/utils_ranges_sycl.h
index cdb43159d15..1821301e911 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/utils_ranges_sycl.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/utils_ranges_sycl.h
@@ -716,6 +716,10 @@ struct __get_sycl_range
     }
 };
 
+//----------------------------------------------------------
+// __select_backend (for the hetero policies)
+//----------------------------------------------------------
+
 //TODO required correct implementation of this __ranges::__select_backend()
 // 1. There is still not RA ranges checks
 // 2. Obviously, a return tag is not necessarily oneapi::dpl::__internal::__hetero_tag

From 1df098adef8d167fcaf47b11580faafb33782c63 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 6 Mar 2024 17:06:49 +0100
Subject: [PATCH 529/566] Fix review comment: Please leave the old semantics
 with a partial specialization. That means keeping enable_if but change the
 condition to check a tag type

---
 include/oneapi/dpl/pstl/algorithm_impl.h | 3 ---
 1 file changed, 3 deletions(-)

diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 5e4490301d2..c5a5aae939a 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -1046,9 +1046,6 @@ __pattern_search_n(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _
 // clear that doing so is worth the trouble and extra layers of call chain.
 // Sometimes a little duplication for sake of regularity is better than the alternative.
 
-//template <class _Tag, typename _ExecutionPolicy, ::std::enable_if_t<oneapi::dpl::__internal::__is_host_dispatch_tag_v<_Tag>>>
-//template <class _Tag, typename _ExecutionPolicy, typename _D> // +
-//template <class _Tag, typename _ExecutionPolicy, typename = void>
 template <class _Tag, typename _ExecutionPolicy>
 struct __brick_copy_n<_Tag, _ExecutionPolicy,
                       ::std::enable_if_t<oneapi::dpl::__internal::__is_host_dispatch_tag_v<_Tag>>>

From ea894931678d153036752f685574d087c2dfefde Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 7 Mar 2024 09:25:10 +0100
Subject: [PATCH 530/566] include/oneapi/dpl/pstl/glue_memory_impl.h - fix
 review comments in get_unvectorized_policy format

---
 include/oneapi/dpl/pstl/glue_memory_impl.h | 18 +++++++++---------
 1 file changed, 9 insertions(+), 9 deletions(-)

diff --git a/include/oneapi/dpl/pstl/glue_memory_impl.h b/include/oneapi/dpl/pstl/glue_memory_impl.h
index f0026f9842e..6c207024777 100644
--- a/include/oneapi/dpl/pstl/glue_memory_impl.h
+++ b/include/oneapi/dpl/pstl/glue_memory_impl.h
@@ -189,23 +189,23 @@ uninitialized_fill_n(_ExecutionPolicy&& __exec, _ForwardIterator __first, _Size
 
 #if (_PSTL_ICPX_OMP_SIMD_DESTROY_WINDOWS_BROKEN || _ONEDPL_ICPX_OMP_SIMD_DESTROY_WINDOWS_BROKEN)
 
-oneapi::dpl::execution::parallel_policy
-get_unvectorized_policy(oneapi::dpl::execution::parallel_unsequenced_policy)
+const oneapi::dpl::execution::parallel_policy&
+get_unvectorized_policy(const oneapi::dpl::execution::parallel_unsequenced_policy&)
 {
-    return {};
+    return oneapi::dpl::execution::par;
 }
 
-oneapi::dpl::execution::sequenced_policy
-get_unvectorized_policy(oneapi::dpl::execution::unsequenced_policy)
+const oneapi::dpl::execution::sequenced_policy&
+get_unvectorized_policy(const oneapi::dpl::execution::unsequenced_policy&)
 {
-    return {};
+    return oneapi::dpl::execution::seq;
 }
 
 template <typename _ExecutionPolicy>
-_ExecutionPolicy&&
-get_unvectorized_policy(_ExecutionPolicy&& __exec)
+const _ExecutionPolicy&
+get_unvectorized_policy(const _ExecutionPolicy& __exec)
 {
-    return ::std::forward<_ExecutionPolicy>(__exec);
+    return __exec;
 }
 
 #endif // (_PSTL_ICPX_OMP_SIMD_DESTROY_WINDOWS_BROKEN || _ONEDPL_ICPX_OMP_SIMD_DESTROY_WINDOWS_BROKEN)

From cab46e16d1654aa11835e634bca76105e8f9dab0 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 7 Mar 2024 18:21:31 +0100
Subject: [PATCH 531/566] 
 include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h - fix error:
 remove __pattern_for_loop for __parallel_tag<_IsVector> +
 __single_stride_type as not existing before

---
 .../dpl/pstl/experimental/internal/for_loop_impl.h  | 13 +------------
 1 file changed, 1 insertion(+), 12 deletions(-)

diff --git a/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h b/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
index 84888832193..d702a3a492e 100644
--- a/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
+++ b/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
@@ -366,7 +366,7 @@ void
 __pattern_for_loop(_Tag, _ExecutionPolicy&& __exec, _Ip __first, _Ip __last, _Function __f, __single_stride_type,
                    _Rest&&... __rest) noexcept
 {
-    static_assert(__is_serial_tag_v<_Tag> || __is_parallel_forward_tag_v<_Tag>);
+    static_assert(__is_host_dispatch_tag_v<_Tag>);
 
     __reduction_pack<_Rest...> __pack{__reduction_pack_tag(), ::std::forward<_Rest>(__rest)...};
 
@@ -424,17 +424,6 @@ __pattern_for_loop_n(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Ip _
     });
 }
 
-template <typename _IsVector, typename _ExecutionPolicy, typename _Ip, typename _Function, typename... _Rest>
-void
-__pattern_for_loop(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _Ip __first, _Ip __last, _Function __f,
-                   __single_stride_type, _Rest&&... __rest)
-{
-    oneapi::dpl::__internal::__pattern_for_loop_n(
-        __tag, ::std::forward<_ExecutionPolicy>(__exec), __first,
-        oneapi::dpl::__internal::__calculate_input_sequence_length(__first, __last, __single_stride_type{}), __f,
-        __single_stride_type{}, ::std::forward<_Rest>(__rest)...);
-}
-
 template <typename _IsVector, typename _ExecutionPolicy, typename _Ip, typename _Size, typename _Function, typename _Sp,
           typename... _Rest>
 void

From d2b82e730754c87e2cb3574da2b90c557a34a29c Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 7 Mar 2024 18:28:13 +0100
Subject: [PATCH 532/566] Revert
 "include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h - fix error:
 remove __pattern_for_loop for __parallel_tag<_IsVector> +
 __single_stride_type as not existing before"

This reverts commit cab46e16d1654aa11835e634bca76105e8f9dab0.
---
 .../dpl/pstl/experimental/internal/for_loop_impl.h  | 13 ++++++++++++-
 1 file changed, 12 insertions(+), 1 deletion(-)

diff --git a/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h b/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
index d702a3a492e..84888832193 100644
--- a/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
+++ b/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
@@ -366,7 +366,7 @@ void
 __pattern_for_loop(_Tag, _ExecutionPolicy&& __exec, _Ip __first, _Ip __last, _Function __f, __single_stride_type,
                    _Rest&&... __rest) noexcept
 {
-    static_assert(__is_host_dispatch_tag_v<_Tag>);
+    static_assert(__is_serial_tag_v<_Tag> || __is_parallel_forward_tag_v<_Tag>);
 
     __reduction_pack<_Rest...> __pack{__reduction_pack_tag(), ::std::forward<_Rest>(__rest)...};
 
@@ -424,6 +424,17 @@ __pattern_for_loop_n(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Ip _
     });
 }
 
+template <typename _IsVector, typename _ExecutionPolicy, typename _Ip, typename _Function, typename... _Rest>
+void
+__pattern_for_loop(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _Ip __first, _Ip __last, _Function __f,
+                   __single_stride_type, _Rest&&... __rest)
+{
+    oneapi::dpl::__internal::__pattern_for_loop_n(
+        __tag, ::std::forward<_ExecutionPolicy>(__exec), __first,
+        oneapi::dpl::__internal::__calculate_input_sequence_length(__first, __last, __single_stride_type{}), __f,
+        __single_stride_type{}, ::std::forward<_Rest>(__rest)...);
+}
+
 template <typename _IsVector, typename _ExecutionPolicy, typename _Ip, typename _Size, typename _Function, typename _Sp,
           typename... _Rest>
 void

From 19665c35a3512e11e41af3fd766d0a0bec576b87 Mon Sep 17 00:00:00 2001
From: Mikhail Dvorskiy <mikhail.dvorskiy@intel.com>
Date: Mon, 11 Mar 2024 20:21:27 +0100
Subject: [PATCH 533/566] [oneDPL][for_loop][tag] + minimal support tag
 dispatching for experimental 'for_loop'

---
 .../dpl/pstl/experimental/internal/for_loop.h |  21 +-
 .../experimental/internal/for_loop_impl.h     | 365 +++++++++---------
 2 files changed, 200 insertions(+), 186 deletions(-)

diff --git a/include/oneapi/dpl/pstl/experimental/internal/for_loop.h b/include/oneapi/dpl/pstl/experimental/internal/for_loop.h
index 46d05a9505b..4a61dd7c09b 100644
--- a/include/oneapi/dpl/pstl/experimental/internal/for_loop.h
+++ b/include/oneapi/dpl/pstl/experimental/internal/for_loop.h
@@ -42,14 +42,20 @@ struct type_identity
 template <typename _Tp>
 using type_identity_t = typename type_identity<_Tp>::type;
 
+//hide vector policy till the algorithms implementation
+// template <>
+// struct __policy_traits<oneapi::dpl::execution::vector_policy>
+// {
+//     typedef ::std::false_type __allow_parallel;
+//     typedef ::std::false_type __allow_unsequenced;
+//     typedef ::std::true_type __allow_vector;
+// };
+
 // TODO: add static asserts for parameters according to the requirements
 template <typename _ExecutionPolicy, typename _Ip, typename... _Rest>
 void
 for_loop(_ExecutionPolicy&& __exec, type_identity_t<_Ip> __start, _Ip __finish, _Rest&&... __rest)
 {
-    static_assert(oneapi::dpl::__internal::__is_host_execution_policy<::std::decay_t<_ExecutionPolicy>>::value,
-                  "for_loop is implemented for the host policies only");
-
     oneapi::dpl::__internal::__for_loop_repack(::std::forward<_ExecutionPolicy>(__exec), __start, __finish,
                                                oneapi::dpl::__internal::__single_stride_type{},
                                                ::std::forward_as_tuple(::std::forward<_Rest>(__rest)...));
@@ -59,9 +65,6 @@ template <typename _ExecutionPolicy, typename _Ip, typename _Sp, typename... _Re
 void
 for_loop_strided(_ExecutionPolicy&& __exec, type_identity_t<_Ip> __start, _Ip __finish, _Sp __stride, _Rest&&... __rest)
 {
-    static_assert(oneapi::dpl::__internal::__is_host_execution_policy<::std::decay_t<_ExecutionPolicy>>::value,
-                  "for_loop_strided is implemented for the host policies only");
-
     oneapi::dpl::__internal::__for_loop_repack(::std::forward<_ExecutionPolicy>(__exec), __start, __finish, __stride,
                                                ::std::forward_as_tuple(::std::forward<_Rest>(__rest)...));
 }
@@ -70,9 +73,6 @@ template <typename _ExecutionPolicy, typename _Ip, typename _Size, typename... _
 void
 for_loop_n(_ExecutionPolicy&& __exec, _Ip __start, _Size __n, _Rest&&... __rest)
 {
-    static_assert(oneapi::dpl::__internal::__is_host_execution_policy<::std::decay_t<_ExecutionPolicy>>::value,
-                  "for_loop_n is implemented for the host policies only");
-
     oneapi::dpl::__internal::__for_loop_repack_n(::std::forward<_ExecutionPolicy>(__exec), __start, __n,
                                                  oneapi::dpl::__internal::__single_stride_type{},
                                                  ::std::forward_as_tuple(::std::forward<_Rest>(__rest)...));
@@ -82,9 +82,6 @@ template <typename _ExecutionPolicy, typename _Ip, typename _Size, typename _Sp,
 void
 for_loop_n_strided(_ExecutionPolicy&& __exec, _Ip __start, _Size __n, _Sp __stride, _Rest&&... __rest)
 {
-    static_assert(oneapi::dpl::__internal::__is_host_execution_policy<::std::decay_t<_ExecutionPolicy>>::value,
-                  "for_loop_n_strided is implemented for the host policies only");
-
     oneapi::dpl::__internal::__for_loop_repack_n(::std::forward<_ExecutionPolicy>(__exec), __start, __n, __stride,
                                                  ::std::forward_as_tuple(::std::forward<_Rest>(__rest)...));
 }
diff --git a/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h b/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
index 84888832193..354a195d404 100644
--- a/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
+++ b/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
@@ -65,7 +65,7 @@ struct __difference<_Ip, ::std::enable_if_t<::std::is_integral_v<_Ip>>>
 template <typename _Ip>
 struct __difference<_Ip, ::std::enable_if_t<!::std::is_integral_v<_Ip>>>
 {
-    using __type = typename ::std::iterator_traits<_Ip>::difference_type;
+    using __type = typename std::iterator_traits<_Ip>::difference_type;
 };
 
 // This type is used as a stride value when it's known that stride == 1 at compile time(the case of for_loop and for_loop_n).
@@ -161,64 +161,38 @@ class __reduction_pack
     }
 };
 
-// Sequenced + vectorized version of for_loop_n
-template <class _Tag, typename _ExecutionPolicy, typename _Ip, typename _Size, typename _Function, typename... _Rest>
+// Sequenced version of for_loop_n
+template <typename _ExecutionPolicy, typename _Ip, typename _Size, typename _Function, typename... _Rest>
 void
-__pattern_for_loop_n(_Tag, _ExecutionPolicy&&, _Ip __first, _Size __n, _Function __f, __single_stride_type,
-                     _Rest&&... __rest) noexcept
+__pattern_for_loop_n(_ExecutionPolicy&&, _Ip __first, _Size __n, _Function __f, __single_stride_type,
+                     /*vector=*/::std::false_type, /*parallel=*/::std::false_type, _Rest&&... __rest) noexcept
 {
-    static_assert(__is_serial_tag_v<_Tag> || __is_parallel_forward_tag_v<_Tag>);
-
     __reduction_pack<_Rest...> __pack{__reduction_pack_tag(), ::std::forward<_Rest>(__rest)...};
 
-    if constexpr (typename _Tag::__is_vector{})
-    {
-        oneapi::dpl::__internal::__brick_walk1(
-            __n, [&__pack, __first, __f](_Size __idx) { __pack.__apply_func(__f, __first + __idx, __idx); },
-            ::std::true_type{});
-    }
-    else
-    {
-        for (_Size __i = 0; __i < __n; ++__i, ++__first)
-            __pack.__apply_func(__f, __first, __i);
-    }
+    for (_Size __i = 0; __i < __n; ++__i, ++__first)
+        __pack.__apply_func(__f, __first, __i);
 
     __pack.__finalize(__n);
 }
 
-template <class _Tag, typename _ExecutionPolicy, typename _Ip, typename _Size, typename _Function, typename _Sp,
-          typename... _Rest>
+template <typename _ExecutionPolicy, typename _Ip, typename _Size, typename _Function, typename _Sp, typename... _Rest>
 void
-__pattern_for_loop_n(_Tag, _ExecutionPolicy&&, _Ip __first, _Size __n, _Function __f, _Sp __stride,
-                     _Rest&&... __rest) noexcept
+__pattern_for_loop_n(_ExecutionPolicy&&, _Ip __first, _Size __n, _Function __f, _Sp __stride,
+                     /*vector=*/::std::false_type, /*parallel=*/::std::false_type, _Rest&&... __rest) noexcept
 {
-    static_assert(__is_serial_tag_v<_Tag> || __is_parallel_forward_tag_v<_Tag>);
-
     __reduction_pack<_Rest...> __pack{__reduction_pack_tag(), ::std::forward<_Rest>(__rest)...};
 
-    if constexpr (typename _Tag::__is_vector{})
-    {
-        oneapi::dpl::__internal::__brick_walk1(
-            __n,
-            [&__pack, __first, __f, __stride](_Size __idx) {
-                __pack.__apply_func(__f, __first + __idx * __stride, __idx);
-            },
-            ::std::true_type{});
-    }
-    else
+    // Simple loop from 0 to __n is not suitable here as we need to ensure that __first is always
+    // <= than the end iterator, even if it's not dereferenced. Some implementation might place
+    // validation checks to enforce this invariant.
+    if (__n > 0)
     {
-        // Simple loop from 0 to __n is not suitable here as we need to ensure that __first is always
-        // <= than the end iterator, even if it's not dereferenced. Some implementation might place
-        // validation checks to enforce this invariant.
-        if (__n > 0)
+        for (_Size __i = 0; __i < __n - 1; ++__i, oneapi::dpl::__internal::__advance(__first, __stride))
         {
-            for (_Size __i = 0; __i < __n - 1; ++__i, oneapi::dpl::__internal::__advance(__first, __stride))
-            {
-                __pack.__apply_func(__f, __first, __i);
-            }
-
-            __pack.__apply_func(__f, __first, __n - 1);
+            __pack.__apply_func(__f, __first, __i);
         }
+
+        __pack.__apply_func(__f, __first, __n - 1);
     }
 
     __pack.__finalize(__n);
@@ -242,68 +216,25 @@ struct __is_random_access_or_integral<_Ip,
 {
 };
 
-template <typename _Ip, typename _Function, typename _Sp, typename _Pack, typename _IndexType>
-::std::enable_if_t<
-    ::std::is_same_v<typename ::std::iterator_traits<_Ip>::iterator_category, ::std::bidirectional_iterator_tag>,
-    _IndexType>
-__execute_loop_strided(_Ip __first, _Ip __last, _Function __f, _Sp __stride, _Pack& __pack, _IndexType) noexcept;
-
-template <typename _Ip, typename _Function, typename _Sp, typename _Pack, typename _IndexType>
-::std::enable_if_t<
-    ::std::is_same_v<typename ::std::iterator_traits<_Ip>::iterator_category, ::std::forward_iterator_tag> ||
-        ::std::is_same_v<typename ::std::iterator_traits<_Ip>::iterator_category, ::std::input_iterator_tag>,
-    _IndexType>
-__execute_loop_strided(_Ip __first, _Ip __last, _Function __f, _Sp __stride, _Pack& __pack, _IndexType) noexcept;
-
 template <typename _Ip>
 inline constexpr bool __is_random_access_or_integral_v = __is_random_access_or_integral<_Ip>::value;
 
 // Sequenced version of for_loop for RAI and integral types
-// Vectorized version of for_loop
-template <class _Tag, typename _ExecutionPolicy, typename _Ip, typename _Function, typename _Sp, typename... _Rest>
-void
-__pattern_for_loop(_Tag __tag, _ExecutionPolicy&& __exec, _Ip __first, _Ip __last, _Function __f, _Sp __stride,
-                   _Rest&&... __rest) noexcept
+template <typename _ExecutionPolicy, typename _Ip, typename _Function, typename _Sp, typename... _Rest>
+::std::enable_if_t<__is_random_access_or_integral_v<_Ip>>
+__pattern_for_loop(_ExecutionPolicy&& __exec, _Ip __first, _Ip __last, _Function __f, _Sp __stride,
+                   /*vector=*/::std::false_type, /*parallel=*/::std::false_type, _Rest&&... __rest) noexcept
 {
-    static_assert(__is_serial_tag_v<_Tag> || __is_parallel_forward_tag_v<_Tag>);
-
-    if constexpr (typename _Tag::__is_vector{} || __is_random_access_or_integral_v<_Ip>)
-    {
-        oneapi::dpl::__internal::__pattern_for_loop_n(
-            __tag, ::std::forward<_ExecutionPolicy>(__exec), __first,
-            oneapi::dpl::__internal::__calculate_input_sequence_length(__first, __last, __stride), __f, __stride,
-            ::std::forward<_Rest>(__rest)...);
-    }
-    else
-    {
-        __reduction_pack<_Rest...> __pack{__reduction_pack_tag(), ::std::forward<_Rest>(__rest)...};
-
-        // Make sure that our index type is able to hold all the possible values
-        using __index_type = typename __difference<_Ip>::__type;
-        __index_type __ordinal_position = 0;
-
-        if (__stride == 1)
-        {
-            // Avoid check for i % stride on each iteration for the most common case.
-            for (; __first != __last; ++__first, ++__ordinal_position)
-                __pack.__apply_func(__f, __first, __ordinal_position);
-        }
-        else
-        {
-            __ordinal_position =
-                oneapi::dpl::__internal::__execute_loop_strided(__first, __last, __f, __stride, __pack,
-                                                                // Only passed to deduce the type for internal counter
-                                                                __index_type{});
-        }
-
-        __pack.__finalize(__ordinal_position);
-    }
+    oneapi::dpl::__internal::__pattern_for_loop_n(
+        ::std::forward<_ExecutionPolicy>(__exec), __first,
+        oneapi::dpl::__internal::__calculate_input_sequence_length(__first, __last, __stride), __f, __stride,
+        ::std::false_type{}, ::std::false_type{}, ::std::forward<_Rest>(__rest)...);
 }
 
 template <typename _Ip, typename _Function, typename _Sp, typename _Pack, typename _IndexType>
-::std::enable_if_t<
-    ::std::is_same_v<typename ::std::iterator_traits<_Ip>::iterator_category, ::std::bidirectional_iterator_tag>,
-    _IndexType>
+::std::enable_if_t<::std::is_same_v<typename std::iterator_traits<_Ip>::iterator_category,
+                                    ::std::bidirectional_iterator_tag>,
+                   _IndexType>
 __execute_loop_strided(_Ip __first, _Ip __last, _Function __f, _Sp __stride, _Pack& __pack, _IndexType) noexcept
 {
     _IndexType __ordinal_position = 0;
@@ -338,10 +269,11 @@ __execute_loop_strided(_Ip __first, _Ip __last, _Function __f, _Sp __stride, _Pa
 }
 
 template <typename _Ip, typename _Function, typename _Sp, typename _Pack, typename _IndexType>
-::std::enable_if_t<
-    ::std::is_same_v<typename ::std::iterator_traits<_Ip>::iterator_category, ::std::forward_iterator_tag> ||
-        ::std::is_same_v<typename ::std::iterator_traits<_Ip>::iterator_category, ::std::input_iterator_tag>,
-    _IndexType>
+::std::enable_if_t<::std::is_same_v<typename std::iterator_traits<_Ip>::iterator_category,
+                                    ::std::forward_iterator_tag> ||
+                       ::std::is_same_v<typename std::iterator_traits<_Ip>::iterator_category,
+                                        ::std::input_iterator_tag>,
+                   _IndexType>
 __execute_loop_strided(_Ip __first, _Ip __last, _Function __f, _Sp __stride, _Pack& __pack, _IndexType) noexcept
 {
     _IndexType __ordinal_position = 0;
@@ -361,13 +293,11 @@ __execute_loop_strided(_Ip __first, _Ip __last, _Function __f, _Sp __stride, _Pa
 }
 
 // Sequenced version of for_loop for non-RAI and non-integral types
-template <class _Tag, typename _ExecutionPolicy, typename _Ip, typename _Function, typename... _Rest>
-void
-__pattern_for_loop(_Tag, _ExecutionPolicy&& __exec, _Ip __first, _Ip __last, _Function __f, __single_stride_type,
-                   _Rest&&... __rest) noexcept
+template <typename _ExecutionPolicy, typename _Ip, typename _Function, typename... _Rest>
+::std::enable_if_t<!__is_random_access_or_integral_v<_Ip>>
+__pattern_for_loop(_ExecutionPolicy&&, _Ip __first, _Ip __last, _Function __f, __single_stride_type,
+                   /*vector=*/::std::false_type, /*parallel=*/::std::false_type, _Rest&&... __rest) noexcept
 {
-    static_assert(__is_serial_tag_v<_Tag> || __is_parallel_forward_tag_v<_Tag>);
-
     __reduction_pack<_Rest...> __pack{__reduction_pack_tag(), ::std::forward<_Rest>(__rest)...};
 
     // Make sure that our index type is able to hold all the possible values
@@ -381,77 +311,135 @@ __pattern_for_loop(_Tag, _ExecutionPolicy&& __exec, _Ip __first, _Ip __last, _Fu
     __pack.__finalize(__ordinal_position);
 }
 
+template <typename _ExecutionPolicy, typename _Ip, typename _Function, typename _Sp, typename... _Rest>
+::std::enable_if_t<!__is_random_access_or_integral_v<_Ip>>
+__pattern_for_loop(_ExecutionPolicy&&, _Ip __first, _Ip __last, _Function __f, _Sp __stride,
+                   /*vector=*/::std::false_type, /*parallel=*/::std::false_type, _Rest&&... __rest) noexcept
+{
+    __reduction_pack<_Rest...> __pack{__reduction_pack_tag(), ::std::forward<_Rest>(__rest)...};
+
+    // Make sure that our index type is able to hold all the possible values
+    using __index_type = typename __difference<_Ip>::__type;
+    __index_type __ordinal_position = 0;
+
+    if (__stride == 1)
+    {
+        // Avoid check for i % stride on each iteration for the most common case.
+        for (; __first != __last; ++__first, ++__ordinal_position)
+            __pack.__apply_func(__f, __first, __ordinal_position);
+    }
+    else
+    {
+        __ordinal_position =
+            oneapi::dpl::__internal::__execute_loop_strided(__first, __last, __f, __stride, __pack,
+                                                            // Only passed to deduce the type for internal counter
+                                                            __index_type{});
+    }
+
+    __pack.__finalize(__ordinal_position);
+}
+
+// Vectorized version of for_loop_n
+template <typename _ExecutionPolicy, typename _Ip, typename _Size, typename _Function, typename... _Rest>
+void
+__pattern_for_loop_n(_ExecutionPolicy&&, _Ip __first, _Size __n, _Function __f, __single_stride_type,
+                     /*vector=*/::std::true_type, /*parallel=*/::std::false_type, _Rest&&... __rest) noexcept
+{
+    __reduction_pack<_Rest...> __pack{__reduction_pack_tag(), ::std::forward<_Rest>(__rest)...};
+
+    oneapi::dpl::__internal::__brick_walk1(
+        __n, [&__pack, __first, __f](_Size __idx) { __pack.__apply_func(__f, __first + __idx, __idx); },
+        ::std::true_type{});
+
+    __pack.__finalize(__n);
+}
+
+template <typename _ExecutionPolicy, typename _Ip, typename _Size, typename _Function, typename _Sp, typename... _Rest>
+void
+__pattern_for_loop_n(_ExecutionPolicy&&, _Ip __first, _Size __n, _Function __f, _Sp __stride,
+                     /*vector=*/::std::true_type, /*parallel=*/::std::false_type, _Rest&&... __rest) noexcept
+{
+    __reduction_pack<_Rest...> __pack{__reduction_pack_tag(), ::std::forward<_Rest>(__rest)...};
+
+    oneapi::dpl::__internal::__brick_walk1(
+        __n,
+        [&__pack, __first, __f, __stride](_Size __idx) { __pack.__apply_func(__f, __first + __idx * __stride, __idx); },
+        ::std::true_type{});
+
+    __pack.__finalize(__n);
+}
+
+// Vectorized version of for_loop
+template <typename _ExecutionPolicy, typename _Ip, typename _Function, typename _Sp, typename... _Rest>
+void
+__pattern_for_loop(_ExecutionPolicy&& __exec, _Ip __first, _Ip __last, _Function __f, _Sp __stride,
+                   /*vector=*/::std::true_type, /*parallel=*/::std::false_type, _Rest&&... __rest) noexcept
+{
+    oneapi::dpl::__internal::__pattern_for_loop_n(
+        ::std::forward<_ExecutionPolicy>(__exec), __first,
+        oneapi::dpl::__internal::__calculate_input_sequence_length(__first, __last, __stride), __f, __stride,
+        ::std::true_type{}, ::std::false_type{}, ::std::forward<_Rest>(__rest)...);
+}
+
 // Parallel version of for_loop_n
 
 // TODO: Using parallel_reduce when we don't have a reduction object in the pack might be ineffective,
 // perhaps it's better to check for presence of reduction object and call parallel_for routine instead.
 // TODO: need to add a static_assert for match between rest and f's arguments, currently there is a lot
 // of unclear error in cast of mismatch.
-template <typename _IsVector, typename _ExecutionPolicy, typename _Ip, typename _Size, typename _Function,
+template <typename _ExecutionPolicy, typename _Ip, typename _Size, typename _Function, typename _IsVector,
           typename... _Rest>
 void
-__pattern_for_loop_n(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Ip __first, _Size __n, _Function __f,
-                     __single_stride_type, _Rest&&... __rest)
+__pattern_for_loop_n(_ExecutionPolicy&& __exec, _Ip __first, _Size __n, _Function __f, __single_stride_type,
+                     _IsVector __is_vector, /*parallel=*/::std::true_type, _Rest&&... __rest)
 {
-    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
-
     using __pack_type = __reduction_pack<_Rest...>;
 
     // Create an identity pack object, operations are done on copies of it.
     const __pack_type __identity{__reduction_pack_tag(), ::std::forward<_Rest>(__rest)...};
 
+    using __backend_tag = typename oneapi::dpl::__internal::__parallel_tag<_IsVector>::__backend_tag;
     oneapi::dpl::__internal::__except_handler([&]() {
-        return __par_backend::__parallel_reduce(
-                   __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), _Size(0), __n, __identity,
-                   [__first, __f](_Size __i, _Size __j, __pack_type __value) {
-                       const auto __subseq_start = __first + __i;
-                       const auto __length = __j - __i;
-
-                       oneapi::dpl::__internal::__brick_walk1(
-                           __length,
-                           [&__value, __f, __i, __subseq_start](_Size __idx) {
-                               __value.__apply_func(__f, __subseq_start + __idx, __i + __idx);
-                           },
-                           _IsVector{});
-
-                       return __value;
-                   },
-                   [](__pack_type __lhs, const __pack_type& __rhs) {
-                       __lhs.__combine(__rhs);
-                       return __lhs;
-                   })
+        return __par_backend::__parallel_reduce(__backend_tag{},
+                                                ::std::forward<_ExecutionPolicy>(__exec), _Size(0), __n, __identity,
+                                                [__is_vector, __first, __f](_Size __i, _Size __j, __pack_type __value) {
+                                                    const auto __subseq_start = __first + __i;
+                                                    const auto __length = __j - __i;
+
+                                                    oneapi::dpl::__internal::__brick_walk1(
+                                                        __length,
+                                                        [&__value, __f, __i, __subseq_start](_Size __idx) {
+                                                            __value.__apply_func(__f, __subseq_start + __idx,
+                                                                                 __i + __idx);
+                                                        },
+                                                        __is_vector);
+
+                                                    return __value;
+                                                },
+                                                [](__pack_type __lhs, const __pack_type& __rhs) {
+                                                    __lhs.__combine(__rhs);
+                                                    return __lhs;
+                                                })
             .__finalize(__n);
     });
 }
 
-template <typename _IsVector, typename _ExecutionPolicy, typename _Ip, typename _Function, typename... _Rest>
-void
-__pattern_for_loop(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _Ip __first, _Ip __last, _Function __f,
-                   __single_stride_type, _Rest&&... __rest)
-{
-    oneapi::dpl::__internal::__pattern_for_loop_n(
-        __tag, ::std::forward<_ExecutionPolicy>(__exec), __first,
-        oneapi::dpl::__internal::__calculate_input_sequence_length(__first, __last, __single_stride_type{}), __f,
-        __single_stride_type{}, ::std::forward<_Rest>(__rest)...);
-}
-
-template <typename _IsVector, typename _ExecutionPolicy, typename _Ip, typename _Size, typename _Function, typename _Sp,
+template <typename _ExecutionPolicy, typename _Ip, typename _Size, typename _Function, typename _Sp, typename _IsVector,
           typename... _Rest>
 void
-__pattern_for_loop_n(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Ip __first, _Size __n, _Function __f,
-                     _Sp __stride, _Rest&&... __rest)
+__pattern_for_loop_n(_ExecutionPolicy&& __exec, _Ip __first, _Size __n, _Function __f, _Sp __stride,
+                     _IsVector __is_vector, /*parallel=*/::std::true_type, _Rest&&... __rest)
 {
-    using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
-
     using __pack_type = __reduction_pack<_Rest...>;
 
     // Create an identity pack object, operations are done on copies of it.
     const __pack_type __identity{__reduction_pack_tag(), ::std::forward<_Rest>(__rest)...};
 
+    using __backend_tag = typename oneapi::dpl::__internal::__parallel_tag<_IsVector>::__backend_tag;
     oneapi::dpl::__internal::__except_handler([&]() {
-        return __par_backend::__parallel_reduce(
-                   __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), _Size(0), __n, __identity,
-                   [__first, __f, __stride](_Size __i, _Size __j, __pack_type __value) {
+        return __par_backend::__parallel_reduce(__backend_tag{},
+                   ::std::forward<_ExecutionPolicy>(__exec), _Size(0), __n, __identity,
+                   [__is_vector, __first, __f, __stride](_Size __i, _Size __j, __pack_type __value) {
                        const auto __subseq_start = __first + __i * __stride;
                        const auto __length = __j - __i;
 
@@ -460,7 +448,7 @@ __pattern_for_loop_n(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Ip _
                            [&__value, __f, __i, __subseq_start, __stride](_Size __idx) {
                                __value.__apply_func(__f, __subseq_start + __idx * __stride, __i + __idx);
                            },
-                           _IsVector{});
+                           __is_vector);
 
                        return __value;
                    },
@@ -472,33 +460,60 @@ __pattern_for_loop_n(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Ip _
     });
 }
 
-template <typename _IsVector, typename _ExecutionPolicy, typename _Ip, typename _Function, typename _Sp,
+template <typename _ExecutionPolicy, typename _Ip, typename _Function, typename _Sp, typename _IsVector,
           typename... _Rest>
 void
-__pattern_for_loop(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _Ip __first, _Ip __last, _Function __f,
-                   _Sp __stride, _Rest&&... __rest)
+__pattern_for_loop(_ExecutionPolicy&& __exec, _Ip __first, _Ip __last, _Function __f, _Sp __stride,
+                   _IsVector __is_vector,
+                   /*parallel=*/::std::true_type, _Rest&&... __rest)
 {
     oneapi::dpl::__internal::__pattern_for_loop_n(
-        __tag, ::std::forward<_ExecutionPolicy>(__exec), __first,
+        ::std::forward<_ExecutionPolicy>(__exec), __first,
         oneapi::dpl::__internal::__calculate_input_sequence_length(__first, __last, __stride), __f, __stride,
-        ::std::forward<_Rest>(__rest)...);
+        __is_vector, ::std::true_type{}, ::std::forward<_Rest>(__rest)...);
 }
 
-// Special versions for for_loop: handles both iterators and integral types(treated as random access iterators)
-template <typename _ExecutionPolicy, typename _Ip>
-auto
-__select_backend_for_loop(_ExecutionPolicy&& exec, _Ip __start)
+// Helper structure to split code functions for integral and iterator types so the return
+// value can be successfully deduced.
+template <typename _Ip>
+struct __use_par_vec_helper
 {
-    static_assert(oneapi::dpl::__internal::__is_host_execution_policy<::std::decay_t<_ExecutionPolicy>>::value);
+    using __it_type = std::conditional_t<std::is_integral_v<_Ip>, _Ip*, _Ip>;
 
-    if constexpr (::std::is_integral_v<_Ip>)
+    template <typename _ExecutionPolicy>
+    static constexpr auto
+    __use_vector(_ExecutionPolicy&& __exec)
     {
-        return __select_backend(exec);
+        using __tag_type = decltype(oneapi::dpl::__internal::__select_backend(__exec, __it_type{}));
+        return typename __tag_type::__is_vector{};
     }
-    else
+
+    template <typename _ExecutionPolicy>
+    static constexpr auto
+    __use_parallel(_ExecutionPolicy&& __exec)
     {
-        return __select_backend(exec, __start);
+        using __tag_type = decltype(oneapi::dpl::__internal::__select_backend(__exec, __it_type{}));
+        using __is_parallel_tag = std::disjunction<
+            std::is_same<oneapi::dpl::__internal::__parallel_tag<std::true_type>, __tag_type>,
+            std::is_same<oneapi::dpl::__internal::__parallel_tag<std::false_type>, __tag_type>>;
+
+        return __is_parallel_tag{};
     }
+};
+
+// Special versions for for_loop: handles both iterators and integral types(treated as random access iterators)
+template <typename _ExecutionPolicy, typename _Ip>
+auto
+__use_vectorization(_ExecutionPolicy&& __exec)
+{
+    return __use_par_vec_helper<_Ip>::__use_vector(::std::forward<_ExecutionPolicy>(__exec));
+}
+
+template <typename _ExecutionPolicy, typename _Ip>
+auto
+__use_parallelization(_ExecutionPolicy&& __exec)
+{
+    return __use_par_vec_helper<_Ip>::__use_parallel(::std::forward<_ExecutionPolicy>(__exec));
 }
 
 // Helper functions to extract to separate a Callable object from the pack of reductions and inductions
@@ -507,10 +522,11 @@ void
 __for_loop_impl(_ExecutionPolicy&& __exec, _Ip __start, _Ip __finish, _Fp&& __f, _Sp __stride,
                 ::std::tuple<_Rest...>&& __t, ::std::index_sequence<_Is...>)
 {
-    const auto __dispatch_tag = __select_backend_for_loop(__exec, __start);
-
-    oneapi::dpl::__internal::__pattern_for_loop(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __start,
-                                                __finish, __f, __stride, ::std::get<_Is>(::std::move(__t))...);
+    oneapi::dpl::__internal::__pattern_for_loop(
+        ::std::forward<_ExecutionPolicy>(__exec), __start, __finish, __f, __stride,
+        oneapi::dpl::__internal::__use_vectorization<_ExecutionPolicy, _Ip>(__exec),
+        oneapi::dpl::__internal::__use_parallelization<_ExecutionPolicy, _Ip>(__exec),
+        ::std::get<_Is>(::std::move(__t))...);
 }
 
 template <typename _ExecutionPolicy, typename _Ip, typename _Size, typename _Fp, typename _Sp, typename... _Rest,
@@ -519,10 +535,11 @@ void
 __for_loop_n_impl(_ExecutionPolicy&& __exec, _Ip __start, _Size __n, _Fp&& __f, _Sp __stride,
                   ::std::tuple<_Rest...>&& __t, ::std::index_sequence<_Is...>)
 {
-    const auto __dispatch_tag = __select_backend_for_loop(__exec, __start);
-
-    oneapi::dpl::__internal::__pattern_for_loop_n(__dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __start,
-                                                  __n, __f, __stride, ::std::get<_Is>(::std::move(__t))...);
+    oneapi::dpl::__internal::__pattern_for_loop_n(
+        ::std::forward<_ExecutionPolicy>(__exec), __start, __n, __f, __stride,
+        oneapi::dpl::__internal::__use_vectorization<_ExecutionPolicy, _Ip>(__exec),
+        oneapi::dpl::__internal::__use_parallelization<_ExecutionPolicy, _Ip>(__exec),
+        ::std::get<_Is>(::std::move(__t))...);
 }
 
 template <typename _ExecutionPolicy, typename _Ip, typename _Sp, typename... _Rest>

From 88c166795771485def53a1e8e5ac5b67a1052cd2 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Tue, 12 Mar 2024 10:05:45 +0100
Subject: [PATCH 534/566] 
 include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h - replace
 std::iterator_traits to ::std::iterator_traits

---
 .../oneapi/dpl/pstl/experimental/internal/for_loop_impl.h | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h b/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
index 354a195d404..329367fb019 100644
--- a/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
+++ b/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
@@ -65,7 +65,7 @@ struct __difference<_Ip, ::std::enable_if_t<::std::is_integral_v<_Ip>>>
 template <typename _Ip>
 struct __difference<_Ip, ::std::enable_if_t<!::std::is_integral_v<_Ip>>>
 {
-    using __type = typename std::iterator_traits<_Ip>::difference_type;
+    using __type = typename ::std::iterator_traits<_Ip>::difference_type;
 };
 
 // This type is used as a stride value when it's known that stride == 1 at compile time(the case of for_loop and for_loop_n).
@@ -232,7 +232,7 @@ __pattern_for_loop(_ExecutionPolicy&& __exec, _Ip __first, _Ip __last, _Function
 }
 
 template <typename _Ip, typename _Function, typename _Sp, typename _Pack, typename _IndexType>
-::std::enable_if_t<::std::is_same_v<typename std::iterator_traits<_Ip>::iterator_category,
+::std::enable_if_t<::std::is_same_v<typename ::std::iterator_traits<_Ip>::iterator_category,
                                     ::std::bidirectional_iterator_tag>,
                    _IndexType>
 __execute_loop_strided(_Ip __first, _Ip __last, _Function __f, _Sp __stride, _Pack& __pack, _IndexType) noexcept
@@ -269,9 +269,9 @@ __execute_loop_strided(_Ip __first, _Ip __last, _Function __f, _Sp __stride, _Pa
 }
 
 template <typename _Ip, typename _Function, typename _Sp, typename _Pack, typename _IndexType>
-::std::enable_if_t<::std::is_same_v<typename std::iterator_traits<_Ip>::iterator_category,
+::std::enable_if_t<::std::is_same_v<typename ::std::iterator_traits<_Ip>::iterator_category,
                                     ::std::forward_iterator_tag> ||
-                       ::std::is_same_v<typename std::iterator_traits<_Ip>::iterator_category,
+                       ::std::is_same_v<typename ::std::iterator_traits<_Ip>::iterator_category,
                                         ::std::input_iterator_tag>,
                    _IndexType>
 __execute_loop_strided(_Ip __first, _Ip __last, _Function __f, _Sp __stride, _Pack& __pack, _IndexType) noexcept

From 343c857cef3e6db721c39ff62c419980b99ddc71 Mon Sep 17 00:00:00 2001
From: Mikhail Dvorskiy <mikhail.dvorskiy@intel.com>
Date: Tue, 12 Mar 2024 16:37:54 +0100
Subject: [PATCH 535/566] [oneDPL][for_loop][tag] + std::declval using

---
 include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h b/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
index 329367fb019..e08777cd402 100644
--- a/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
+++ b/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
@@ -484,7 +484,7 @@ struct __use_par_vec_helper
     static constexpr auto
     __use_vector(_ExecutionPolicy&& __exec)
     {
-        using __tag_type = decltype(oneapi::dpl::__internal::__select_backend(__exec, __it_type{}));
+        using __tag_type = decltype(oneapi::dpl::__internal::__select_backend(__exec, std::declval<__it_type>()));
         return typename __tag_type::__is_vector{};
     }
 
@@ -492,7 +492,7 @@ struct __use_par_vec_helper
     static constexpr auto
     __use_parallel(_ExecutionPolicy&& __exec)
     {
-        using __tag_type = decltype(oneapi::dpl::__internal::__select_backend(__exec, __it_type{}));
+        using __tag_type = decltype(oneapi::dpl::__internal::__select_backend(__exec, std::declval<__it_type>()));
         using __is_parallel_tag = std::disjunction<
             std::is_same<oneapi::dpl::__internal::__parallel_tag<std::true_type>, __tag_type>,
             std::is_same<oneapi::dpl::__internal::__parallel_tag<std::false_type>, __tag_type>>;

From 0667c9eeca33180c91d23b35f66ddc3ca2415742 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 13 Mar 2024 09:44:58 +0100
Subject: [PATCH 536/566] 
 include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h - fix review
 comment: seems like it can be done even simple because we already have
 __is_parallel_tag, implemented in this PR. Let's use that.

---
 .../oneapi/dpl/pstl/experimental/internal/for_loop_impl.h   | 6 +-----
 1 file changed, 1 insertion(+), 5 deletions(-)

diff --git a/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h b/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
index e08777cd402..79e0f3eb855 100644
--- a/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
+++ b/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
@@ -493,11 +493,7 @@ struct __use_par_vec_helper
     __use_parallel(_ExecutionPolicy&& __exec)
     {
         using __tag_type = decltype(oneapi::dpl::__internal::__select_backend(__exec, std::declval<__it_type>()));
-        using __is_parallel_tag = std::disjunction<
-            std::is_same<oneapi::dpl::__internal::__parallel_tag<std::true_type>, __tag_type>,
-            std::is_same<oneapi::dpl::__internal::__parallel_tag<std::false_type>, __tag_type>>;
-
-        return __is_parallel_tag{};
+        return oneapi::dpl::__internal::__is_parallel_tag<__tag_type>{};
     }
 };
 

From 1ed29ebcfe10214833126926eb71b79cd46f7546 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 13 Mar 2024 11:12:05 +0100
Subject: [PATCH 537/566] Fix review comment: Seems like _BackendTag should be
 the part of operator(), not the part of a class template

---
 .../pstl/hetero/dpcpp/parallel_backend_sycl.h | 16 ++++++-------
 .../dpcpp/parallel_backend_sycl_histogram.h   | 23 +++++++++----------
 2 files changed, 19 insertions(+), 20 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
index e3c3f62d85a..2bca3eb165b 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
@@ -1685,16 +1685,16 @@ __parallel_sort_impl(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPo
 }
 
 // Please see the comment for __parallel_for_submitter for optional kernel name explanation
-template <typename _BackendTag, typename _GlobalSortName, typename _CopyBackName>
+template <typename _GlobalSortName, typename _CopyBackName>
 struct __parallel_partial_sort_submitter;
 
-template <typename _BackendTag, typename... _GlobalSortName, typename... _CopyBackName>
-struct __parallel_partial_sort_submitter<_BackendTag, __internal::__optional_kernel_name<_GlobalSortName...>,
+template <typename... _GlobalSortName, typename... _CopyBackName>
+struct __parallel_partial_sort_submitter<__internal::__optional_kernel_name<_GlobalSortName...>,
                                          __internal::__optional_kernel_name<_CopyBackName...>>
 {
-    template <typename _ExecutionPolicy, typename _Range, typename _Merge, typename _Compare>
+    template <typename _BackendTag, typename _ExecutionPolicy, typename _Range, typename _Merge, typename _Compare>
     auto
-    operator()(_ExecutionPolicy&& __exec, _Range&& __rng, _Merge __merge, _Compare __comp) const
+    operator()(_BackendTag, _ExecutionPolicy&& __exec, _Range&& __rng, _Merge __merge, _Compare __comp) const
     {
         using _Tp = oneapi::dpl::__internal::__value_t<_Range>;
         using _Size = oneapi::dpl::__internal::__difference_t<_Range>;
@@ -1768,9 +1768,9 @@ __parallel_partial_sort_impl(oneapi::dpl::__internal::__device_backend_tag, _Exe
     using _CopyBackKernel =
         oneapi::dpl::__par_backend_hetero::__internal::__kernel_name_provider<__sort_copy_back_kernel<_CustomName>>;
 
-    return __parallel_partial_sort_submitter<oneapi::dpl::__internal::__device_backend_tag, _GlobalSortKernel,
-                                             _CopyBackKernel>()(::std::forward<_ExecutionPolicy>(__exec),
-                                                                ::std::forward<_Range>(__rng), __merge, __comp);
+    return __parallel_partial_sort_submitter<_GlobalSortKernel, _CopyBackKernel>()(
+        oneapi::dpl::__internal::__device_backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
+        ::std::forward<_Range>(__rng), __merge, __comp);
 }
 
 //------------------------------------------------------------------------
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h
index 083bffcefc0..e5d2c5a3a15 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h
@@ -399,18 +399,17 @@ __histogram_general_local_atomics(oneapi::dpl::__internal::__device_backend_tag,
         ::std::forward<_Range2>(__bins), __binhash_manager);
 }
 
-template <typename _BackendTag, typename _KernelName>
+template <typename _KernelName>
 struct __histogram_general_private_global_atomics_submitter;
 
-template <typename _BackendTag, typename... _KernelName>
-struct __histogram_general_private_global_atomics_submitter<_BackendTag,
-                                                            __internal::__optional_kernel_name<_KernelName...>>
+template <typename... _KernelName>
+struct __histogram_general_private_global_atomics_submitter<__internal::__optional_kernel_name<_KernelName...>>
 {
-    template <typename _ExecutionPolicy, typename _Range1, typename _Range2, typename _BinHashMgr>
+    template <typename _BackendTag, typename _ExecutionPolicy, typename _Range1, typename _Range2, typename _BinHashMgr>
     auto
-    operator()(_ExecutionPolicy&& __exec, const sycl::event& __init_event, ::std::uint16_t __min_iters_per_work_item,
-               ::std::uint16_t __work_group_size, _Range1&& __input, _Range2&& __bins,
-               const _BinHashMgr& __binhash_manager)
+    operator()(_BackendTag, _ExecutionPolicy&& __exec, const sycl::event& __init_event,
+               ::std::uint16_t __min_iters_per_work_item, ::std::uint16_t __work_group_size, _Range1&& __input,
+               _Range2&& __bins, const _BinHashMgr& __binhash_manager)
     {
         const ::std::size_t __n = __input.size();
         const ::std::size_t __num_bins = __bins.size();
@@ -489,10 +488,10 @@ __histogram_general_private_global_atomics(oneapi::dpl::__internal::__device_bac
     using _global_atomics_name = oneapi::dpl::__par_backend_hetero::__internal::__kernel_name_provider<
         __histo_kernel_private_glocal_atomics<_kernel_base_name>>;
 
-    return __histogram_general_private_global_atomics_submitter<oneapi::dpl::__internal::__device_backend_tag,
-                                                                _global_atomics_name>()(
-        ::std::forward<_ExecutionPolicy>(__exec), __init_event, __min_iters_per_work_item, __work_group_size,
-        ::std::forward<_Range1>(__input), ::std::forward<_Range2>(__bins), __binhash_manager);
+    return __histogram_general_private_global_atomics_submitter<_global_atomics_name>()(
+        oneapi::dpl::__internal::__device_backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __init_event,
+        __min_iters_per_work_item, __work_group_size, ::std::forward<_Range1>(__input), ::std::forward<_Range2>(__bins),
+        __binhash_manager);
 }
 
 template <::std::uint16_t __iters_per_work_item, typename _ExecutionPolicy, typename _Range1, typename _Range2,

From 0900c07bb04564ae8361e878ddcbc319766efcaa Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 13 Mar 2024 11:13:29 +0100
Subject: [PATCH 538/566] include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
 - fix review comment: it is logical to capture __tag first, before __exec

---
 include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
index 04f90cbdc04..77fd740c563 100644
--- a/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
@@ -243,7 +243,7 @@ __pattern_adjacent_difference(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&
     // if we have the only element, just copy it according to the specification
     if (__n == 1)
     {
-        return __internal::__except_handler([&__exec, __first, __last, __d_first, __d_last, &__op, __tag]() {
+        return __internal::__except_handler([__tag, &__exec, __first, __last, __d_first, __d_last, &__op]() {
             auto __wrapped_policy = __par_backend_hetero::make_wrapped_policy<adjacent_difference_wrapper>(
                 ::std::forward<_ExecutionPolicy>(__exec));
 

From ffe92fe0cc58b6d534be9b3337b9125f322f5c0c Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 13 Mar 2024 11:23:01 +0100
Subject: [PATCH 539/566] include/oneapi/dpl/pstl/execution_impl.h - fix review
 comment: The change it pretty localized and we can always upstream it back.
 Though, I don't want to insist. It up to you guys to decide. And I agree
 __parallel_policy_tag_selector_t is a good name (_t is essential)

---
 include/oneapi/dpl/pstl/execution_impl.h | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/include/oneapi/dpl/pstl/execution_impl.h b/include/oneapi/dpl/pstl/execution_impl.h
index dc653a3ec88..f34ac0c5434 100644
--- a/include/oneapi/dpl/pstl/execution_impl.h
+++ b/include/oneapi/dpl/pstl/execution_impl.h
@@ -68,7 +68,7 @@ struct __parallel_forward_tag
 //----------------------------------------------------------
 
 template <class _IsVector, class... _IteratorTypes>
-using __tag_type = ::std::conditional_t<
+using __parallel_policy_tag_selector_t = ::std::conditional_t<
     __internal::__is_random_access_iterator_v<_IteratorTypes...>, __parallel_tag<_IsVector>,
     ::std::conditional_t<__is_forward_iterator_v<_IteratorTypes...>, __parallel_forward_tag, __serial_tag<_IsVector>>>;
 
@@ -87,14 +87,14 @@ __select_backend(oneapi::dpl::execution::unsequenced_policy, _IteratorTypes&&...
 }
 
 template <class... _IteratorTypes>
-__tag_type<std::false_type, _IteratorTypes...>
+__parallel_policy_tag_selector_t<std::false_type, _IteratorTypes...>
 __select_backend(oneapi::dpl::execution::parallel_policy, _IteratorTypes&&...)
 {
     return {};
 }
 
 template <class... _IteratorTypes>
-__tag_type<__internal::__is_random_access_iterator<_IteratorTypes...>, _IteratorTypes...>
+__parallel_policy_tag_selector_t<__internal::__is_random_access_iterator<_IteratorTypes...>, _IteratorTypes...>
 __select_backend(oneapi::dpl::execution::parallel_unsequenced_policy, _IteratorTypes&&...)
 {
     return {};

From 5b9818eed389812cf6676d5b5c3897f6df347a09 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 13 Mar 2024 11:28:13 +0100
Subject: [PATCH 540/566] 
 include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h - fix review
 comment: I am just curious why that line was changed? clang-format should not
 touch that because as I can see it's unmodified

---
 include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h | 5 ++---
 1 file changed, 2 insertions(+), 3 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
index 2bca3eb165b..bc23a9e51ad 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
@@ -1255,9 +1255,8 @@ __parallel_find(oneapi::dpl::__internal::__device_backend_tag __backend_tag, _Ex
     auto __keep = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _Iterator>();
     auto __buf = __keep(__first, __last);
 
-    using _TagType =
-        typename ::std::conditional<_IsFirst::value, __parallel_find_forward_tag<decltype(__buf.all_view())>,
-                                    __parallel_find_backward_tag<decltype(__buf.all_view())>>::type;
+    using _TagType = ::std::conditional_t<_IsFirst::value, __parallel_find_forward_tag<decltype(__buf.all_view())>,
+                                          __parallel_find_backward_tag<decltype(__buf.all_view())>>;
     return __first + oneapi::dpl::__par_backend_hetero::__parallel_find_or(
                          __backend_tag,
                          __par_backend_hetero::make_wrapped_policy<__find_policy_wrapper>(

From e9529f2b32c991aacc2c8ee04aad2a7c7d697f46 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 13 Mar 2024 11:31:43 +0100
Subject: [PATCH 541/566] 
 include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h - fix review
 comment: I put similar comment somewhere else. I don't understand why
 _BackendTag should be a part of class template, not the part of operator()

---
 .../pstl/hetero/dpcpp/parallel_backend_sycl.h | 24 +++++++++----------
 1 file changed, 11 insertions(+), 13 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
index bc23a9e51ad..853372c348c 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
@@ -1545,19 +1545,17 @@ struct __leaf_sort_kernel
 };
 
 // Please see the comment for __parallel_for_submitter for optional kernel name explanation
-template <typename _BackendTag, typename _IdType, typename _LeafSortName, typename _GlobalSortName,
-          typename _CopyBackName>
+template <typename _IdType, typename _LeafSortName, typename _GlobalSortName, typename _CopyBackName>
 struct __parallel_sort_submitter;
 
-template <typename _BackendTag, typename _IdType, typename... _LeafSortName, typename... _GlobalSortName,
-          typename... _CopyBackName>
-struct __parallel_sort_submitter<_BackendTag, _IdType, __internal::__optional_kernel_name<_LeafSortName...>,
+template <typename _IdType, typename... _LeafSortName, typename... _GlobalSortName, typename... _CopyBackName>
+struct __parallel_sort_submitter<_IdType, __internal::__optional_kernel_name<_LeafSortName...>,
                                  __internal::__optional_kernel_name<_GlobalSortName...>,
                                  __internal::__optional_kernel_name<_CopyBackName...>>
 {
-    template <typename _ExecutionPolicy, typename _Range, typename _Compare>
+    template <typename _BackendTag, typename _ExecutionPolicy, typename _Range, typename _Compare>
     auto
-    operator()(_ExecutionPolicy&& __exec, _Range&& __rng, _Compare __comp) const
+    operator()(_BackendTag, _ExecutionPolicy&& __exec, _Range&& __rng, _Compare __comp) const
     {
         using _Tp = oneapi::dpl::__internal::__value_t<_Range>;
         using _Size = oneapi::dpl::__internal::__difference_t<_Range>;
@@ -1664,9 +1662,9 @@ __parallel_sort_impl(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPo
             oneapi::dpl::__par_backend_hetero::__internal::__kernel_name_provider<__sort_global_kernel<_CustomName, _wi_index_type>>;
         using _CopyBackKernel =
             oneapi::dpl::__par_backend_hetero::__internal::__kernel_name_provider<__sort_copy_back_kernel<_CustomName, _wi_index_type>>;
-        return __parallel_sort_submitter<oneapi::dpl::__internal::__device_backend_tag, _wi_index_type, _LeafSortKernel,
-                                         _GlobalSortKernel, _CopyBackKernel>()(::std::forward<_ExecutionPolicy>(__exec),
-                                                                               ::std::forward<_Range>(__rng), __comp);
+        return __parallel_sort_submitter<_wi_index_type, _LeafSortKernel, _GlobalSortKernel, _CopyBackKernel>()(
+            oneapi::dpl::__internal::__device_backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
+            ::std::forward<_Range>(__rng), __comp);
     }
     else
     {
@@ -1677,9 +1675,9 @@ __parallel_sort_impl(oneapi::dpl::__internal::__device_backend_tag, _ExecutionPo
             oneapi::dpl::__par_backend_hetero::__internal::__kernel_name_provider<__sort_global_kernel<_CustomName, _wi_index_type>>;
         using _CopyBackKernel =
             oneapi::dpl::__par_backend_hetero::__internal::__kernel_name_provider<__sort_copy_back_kernel<_CustomName, _wi_index_type>>;
-        return __parallel_sort_submitter<oneapi::dpl::__internal::__device_backend_tag, _wi_index_type, _LeafSortKernel,
-                                         _GlobalSortKernel, _CopyBackKernel>()(::std::forward<_ExecutionPolicy>(__exec),
-                                                                               ::std::forward<_Range>(__rng), __comp);
+        return __parallel_sort_submitter<_wi_index_type, _LeafSortKernel, _GlobalSortKernel, _CopyBackKernel>()(
+            oneapi::dpl::__internal::__device_backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec),
+            ::std::forward<_Range>(__rng), __comp);
     }
 }
 

From 4a211dde0641838b35a60c4779946f4a06c46cd0 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 13 Mar 2024 11:37:58 +0100
Subject: [PATCH 542/566] Fix review comment: why do you need having
 _BackendTag as template argument? If it's because of _IsSync then we need to
 swap template parameters order.

---
 .../dpl/internal/async_impl/async_impl_hetero.h |  5 +++--
 .../dpl/pstl/hetero/algorithm_impl_hetero.h     | 17 +++++++++--------
 2 files changed, 12 insertions(+), 10 deletions(-)

diff --git a/include/oneapi/dpl/internal/async_impl/async_impl_hetero.h b/include/oneapi/dpl/internal/async_impl/async_impl_hetero.h
index 02ee5f852f1..f053974f7b2 100644
--- a/include/oneapi/dpl/internal/async_impl/async_impl_hetero.h
+++ b/include/oneapi/dpl/internal/async_impl/async_impl_hetero.h
@@ -48,10 +48,11 @@ __pattern_walk1_async(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _For
     return __future_obj;
 }
 
-template <typename _BackendTag, typename _IsSync = ::std::false_type,
+template <typename _IsSync = ::std::false_type,
           __par_backend_hetero::access_mode __acc_mode1 = __par_backend_hetero::access_mode::read,
           __par_backend_hetero::access_mode __acc_mode2 = __par_backend_hetero::access_mode::write,
-          typename _ExecutionPolicy, typename _ForwardIterator1, typename _ForwardIterator2, typename _Function>
+          typename _BackendTag, typename _ExecutionPolicy, typename _ForwardIterator1, typename _ForwardIterator2,
+          typename _Function>
 auto
 __pattern_walk2_async(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _ForwardIterator1 __first1,
                       _ForwardIterator1 __last1, _ForwardIterator2 __first2, _Function __f)
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 66166a01f8b..797a44cc7c3 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -78,10 +78,11 @@ __pattern_walk1_n(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _F
 // TODO: A tag _IsSync is used for provide a patterns call pipeline, where the last one should be synchronous
 // Probably it should be re-designed by a pipeline approach, when a pattern returns some sync obejects
 // and ones are combined into a "pipeline" (probably like Range pipeline)
-template <typename _BackendTag, typename _IsSync = ::std::true_type,
+template <typename _IsSync = ::std::true_type,
           __par_backend_hetero::access_mode __acc_mode1 = __par_backend_hetero::access_mode::read,
           __par_backend_hetero::access_mode __acc_mode2 = __par_backend_hetero::access_mode::write,
-          typename _ExecutionPolicy, typename _ForwardIterator1, typename _ForwardIterator2, typename _Function>
+          typename _BackendTag, typename _ExecutionPolicy, typename _ForwardIterator1, typename _ForwardIterator2,
+          typename _Function>
 _ForwardIterator2
 __pattern_walk2(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _ForwardIterator1 __first1,
                 _ForwardIterator1 __last1, _ForwardIterator2 __first2, _Function __f)
@@ -125,7 +126,7 @@ _ForwardIterator2
 __pattern_swap(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _ForwardIterator1 __first1,
                _ForwardIterator1 __last1, _ForwardIterator2 __first2, _Function __f)
 {
-    return __pattern_walk2<_BackendTag, /*_IsSync=*/::std::true_type, __par_backend_hetero::access_mode::read_write,
+    return __pattern_walk2</*_IsSync=*/::std::true_type, __par_backend_hetero::access_mode::read_write,
                            __par_backend_hetero::access_mode::read_write>(
         __tag, ::std::forward<_ExecutionPolicy>(__exec), __first1, __last1, __first2, __f);
 }
@@ -258,7 +259,7 @@ __pattern_walk2_transform_if(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&&
 {
     // Require `read_write` access mode for output sequence to force a copy in for host iterators to capture incoming
     // values of the output sequence for elements where the predicate is false.
-    return __pattern_walk2<_BackendTag, /*_IsSync=*/::std::true_type, __par_backend_hetero::access_mode::read,
+    return __pattern_walk2</*_IsSync=*/::std::true_type, __par_backend_hetero::access_mode::read,
                            __par_backend_hetero::access_mode::read_write>(
         __tag,
         __par_backend_hetero::make_wrapped_policy<__walk2_transform_if_wrapper>(
@@ -1060,7 +1061,7 @@ __pattern_unique(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _It
     auto __copy_last = __pattern_unique_copy(__tag, __exec, __first, __last, __copy_first, __pred);
 
     //TODO: optimize copy back depending on Iterator, i.e. set_final_data for host iterator/pointer
-    return __pattern_walk2<_BackendTag, /*_IsSync=*/::std::true_type, __par_backend_hetero::access_mode::read_write,
+    return __pattern_walk2</*_IsSync=*/::std::true_type, __par_backend_hetero::access_mode::read_write,
                            __par_backend_hetero::access_mode::read_write>(
         __tag, __par_backend_hetero::make_wrapped_policy<copy_back_wrapper>(::std::forward<_ExecutionPolicy>(__exec)),
         __copy_first, __copy_last, __first, __brick_copy<__hetero_tag<_BackendTag>, _ExecutionPolicy>{});
@@ -1326,7 +1327,7 @@ __pattern_stable_partition(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& _
     auto true_count = copy_result.first - __true_result;
 
     //TODO: optimize copy back if possible (inplace, decrease number of submits)
-    __pattern_walk2<_BackendTag, /*_IsSync=*/::std::false_type>(
+    __pattern_walk2</*_IsSync=*/::std::false_type>(
         __tag, __par_backend_hetero::make_wrapped_policy<copy_back_wrapper>(__exec), __true_result, copy_result.first,
         __first, __brick_move<__hetero_tag<_BackendTag>, _ExecutionPolicy>{});
 
@@ -1497,7 +1498,7 @@ __pattern_partial_sort_copy(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&&
     {
         // If our output buffer is larger than the input buffer, simply copy elements to the output and use
         // full sort on them.
-        auto __out_end = __pattern_walk2<_BackendTag, /*_IsSync=*/::std::false_type>(
+        auto __out_end = __pattern_walk2</*_IsSync=*/::std::false_type>(
             __tag, __par_backend_hetero::make_wrapped_policy<__initial_copy_1>(__exec), __first, __last, __out_first,
             __brick_copy<__hetero_tag<_BackendTag>, _ExecutionPolicy>{});
 
@@ -1519,7 +1520,7 @@ __pattern_partial_sort_copy(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&&
 
         auto __buf_first = __buf.get();
 
-        auto __buf_last = __pattern_walk2<_BackendTag, /*_IsSync=*/::std::false_type>(
+        auto __buf_last = __pattern_walk2</*_IsSync=*/::std::false_type>(
             __tag, __par_backend_hetero::make_wrapped_policy<__initial_copy_2>(__exec), __first, __last, __buf_first,
             __brick_copy<__hetero_tag<_BackendTag>, _ExecutionPolicy>{});
 

From b9bd47553ceeaeb02dd16eebc159c967b30b995f Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 13 Mar 2024 12:51:34 +0100
Subject: [PATCH 543/566] include/oneapi/dpl/pstl/execution_impl.h - fix review
 comment: What? why typename = void?

---
 include/oneapi/dpl/pstl/execution_impl.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/include/oneapi/dpl/pstl/execution_impl.h b/include/oneapi/dpl/pstl/execution_impl.h
index f34ac0c5434..1a7b2aa048a 100644
--- a/include/oneapi/dpl/pstl/execution_impl.h
+++ b/include/oneapi/dpl/pstl/execution_impl.h
@@ -155,7 +155,7 @@ inline constexpr bool __is_parallel_tag_v = __is_parallel_tag<_Tag>::value;
 // __is_host_dispatch_tag_v
 //----------------------------------------------------------
 
-template <class _Tag, typename = void>
+template <class _Tag>
 inline constexpr bool __is_host_dispatch_tag_v =
     __is_serial_tag_v<_Tag> || __is_parallel_forward_tag_v<_Tag> || __is_parallel_tag_v<_Tag>;
 

From 3b3a4d227a69c326a62d77bdf26d5f4bf98a1b38 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 13 Mar 2024 12:57:49 +0100
Subject: [PATCH 544/566] 
 include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h - fix
 review comment: What's the purpose of those changes?

---
 .../oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
index 0aea9efd28a..95d23fc16e9 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_fpga.h
@@ -82,8 +82,8 @@ auto
 __parallel_for(oneapi::dpl::__internal::__fpga_backend_tag, _ExecutionPolicy&& __exec, _Fp __brick, _Index __count,
                _Ranges&&... __rngs)
 {
-    using _Policy = ::std::decay_t<_ExecutionPolicy>;
-    using __parallel_for_name = __internal::__kernel_name_provider<typename _Policy::kernel_name>;
+    using _CustomName = oneapi::dpl::__internal::__policy_kernel_name<_ExecutionPolicy>;
+    using __parallel_for_name = __internal::__kernel_name_provider<_CustomName>;
 
     return __parallel_for_fpga_submitter<__parallel_for_name>()(std::forward<_ExecutionPolicy>(__exec), __brick,
                                                                 __count, std::forward<_Ranges>(__rngs)...);

From 2420ee759fa3b87af76ce20c66168d1144204ab4 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 13 Mar 2024 13:55:06 +0100
Subject: [PATCH 545/566] fix review comment: Is this void at the very end
 because we now have typename = void in a primary template?

---
 include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index 797a44cc7c3..ac416480590 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -388,7 +388,7 @@ struct __brick_copy_n<__hetero_tag<_BackendTag>, _ExecutionPolicy>
 };
 
 template <typename _BackendTag, typename _ExecutionPolicy>
-struct __brick_copy<__hetero_tag<_BackendTag>, _ExecutionPolicy, void>
+struct __brick_copy<__hetero_tag<_BackendTag>, _ExecutionPolicy>
 {
     template <typename _SourceT, typename _TargetT>
     void
@@ -399,7 +399,7 @@ struct __brick_copy<__hetero_tag<_BackendTag>, _ExecutionPolicy, void>
 };
 
 template <typename _BackendTag, typename _ExecutionPolicy>
-struct __brick_move<__hetero_tag<_BackendTag>, _ExecutionPolicy, void>
+struct __brick_move<__hetero_tag<_BackendTag>, _ExecutionPolicy>
 {
     template <typename _SourceT, typename _TargetT>
     void
@@ -410,7 +410,7 @@ struct __brick_move<__hetero_tag<_BackendTag>, _ExecutionPolicy, void>
 };
 
 template <typename _BackendTag, typename _ExecutionPolicy, typename _SourceT>
-struct __brick_fill<__hetero_tag<_BackendTag>, _ExecutionPolicy, _SourceT, void>
+struct __brick_fill<__hetero_tag<_BackendTag>, _ExecutionPolicy, _SourceT>
 {
     _SourceT __value;
     template <typename _TargetT>
@@ -422,7 +422,7 @@ struct __brick_fill<__hetero_tag<_BackendTag>, _ExecutionPolicy, _SourceT, void>
 };
 
 template <typename _BackendTag, typename _ExecutionPolicy, typename _SourceT>
-struct __brick_fill_n<__hetero_tag<_BackendTag>, _ExecutionPolicy, _SourceT, void>
+struct __brick_fill_n<__hetero_tag<_BackendTag>, _ExecutionPolicy, _SourceT>
 {
     _SourceT __value;
     template <typename _TargetT>

From a642e397378c2096c65e1a128839159034685c67 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 13 Mar 2024 14:36:02 +0100
Subject: [PATCH 546/566] Apply GitHUB clang format

---
 include/oneapi/dpl/pstl/algorithm_impl.h      |  2 +-
 .../experimental/internal/for_loop_impl.h     | 58 +++++++++----------
 2 files changed, 29 insertions(+), 31 deletions(-)

diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 43c9eb5aab2..f8d7c10c7a3 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -1717,7 +1717,7 @@ __pattern_reverse(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomA
     using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
 
     if (__first == __last)
-        return;    
+        return;
 
     __par_backend::__parallel_for(
         __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __first + (__last - __first) / 2,
diff --git a/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h b/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
index 79e0f3eb855..47769c4645d 100644
--- a/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
+++ b/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
@@ -232,9 +232,9 @@ __pattern_for_loop(_ExecutionPolicy&& __exec, _Ip __first, _Ip __last, _Function
 }
 
 template <typename _Ip, typename _Function, typename _Sp, typename _Pack, typename _IndexType>
-::std::enable_if_t<::std::is_same_v<typename ::std::iterator_traits<_Ip>::iterator_category,
-                                    ::std::bidirectional_iterator_tag>,
-                   _IndexType>
+::std::enable_if_t<
+    ::std::is_same_v<typename ::std::iterator_traits<_Ip>::iterator_category, ::std::bidirectional_iterator_tag>,
+    _IndexType>
 __execute_loop_strided(_Ip __first, _Ip __last, _Function __f, _Sp __stride, _Pack& __pack, _IndexType) noexcept
 {
     _IndexType __ordinal_position = 0;
@@ -269,11 +269,10 @@ __execute_loop_strided(_Ip __first, _Ip __last, _Function __f, _Sp __stride, _Pa
 }
 
 template <typename _Ip, typename _Function, typename _Sp, typename _Pack, typename _IndexType>
-::std::enable_if_t<::std::is_same_v<typename ::std::iterator_traits<_Ip>::iterator_category,
-                                    ::std::forward_iterator_tag> ||
-                       ::std::is_same_v<typename ::std::iterator_traits<_Ip>::iterator_category,
-                                        ::std::input_iterator_tag>,
-                   _IndexType>
+::std::enable_if_t<
+    ::std::is_same_v<typename ::std::iterator_traits<_Ip>::iterator_category, ::std::forward_iterator_tag> ||
+        ::std::is_same_v<typename ::std::iterator_traits<_Ip>::iterator_category, ::std::input_iterator_tag>,
+    _IndexType>
 __execute_loop_strided(_Ip __first, _Ip __last, _Function __f, _Sp __stride, _Pack& __pack, _IndexType) noexcept
 {
     _IndexType __ordinal_position = 0;
@@ -400,26 +399,25 @@ __pattern_for_loop_n(_ExecutionPolicy&& __exec, _Ip __first, _Size __n, _Functio
 
     using __backend_tag = typename oneapi::dpl::__internal::__parallel_tag<_IsVector>::__backend_tag;
     oneapi::dpl::__internal::__except_handler([&]() {
-        return __par_backend::__parallel_reduce(__backend_tag{},
-                                                ::std::forward<_ExecutionPolicy>(__exec), _Size(0), __n, __identity,
-                                                [__is_vector, __first, __f](_Size __i, _Size __j, __pack_type __value) {
-                                                    const auto __subseq_start = __first + __i;
-                                                    const auto __length = __j - __i;
-
-                                                    oneapi::dpl::__internal::__brick_walk1(
-                                                        __length,
-                                                        [&__value, __f, __i, __subseq_start](_Size __idx) {
-                                                            __value.__apply_func(__f, __subseq_start + __idx,
-                                                                                 __i + __idx);
-                                                        },
-                                                        __is_vector);
-
-                                                    return __value;
-                                                },
-                                                [](__pack_type __lhs, const __pack_type& __rhs) {
-                                                    __lhs.__combine(__rhs);
-                                                    return __lhs;
-                                                })
+        return __par_backend::__parallel_reduce(
+                   __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), _Size(0), __n, __identity,
+                   [__is_vector, __first, __f](_Size __i, _Size __j, __pack_type __value) {
+                       const auto __subseq_start = __first + __i;
+                       const auto __length = __j - __i;
+
+                       oneapi::dpl::__internal::__brick_walk1(
+                           __length,
+                           [&__value, __f, __i, __subseq_start](_Size __idx) {
+                               __value.__apply_func(__f, __subseq_start + __idx, __i + __idx);
+                           },
+                           __is_vector);
+
+                       return __value;
+                   },
+                   [](__pack_type __lhs, const __pack_type& __rhs) {
+                       __lhs.__combine(__rhs);
+                       return __lhs;
+                   })
             .__finalize(__n);
     });
 }
@@ -437,8 +435,8 @@ __pattern_for_loop_n(_ExecutionPolicy&& __exec, _Ip __first, _Size __n, _Functio
 
     using __backend_tag = typename oneapi::dpl::__internal::__parallel_tag<_IsVector>::__backend_tag;
     oneapi::dpl::__internal::__except_handler([&]() {
-        return __par_backend::__parallel_reduce(__backend_tag{},
-                   ::std::forward<_ExecutionPolicy>(__exec), _Size(0), __n, __identity,
+        return __par_backend::__parallel_reduce(
+                   __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), _Size(0), __n, __identity,
                    [__is_vector, __first, __f, __stride](_Size __i, _Size __j, __pack_type __value) {
                        const auto __subseq_start = __first + __i * __stride;
                        const auto __length = __j - __i;

From b853589172f372ae9c394a30a4fbdd4b2b9d7130 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 13 Mar 2024 17:55:31 +0100
Subject: [PATCH 547/566] Fix review comment: replace
 ::std::decay_t<decltype(__dispatch_tag)> to decltype(__dispatch_tag)

---
 .../dpl/internal/async_impl/glue_async_impl.h |  4 +-
 include/oneapi/dpl/pstl/glue_algorithm_impl.h | 12 +++---
 .../dpl/pstl/glue_algorithm_ranges_impl.h     |  8 ++--
 include/oneapi/dpl/pstl/glue_memory_impl.h    | 42 +++++++++----------
 4 files changed, 32 insertions(+), 34 deletions(-)

diff --git a/include/oneapi/dpl/internal/async_impl/glue_async_impl.h b/include/oneapi/dpl/internal/async_impl/glue_async_impl.h
index 194f6a84f08..dfd4a969ec8 100644
--- a/include/oneapi/dpl/internal/async_impl/glue_async_impl.h
+++ b/include/oneapi/dpl/internal/async_impl/glue_async_impl.h
@@ -77,12 +77,12 @@ auto
 copy_async(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _ForwardIterator1 __last, _ForwardIterator2 __result,
            _Events&&... __dependencies)
 {
-    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first, __result);
+    auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first, __result);
 
     wait_for_all(::std::forward<_Events>(__dependencies)...);
     auto ret_val = oneapi::dpl::__internal::__pattern_walk2_brick_async(
         __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result,
-        oneapi::dpl::__internal::__brick_copy<::std::decay_t<decltype(__dispatch_tag)>, _ExecutionPolicy>{});
+        oneapi::dpl::__internal::__brick_copy<decltype(__dispatch_tag), _ExecutionPolicy>{});
     return ret_val;
 }
 
diff --git a/include/oneapi/dpl/pstl/glue_algorithm_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
index fa9bbb43051..0d17726cc24 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_impl.h
@@ -267,11 +267,11 @@ template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterato
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy, _ForwardIterator2>
 copy(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _ForwardIterator1 __last, _ForwardIterator2 __result)
 {
-    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first, __result);
+    auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first, __result);
 
     return oneapi::dpl::__internal::__pattern_walk2_brick(
         __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result,
-        oneapi::dpl::__internal::__brick_copy<::std::decay_t<decltype(__dispatch_tag)>, _ExecutionPolicy>{});
+        oneapi::dpl::__internal::__brick_copy<decltype(__dispatch_tag), _ExecutionPolicy>{});
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator1, class _Size, class _ForwardIterator2>
@@ -280,11 +280,11 @@ copy_n(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _Size __n, _Forward
 {
     using _DecayedExecutionPolicy = ::std::decay_t<_ExecutionPolicy>;
 
-    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first, __result);
+    auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first, __result);
 
     return oneapi::dpl::__internal::__pattern_walk2_brick_n(
         __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __n, __result,
-        oneapi::dpl::__internal::__brick_copy_n<::std::decay_t<decltype(__dispatch_tag)>, _DecayedExecutionPolicy>{});
+        oneapi::dpl::__internal::__brick_copy_n<decltype(__dispatch_tag), _DecayedExecutionPolicy>{});
 }
 
 template <class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Predicate>
@@ -814,11 +814,11 @@ move(_ExecutionPolicy&& __exec, _ForwardIterator1 __first, _ForwardIterator1 __l
 {
     using _DecayedExecutionPolicy = ::std::decay_t<_ExecutionPolicy>;
 
-    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first, __d_first);
+    auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first, __d_first);
 
     return oneapi::dpl::__internal::__pattern_walk2_brick(
         __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __d_first,
-        oneapi::dpl::__internal::__brick_move<::std::decay_t<decltype(__dispatch_tag)>, _DecayedExecutionPolicy>{});
+        oneapi::dpl::__internal::__brick_move<decltype(__dispatch_tag), _DecayedExecutionPolicy>{});
 }
 
 // [partial.sort]
diff --git a/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h b/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
index 3796812f620..4e51d0e4e2a 100644
--- a/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
+++ b/include/oneapi/dpl/pstl/glue_algorithm_ranges_impl.h
@@ -249,11 +249,11 @@ template <typename _ExecutionPolicy, typename _Range1, typename _Range2>
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy>
 copy(_ExecutionPolicy&& __exec, _Range1&& __rng, _Range2&& __result)
 {
-    const auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend(__exec, __rng, __result);
+    auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend(__exec, __rng, __result);
 
     oneapi::dpl::__internal::__ranges::__pattern_walk_n(
         __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
-        oneapi::dpl::__internal::__brick_copy<::std::decay_t<decltype(__dispatch_tag)>, _ExecutionPolicy>{},
+        oneapi::dpl::__internal::__brick_copy<decltype(__dispatch_tag), _ExecutionPolicy>{},
         views::all_read(::std::forward<_Range1>(__rng)), views::all_write(::std::forward<_Range2>(__result)));
 }
 
@@ -606,13 +606,13 @@ template <typename _ExecutionPolicy, typename _Range1, typename _Range2>
 oneapi::dpl::__internal::__enable_if_execution_policy<_ExecutionPolicy>
 move(_ExecutionPolicy&& __exec, _Range1&& __rng1, _Range2&& __rng2)
 {
-    const auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend(__exec, __rng1, __rng2);
+    auto __dispatch_tag = oneapi::dpl::__ranges::__select_backend(__exec, __rng1, __rng2);
 
     using _DecayedExecutionPolicy = ::std::decay_t<_ExecutionPolicy>;
 
     oneapi::dpl::__internal::__ranges::__pattern_walk_n(
         __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec),
-        oneapi::dpl::__internal::__brick_move<::std::decay_t<decltype(__dispatch_tag)>, _DecayedExecutionPolicy>{},
+        oneapi::dpl::__internal::__brick_move<decltype(__dispatch_tag), _DecayedExecutionPolicy>{},
         views::all_read(::std::forward<_Range1>(__rng1)), views::all_write(::std::forward<_Range2>(__rng2)));
 }
 
diff --git a/include/oneapi/dpl/pstl/glue_memory_impl.h b/include/oneapi/dpl/pstl/glue_memory_impl.h
index 6c207024777..fac93889dfb 100644
--- a/include/oneapi/dpl/pstl/glue_memory_impl.h
+++ b/include/oneapi/dpl/pstl/glue_memory_impl.h
@@ -45,13 +45,13 @@ uninitialized_copy(_ExecutionPolicy&& __exec, _InputIterator __first, _InputIter
     typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _ValueType2;
     typedef ::std::decay_t<_ExecutionPolicy> _DecayedExecutionPolicy;
 
-    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first, __result);
+    auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first, __result);
 
     if constexpr (::std::is_trivial_v<_ValueType1> && ::std::is_trivial_v<_ValueType2>)
     {
         return oneapi::dpl::__internal::__pattern_walk2_brick(
             __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result,
-            oneapi::dpl::__internal::__brick_copy<::std::decay_t<decltype(__dispatch_tag)>, _DecayedExecutionPolicy>{});
+            oneapi::dpl::__internal::__brick_copy<decltype(__dispatch_tag), _DecayedExecutionPolicy>{});
     }
     else
     {
@@ -69,14 +69,13 @@ uninitialized_copy_n(_ExecutionPolicy&& __exec, _InputIterator __first, _Size __
     typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _ValueType2;
     typedef ::std::decay_t<_ExecutionPolicy> _DecayedExecutionPolicy;
 
-    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first, __result);
+    auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first, __result);
 
     if constexpr (::std::is_trivial_v<_ValueType1> && ::std::is_trivial_v<_ValueType2>)
     {
         return oneapi::dpl::__internal::__pattern_walk2_brick_n(
             __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __n, __result,
-            oneapi::dpl::__internal::__brick_copy_n<::std::decay_t<decltype(__dispatch_tag)>,
-                                                    _DecayedExecutionPolicy>{});
+            oneapi::dpl::__internal::__brick_copy_n<decltype(__dispatch_tag), _DecayedExecutionPolicy>{});
     }
     else
     {
@@ -96,13 +95,13 @@ uninitialized_move(_ExecutionPolicy&& __exec, _InputIterator __first, _InputIter
     typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _ValueType2;
     typedef ::std::decay_t<_ExecutionPolicy> _DecayedExecutionPolicy;
 
-    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first, __result);
+    auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first, __result);
 
     if constexpr (::std::is_trivial_v<_ValueType1> && ::std::is_trivial_v<_ValueType2>)
     {
         return oneapi::dpl::__internal::__pattern_walk2_brick(
             __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last, __result,
-            oneapi::dpl::__internal::__brick_copy<::std::decay_t<decltype(__dispatch_tag)>, _DecayedExecutionPolicy>{});
+            oneapi::dpl::__internal::__brick_copy<decltype(__dispatch_tag), _DecayedExecutionPolicy>{});
     }
     else
     {
@@ -120,14 +119,13 @@ uninitialized_move_n(_ExecutionPolicy&& __exec, _InputIterator __first, _Size __
     typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _ValueType2;
     typedef ::std::decay_t<_ExecutionPolicy> _DecayedExecutionPolicy;
 
-    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first, __result);
+    auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first, __result);
 
     if constexpr (::std::is_trivial_v<_ValueType1> && ::std::is_trivial_v<_ValueType2>)
     {
         return oneapi::dpl::__internal::__pattern_walk2_brick_n(
             __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __n, __result,
-            oneapi::dpl::__internal::__brick_copy_n<::std::decay_t<decltype(__dispatch_tag)>,
-                                                    _DecayedExecutionPolicy>{});
+            oneapi::dpl::__internal::__brick_copy_n<decltype(__dispatch_tag), _DecayedExecutionPolicy>{});
     }
     else
     {
@@ -146,14 +144,14 @@ uninitialized_fill(_ExecutionPolicy&& __exec, _ForwardIterator __first, _Forward
     typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _ValueType;
     typedef ::std::decay_t<_ExecutionPolicy> _DecayedExecutionPolicy;
 
-    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first);
+    auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first);
 
     if constexpr (::std::is_arithmetic_v<_ValueType>)
     {
         oneapi::dpl::__internal::__pattern_walk_brick(
             __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
-            oneapi::dpl::__internal::__brick_fill<::std::decay_t<decltype(__dispatch_tag)>, _DecayedExecutionPolicy,
-                                                  _ValueType>{_ValueType(__value)});
+            oneapi::dpl::__internal::__brick_fill<decltype(__dispatch_tag), _DecayedExecutionPolicy, _ValueType>{
+                _ValueType(__value)});
     }
     else
     {
@@ -170,14 +168,14 @@ uninitialized_fill_n(_ExecutionPolicy&& __exec, _ForwardIterator __first, _Size
     typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _ValueType;
     typedef ::std::decay_t<_ExecutionPolicy> _DecayedExecutionPolicy;
 
-    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first);
+    auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first);
 
     if constexpr (::std::is_arithmetic_v<_ValueType>)
     {
         return oneapi::dpl::__internal::__pattern_walk_brick_n(
             __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __n,
-            oneapi::dpl::__internal::__brick_fill_n<::std::decay_t<decltype(__dispatch_tag)>, _DecayedExecutionPolicy,
-                                                    _ValueType>{_ValueType(__value)});
+            oneapi::dpl::__internal::__brick_fill_n<decltype(__dispatch_tag), _DecayedExecutionPolicy, _ValueType>{
+                _ValueType(__value)});
     }
     else
     {
@@ -308,14 +306,14 @@ uninitialized_value_construct(_ExecutionPolicy&& __exec, _ForwardIterator __firs
     typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _ValueType;
     typedef ::std::decay_t<_ExecutionPolicy> _DecayedExecutionPolicy;
 
-    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first);
+    auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first);
 
     if constexpr (::std::is_trivial_v<_ValueType>)
     {
         oneapi::dpl::__internal::__pattern_walk_brick(
             __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
-            oneapi::dpl::__internal::__brick_fill<::std::decay_t<decltype(__dispatch_tag)>, _DecayedExecutionPolicy,
-                                                  _ValueType>{_ValueType()});
+            oneapi::dpl::__internal::__brick_fill<decltype(__dispatch_tag), _DecayedExecutionPolicy, _ValueType>{
+                _ValueType()});
     }
     else
     {
@@ -332,14 +330,14 @@ uninitialized_value_construct_n(_ExecutionPolicy&& __exec, _ForwardIterator __fi
     typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _ValueType;
     typedef ::std::decay_t<_ExecutionPolicy> _DecayedExecutionPolicy;
 
-    const auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first);
+    auto __dispatch_tag = oneapi::dpl::__internal::__select_backend(__exec, __first);
 
     if constexpr (::std::is_trivial_v<_ValueType>)
     {
         return oneapi::dpl::__internal::__pattern_walk_brick_n(
             __dispatch_tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __n,
-            oneapi::dpl::__internal::__brick_fill_n<::std::decay_t<decltype(__dispatch_tag)>, _DecayedExecutionPolicy,
-                                                    _ValueType>{_ValueType()});
+            oneapi::dpl::__internal::__brick_fill_n<decltype(__dispatch_tag), _DecayedExecutionPolicy, _ValueType>{
+                _ValueType()});
     }
     else
     {

From 0146aa6a315c682db49ac17ec31487e8722758d5 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 13 Mar 2024 18:43:43 +0100
Subject: [PATCH 548/566] 
 include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h - fix review comment:
 implement __pattern_fill_n through __pattern_fill

---
 include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h | 7 +------
 1 file changed, 1 insertion(+), 6 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index ac416480590..eda8ba781af 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -322,12 +322,7 @@ _ForwardIterator
 __pattern_fill_n(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _ForwardIterator __first, _Size __count,
                  const _T& __value)
 {
-    // TODO: is this new implementation are correct?
-    // Previously we hadn't hetero impl for __pattern_fill_n
-
-    return __pattern_walk1_n(__tag, ::std::forward<_ExecutionPolicy>(__exec),
-                             __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::write>(__first),
-                             __count, fill_functor<_T>{__value});
+    return __pattern_fill(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __first + __count, __value);
 }
 
 //------------------------------------------------------------------------

From 726e5e304ee766d80df3591ca485822e5ee09b04 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 13 Mar 2024 18:44:55 +0100
Subject: [PATCH 549/566] 
 include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h - fix review comment:
 implement __pattern_generate_n through __pattern_generate

---
 include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h | 7 +------
 1 file changed, 1 insertion(+), 6 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index eda8ba781af..e07fd975db3 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -359,12 +359,7 @@ _ForwardIterator
 __pattern_generate_n(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _ForwardIterator __first,
                      _Size __count, _Generator __g)
 {
-    // TODO: is this new implementation are correct?
-    // Previously we hadn't hetero impl for __pattern_generate_n
-
-    return __pattern_walk1_n(__tag, ::std::forward<_ExecutionPolicy>(__exec),
-                             __par_backend_hetero::make_iter_mode<__par_backend_hetero::access_mode::write>(__first),
-                             __count, generate_functor<_Generator>{__g});
+    return __pattern_generate(__tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __first + __count, __g);
 }
 
 //------------------------------------------------------------------------

From cd807f1413fe4de8823edb8d57e24493edbaab7a Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 13 Mar 2024 18:53:25 +0100
Subject: [PATCH 550/566] include/oneapi/dpl/pstl/omp/parallel_scan.h - fix
 review comment

---
 include/oneapi/dpl/pstl/omp/parallel_scan.h | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/include/oneapi/dpl/pstl/omp/parallel_scan.h b/include/oneapi/dpl/pstl/omp/parallel_scan.h
index 4c2ff2bc9e3..75525246092 100644
--- a/include/oneapi/dpl/pstl/omp/parallel_scan.h
+++ b/include/oneapi/dpl/pstl/omp/parallel_scan.h
@@ -89,7 +89,8 @@ __parallel_strict_scan_body(_ExecutionPolicy&& __exec, _Index __n, _Tp __initial
     const _Index __slack = 4;
     _Index __tilesize = (__n - 1) / (__slack * __p) + 1;
     _Index __m = (__n - 1) / __tilesize;
-    __buffer<oneapi::dpl::__internal::__omp_backend_tag, _ExecutionPolicy, _Tp> __buf(__exec, __m + 1);
+    __buffer<oneapi::dpl::__internal::__omp_backend_tag, _ExecutionPolicy, _Tp> __buf(
+        ::std::forward<_ExecutionPolicy>(__exec), __m + 1);
     _Tp* __r = __buf.get();
 
     oneapi::dpl::__omp_backend::__upsweep(_Index(0), _Index(__m + 1), __tilesize, __r, __n - __m * __tilesize, __reduce,

From 96621e99bd5cee2320e3d45780f7e6af44b7e738 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 13 Mar 2024 18:55:44 +0100
Subject: [PATCH 551/566] 
 test/parallel_api/experimental/for_loop_reduction.pass.cpp - revert changes
 from test

---
 .../experimental/for_loop_reduction.pass.cpp     | 16 ++++++----------
 1 file changed, 6 insertions(+), 10 deletions(-)

diff --git a/test/parallel_api/experimental/for_loop_reduction.pass.cpp b/test/parallel_api/experimental/for_loop_reduction.pass.cpp
index 7f6b2b9e7f5..cfd918ff2bd 100644
--- a/test/parallel_api/experimental/for_loop_reduction.pass.cpp
+++ b/test/parallel_api/experimental/for_loop_reduction.pass.cpp
@@ -77,10 +77,8 @@ test()
     {
         Sequence<T> in_out(n, [](long int k) { return T(k % 5 != 1 ? 3 * k - 7 : 0); });
         Sequence<T> expected = in_out;
-
-        // for_loop staff is implemented for the host policies only
-        invoke_on_all_host_policies()(test_body(), in_out.begin(), in_out.end(), expected.begin(), expected.end(),
-                                      in_out.size());
+        invoke_on_all_policies<>()(test_body(), in_out.begin(), in_out.end(), expected.begin(), expected.end(),
+                                   in_out.size());
     }
 }
 
@@ -178,12 +176,10 @@ test_predefined(::std::initializer_list<T> init_list)
     // Just arbitrary numbers
     Sequence<T> in_out = init_list;
     Sequence<T> expected = in_out;
-
-    // for_loop staff is implemented for the host policies only
-    invoke_on_all_host_policies()(test_body_predefined(), in_out.begin(), in_out.end(), expected.begin(),
-                                  expected.end(), in_out.size());
-    invoke_on_all_host_policies()(test_body_predefined_bits(), in_out.begin(), in_out.end(), expected.begin(),
-                                  expected.end(), in_out.size());
+    invoke_on_all_policies<>()(test_body_predefined(), in_out.begin(), in_out.end(), expected.begin(), expected.end(),
+                               in_out.size());
+    invoke_on_all_policies<>()(test_body_predefined_bits(), in_out.begin(), in_out.end(), expected.begin(),
+                               expected.end(), in_out.size());
 }
 
 void

From d97ff4782b7347ba6c1ab24905d2b77b7745244f Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 13 Mar 2024 19:00:30 +0100
Subject: [PATCH 552/566] test/parallel_api/experimental/for_loop.pass.cpp -
 revert changes from test

---
 test/parallel_api/experimental/for_loop.pass.cpp | 10 ++++------
 1 file changed, 4 insertions(+), 6 deletions(-)

diff --git a/test/parallel_api/experimental/for_loop.pass.cpp b/test/parallel_api/experimental/for_loop.pass.cpp
index 725d2bc9747..8c59ffe1afd 100644
--- a/test/parallel_api/experimental/for_loop.pass.cpp
+++ b/test/parallel_api/experimental/for_loop.pass.cpp
@@ -296,9 +296,8 @@ test_for_loop()
         Sequence<T> in_out(n, Gen<T>());
         Sequence<T> expected = in_out;
 
-        // for_loop staff is implemented for the host policies only
-        invoke_on_all_host_policies()(test_for_loop_impl(), in_out.begin(), in_out.end(), expected.begin(),
-                                      expected.end(), in_out.size());
+        invoke_on_all_policies<>()(test_for_loop_impl(), in_out.begin(), in_out.end(), expected.begin(), expected.end(),
+                                   in_out.size());
     }
 }
 
@@ -314,9 +313,8 @@ test_for_loop_strided()
         ::std::vector<size_t> strides = {1, 2, 10, n > 1 ? n - 1 : 1, n > 0 ? n : 1, n + 1};
         for (size_t stride : strides)
         {
-            // for_loop staff is implemented for the host policies only
-            invoke_on_all_host_policies()(test_for_loop_strided_impl(), in_out.begin(), in_out.end(), expected.begin(),
-                                          expected.end(), in_out.size(), stride);
+            invoke_on_all_policies<>()(test_for_loop_strided_impl(), in_out.begin(), in_out.end(), expected.begin(),
+                                       expected.end(), in_out.size(), stride);
         }
     }
 }

From 079a7f5f60a5fd95b727f09821780aff389f9bde Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 13 Mar 2024 19:00:39 +0100
Subject: [PATCH 553/566] 
 test/parallel_api/experimental/for_loop_induction.pass.cpp - revert changes
 from test

---
 test/parallel_api/experimental/for_loop_induction.pass.cpp | 6 ++----
 1 file changed, 2 insertions(+), 4 deletions(-)

diff --git a/test/parallel_api/experimental/for_loop_induction.pass.cpp b/test/parallel_api/experimental/for_loop_induction.pass.cpp
index e01569c58a1..505ceb54e1e 100644
--- a/test/parallel_api/experimental/for_loop_induction.pass.cpp
+++ b/test/parallel_api/experimental/for_loop_induction.pass.cpp
@@ -153,10 +153,8 @@ test()
     {
         Sequence<T> in_out(n, [](long int k) { return T(k % 5 != 1 ? 3 * k - 7 : 0); });
         Sequence<T> expected = in_out;
-
-        // for_loop staff is implemented for the host policies only
-        invoke_on_all_host_policies()(test_body(), in_out.begin(), in_out.end(), expected.begin(), expected.end(),
-                                      in_out.size());
+        invoke_on_all_policies<>()(test_body(), in_out.begin(), in_out.end(), expected.begin(), expected.end(),
+                                   in_out.size());
     }
 }
 

From ee0433e855f36f96cf14f006508e21769bfe7645 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 14 Mar 2024 09:36:08 +0100
Subject: [PATCH 554/566] Revert changes from __buffer_impl classes (partial)

---
 .../hetero/dpcpp/parallel_backend_sycl_utils.h   | 16 +++++++---------
 include/oneapi/dpl/pstl/omp/util.h               | 14 ++++++--------
 .../oneapi/dpl/pstl/parallel_backend_serial.h    | 14 ++++++--------
 include/oneapi/dpl/pstl/parallel_backend_tbb.h   | 14 ++++++--------
 4 files changed, 25 insertions(+), 33 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_utils.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_utils.h
index 40b4b584c04..62025411b24 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_utils.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_utils.h
@@ -376,20 +376,18 @@ struct __local_buffer<sycl::buffer<::std::tuple<_T...>, __dim, _AllocT>>
     using type = sycl::buffer<oneapi::dpl::__internal::tuple<_T...>, __dim, _AllocT>;
 };
 
-template <typename _BackendTag, typename _ExecutionPolicy, typename _Tp, typename = void>
-class __buffer_impl;
-
 // impl for sycl::buffer<...>
-template <typename _BackendTag, typename _ExecutionPolicy, typename _Tp>
-class __buffer_impl<_BackendTag, _ExecutionPolicy, _Tp,
-                    ::std::enable_if_t<oneapi::dpl::__internal::__is_hetero_backend_tag_v<_BackendTag>>>
+template <typename _ExecutionPolicy, typename _T>
+class __buffer_impl
 {
   private:
-    using __container_t = typename __local_buffer<sycl::buffer<_Tp>>::type;
+    using __container_t = typename __local_buffer<sycl::buffer<_T>>::type;
 
     __container_t __container;
 
   public:
+    static_assert(::std::is_same_v<_ExecutionPolicy, ::std::decay_t<_ExecutionPolicy>>);
+
     __buffer_impl(_ExecutionPolicy /*__exec*/, ::std::size_t __n_elements) : __container{sycl::range<1>(__n_elements)}
     {
     }
@@ -450,8 +448,8 @@ struct __memobj_traits<_T*>
 
 } // namespace __internal
 
-template <typename _BackendTag, typename _ExecutionPolicy, typename _T>
-using __buffer = __internal::__buffer_impl<_BackendTag, ::std::decay_t<_ExecutionPolicy>, _T>;
+template <typename _ExecutionPolicy, typename _T>
+using __buffer = __internal::__buffer_impl<::std::decay_t<_ExecutionPolicy>, _T>;
 
 template <typename T>
 struct __repacked_tuple
diff --git a/include/oneapi/dpl/pstl/omp/util.h b/include/oneapi/dpl/pstl/omp/util.h
index c616aedc206..92689638121 100644
--- a/include/oneapi/dpl/pstl/omp/util.h
+++ b/include/oneapi/dpl/pstl/omp/util.h
@@ -57,12 +57,8 @@ __cancel_execution(oneapi::dpl::__internal::__omp_backend_tag)
 // raw buffer
 //------------------------------------------------------------------------
 
-template <typename _BackendTag, typename _ExecutionPolicy, typename _Tp, typename = void>
-class __buffer_impl;
-
-template <typename _BackendTag, typename _ExecutionPolicy, typename _Tp>
-class __buffer_impl<_BackendTag, _ExecutionPolicy, _Tp,
-                    ::std::enable_if_t<::std::is_same_v<_BackendTag, oneapi::dpl::__internal::__omp_backend_tag>>>
+template <typename _ExecutionPolicy, typename _Tp>
+class __buffer_impl
 {
     std::allocator<_Tp> __allocator_;
     _Tp* __ptr_;
@@ -72,6 +68,8 @@ class __buffer_impl<_BackendTag, _ExecutionPolicy, _Tp,
     operator=(const __buffer_impl&) = delete;
 
   public:
+    static_assert(::std::is_same_v<_ExecutionPolicy, ::std::decay_t<_ExecutionPolicy>>);
+
     __buffer_impl(_ExecutionPolicy /*__exec*/, std::size_t __n)
         : __allocator_(), __ptr_(__allocator_.allocate(__n)), __buf_size_(__n)
     {
@@ -87,8 +85,8 @@ class __buffer_impl<_BackendTag, _ExecutionPolicy, _Tp,
     ~__buffer_impl() { __allocator_.deallocate(__ptr_, __buf_size_); }
 };
 
-template <typename _BackendTag, typename _ExecutionPolicy, typename _Tp>
-using __buffer = __buffer_impl<_BackendTag, ::std::decay_t<_ExecutionPolicy>, _Tp>;
+template <typename _ExecutionPolicy, typename _Tp>
+using __buffer = __buffer_impl<::std::decay_t<_ExecutionPolicy>, _Tp>;
 
 // Preliminary size of each chunk: requires further discussion
 constexpr std::size_t __default_chunk_size = 2048;
diff --git a/include/oneapi/dpl/pstl/parallel_backend_serial.h b/include/oneapi/dpl/pstl/parallel_backend_serial.h
index 867b925ff91..b1fb6d9f39c 100644
--- a/include/oneapi/dpl/pstl/parallel_backend_serial.h
+++ b/include/oneapi/dpl/pstl/parallel_backend_serial.h
@@ -32,12 +32,8 @@ namespace dpl
 namespace __serial_backend
 {
 
-template <typename _BackendTag, typename _ExecutionPolicy, typename _Tp, typename = void>
-class __buffer_impl;
-
-template <typename _BackendTag, typename _ExecutionPolicy, typename _Tp>
-class __buffer_impl<_BackendTag, _ExecutionPolicy, _Tp,
-                    ::std::enable_if_t<::std::is_same_v<_BackendTag, oneapi::dpl::__internal::__serial_backend_tag>>>
+template <typename _ExecutionPolicy, typename _Tp>
+class __buffer_impl
 {
     ::std::allocator<_Tp> __allocator_;
     _Tp* __ptr_;
@@ -47,6 +43,8 @@ class __buffer_impl<_BackendTag, _ExecutionPolicy, _Tp,
     operator=(const __buffer_impl&) = delete;
 
   public:
+    static_assert(::std::is_same_v<_ExecutionPolicy, ::std::decay_t<_ExecutionPolicy>>);
+
     __buffer_impl(_ExecutionPolicy /*__exec*/, ::std::size_t __n)
         : __allocator_(), __ptr_(__allocator_.allocate(__n)), __buf_size_(__n)
     {
@@ -61,8 +59,8 @@ class __buffer_impl<_BackendTag, _ExecutionPolicy, _Tp,
     ~__buffer_impl() { __allocator_.deallocate(__ptr_, __buf_size_); }
 };
 
-template <typename _BackendTag, typename _ExecutionPolicy, typename _Tp>
-using __buffer = __buffer_impl<_BackendTag, ::std::decay_t<_ExecutionPolicy>, _Tp>;
+template <typename _ExecutionPolicy, typename _Tp>
+using __buffer = __buffer_impl<::std::decay_t<_ExecutionPolicy>, _Tp>;
 
 inline void
 __cancel_execution(oneapi::dpl::__internal::__serial_backend_tag)
diff --git a/include/oneapi/dpl/pstl/parallel_backend_tbb.h b/include/oneapi/dpl/pstl/parallel_backend_tbb.h
index c9cc0a8ff52..81707201221 100644
--- a/include/oneapi/dpl/pstl/parallel_backend_tbb.h
+++ b/include/oneapi/dpl/pstl/parallel_backend_tbb.h
@@ -49,17 +49,13 @@ namespace dpl
 namespace __tbb_backend
 {
 
-template <typename _BackendTag, typename _ExecutionPolicy, typename _Tp, typename = void>
-class __buffer_impl;
-
 //! Raw memory buffer with automatic freeing and no exceptions.
 /** Some of our algorithms need to start with raw memory buffer,
 not an initialize array, because initialization/destruction
 would make the span be at least O(N). */
 // tbb::allocator can improve performance in some cases.
-template <typename _BackendTag, typename _ExecutionPolicy, typename _Tp>
-class __buffer_impl<_BackendTag, _ExecutionPolicy, _Tp,
-                    ::std::enable_if_t<::std::is_same_v<_BackendTag, oneapi::dpl::__internal::__tbb_backend_tag>>>
+template <typename _ExecutionPolicy, typename _Tp>
+class __buffer_impl
 {
     tbb::tbb_allocator<_Tp> _M_allocator;
     _Tp* _M_ptr;
@@ -69,6 +65,8 @@ class __buffer_impl<_BackendTag, _ExecutionPolicy, _Tp,
     operator=(const __buffer_impl&) = delete;
 
   public:
+    static_assert(::std::is_same_v<_ExecutionPolicy, ::std::decay_t<_ExecutionPolicy>>);
+
     //! Try to obtain buffer of given size to store objects of _Tp type
     __buffer_impl(_ExecutionPolicy /*__exec*/, const ::std::size_t __n)
         : _M_allocator(), _M_ptr(_M_allocator.allocate(__n)), _M_buf_size(__n)
@@ -86,8 +84,8 @@ class __buffer_impl<_BackendTag, _ExecutionPolicy, _Tp,
     ~__buffer_impl() { _M_allocator.deallocate(_M_ptr, _M_buf_size); }
 };
 
-template <typename _BackendTag, typename _ExecutionPolicy, typename _Tp>
-using __buffer = __buffer_impl<_BackendTag, ::std::decay_t<_ExecutionPolicy>, _Tp>;
+template <typename _ExecutionPolicy, typename _Tp>
+using __buffer = __buffer_impl<::std::decay_t<_ExecutionPolicy>, _Tp>;
 
 // Wrapper for tbb::task
 inline void

From 02a6bae0faa595c4120668151190ea67f44a74a0 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 14 Mar 2024 09:48:45 +0100
Subject: [PATCH 555/566] include/oneapi/dpl/pstl/execution_impl.h - revert
 changes from __buffer_impl classes (partial) : __buffer_backend_tag_type
 trait net required anymore

---
 include/oneapi/dpl/pstl/execution_impl.h | 16 ----------------
 1 file changed, 16 deletions(-)

diff --git a/include/oneapi/dpl/pstl/execution_impl.h b/include/oneapi/dpl/pstl/execution_impl.h
index 1a7b2aa048a..133717bf68e 100644
--- a/include/oneapi/dpl/pstl/execution_impl.h
+++ b/include/oneapi/dpl/pstl/execution_impl.h
@@ -159,22 +159,6 @@ template <class _Tag>
 inline constexpr bool __is_host_dispatch_tag_v =
     __is_serial_tag_v<_Tag> || __is_parallel_forward_tag_v<_Tag> || __is_parallel_tag_v<_Tag>;
 
-//----------------------------------------------------------
-// __buffer_backend_tag_type
-//----------------------------------------------------------
-
-template <class _Tag>
-struct __buffer_backend_tag_type
-{
-    using type = typename _Tag::__backend_tag;
-};
-
-template <class _IsVector>
-struct __buffer_backend_tag_type<__serial_tag<_IsVector>>
-{
-    using type = __par_backend_tag;
-};
-
 } // namespace __internal
 } // namespace dpl
 } // namespace oneapi

From 48aec87fdf1b49dbcc5785172274d300afcdea47 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 14 Mar 2024 09:48:59 +0100
Subject: [PATCH 556/566] Revert changes from __buffer_impl classes (partial) -
 calls

---
 .../internal/exclusive_scan_by_segment_impl.h | 10 ++++----
 .../internal/inclusive_scan_by_segment_impl.h |  6 ++---
 .../dpl/internal/reduce_by_segment_impl.h     | 20 ++++++----------
 .../dpl/internal/scan_by_segment_impl.h       |  6 ++---
 include/oneapi/dpl/pstl/algorithm_impl.h      | 20 ++++++++--------
 .../dpl/pstl/hetero/algorithm_impl_hetero.h   | 24 +++++++++----------
 .../hetero/algorithm_ranges_impl_hetero.h     | 16 +++++--------
 .../pstl/hetero/dpcpp/parallel_backend_sycl.h |  8 +++----
 .../dpcpp/parallel_backend_sycl_histogram.h   |  3 +--
 .../dpcpp/parallel_backend_sycl_radix_sort.h  |  4 +---
 .../dpl/pstl/hetero/numeric_impl_hetero.h     |  2 +-
 include/oneapi/dpl/pstl/omp/parallel_scan.h   |  3 +--
 .../oneapi/dpl/pstl/parallel_backend_tbb.h    |  6 ++---
 13 files changed, 51 insertions(+), 77 deletions(-)

diff --git a/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h b/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h
index 30a9e9fd3a3..cd33872aa0d 100644
--- a/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/exclusive_scan_by_segment_impl.h
@@ -46,8 +46,6 @@ pattern_exclusive_scan_by_segment(_Tag, Policy&& policy, InputIterator1 first1,
 {
     static_assert(__internal::__is_host_dispatch_tag_v<_Tag>);
 
-    using __buffer_backend_tag = typename __internal::__buffer_backend_tag_type<_Tag>::type;
-
     const auto n = ::std::distance(first1, last1);
 
     // Check for empty and single element ranges
@@ -66,7 +64,7 @@ pattern_exclusive_scan_by_segment(_Tag, Policy&& policy, InputIterator1 first1,
     InputIterator2 last2 = first2 + n;
 
     // compute head flags
-    oneapi::dpl::__par_backend::__buffer<__buffer_backend_tag, Policy, FlagType> _flags(policy, n);
+    oneapi::dpl::__par_backend::__buffer<Policy, FlagType> _flags(policy, n);
     auto flags = _flags.get();
     flags[0] = 1;
 
@@ -74,7 +72,7 @@ pattern_exclusive_scan_by_segment(_Tag, Policy&& policy, InputIterator1 first1,
               oneapi::dpl::__internal::__not_pred<BinaryPredicate>(binary_pred));
 
     // shift input one to the right and initialize segments with init
-    oneapi::dpl::__par_backend::__buffer<__buffer_backend_tag, Policy, OutputType> _temp(policy, n);
+    oneapi::dpl::__par_backend::__buffer<Policy, OutputType> _temp(policy, n);
     auto temp = _temp.get();
 
     temp[0] = init;
@@ -125,7 +123,7 @@ exclusive_scan_by_segment_impl(__internal::__hetero_tag<_BackendTag>, Policy&& p
     InputIterator2 last2 = first2 + n;
 
     // compute head flags
-    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, Policy, FlagType> _flags(policy, n);
+    oneapi::dpl::__par_backend_hetero::__buffer<Policy, FlagType> _flags(policy, n);
     {
         auto flag_buf = _flags.get_buffer();
         auto flags = flag_buf.get_host_access(sycl::read_write);
@@ -136,7 +134,7 @@ exclusive_scan_by_segment_impl(__internal::__hetero_tag<_BackendTag>, Policy&& p
               oneapi::dpl::__internal::__not_pred<BinaryPredicate>(binary_pred));
 
     // shift input one to the right and initialize segments with init
-    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, Policy, OutputType> _temp(policy, n);
+    oneapi::dpl::__par_backend_hetero::__buffer<Policy, OutputType> _temp(policy, n);
     {
         auto temp_buf = _temp.get_buffer();
         auto temp = temp_buf.get_host_access(sycl::read_write);
diff --git a/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h b/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h
index 4c7405d4c5d..22dc5cbf46b 100644
--- a/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/inclusive_scan_by_segment_impl.h
@@ -45,8 +45,6 @@ pattern_inclusive_scan_by_segment(_Tag, Policy&& policy, InputIterator1 first1,
 {
     static_assert(__internal::__is_host_dispatch_tag_v<_Tag>);
 
-    using __buffer_backend_tag = typename __internal::__buffer_backend_tag_type<_Tag>::type;
-
     const auto n = ::std::distance(first1, last1);
 
     // Check for empty and single element ranges
@@ -61,7 +59,7 @@ pattern_inclusive_scan_by_segment(_Tag, Policy&& policy, InputIterator1 first1,
     typedef unsigned int FlagType;
     typedef typename ::std::iterator_traits<InputIterator2>::value_type ValueType;
 
-    oneapi::dpl::__par_backend::__buffer<__buffer_backend_tag, Policy, FlagType> _mask(policy, n);
+    oneapi::dpl::__par_backend::__buffer<Policy, FlagType> _mask(policy, n);
     auto mask = _mask.get();
 
     mask[0] = 1;
@@ -111,7 +109,7 @@ inclusive_scan_by_segment_impl(__internal::__hetero_tag<_BackendTag>, Policy&& p
 
     FlagType initial_mask = 1;
 
-    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, Policy, FlagType> _mask(policy, n);
+    oneapi::dpl::__par_backend_hetero::__buffer<Policy, FlagType> _mask(policy, n);
     {
         auto mask_buf = _mask.get_buffer();
         auto mask = mask_buf.get_host_access(sycl::read_write);
diff --git a/include/oneapi/dpl/internal/reduce_by_segment_impl.h b/include/oneapi/dpl/internal/reduce_by_segment_impl.h
index 86252f6fe54..78986bca302 100644
--- a/include/oneapi/dpl/internal/reduce_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/reduce_by_segment_impl.h
@@ -87,8 +87,6 @@ reduce_by_segment_impl(_Tag, Policy&& policy, InputIterator1 first1, InputIterat
 {
     static_assert(__internal::__is_host_dispatch_tag_v<_Tag>);
 
-    using __buffer_backend_tag = typename __internal::__buffer_backend_tag_type<_Tag>::type;
-
     // The algorithm reduces values in [first2, first2 + (last1-first1)) where the associated
     // keys for the values are equal to the adjacent key. This function's implementation is a derivative work
     // and responsible for the second copyright notice in this header.
@@ -116,7 +114,7 @@ reduce_by_segment_impl(_Tag, Policy&& policy, InputIterator1 first1, InputIterat
 
     // buffer that is used to store a flag indicating if the associated key is not equal to
     // the next key, and thus its associated sum should be part of the final result
-    oneapi::dpl::__par_backend::__buffer<__buffer_backend_tag, Policy, FlagType> _mask(policy, n + 1);
+    oneapi::dpl::__par_backend::__buffer<Policy, FlagType> _mask(policy, n + 1);
     auto mask = _mask.get();
     mask[0] = 1;
 
@@ -132,11 +130,11 @@ reduce_by_segment_impl(_Tag, Policy&& policy, InputIterator1 first1, InputIterat
     // buffer stores the sums of values associated with a given key. Sums are copied with
     // a shift into result2, and the shift is computed at the same time as the sums, so the
     // sums can't be written to result2 directly.
-    oneapi::dpl::__par_backend::__buffer<__buffer_backend_tag, Policy, ValueType> _scanned_values(policy, n);
+    oneapi::dpl::__par_backend::__buffer<Policy, ValueType> _scanned_values(policy, n);
 
     // Buffer is used to store results of the scan of the mask. Values indicate which position
     // in result2 needs to be written with the scanned_values element.
-    oneapi::dpl::__par_backend::__buffer<__buffer_backend_tag, Policy, FlagType> _scanned_tail_flags(policy, n);
+    oneapi::dpl::__par_backend::__buffer<Policy, FlagType> _scanned_tail_flags(policy, n);
 
     // Compute the sum of the segments. scanned_tail_flags values are not used.
     inclusive_scan(policy, make_zip_iterator(first2, _mask.get()), make_zip_iterator(first2, _mask.get()) + n,
@@ -265,21 +263,17 @@ __sycl_reduce_by_segment(__internal::__hetero_tag<_BackendTag>, _ExecutionPolicy
 
     // intermediate reductions within a workgroup
     auto __partials =
-        oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, __val_type>(__exec, __n_groups)
-            .get_buffer();
+        oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, __val_type>(__exec, __n_groups).get_buffer();
 
-    auto __end_idx =
-        oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, __diff_type>(__exec, 1).get_buffer();
+    auto __end_idx = oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, __diff_type>(__exec, 1).get_buffer();
 
     // the number of segment ends found in each work group
     auto __seg_ends =
-        oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, __diff_type>(__exec, __n_groups)
-            .get_buffer();
+        oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, __diff_type>(__exec, __n_groups).get_buffer();
 
     // buffer that stores an exclusive scan of the results
     auto __seg_ends_scanned =
-        oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, __diff_type>(__exec, __n_groups)
-            .get_buffer();
+        oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, __diff_type>(__exec, __n_groups).get_buffer();
 
     // 1. Count the segment ends in each workgroup
     auto __seg_end_identification = __exec.queue().submit([&](sycl::handler& __cgh) {
diff --git a/include/oneapi/dpl/internal/scan_by_segment_impl.h b/include/oneapi/dpl/internal/scan_by_segment_impl.h
index 85903a93fdf..15606d05f8e 100644
--- a/include/oneapi/dpl/internal/scan_by_segment_impl.h
+++ b/include/oneapi/dpl/internal/scan_by_segment_impl.h
@@ -145,13 +145,11 @@ struct __sycl_scan_by_segment_impl
         ::std::size_t __n_groups = __internal::__dpl_ceiling_div(__n, __wgroup_size * __vals_per_item);
 
         auto __partials =
-            oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, __val_type>(__exec, __n_groups)
-                .get_buffer();
+            oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, __val_type>(__exec, __n_groups).get_buffer();
 
         // the number of segment ends found in each work group
         auto __seg_ends =
-            oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, bool>(__exec, __n_groups)
-                .get_buffer();
+            oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, bool>(__exec, __n_groups).get_buffer();
 
         // 1. Work group reduction
         auto __wg_scan = __exec.queue().submit([&](sycl::handler& __cgh) {
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index f8d7c10c7a3..1992b06cf0c 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -1323,7 +1323,7 @@ __pattern_copy_if(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomA
     const _DifferenceType __n = __last - __first;
     if (_DifferenceType(1) < __n)
     {
-        __par_backend::__buffer<__backend_tag, _ExecutionPolicy, bool> __mask_buf(__exec, __n);
+        __par_backend::__buffer<_ExecutionPolicy, bool> __mask_buf(__exec, __n);
         return __internal::__except_handler([&__exec, __n, __first, __result, __pred, &__mask_buf]() {
             bool* __mask = __mask_buf.get();
             _DifferenceType __m{};
@@ -1442,7 +1442,7 @@ __remove_elements(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Forward
     typedef typename ::std::iterator_traits<_ForwardIterator>::difference_type _DifferenceType;
     typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _Tp;
     _DifferenceType __n = __last - __first;
-    __par_backend::__buffer<__backend_tag, _ExecutionPolicy, bool> __mask_buf(__exec, __n);
+    __par_backend::__buffer<_ExecutionPolicy, bool> __mask_buf(__exec, __n);
     // 1. find a first iterator that should be removed
     return __internal::__except_handler([&]() {
         bool* __mask = __mask_buf.get();
@@ -1479,7 +1479,7 @@ __remove_elements(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Forward
         __n -= __min;
         __first += __min;
 
-        __par_backend::__buffer<__backend_tag, _ExecutionPolicy, _Tp> __buf(__exec, __n);
+        __par_backend::__buffer<_ExecutionPolicy, _Tp> __buf(__exec, __n);
         _Tp* __result = __buf.get();
         __mask += __min;
         _DifferenceType __m{};
@@ -1607,7 +1607,7 @@ __pattern_unique_copy(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Ran
     const _DifferenceType __n = __last - __first;
     if (_DifferenceType(2) < __n)
     {
-        __par_backend::__buffer<__backend_tag, _ExecutionPolicy, bool> __mask_buf(__exec, __n);
+        __par_backend::__buffer<_ExecutionPolicy, bool> __mask_buf(__exec, __n);
         if (_DifferenceType(2) < __n)
         {
             return __internal::__except_handler([&__exec, __n, __first, __result, __pred, &__mask_buf]() {
@@ -1855,7 +1855,7 @@ __pattern_rotate(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAc
     auto __m = __middle - __first;
     if (__m <= __n / 2)
     {
-        __par_backend::__buffer<__backend_tag, _ExecutionPolicy, _Tp> __buf(__exec, __n - __m);
+        __par_backend::__buffer<_ExecutionPolicy, _Tp> __buf(__exec, __n - __m);
         return __internal::__except_handler([&__exec, __n, __m, __first, __middle, __last, &__buf]() {
             _Tp* __result = __buf.get();
             __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __middle, __last,
@@ -1881,7 +1881,7 @@ __pattern_rotate(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAc
     }
     else
     {
-        __par_backend::__buffer<__backend_tag, _ExecutionPolicy, _Tp> __buf(__exec, __m);
+        __par_backend::__buffer<_ExecutionPolicy, _Tp> __buf(__exec, __m);
         return __internal::__except_handler([&__exec, __n, __m, __first, __middle, __last, &__buf]() {
             _Tp* __result = __buf.get();
             __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __middle,
@@ -2359,7 +2359,7 @@ __pattern_partition_copy(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _
     const _DifferenceType __n = __last - __first;
     if (_DifferenceType(1) < __n)
     {
-        __par_backend::__buffer<__backend_tag, _ExecutionPolicy, bool> __mask_buf(__exec, __n);
+        __par_backend::__buffer<_ExecutionPolicy, bool> __mask_buf(__exec, __n);
         return __internal::__except_handler([&__exec, __n, __first, __out_true, __out_false, __pred, &__mask_buf]() {
             bool* __mask = __mask_buf.get();
             _ReturnType __m{};
@@ -2585,7 +2585,7 @@ __pattern_partial_sort_copy(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec
         {
             typedef typename ::std::iterator_traits<_RandomAccessIterator1>::value_type _T1;
             typedef typename ::std::iterator_traits<_RandomAccessIterator2>::value_type _T2;
-            __par_backend::__buffer<__backend_tag, _ExecutionPolicy, _T1> __buf(__exec, __n1);
+            __par_backend::__buffer<_ExecutionPolicy, _T1> __buf(__exec, __n1);
             _T1* __r = __buf.get();
 
             __par_backend::__parallel_stable_sort(
@@ -3078,7 +3078,7 @@ __pattern_inplace_merge(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _R
 
     typedef typename ::std::iterator_traits<_RandomAccessIterator>::value_type _Tp;
     auto __n = __last - __first;
-    __par_backend::__buffer<__backend_tag, _ExecutionPolicy, _Tp> __buf(__exec, __n);
+    __par_backend::__buffer<_ExecutionPolicy, _Tp> __buf(__exec, __n);
     _Tp* __r = __buf.get();
     __internal::__except_handler([&]() {
         auto __move_values = [](_RandomAccessIterator __x, _Tp* __z) {
@@ -3212,7 +3212,7 @@ __parallel_set_op(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Forward
     const _DifferenceType __n1 = __last1 - __first1;
     const _DifferenceType __n2 = __last2 - __first2;
 
-    __par_backend::__buffer<__backend_tag, _ExecutionPolicy, _T> __buf(__exec, __size_func(__n1, __n2));
+    __par_backend::__buffer<_ExecutionPolicy, _T> __buf(__exec, __size_func(__n1, __n2));
 
     return __internal::__except_handler([&__exec, __n1, __first1, __last1, __first2, __last2, __result, __comp,
                                          __size_func, __set_op, &__buf]() {
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
index e07fd975db3..a630c1cc2dd 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h
@@ -1023,8 +1023,7 @@ __pattern_remove_if(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec,
 
     using _ValueType = typename ::std::iterator_traits<_Iterator>::value_type;
 
-    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _ValueType> __buf(__exec,
-                                                                                                 __last - __first);
+    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __buf(__exec, __last - __first);
     auto __copy_first = __buf.get();
 
     auto __copy_last = __pattern_copy_if(__tag, __exec, __first, __last, __copy_first, __not_pred<_Predicate>{__pred});
@@ -1045,8 +1044,7 @@ __pattern_unique(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _It
 
     using _ValueType = typename ::std::iterator_traits<_Iterator>::value_type;
 
-    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _ValueType> __buf(__exec,
-                                                                                                 __last - __first);
+    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __buf(__exec, __last - __first);
     auto __copy_first = __buf.get();
     auto __copy_last = __pattern_unique_copy(__tag, __exec, __first, __last, __copy_first, __pred);
 
@@ -1220,7 +1218,7 @@ __pattern_inplace_merge(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __ex
     assert(__first < __middle && __middle < __last);
 
     auto __n = __last - __first;
-    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _ValueType> __buf(__exec, __n);
+    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __buf(__exec, __n);
     auto __copy_first = __buf.get();
     auto __copy_last = __copy_first + __n;
 
@@ -1308,8 +1306,8 @@ __pattern_stable_partition(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& _
 
     auto __n = __last - __first;
 
-    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _ValueType> __true_buf(__exec, __n);
-    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _ValueType> __false_buf(__exec, __n);
+    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __true_buf(__exec, __n);
+    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __false_buf(__exec, __n);
     auto __true_result = __true_buf.get();
     auto __false_result = __false_buf.get();
 
@@ -1506,7 +1504,7 @@ __pattern_partial_sort_copy(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&&
         // - create a temporary buffer and copy all the elements from the input buffer there
         // - run partial sort on the temporary buffer
         // - copy k elements from the temporary buffer to the output buffer.
-        oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _ValueType> __buf(__exec, __in_size);
+        oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __buf(__exec, __in_size);
 
         auto __buf_first = __buf.get();
 
@@ -1621,7 +1619,7 @@ __pattern_rotate(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Iterator
 
     auto __keep = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read_write, _Iterator>();
     auto __buf = __keep(__first, __last);
-    auto __temp_buf = oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _Tp>(__exec, __n);
+    auto __temp_buf = oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _Tp>(__exec, __n);
 
     auto __temp_rng =
         oneapi::dpl::__ranges::all_view<_Tp, __par_backend_hetero::access_mode::write>(__temp_buf.get_buffer());
@@ -1704,7 +1702,7 @@ __pattern_hetero_set_op(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _F
         __comp, __n1, __n2};
 
     // temporary buffer to store boolean mask
-    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, int32_t> __mask_buf(__exec, __n1);
+    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, int32_t> __mask_buf(__exec, __n1);
 
     auto __keep1 =
         oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::read, _ForwardIterator1>();
@@ -1833,7 +1831,7 @@ __pattern_set_union(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec,
 
     // temporary buffer to store intermediate result
     const auto __n2 = __last2 - __first2;
-    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _ValueType> __diff(__exec, __n2);
+    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __diff(__exec, __n2);
     auto __buf = __diff.get();
 
     //1. Calc difference {2} \ {1}
@@ -1914,10 +1912,10 @@ __pattern_set_symmetric_difference(__hetero_tag<_BackendTag> __tag, _ExecutionPo
 
     // temporary buffers to store intermediate result
     const auto __n1 = __last1 - __first1;
-    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _ValueType> __diff_1(__exec, __n1);
+    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __diff_1(__exec, __n1);
     auto __buf_1 = __diff_1.get();
     const auto __n2 = __last2 - __first2;
-    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _ValueType> __diff_2(__exec, __n2);
+    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __diff_2(__exec, __n2);
     auto __buf_2 = __diff_2.get();
 
     //1. Calc difference {1} \ {2}
diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
index 227ff9761b5..f9f39a41b49 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
@@ -360,8 +360,7 @@ __pattern_scan_copy(__hetero_tag<_BackendTag>, _ExecutionPolicy&& __exec, _Range
     _DataAcc __get_data_op;
     _MaskAssigner __add_mask_op;
 
-    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, int32_t> __mask_buf(__exec,
-                                                                                                   __rng1.size());
+    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, int32_t> __mask_buf(__exec, __rng1.size());
 
     auto __res =
         __par_backend_hetero::__parallel_transform_scan_base(
@@ -414,7 +413,7 @@ __pattern_remove_if(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec,
 
     using _ValueType = oneapi::dpl::__internal::__value_t<_Range>;
 
-    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _ValueType> __buf(__exec, __rng.size());
+    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __buf(__exec, __rng.size());
     auto __copy_rng = oneapi::dpl::__ranges::views::all(__buf.get_buffer());
 
     auto __copy_last_id = __ranges::__pattern_copy_if(__tag, __exec, __rng, __copy_rng, __not_pred<_Predicate>{__pred},
@@ -462,7 +461,7 @@ __pattern_unique(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Ra
 
     using _ValueType = oneapi::dpl::__internal::__value_t<_Range>;
 
-    oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _ValueType> __buf(__exec, __rng.size());
+    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __buf(__exec, __rng.size());
     auto res_rng = oneapi::dpl::__ranges::views::all(__buf.get_buffer());
     auto res = __ranges::__pattern_unique_copy(__tag, __exec, __rng, res_rng, __pred,
                                                oneapi::dpl::__internal::__pstl_assign());
@@ -716,14 +715,11 @@ __pattern_reduce_by_segment(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&&
     // Round 1: reduce with extra indices added to avoid long segments
     // TODO: At threshold points check if the key is equal to the key at the previous threshold point, indicating a long sequence.
     // Skip a round of copy_if and reduces if there are none.
-    auto __idx = oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, __diff_type>(__exec, __n)
-                     .get_buffer();
+    auto __idx = oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, __diff_type>(__exec, __n).get_buffer();
     auto __tmp_out_keys =
-        oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, __key_type>(__exec, __n)
-            .get_buffer();
+        oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, __key_type>(__exec, __n).get_buffer();
     auto __tmp_out_values =
-        oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, __val_type>(__exec, __n)
-            .get_buffer();
+        oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, __val_type>(__exec, __n).get_buffer();
 
     // Replicating first element of keys view to be able to compare (i-1)-th and (i)-th key with aligned sequences,
     //  dropping the last key for the i-1 sequence.
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
index 853372c348c..d60d5b3626e 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h
@@ -871,9 +871,7 @@ __parallel_scan_copy(oneapi::dpl::__internal::__device_backend_tag __backend_tag
     _MaskAssigner __add_mask_op;
 
     // temporary buffer to store boolean mask
-    oneapi::dpl::__par_backend_hetero::__buffer<oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy,
-                                                int32_t>
-        __mask_buf(__exec, __n);
+    oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, int32_t> __mask_buf(__exec, __n);
 
     return __parallel_transform_scan_base(
         __backend_tag, ::std::forward<_ExecutionPolicy>(__exec),
@@ -1578,7 +1576,7 @@ struct __parallel_sort_submitter<_IdType, __internal::__optional_kernel_name<_Le
         });
 
         // 2. Merge sorting
-        oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _Tp> __temp_buf(__exec, __n);
+        oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _Tp> __temp_buf(__exec, __n);
         auto __temp = __temp_buf.get_buffer();
         bool __data_in_temp = false;
         _IdType __n_sorted = __leaf;
@@ -1699,7 +1697,7 @@ struct __parallel_partial_sort_submitter<__internal::__optional_kernel_name<_Glo
         _Size __n = __rng.size();
         assert(__n > 1);
 
-        oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _Tp> __temp_buf(__exec, __n);
+        oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _Tp> __temp_buf(__exec, __n);
         auto __temp = __temp_buf.get_buffer();
         _PRINT_INFO_IN_DEBUG_MODE(__exec);
 
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h
index e5d2c5a3a15..e7ac2ba50e1 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_histogram.h
@@ -426,8 +426,7 @@ struct __histogram_general_private_global_atomics_submitter<__internal::__option
             oneapi::dpl::__internal::__dpl_ceiling_div(__n, __work_group_size * __iters_per_work_item);
 
         auto __private_histograms =
-            oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _ExecutionPolicy, _bin_type>(
-                __exec, __segments * __num_bins)
+            oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _bin_type>(__exec, __segments * __num_bins)
                 .get_buffer();
 
         return __exec.queue().submit([&](auto& __h) {
diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_radix_sort.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_radix_sort.h
index 05a1e8ccfc8..4fd1aa1bef9 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_radix_sort.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_radix_sort.h
@@ -827,9 +827,7 @@ __parallel_radix_sort(oneapi::dpl::__internal::__device_backend_tag, _ExecutionP
         sycl::buffer<::std::uint32_t, 1> __tmp_buf{sycl::range<1>(__tmp_buf_size)};
 
         // memory for storing values sorted for an iteration
-        oneapi::dpl::__par_backend_hetero::__buffer<oneapi::dpl::__internal::__device_backend_tag, _ExecutionPolicy,
-                                                    _ValueT>
-            __out_buffer_holder{__exec, __n};
+        oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueT> __out_buffer_holder{__exec, __n};
         auto __out_rng = oneapi::dpl::__ranges::all_view<_ValueT, __par_backend_hetero::access_mode::read_write>(
             __out_buffer_holder.get_buffer());
 
diff --git a/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
index 77fd740c563..23beae41c10 100644
--- a/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/numeric_impl_hetero.h
@@ -162,7 +162,7 @@ __pattern_transform_scan_base(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&
         using _NewExecutionPolicy = decltype(__policy);
 
         // Create temporary buffer
-        oneapi::dpl::__par_backend_hetero::__buffer<_BackendTag, _NewExecutionPolicy, _Type> __tmp_buf(__policy, __n);
+        oneapi::dpl::__par_backend_hetero::__buffer<_NewExecutionPolicy, _Type> __tmp_buf(__policy, __n);
         auto __first_tmp = __tmp_buf.get();
         auto __last_tmp = __first_tmp + __n;
         auto __keep2 = oneapi::dpl::__ranges::__get_sycl_range<__par_backend_hetero::access_mode::write, _Iterator2>();
diff --git a/include/oneapi/dpl/pstl/omp/parallel_scan.h b/include/oneapi/dpl/pstl/omp/parallel_scan.h
index 75525246092..c3bc022cb2e 100644
--- a/include/oneapi/dpl/pstl/omp/parallel_scan.h
+++ b/include/oneapi/dpl/pstl/omp/parallel_scan.h
@@ -89,8 +89,7 @@ __parallel_strict_scan_body(_ExecutionPolicy&& __exec, _Index __n, _Tp __initial
     const _Index __slack = 4;
     _Index __tilesize = (__n - 1) / (__slack * __p) + 1;
     _Index __m = (__n - 1) / __tilesize;
-    __buffer<oneapi::dpl::__internal::__omp_backend_tag, _ExecutionPolicy, _Tp> __buf(
-        ::std::forward<_ExecutionPolicy>(__exec), __m + 1);
+    __buffer<_ExecutionPolicy, _Tp> __buf(::std::forward<_ExecutionPolicy>(__exec), __m + 1);
     _Tp* __r = __buf.get();
 
     oneapi::dpl::__omp_backend::__upsweep(_Index(0), _Index(__m + 1), __tilesize, __r, __n - __m * __tilesize, __reduce,
diff --git a/include/oneapi/dpl/pstl/parallel_backend_tbb.h b/include/oneapi/dpl/pstl/parallel_backend_tbb.h
index 81707201221..4f2696eff3f 100644
--- a/include/oneapi/dpl/pstl/parallel_backend_tbb.h
+++ b/include/oneapi/dpl/pstl/parallel_backend_tbb.h
@@ -393,8 +393,7 @@ __parallel_strict_scan(oneapi::dpl::__internal::__tbb_backend_tag, _ExecutionPol
             const _Index __slack = 4;
             _Index __tilesize = (__n - 1) / (__slack * __p) + 1;
             _Index __m = (__n - 1) / __tilesize;
-            __tbb_backend::__buffer<oneapi::dpl::__internal::__tbb_backend_tag, _ExecutionPolicy, _Tp> __buf(__exec,
-                                                                                                             __m + 1);
+            __tbb_backend::__buffer<_ExecutionPolicy, _Tp> __buf(__exec, __m + 1);
             _Tp* __r = __buf.get();
             __tbb_backend::__upsweep(_Index(0), _Index(__m + 1), __tilesize, __r, __n - __m * __tilesize, __reduce,
                                      __combine);
@@ -1200,8 +1199,7 @@ __parallel_stable_sort(oneapi::dpl::__internal::__tbb_backend_tag, _ExecutionPol
         const _DifferenceType __sort_cut_off = _ONEDPL_STABLE_SORT_CUT_OFF;
         if (__n > __sort_cut_off)
         {
-            __tbb_backend::__buffer<oneapi::dpl::__internal::__tbb_backend_tag, _ExecutionPolicy, _ValueType> __buf(
-                __exec, __n);
+            __tbb_backend::__buffer<_ExecutionPolicy, _ValueType> __buf(__exec, __n);
             __root_task<__stable_sort_func<_RandomAccessIterator, _ValueType*, _Compare, _LeafSort>> __root{
                 __xs, __xe, __buf.get(), true, __comp, __leaf_sort, __nsort, __xs, __buf.get()};
             __task::spawn_root_and_wait(__root);

From 0c8fdbe79e3237e0f8c4e80b577cc00f68f77c4a Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 14 Mar 2024 10:34:46 +0100
Subject: [PATCH 557/566] 
 include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h - fix compile
 errors in for_loop tests

---
 .../experimental/internal/for_loop_impl.h     | 31 +++++++++++++++++--
 1 file changed, 29 insertions(+), 2 deletions(-)

diff --git a/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h b/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
index 47769c4645d..103f4ee594b 100644
--- a/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
+++ b/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
@@ -476,6 +476,33 @@ __pattern_for_loop(_ExecutionPolicy&& __exec, _Ip __first, _Ip __last, _Function
 template <typename _Ip>
 struct __use_par_vec_helper
 {
+  protected:
+
+    template <typename _Tag>
+    static constexpr auto __get_is_vector(_Tag)
+    {
+        return typename _Tag::__is_vector{};
+    }
+
+    template <typename _BackendTag>
+    static constexpr auto __get_is_vector(__hetero_tag<_BackendTag>)
+    {
+        return ::std::true_type{};
+    }
+
+    template <typename _Tag>
+    static constexpr auto __get_is_parallel(_Tag)
+    {
+        return oneapi::dpl::__internal::__is_parallel_tag<_Tag>{};
+    }
+
+    template <typename _BackendTag>
+    static constexpr auto __get_is_parallel(__hetero_tag<_BackendTag>)
+    {
+        return ::std::true_type{};
+    }
+
+  public:
     using __it_type = std::conditional_t<std::is_integral_v<_Ip>, _Ip*, _Ip>;
 
     template <typename _ExecutionPolicy>
@@ -483,7 +510,7 @@ struct __use_par_vec_helper
     __use_vector(_ExecutionPolicy&& __exec)
     {
         using __tag_type = decltype(oneapi::dpl::__internal::__select_backend(__exec, std::declval<__it_type>()));
-        return typename __tag_type::__is_vector{};
+        return __get_is_vector(__tag_type{});
     }
 
     template <typename _ExecutionPolicy>
@@ -491,7 +518,7 @@ struct __use_par_vec_helper
     __use_parallel(_ExecutionPolicy&& __exec)
     {
         using __tag_type = decltype(oneapi::dpl::__internal::__select_backend(__exec, std::declval<__it_type>()));
-        return oneapi::dpl::__internal::__is_parallel_tag<__tag_type>{};
+        return __get_is_parallel(__tag_type{});
     }
 };
 

From 2bc44ca45a0936fcc601a1d29cd9a6e05905dc8d Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 14 Mar 2024 10:55:43 +0100
Subject: [PATCH 558/566] 
 include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h - fix compile
 errors in for_loop tests (ONEDPL_BACKEND=serial)

---
 include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h b/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
index 103f4ee594b..17e5170e47b 100644
--- a/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
+++ b/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
@@ -471,6 +471,9 @@ __pattern_for_loop(_ExecutionPolicy&& __exec, _Ip __first, _Ip __last, _Function
         __is_vector, ::std::true_type{}, ::std::forward<_Rest>(__rest)...);
 }
 
+template <typename _BackendTag>
+struct __hetero_tag;
+
 // Helper structure to split code functions for integral and iterator types so the return
 // value can be successfully deduced.
 template <typename _Ip>

From dd72781a43a164d4cc484f795221255c570a740e Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Thu, 14 Mar 2024 11:56:07 +0100
Subject: [PATCH 559/566] 
 include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h - fix compile
 errors in for_loop tests (ONEDPL_BACKEND=serial)

---
 .../oneapi/dpl/pstl/experimental/internal/for_loop_impl.h  | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)

diff --git a/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h b/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
index 17e5170e47b..8dfc4f523b7 100644
--- a/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
+++ b/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
@@ -471,9 +471,6 @@ __pattern_for_loop(_ExecutionPolicy&& __exec, _Ip __first, _Ip __last, _Function
         __is_vector, ::std::true_type{}, ::std::forward<_Rest>(__rest)...);
 }
 
-template <typename _BackendTag>
-struct __hetero_tag;
-
 // Helper structure to split code functions for integral and iterator types so the return
 // value can be successfully deduced.
 template <typename _Ip>
@@ -487,11 +484,13 @@ struct __use_par_vec_helper
         return typename _Tag::__is_vector{};
     }
 
+#if _ONEDPL_BACKEND_SYCL
     template <typename _BackendTag>
     static constexpr auto __get_is_vector(__hetero_tag<_BackendTag>)
     {
         return ::std::true_type{};
     }
+#endif // _ONEDPL_BACKEND_SYCL
 
     template <typename _Tag>
     static constexpr auto __get_is_parallel(_Tag)
@@ -499,11 +498,13 @@ struct __use_par_vec_helper
         return oneapi::dpl::__internal::__is_parallel_tag<_Tag>{};
     }
 
+#if _ONEDPL_BACKEND_SYCL
     template <typename _BackendTag>
     static constexpr auto __get_is_parallel(__hetero_tag<_BackendTag>)
     {
         return ::std::true_type{};
     }
+#endif // _ONEDPL_BACKEND_SYCL
 
   public:
     using __it_type = std::conditional_t<std::is_integral_v<_Ip>, _Ip*, _Ip>;

From 0bd539cad073a0344767f469c35e88bf297d6fd8 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <sergey.kopienko@intel.com>
Date: Thu, 14 Mar 2024 16:58:56 +0100
Subject: [PATCH 560/566] Disable `for_loop` staff for hetero policies (#1447)

---
 .../dpl/pstl/experimental/internal/for_loop.h | 12 +++++++
 .../experimental/internal/for_loop_impl.h     | 35 ++-----------------
 .../experimental/for_loop.pass.cpp            | 10 +++---
 .../experimental/for_loop_induction.pass.cpp  |  6 ++--
 .../experimental/for_loop_reduction.pass.cpp  | 16 +++++----
 5 files changed, 34 insertions(+), 45 deletions(-)

diff --git a/include/oneapi/dpl/pstl/experimental/internal/for_loop.h b/include/oneapi/dpl/pstl/experimental/internal/for_loop.h
index 4a61dd7c09b..5d8802083d3 100644
--- a/include/oneapi/dpl/pstl/experimental/internal/for_loop.h
+++ b/include/oneapi/dpl/pstl/experimental/internal/for_loop.h
@@ -56,6 +56,9 @@ template <typename _ExecutionPolicy, typename _Ip, typename... _Rest>
 void
 for_loop(_ExecutionPolicy&& __exec, type_identity_t<_Ip> __start, _Ip __finish, _Rest&&... __rest)
 {
+    static_assert(oneapi::dpl::__internal::__is_host_execution_policy<::std::decay_t<_ExecutionPolicy>>::value,
+                  "for_loop is implemented for the host policies only");
+
     oneapi::dpl::__internal::__for_loop_repack(::std::forward<_ExecutionPolicy>(__exec), __start, __finish,
                                                oneapi::dpl::__internal::__single_stride_type{},
                                                ::std::forward_as_tuple(::std::forward<_Rest>(__rest)...));
@@ -65,6 +68,9 @@ template <typename _ExecutionPolicy, typename _Ip, typename _Sp, typename... _Re
 void
 for_loop_strided(_ExecutionPolicy&& __exec, type_identity_t<_Ip> __start, _Ip __finish, _Sp __stride, _Rest&&... __rest)
 {
+    static_assert(oneapi::dpl::__internal::__is_host_execution_policy<::std::decay_t<_ExecutionPolicy>>::value,
+                  "for_loop_strided is implemented for the host policies only");
+
     oneapi::dpl::__internal::__for_loop_repack(::std::forward<_ExecutionPolicy>(__exec), __start, __finish, __stride,
                                                ::std::forward_as_tuple(::std::forward<_Rest>(__rest)...));
 }
@@ -73,6 +79,9 @@ template <typename _ExecutionPolicy, typename _Ip, typename _Size, typename... _
 void
 for_loop_n(_ExecutionPolicy&& __exec, _Ip __start, _Size __n, _Rest&&... __rest)
 {
+    static_assert(oneapi::dpl::__internal::__is_host_execution_policy<::std::decay_t<_ExecutionPolicy>>::value,
+                  "for_loop_n is implemented for the host policies only");
+
     oneapi::dpl::__internal::__for_loop_repack_n(::std::forward<_ExecutionPolicy>(__exec), __start, __n,
                                                  oneapi::dpl::__internal::__single_stride_type{},
                                                  ::std::forward_as_tuple(::std::forward<_Rest>(__rest)...));
@@ -82,6 +91,9 @@ template <typename _ExecutionPolicy, typename _Ip, typename _Size, typename _Sp,
 void
 for_loop_n_strided(_ExecutionPolicy&& __exec, _Ip __start, _Size __n, _Sp __stride, _Rest&&... __rest)
 {
+    static_assert(oneapi::dpl::__internal::__is_host_execution_policy<::std::decay_t<_ExecutionPolicy>>::value,
+                  "for_loop_n_strided is implemented for the host policies only");
+
     oneapi::dpl::__internal::__for_loop_repack_n(::std::forward<_ExecutionPolicy>(__exec), __start, __n, __stride,
                                                  ::std::forward_as_tuple(::std::forward<_Rest>(__rest)...));
 }
diff --git a/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h b/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
index 8dfc4f523b7..47769c4645d 100644
--- a/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
+++ b/include/oneapi/dpl/pstl/experimental/internal/for_loop_impl.h
@@ -476,37 +476,6 @@ __pattern_for_loop(_ExecutionPolicy&& __exec, _Ip __first, _Ip __last, _Function
 template <typename _Ip>
 struct __use_par_vec_helper
 {
-  protected:
-
-    template <typename _Tag>
-    static constexpr auto __get_is_vector(_Tag)
-    {
-        return typename _Tag::__is_vector{};
-    }
-
-#if _ONEDPL_BACKEND_SYCL
-    template <typename _BackendTag>
-    static constexpr auto __get_is_vector(__hetero_tag<_BackendTag>)
-    {
-        return ::std::true_type{};
-    }
-#endif // _ONEDPL_BACKEND_SYCL
-
-    template <typename _Tag>
-    static constexpr auto __get_is_parallel(_Tag)
-    {
-        return oneapi::dpl::__internal::__is_parallel_tag<_Tag>{};
-    }
-
-#if _ONEDPL_BACKEND_SYCL
-    template <typename _BackendTag>
-    static constexpr auto __get_is_parallel(__hetero_tag<_BackendTag>)
-    {
-        return ::std::true_type{};
-    }
-#endif // _ONEDPL_BACKEND_SYCL
-
-  public:
     using __it_type = std::conditional_t<std::is_integral_v<_Ip>, _Ip*, _Ip>;
 
     template <typename _ExecutionPolicy>
@@ -514,7 +483,7 @@ struct __use_par_vec_helper
     __use_vector(_ExecutionPolicy&& __exec)
     {
         using __tag_type = decltype(oneapi::dpl::__internal::__select_backend(__exec, std::declval<__it_type>()));
-        return __get_is_vector(__tag_type{});
+        return typename __tag_type::__is_vector{};
     }
 
     template <typename _ExecutionPolicy>
@@ -522,7 +491,7 @@ struct __use_par_vec_helper
     __use_parallel(_ExecutionPolicy&& __exec)
     {
         using __tag_type = decltype(oneapi::dpl::__internal::__select_backend(__exec, std::declval<__it_type>()));
-        return __get_is_parallel(__tag_type{});
+        return oneapi::dpl::__internal::__is_parallel_tag<__tag_type>{};
     }
 };
 
diff --git a/test/parallel_api/experimental/for_loop.pass.cpp b/test/parallel_api/experimental/for_loop.pass.cpp
index 8c59ffe1afd..725d2bc9747 100644
--- a/test/parallel_api/experimental/for_loop.pass.cpp
+++ b/test/parallel_api/experimental/for_loop.pass.cpp
@@ -296,8 +296,9 @@ test_for_loop()
         Sequence<T> in_out(n, Gen<T>());
         Sequence<T> expected = in_out;
 
-        invoke_on_all_policies<>()(test_for_loop_impl(), in_out.begin(), in_out.end(), expected.begin(), expected.end(),
-                                   in_out.size());
+        // for_loop staff is implemented for the host policies only
+        invoke_on_all_host_policies()(test_for_loop_impl(), in_out.begin(), in_out.end(), expected.begin(),
+                                      expected.end(), in_out.size());
     }
 }
 
@@ -313,8 +314,9 @@ test_for_loop_strided()
         ::std::vector<size_t> strides = {1, 2, 10, n > 1 ? n - 1 : 1, n > 0 ? n : 1, n + 1};
         for (size_t stride : strides)
         {
-            invoke_on_all_policies<>()(test_for_loop_strided_impl(), in_out.begin(), in_out.end(), expected.begin(),
-                                       expected.end(), in_out.size(), stride);
+            // for_loop staff is implemented for the host policies only
+            invoke_on_all_host_policies()(test_for_loop_strided_impl(), in_out.begin(), in_out.end(), expected.begin(),
+                                          expected.end(), in_out.size(), stride);
         }
     }
 }
diff --git a/test/parallel_api/experimental/for_loop_induction.pass.cpp b/test/parallel_api/experimental/for_loop_induction.pass.cpp
index 505ceb54e1e..e01569c58a1 100644
--- a/test/parallel_api/experimental/for_loop_induction.pass.cpp
+++ b/test/parallel_api/experimental/for_loop_induction.pass.cpp
@@ -153,8 +153,10 @@ test()
     {
         Sequence<T> in_out(n, [](long int k) { return T(k % 5 != 1 ? 3 * k - 7 : 0); });
         Sequence<T> expected = in_out;
-        invoke_on_all_policies<>()(test_body(), in_out.begin(), in_out.end(), expected.begin(), expected.end(),
-                                   in_out.size());
+
+        // for_loop staff is implemented for the host policies only
+        invoke_on_all_host_policies()(test_body(), in_out.begin(), in_out.end(), expected.begin(), expected.end(),
+                                      in_out.size());
     }
 }
 
diff --git a/test/parallel_api/experimental/for_loop_reduction.pass.cpp b/test/parallel_api/experimental/for_loop_reduction.pass.cpp
index cfd918ff2bd..7f6b2b9e7f5 100644
--- a/test/parallel_api/experimental/for_loop_reduction.pass.cpp
+++ b/test/parallel_api/experimental/for_loop_reduction.pass.cpp
@@ -77,8 +77,10 @@ test()
     {
         Sequence<T> in_out(n, [](long int k) { return T(k % 5 != 1 ? 3 * k - 7 : 0); });
         Sequence<T> expected = in_out;
-        invoke_on_all_policies<>()(test_body(), in_out.begin(), in_out.end(), expected.begin(), expected.end(),
-                                   in_out.size());
+
+        // for_loop staff is implemented for the host policies only
+        invoke_on_all_host_policies()(test_body(), in_out.begin(), in_out.end(), expected.begin(), expected.end(),
+                                      in_out.size());
     }
 }
 
@@ -176,10 +178,12 @@ test_predefined(::std::initializer_list<T> init_list)
     // Just arbitrary numbers
     Sequence<T> in_out = init_list;
     Sequence<T> expected = in_out;
-    invoke_on_all_policies<>()(test_body_predefined(), in_out.begin(), in_out.end(), expected.begin(), expected.end(),
-                               in_out.size());
-    invoke_on_all_policies<>()(test_body_predefined_bits(), in_out.begin(), in_out.end(), expected.begin(),
-                               expected.end(), in_out.size());
+
+    // for_loop staff is implemented for the host policies only
+    invoke_on_all_host_policies()(test_body_predefined(), in_out.begin(), in_out.end(), expected.begin(),
+                                  expected.end(), in_out.size());
+    invoke_on_all_host_policies()(test_body_predefined_bits(), in_out.begin(), in_out.end(), expected.begin(),
+                                  expected.end(), in_out.size());
 }
 
 void

From 68fdd8f1d699c3c85d0d9bcd1f9f08135c76e35d Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 15 Mar 2024 09:29:26 +0100
Subject: [PATCH 561/566] Fix review comment: extra static_assert checks
 removed.

---
 .../oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_utils.h  | 2 --
 include/oneapi/dpl/pstl/omp/util.h                              | 2 --
 include/oneapi/dpl/pstl/parallel_backend_serial.h               | 2 --
 include/oneapi/dpl/pstl/parallel_backend_tbb.h                  | 2 --
 4 files changed, 8 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_utils.h b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_utils.h
index 62025411b24..4f1cdf50064 100644
--- a/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_utils.h
+++ b/include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_utils.h
@@ -386,8 +386,6 @@ class __buffer_impl
     __container_t __container;
 
   public:
-    static_assert(::std::is_same_v<_ExecutionPolicy, ::std::decay_t<_ExecutionPolicy>>);
-
     __buffer_impl(_ExecutionPolicy /*__exec*/, ::std::size_t __n_elements) : __container{sycl::range<1>(__n_elements)}
     {
     }
diff --git a/include/oneapi/dpl/pstl/omp/util.h b/include/oneapi/dpl/pstl/omp/util.h
index 92689638121..e7c4e3cbc40 100644
--- a/include/oneapi/dpl/pstl/omp/util.h
+++ b/include/oneapi/dpl/pstl/omp/util.h
@@ -68,8 +68,6 @@ class __buffer_impl
     operator=(const __buffer_impl&) = delete;
 
   public:
-    static_assert(::std::is_same_v<_ExecutionPolicy, ::std::decay_t<_ExecutionPolicy>>);
-
     __buffer_impl(_ExecutionPolicy /*__exec*/, std::size_t __n)
         : __allocator_(), __ptr_(__allocator_.allocate(__n)), __buf_size_(__n)
     {
diff --git a/include/oneapi/dpl/pstl/parallel_backend_serial.h b/include/oneapi/dpl/pstl/parallel_backend_serial.h
index b1fb6d9f39c..edd6652d359 100644
--- a/include/oneapi/dpl/pstl/parallel_backend_serial.h
+++ b/include/oneapi/dpl/pstl/parallel_backend_serial.h
@@ -43,8 +43,6 @@ class __buffer_impl
     operator=(const __buffer_impl&) = delete;
 
   public:
-    static_assert(::std::is_same_v<_ExecutionPolicy, ::std::decay_t<_ExecutionPolicy>>);
-
     __buffer_impl(_ExecutionPolicy /*__exec*/, ::std::size_t __n)
         : __allocator_(), __ptr_(__allocator_.allocate(__n)), __buf_size_(__n)
     {
diff --git a/include/oneapi/dpl/pstl/parallel_backend_tbb.h b/include/oneapi/dpl/pstl/parallel_backend_tbb.h
index 4f2696eff3f..2ddfa61007a 100644
--- a/include/oneapi/dpl/pstl/parallel_backend_tbb.h
+++ b/include/oneapi/dpl/pstl/parallel_backend_tbb.h
@@ -65,8 +65,6 @@ class __buffer_impl
     operator=(const __buffer_impl&) = delete;
 
   public:
-    static_assert(::std::is_same_v<_ExecutionPolicy, ::std::decay_t<_ExecutionPolicy>>);
-
     //! Try to obtain buffer of given size to store objects of _Tp type
     __buffer_impl(_ExecutionPolicy /*__exec*/, const ::std::size_t __n)
         : _M_allocator(), _M_ptr(_M_allocator.allocate(__n)), _M_buf_size(__n)

From dae09182849da676af7646edd2800d8b88b7fb97 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 15 Mar 2024 09:31:20 +0100
Subject: [PATCH 562/566] include/oneapi/dpl/pstl/algorithm_impl.h - fix review
 comment: capturing tag first is more logical

---
 include/oneapi/dpl/pstl/algorithm_impl.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 1992b06cf0c..51f72ce9af6 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -1028,7 +1028,7 @@ __pattern_search_n(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _
     }
     else
     {
-        return __internal::__except_handler([&__exec, __first, __last, __count, &__value, __pred, __tag]() {
+        return __internal::__except_handler([__tag, &__exec, __first, __last, __count, &__value, __pred]() {
             return __internal::__parallel_find(
                 __tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
                 [__last, __count, &__value, __pred](_RandomAccessIterator __i, _RandomAccessIterator __j) {

From ff10491c1ba8230b90146f264bf4cc27ced7aa15 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Fri, 15 Mar 2024 09:39:01 +0100
Subject: [PATCH 563/566] 
 include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h - fix review
 comment: why do we add ranges namespace?

---
 .../pstl/hetero/algorithm_ranges_impl_hetero.h  | 17 ++++++++---------
 1 file changed, 8 insertions(+), 9 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
index f9f39a41b49..1fd7d5be169 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
@@ -148,8 +148,8 @@ __pattern_find_end(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _
 
     if (__rng1.size() == __rng2.size())
     {
-        const bool __res = __ranges::__pattern_equal(__tag, ::std::forward<_ExecutionPolicy>(__exec), __rng1,
-                                                     ::std::forward<_Range2>(__rng2), __pred);
+        const bool __res = __pattern_equal(__tag, ::std::forward<_ExecutionPolicy>(__exec), __rng1,
+                                           ::std::forward<_Range2>(__rng2), __pred);
         return __res ? 0 : __rng1.size();
     }
 
@@ -228,7 +228,7 @@ __pattern_search(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Ra
 
     if (__rng1.size() == __rng2.size())
     {
-        const bool __res = __ranges::__pattern_equal(
+        const bool __res = __pattern_equal(
             __tag, __par_backend_hetero::make_wrapped_policy<equal_wrapper>(::std::forward<_ExecutionPolicy>(__exec)),
             __rng1, ::std::forward<_Range2>(__rng2), __pred);
         return __res ? 0 : __rng1.size();
@@ -416,8 +416,8 @@ __pattern_remove_if(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec,
     oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __buf(__exec, __rng.size());
     auto __copy_rng = oneapi::dpl::__ranges::views::all(__buf.get_buffer());
 
-    auto __copy_last_id = __ranges::__pattern_copy_if(__tag, __exec, __rng, __copy_rng, __not_pred<_Predicate>{__pred},
-                                                      oneapi::dpl::__internal::__pstl_assign());
+    auto __copy_last_id = __pattern_copy_if(__tag, __exec, __rng, __copy_rng, __not_pred<_Predicate>{__pred},
+                                            oneapi::dpl::__internal::__pstl_assign());
     auto __copy_rng_truncated = __copy_rng | oneapi::dpl::experimental::ranges::views::take(__copy_last_id);
 
     oneapi::dpl::__internal::__ranges::__pattern_walk_n(
@@ -463,8 +463,7 @@ __pattern_unique(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Ra
 
     oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __buf(__exec, __rng.size());
     auto res_rng = oneapi::dpl::__ranges::views::all(__buf.get_buffer());
-    auto res = __ranges::__pattern_unique_copy(__tag, __exec, __rng, res_rng, __pred,
-                                               oneapi::dpl::__internal::__pstl_assign());
+    auto res = __pattern_unique_copy(__tag, __exec, __rng, res_rng, __pred, oneapi::dpl::__internal::__pstl_assign());
 
     __pattern_walk_n(__tag, ::std::forward<_ExecutionPolicy>(__exec),
                      __brick_copy<__hetero_tag<_BackendTag>, _ExecutionPolicy>{}, res_rng,
@@ -740,7 +739,7 @@ __pattern_reduce_by_segment(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&&
     // evenly divisible by wg size (ensures segments are not long), or has a key not equal to the
     // adjacent element (marks end of real segments)
     // TODO: replace wgroup size with segment size based on platform specifics.
-    auto __intermediate_result_end = __ranges::__pattern_copy_if(
+    auto __intermediate_result_end = __pattern_copy_if(
         __tag, oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__assign_key1_wrapper>(__exec), __view1, __view2,
         [__n, __binary_pred, __wgroup_size](const auto& __a) {
             // The size of key range for the (i-1) view is one less, so for the 0th index we do not check the keys
@@ -783,7 +782,7 @@ __pattern_reduce_by_segment(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&&
 
     // element is copied if it is the 0th element (marks beginning of first segment), or has a key not equal to
     // the adjacent element (end of a segment). Artificial segments based on wg size are not created.
-    auto __result_end = __ranges::__pattern_copy_if(
+    auto __result_end = __pattern_copy_if(
         __tag, oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__assign_key2_wrapper>(__exec), __view3, __view4,
         [__binary_pred](const auto& __a) {
             // The size of key range for the (i-1) view is one less, so for the 0th index we do not check the keys

From 86a59e61d0b7694439c7ac2f6e88fbd1fac92610 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <sergey.kopienko@intel.com>
Date: Mon, 18 Mar 2024 16:36:02 +0100
Subject: [PATCH 564/566] Fix compile errors in `ranges` tests (#1450)

---
 .../hetero/algorithm_ranges_impl_hetero.h     | 37 ++++++++++---------
 1 file changed, 20 insertions(+), 17 deletions(-)

diff --git a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
index 1fd7d5be169..25ce34a8680 100644
--- a/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
+++ b/include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h
@@ -148,8 +148,8 @@ __pattern_find_end(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _
 
     if (__rng1.size() == __rng2.size())
     {
-        const bool __res = __pattern_equal(__tag, ::std::forward<_ExecutionPolicy>(__exec), __rng1,
-                                           ::std::forward<_Range2>(__rng2), __pred);
+        const bool __res = __ranges::__pattern_equal(__tag, ::std::forward<_ExecutionPolicy>(__exec), __rng1,
+                                                     ::std::forward<_Range2>(__rng2), __pred);
         return __res ? 0 : __rng1.size();
     }
 
@@ -228,7 +228,7 @@ __pattern_search(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Ra
 
     if (__rng1.size() == __rng2.size())
     {
-        const bool __res = __pattern_equal(
+        const bool __res = __ranges::__pattern_equal(
             __tag, __par_backend_hetero::make_wrapped_policy<equal_wrapper>(::std::forward<_ExecutionPolicy>(__exec)),
             __rng1, ::std::forward<_Range2>(__rng2), __pred);
         return __res ? 0 : __rng1.size();
@@ -259,8 +259,8 @@ __pattern_search_n(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _
     auto __s_rng = oneapi::dpl::experimental::ranges::views::iota(0, __count) |
                    oneapi::dpl::experimental::ranges::views::transform([__value](auto) { return __value; });
 
-    return __pattern_search(__tag, ::std::forward<_ExecutionPolicy>(__exec), ::std::forward<_Range>(__rng), __s_rng,
-                            __pred);
+    return __ranges::__pattern_search(__tag, ::std::forward<_ExecutionPolicy>(__exec), ::std::forward<_Range>(__rng),
+                                      __s_rng, __pred);
 }
 
 template <typename _Size>
@@ -396,8 +396,9 @@ __pattern_copy_if(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _R
     unseq_backend::__create_mask<_Predicate, _SizeType> __create_mask_op{__pred};
     unseq_backend::__copy_by_mask<_ReduceOp, _Assign, /*inclusive*/ ::std::true_type, 1> __copy_by_mask_op;
 
-    return __pattern_scan_copy(__tag, ::std::forward<_ExecutionPolicy>(__exec), ::std::forward<_Range1>(__rng1),
-                               ::std::forward<_Range2>(__rng2), __create_mask_op, __copy_by_mask_op);
+    return __ranges::__pattern_scan_copy(__tag, ::std::forward<_ExecutionPolicy>(__exec),
+                                         ::std::forward<_Range1>(__rng1), ::std::forward<_Range2>(__rng2),
+                                         __create_mask_op, __copy_by_mask_op);
 }
 
 //------------------------------------------------------------------------
@@ -416,8 +417,8 @@ __pattern_remove_if(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec,
     oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __buf(__exec, __rng.size());
     auto __copy_rng = oneapi::dpl::__ranges::views::all(__buf.get_buffer());
 
-    auto __copy_last_id = __pattern_copy_if(__tag, __exec, __rng, __copy_rng, __not_pred<_Predicate>{__pred},
-                                            oneapi::dpl::__internal::__pstl_assign());
+    auto __copy_last_id = __ranges::__pattern_copy_if(__tag, __exec, __rng, __copy_rng, __not_pred<_Predicate>{__pred},
+                                                      oneapi::dpl::__internal::__pstl_assign());
     auto __copy_rng_truncated = __copy_rng | oneapi::dpl::experimental::ranges::views::take(__copy_last_id);
 
     oneapi::dpl::__internal::__ranges::__pattern_walk_n(
@@ -444,8 +445,9 @@ __pattern_unique_copy(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec
     __create_mask_unique_copy<__not_pred<_BinaryPredicate>, _It1DifferenceType> __create_mask_op{
         __not_pred<_BinaryPredicate>{__pred}};
 
-    return __pattern_scan_copy(__tag, ::std::forward<_ExecutionPolicy>(__exec), ::std::forward<_Range1>(__rng),
-                               ::std::forward<_Range2>(__result), __create_mask_op, __copy_by_mask_op);
+    return __ranges::__pattern_scan_copy(__tag, ::std::forward<_ExecutionPolicy>(__exec),
+                                         ::std::forward<_Range1>(__rng), ::std::forward<_Range2>(__result),
+                                         __create_mask_op, __copy_by_mask_op);
 }
 
 //------------------------------------------------------------------------
@@ -463,11 +465,12 @@ __pattern_unique(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&& __exec, _Ra
 
     oneapi::dpl::__par_backend_hetero::__buffer<_ExecutionPolicy, _ValueType> __buf(__exec, __rng.size());
     auto res_rng = oneapi::dpl::__ranges::views::all(__buf.get_buffer());
-    auto res = __pattern_unique_copy(__tag, __exec, __rng, res_rng, __pred, oneapi::dpl::__internal::__pstl_assign());
+    auto res = __ranges::__pattern_unique_copy(__tag, __exec, __rng, res_rng, __pred,
+                                               oneapi::dpl::__internal::__pstl_assign());
 
-    __pattern_walk_n(__tag, ::std::forward<_ExecutionPolicy>(__exec),
-                     __brick_copy<__hetero_tag<_BackendTag>, _ExecutionPolicy>{}, res_rng,
-                     ::std::forward<_Range>(__rng));
+    __ranges::__pattern_walk_n(__tag, ::std::forward<_ExecutionPolicy>(__exec),
+                               __brick_copy<__hetero_tag<_BackendTag>, _ExecutionPolicy>{}, res_rng,
+                               ::std::forward<_Range>(__rng));
     return res;
 }
 
@@ -739,7 +742,7 @@ __pattern_reduce_by_segment(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&&
     // evenly divisible by wg size (ensures segments are not long), or has a key not equal to the
     // adjacent element (marks end of real segments)
     // TODO: replace wgroup size with segment size based on platform specifics.
-    auto __intermediate_result_end = __pattern_copy_if(
+    auto __intermediate_result_end = __ranges::__pattern_copy_if(
         __tag, oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__assign_key1_wrapper>(__exec), __view1, __view2,
         [__n, __binary_pred, __wgroup_size](const auto& __a) {
             // The size of key range for the (i-1) view is one less, so for the 0th index we do not check the keys
@@ -782,7 +785,7 @@ __pattern_reduce_by_segment(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&&
 
     // element is copied if it is the 0th element (marks beginning of first segment), or has a key not equal to
     // the adjacent element (end of a segment). Artificial segments based on wg size are not created.
-    auto __result_end = __pattern_copy_if(
+    auto __result_end = __ranges::__pattern_copy_if(
         __tag, oneapi::dpl::__par_backend_hetero::make_wrapped_policy<__assign_key2_wrapper>(__exec), __view3, __view4,
         [__binary_pred](const auto& __a) {
             // The size of key range for the (i-1) view is one less, so for the 0th index we do not check the keys

From ea59cb396cdeec94649cacb9501cbb19b31b05c5 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <sergey.kopienko@intel.com>
Date: Mon, 18 Mar 2024 18:31:40 +0100
Subject: [PATCH 565/566] Fix review comments for tag dispatching (#1451)

---
 include/oneapi/dpl/pstl/algorithm_fwd.h  | 39 ++++++-----
 include/oneapi/dpl/pstl/algorithm_impl.h | 83 ++++++++++++------------
 2 files changed, 63 insertions(+), 59 deletions(-)

diff --git a/include/oneapi/dpl/pstl/algorithm_fwd.h b/include/oneapi/dpl/pstl/algorithm_fwd.h
index cada84d0f22..5af00e5425c 100644
--- a/include/oneapi/dpl/pstl/algorithm_fwd.h
+++ b/include/oneapi/dpl/pstl/algorithm_fwd.h
@@ -71,13 +71,13 @@ template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Fun
 void
 __pattern_walk1(_Tag, _ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _Function) noexcept;
 
-template <class _ExecutionPolicy, class _RandomAccessIterator, class _Function>
+template <class _ExecutionPolicy, class _ForwardIterator, class _Function>
 void
-__pattern_walk1(__parallel_forward_tag, _ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _Function);
+__pattern_walk1(__parallel_forward_tag, _ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _Function);
 
-template <class _IsVector, class _ExecutionPolicy, class _ForwardIterator, class _Function>
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Function>
 void
-__pattern_walk1(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _Function);
+__pattern_walk1(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator, _Function);
 
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Brick>
 void
@@ -298,9 +298,10 @@ template <class _Tag, class _ExecutionPolicy, class _ForwardIterator, class _Pre
 _ForwardIterator
 __pattern_find_if(_Tag, _ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _Predicate) noexcept;
 
-template <class _IsVector, class _ExecutionPolicy, class _ForwardIterator, class _Predicate>
-_ForwardIterator
-__pattern_find_if(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _ForwardIterator, _ForwardIterator, _Predicate);
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Predicate>
+_RandomAccessIterator
+__pattern_find_if(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator,
+                  _Predicate);
 
 //------------------------------------------------------------------------
 // find_end
@@ -346,11 +347,11 @@ _ForwardIterator1
 __pattern_find_first_of(_Tag, _ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2,
                         _ForwardIterator2, _BinaryPredicate) noexcept;
 
-template <class _IsVector, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2,
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
           class _BinaryPredicate>
-_ForwardIterator1
-__pattern_find_first_of(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1,
-                        _ForwardIterator2, _ForwardIterator2, _BinaryPredicate);
+_RandomAccessIterator1
+__pattern_find_first_of(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator1, _RandomAccessIterator1,
+                        _RandomAccessIterator2, _RandomAccessIterator2, _BinaryPredicate);
 
 //------------------------------------------------------------------------
 // search
@@ -1231,10 +1232,12 @@ bool
 __pattern_lexicographical_compare(_Tag, _ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1, _ForwardIterator2,
                                   _ForwardIterator2, _Compare) noexcept;
 
-template <class _IsVector, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Compare>
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
+          class _Compare>
 bool
-__pattern_lexicographical_compare(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _ForwardIterator1, _ForwardIterator1,
-                                  _ForwardIterator2, _ForwardIterator2, _Compare) noexcept;
+__pattern_lexicographical_compare(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator1,
+                                  _RandomAccessIterator1, _RandomAccessIterator2, _RandomAccessIterator2,
+                                  _Compare) noexcept;
 
 template <class _Tag, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2, class _Function>
 _ForwardIterator2
@@ -1259,10 +1262,10 @@ _ForwardIterator
 __pattern_shift_left(_Tag, _ExecutionPolicy&&, _ForwardIterator, _ForwardIterator,
                      typename ::std::iterator_traits<_ForwardIterator>::difference_type) noexcept;
 
-template <class _IsVector, class _ExecutionPolicy, class _ForwardIterator>
-_ForwardIterator
-__pattern_shift_left(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _ForwardIterator, _ForwardIterator,
-                     typename ::std::iterator_traits<_ForwardIterator>::difference_type);
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator>
+_RandomAccessIterator
+__pattern_shift_left(__parallel_tag<_IsVector>, _ExecutionPolicy&&, _RandomAccessIterator, _RandomAccessIterator,
+                     typename ::std::iterator_traits<_RandomAccessIterator>::difference_type);
 
 //------------------------------------------------------------------------
 // shift_right
diff --git a/include/oneapi/dpl/pstl/algorithm_impl.h b/include/oneapi/dpl/pstl/algorithm_impl.h
index 51f72ce9af6..d0be4420f8c 100644
--- a/include/oneapi/dpl/pstl/algorithm_impl.h
+++ b/include/oneapi/dpl/pstl/algorithm_impl.h
@@ -159,16 +159,16 @@ __pattern_walk1(__parallel_forward_tag, _ExecutionPolicy&& __exec, _ForwardItera
     });
 }
 
-template <class _IsVector, class _ExecutionPolicy, class _ForwardIterator, class _Function>
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Function>
 void
-__pattern_walk1(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _ForwardIterator __first, _ForwardIterator __last,
-                _Function __f)
+__pattern_walk1(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
+                _RandomAccessIterator __last, _Function __f)
 {
     using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
 
     __internal::__except_handler([&]() {
         __par_backend::__parallel_for(__backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
-                                      [__f](_ForwardIterator __i, _ForwardIterator __j) {
+                                      [__f](_RandomAccessIterator __i, _RandomAccessIterator __j) {
                                           __internal::__brick_walk1(__i, __j, __f, _IsVector{});
                                       });
     });
@@ -711,15 +711,15 @@ __pattern_find_if(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIt
     return __internal::__brick_find_if(__first, __last, __pred, typename _Tag::__is_vector{});
 }
 
-template <class _IsVector, class _ExecutionPolicy, class _ForwardIterator, class _Predicate>
-_ForwardIterator
-__pattern_find_if(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _ForwardIterator __first,
-                  _ForwardIterator __last, _Predicate __pred)
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _Predicate>
+_RandomAccessIterator
+__pattern_find_if(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
+                  _RandomAccessIterator __last, _Predicate __pred)
 {
     return __except_handler([&]() {
         return __parallel_find(
             __tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
-            [__pred](_ForwardIterator __i, _ForwardIterator __j) {
+            [__pred](_RandomAccessIterator __i, _RandomAccessIterator __j) {
                 return __brick_find_if(__i, __j, __pred, _IsVector{});
             },
             ::std::true_type{});
@@ -911,17 +911,17 @@ __pattern_find_first_of(_Tag, _ExecutionPolicy&&, _ForwardIterator1 __first, _Fo
                                              typename _Tag::__is_vector{});
 }
 
-template <class _IsVector, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2,
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
           class _BinaryPredicate>
-_ForwardIterator1
-__pattern_find_first_of(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _ForwardIterator1 __first,
-                        _ForwardIterator1 __last, _ForwardIterator2 __s_first, _ForwardIterator2 __s_last,
-                        _BinaryPredicate __pred)
+_RandomAccessIterator1
+__pattern_find_first_of(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _RandomAccessIterator1 __first,
+                        _RandomAccessIterator1 __last, _RandomAccessIterator2 __s_first,
+                        _RandomAccessIterator2 __s_last, _BinaryPredicate __pred)
 {
     return __internal::__except_handler([&]() {
         return __internal::__parallel_find(
             __tag, ::std::forward<_ExecutionPolicy>(__exec), __first, __last,
-            [__s_first, __s_last, &__pred](_ForwardIterator1 __i, _ForwardIterator1 __j) {
+            [__s_first, __s_last, &__pred](_RandomAccessIterator1 __i, _RandomAccessIterator1 __j) {
                 return __internal::__brick_find_first_of(__i, __j, __s_first, __s_last, __pred, _IsVector{});
             },
             ::std::true_type{});
@@ -1432,15 +1432,15 @@ __pattern_unique(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _ForwardIte
 
 // That function is shared between two algorithms - remove_if (__pattern_remove_if) and unique (pattern unique). But a mask calculation is different.
 // So, a caller passes _CalcMask brick into remove_elements.
-template <class _IsVector, class _ExecutionPolicy, class _ForwardIterator, class _CalcMask>
-_ForwardIterator
-__remove_elements(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _ForwardIterator __first,
-                  _ForwardIterator __last, _CalcMask __calc_mask)
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator, class _CalcMask>
+_RandomAccessIterator
+__remove_elements(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
+                  _RandomAccessIterator __last, _CalcMask __calc_mask)
 {
     using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
 
-    typedef typename ::std::iterator_traits<_ForwardIterator>::difference_type _DifferenceType;
-    typedef typename ::std::iterator_traits<_ForwardIterator>::value_type _Tp;
+    typedef typename ::std::iterator_traits<_RandomAccessIterator>::difference_type _DifferenceType;
+    typedef typename ::std::iterator_traits<_RandomAccessIterator>::value_type _Tp;
     _DifferenceType __n = __last - __first;
     __par_backend::__buffer<_ExecutionPolicy, bool> __mask_buf(__exec, __n);
     // 1. find a first iterator that should be removed
@@ -1494,7 +1494,7 @@ __remove_elements(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Forward
             [=](_DifferenceType __i, _DifferenceType __len, _DifferenceType __initial) {
                 __internal::__brick_copy_by_mask(
                     __first + __i, __first + __i + __len, __result + __initial, __mask + __i,
-                    [](_ForwardIterator __x, _Tp* __z) {
+                    [](_RandomAccessIterator __x, _Tp* __z) {
                         if constexpr (::std::is_trivial_v<_Tp>)
                             *__z = ::std::move(*__x);
                         else
@@ -3187,16 +3187,16 @@ __pattern_includes(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _
 
 inline constexpr auto __set_algo_cut_off = 1000;
 
-template <class _IsVector, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2,
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
           class _OutputIterator, class _Compare, class _SizeFunction, class _SetOP>
 _OutputIterator
-__parallel_set_op(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _ForwardIterator1 __first1,
-                  _ForwardIterator1 __last1, _ForwardIterator2 __first2, _ForwardIterator2 __last2,
+__parallel_set_op(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1,
+                  _RandomAccessIterator1 __last1, _RandomAccessIterator2 __first2, _RandomAccessIterator2 __last2,
                   _OutputIterator __result, _Compare __comp, _SizeFunction __size_func, _SetOP __set_op)
 {
     using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
 
-    typedef typename ::std::iterator_traits<_ForwardIterator1>::difference_type _DifferenceType;
+    typedef typename ::std::iterator_traits<_RandomAccessIterator1>::difference_type _DifferenceType;
     typedef typename ::std::iterator_traits<_OutputIterator>::value_type _T;
 
     struct _SetRange
@@ -3228,7 +3228,7 @@ __parallel_set_op(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Forward
             __backend_tag{}, ::std::forward<_ExecutionPolicy>(__exec), __n1, _SetRange{0, 0, 0}, //-1, 0},
             [=](_DifferenceType __i, _DifferenceType __len) {                                    // Reduce
                 //[__b; __e) - a subrange of the first sequence, to reduce
-                _ForwardIterator1 __b = __first1 + __i, __e = __first1 + (__i + __len);
+                _RandomAccessIterator1 __b = __first1 + __i, __e = __first1 + (__i + __len);
 
                 //try searching for the first element which not equal to *__b
                 if (__b != __first1)
@@ -3241,7 +3241,7 @@ __parallel_set_op(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Forward
                 //check is [__b; __e) empty
                 if (__e - __b < 1)
                 {
-                    _ForwardIterator2 __bb = __last2;
+                    _RandomAccessIterator2 __bb = __last2;
                     if (__b != __last1)
                         __bb = ::std::lower_bound(__first2, __last2, *__b, __comp);
 
@@ -3250,11 +3250,11 @@ __parallel_set_op(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Forward
                 }
 
                 //try searching for "corresponding" subrange [__bb; __ee) in the second sequence
-                _ForwardIterator2 __bb = __first2;
+                _RandomAccessIterator2 __bb = __first2;
                 if (__b != __first1)
                     __bb = ::std::lower_bound(__first2, __last2, *__b, __comp);
 
-                _ForwardIterator2 __ee = __last2;
+                _RandomAccessIterator2 __ee = __last2;
                 if (__e != __last1)
                     __ee = ::std::lower_bound(__bb, __last2, *__e, __comp);
 
@@ -3280,16 +3280,16 @@ __parallel_set_op(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Forward
 }
 
 //a shared parallel pattern for '__pattern_set_union' and '__pattern_set_symmetric_difference'
-template <class _IsVector, class _ExecutionPolicy, class _ForwardIterator1, class _ForwardIterator2,
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator1, class _RandomAccessIterator2,
           class _OutputIterator, class _Compare, class _SetUnionOp>
 _OutputIterator
-__parallel_set_union_op(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _ForwardIterator1 __first1,
-                        _ForwardIterator1 __last1, _ForwardIterator2 __first2, _ForwardIterator2 __last2,
+__parallel_set_union_op(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __exec, _RandomAccessIterator1 __first1,
+                        _RandomAccessIterator1 __last1, _RandomAccessIterator2 __first2, _RandomAccessIterator2 __last2,
                         _OutputIterator __result, _Compare __comp, _SetUnionOp __set_union_op)
 {
     using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
 
-    typedef typename ::std::iterator_traits<_ForwardIterator1>::difference_type _DifferenceType;
+    typedef typename ::std::iterator_traits<_RandomAccessIterator1>::difference_type _DifferenceType;
 
     const auto __n1 = __last1 - __first1;
     const auto __n2 = __last2 - __first2;
@@ -3307,7 +3307,7 @@ __parallel_set_union_op(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __ex
                                                  __result, __copy_range);
 
     // testing  whether the sequences are intersected
-    _ForwardIterator1 __left_bound_seq_1 = ::std::lower_bound(__first1, __last1, *__first2, __comp);
+    _RandomAccessIterator1 __left_bound_seq_1 = ::std::lower_bound(__first1, __last1, *__first2, __comp);
 
     if (__left_bound_seq_1 == __last1)
     {
@@ -3326,7 +3326,7 @@ __parallel_set_union_op(__parallel_tag<_IsVector> __tag, _ExecutionPolicy&& __ex
     }
 
     // testing  whether the sequences are intersected
-    _ForwardIterator2 __left_bound_seq_2 = ::std::lower_bound(__first2, __last2, *__first1, __comp);
+    _RandomAccessIterator2 __left_bound_seq_2 = ::std::lower_bound(__first2, __last2, *__first1, __comp);
 
     if (__left_bound_seq_2 == __last2)
     {
@@ -4248,10 +4248,11 @@ __pattern_shift_left(_Tag, _ExecutionPolicy&&, _ForwardIterator __first, _Forwar
     return __brick_shift_left(__first, __last, __n, typename _Tag::__is_vector{});
 }
 
-template <class _IsVector, class _ExecutionPolicy, class _ForwardIterator>
-_ForwardIterator
-__pattern_shift_left(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _ForwardIterator __first,
-                     _ForwardIterator __last, typename ::std::iterator_traits<_ForwardIterator>::difference_type __n)
+template <class _IsVector, class _ExecutionPolicy, class _RandomAccessIterator>
+_RandomAccessIterator
+__pattern_shift_left(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _RandomAccessIterator __first,
+                     _RandomAccessIterator __last,
+                     typename ::std::iterator_traits<_RandomAccessIterator>::difference_type __n)
 {
     using __backend_tag = typename __parallel_tag<_IsVector>::__backend_tag;
 
@@ -4262,7 +4263,7 @@ __pattern_shift_left(__parallel_tag<_IsVector>, _ExecutionPolicy&& __exec, _Forw
     if (__n >= __size)
         return __first;
 
-    using _DiffType = typename ::std::iterator_traits<_ForwardIterator>::difference_type;
+    using _DiffType = typename ::std::iterator_traits<_RandomAccessIterator>::difference_type;
 
     _DiffType __mid = __size / 2 + __size % 2;
     _DiffType __size_res = __size - __n;

From e26873bf1cbeb90c988bc6c220a2ceb19ef15f16 Mon Sep 17 00:00:00 2001
From: Sergey Kopienko <>
Date: Wed, 20 Mar 2024 12:43:01 +0100
Subject: [PATCH 566/566] include/oneapi/dpl/pstl/iterator_defs.h - remove
 template alias __is_random_access_iterator_t as unused anymore

---
 include/oneapi/dpl/pstl/iterator_defs.h | 3 ---
 1 file changed, 3 deletions(-)

diff --git a/include/oneapi/dpl/pstl/iterator_defs.h b/include/oneapi/dpl/pstl/iterator_defs.h
index f757770f28f..1b85d74fc40 100644
--- a/include/oneapi/dpl/pstl/iterator_defs.h
+++ b/include/oneapi/dpl/pstl/iterator_defs.h
@@ -49,9 +49,6 @@ struct __is_forward_iterator : decltype(__is_iterator_of<::std::forward_iterator
 {
 };
 
-template <typename... _IteratorTypes>
-using __is_random_access_iterator_t = typename __is_random_access_iterator<_IteratorTypes...>::type;
-
 template <typename... _IteratorTypes>
 inline constexpr bool __is_random_access_iterator_v = __is_random_access_iterator<_IteratorTypes...>::value;