Fixed failing iamin tests

codeplaysoftware · Nov 27, 2023 · 3219bcb · 3219bcb
1 parent 773595c
commit 3219bcb
Show file tree

Hide file tree

Showing 4 changed files with 12 additions and 12 deletions.
diff --git a/src/interface/blas1/backend/amd_gpu.hpp b/src/interface/blas1/backend/amd_gpu.hpp
@@ -86,8 +86,9 @@ typename sb_handle_t::event_t _iamin(
     constexpr int localSize = 512;
     const index_t nWG = std::min((_N + localSize - 1) / (localSize * 4),
                                  static_cast<index_t>(512));
-    return blas::internal::_iamax_iamin_impl<localSize, 0, false, false>(
-        sb_handle, _N, _vx, _incx, _rs, nWG, _dependencies);
+    return blas::internal::_iamax_iamin_impl<localSize, localSize, false,
+                                             false>(sb_handle, _N, _vx, _incx,
+                                                    _rs, nWG, _dependencies);
   }
 }
 }  // namespace backend

diff --git a/src/interface/blas1/backend/default_cpu.hpp b/src/interface/blas1/backend/default_cpu.hpp
@@ -49,12 +49,11 @@ template <typename sb_handle_t, typename container_0_t, typename container_1_t,
 typename sb_handle_t::event_t _iamax(
     sb_handle_t& sb_handle, index_t _N, container_0_t _vx, increment_t _incx,
     container_1_t _rs, const typename sb_handle_t::event_t& _dependencies) {
+  constexpr int localSize = 128;
   if (_N < 8192) {
-    constexpr int localSize = 128;
     return blas::internal::_iamax_iamin_impl<localSize, 0, true, true>(
         sb_handle, _N, _vx, _incx, _rs, static_cast<index_t>(1), _dependencies);
   } else {
-    constexpr int localSize = 128;
     const index_t nWG = std::min((_N + localSize - 1) / (localSize * 4),
                                  static_cast<index_t>(512));
     return blas::internal::_iamax_iamin_impl<localSize, 0, true, false>(
@@ -71,12 +70,11 @@ template <typename sb_handle_t, typename container_0_t, typename container_1_t,
 typename sb_handle_t::event_t _iamin(
     sb_handle_t& sb_handle, index_t _N, container_0_t _vx, increment_t _incx,
     container_1_t _rs, const typename sb_handle_t::event_t& _dependencies) {
+  constexpr int localSize = 128;
   if (_N < 8192) {
-    constexpr int localSize = 128;
     return blas::internal::_iamax_iamin_impl<localSize, 0, false, true>(
         sb_handle, _N, _vx, _incx, _rs, static_cast<index_t>(1), _dependencies);
   } else {
-    constexpr int localSize = 128;
     const index_t nWG = std::min((_N + localSize - 1) / (localSize * 4),
                                  static_cast<index_t>(512));
     return blas::internal::_iamax_iamin_impl<localSize, 0, false, false>(

diff --git a/src/interface/blas1/backend/intel_gpu.hpp b/src/interface/blas1/backend/intel_gpu.hpp
@@ -50,12 +50,11 @@ template <typename sb_handle_t, typename container_0_t, typename container_1_t,
 typename sb_handle_t::event_t _iamax(
     sb_handle_t& sb_handle, index_t _N, container_0_t _vx, increment_t _incx,
     container_1_t _rs, const typename sb_handle_t::event_t& _dependencies) {
+  constexpr int localSize = 128;
   if (_N < 8192) {
-    constexpr int localSize = 128;
     return blas::internal::_iamax_iamin_impl<localSize, localSize, true, true>(
         sb_handle, _N, _vx, _incx, _rs, static_cast<index_t>(1), _dependencies);
   } else {
-    constexpr int localSize = 128;
     const index_t nWG = std::min((_N + localSize - 1) / (localSize * 4),
                                  static_cast<index_t>(512));
     return blas::internal::_iamax_iamin_impl<localSize, localSize, true, false>(
@@ -72,12 +71,11 @@ template <typename sb_handle_t, typename container_0_t, typename container_1_t,
 typename sb_handle_t::event_t _iamin(
     sb_handle_t& sb_handle, index_t _N, container_0_t _vx, increment_t _incx,
     container_1_t _rs, const typename sb_handle_t::event_t& _dependencies) {
+  constexpr int localSize = 128;
   if (_N < 8192) {
-    constexpr int localSize = 128;
     return blas::internal::_iamax_iamin_impl<localSize, localSize, false, true>(
         sb_handle, _N, _vx, _incx, _rs, static_cast<index_t>(1), _dependencies);
   } else {
-    constexpr int localSize = 128;
     const index_t nWG = std::min((_N + localSize - 1) / (localSize * 4),
                                  static_cast<index_t>(512));
     return blas::internal::_iamax_iamin_impl<localSize, localSize, false,

diff --git a/src/interface/blas1_interface.hpp b/src/interface/blas1_interface.hpp
@@ -257,7 +257,7 @@ typename sb_handle_t::event_t _asum_impl(
  *
  * @tparam localSize value to indicate work group size
  * @tparam localMemSize value to indicate size of local memory required
- * @tparam is_max boolean variable to indicate if required operation is 
+ * @tparam is_max boolean variable to indicate if required operation is
  * iamax or not
  * @tparam single boolean variable to indicate whether to execute a single
  * step reduction or a two step reduction
@@ -322,7 +322,10 @@ typename sb_handle_t::event_t _iamax_iamin_impl(
     auto step0 = make_integer_max_min<is_max, true>(gpu_res_vec, tupOp);
     auto step1 = make_integer_max_min<is_max, false>(rs, gpu_res_vec);
     if constexpr (localMemSize == 0) {
-      tuple_t init{(index_t)0, (scalar_t)0.f};
+      const scalar_t val = is_max ? static_cast<scalar_t>(0)
+                                  : std::numeric_limits<scalar_t>::max();
+      const index_t idx = std::numeric_limits<index_t>::max();
+      const tuple_t init{idx, val};
       ret = typename sb_handle_t::event_t{
           helper::fill(q, gpu_res, init, memory_size, _dependencies)};
       ret = concatenate_vectors(