computeTFactor parallelise computation with bulk (MC Local and Distributed)

include/dlaf/eigensolver/eigensolver/impl.h

@@ -65,6 +65,7 @@ void Eigensolver<B, D, T>::call(comm::CommunicatorGrid grid, blas::Uplo uplo, Ma
   eigensolver::tridiagSolver<B>(grid, ret.tridiagonal, evals, mat_e);
 
   backTransformationBandToTridiag<B>(grid, band_size, mat_e, ret.hh_reflectors);
+  pika::threads::get_thread_manager().wait();
   backTransformationReductionToBand<B>(grid, band_size, mat_e, mat_a, taus);
 }
 }

include/dlaf/eigensolver/get_tfactor_nworkers.h

@@ -0,0 +1,32 @@
+//
+// Distributed Linear Algebra with Future (DLAF)
+//
+// Copyright (c) 2018-2023, ETH Zurich
+// All rights reserved.
+//
+// Please, refer to the LICENSE file in the root directory.
+// SPDX-License-Identifier: BSD-3-Clause
+//
+#pragma once
+
+#include <algorithm>
+#include <cmath>
+
+#include <pika/runtime.hpp>
+
+#include "dlaf/common/assert.h"
+#include "dlaf/tune.h"
+
+namespace dlaf::factorization::internal {
+
+inline size_t getTFactorNWorkers() noexcept {
+  const size_t nworkers = getTuneParameters().tfactor_nworkers;
+
+  // Note: precautionarily we leave at least 1 thread "free" to do other stuff
+  const size_t max_workers = pika::resource::get_thread_pool("default").get_os_thread_count() - 1;
+
+  // 1 <= number of workers < max_workers
+  return std::max<std::size_t>(1, std::min<std::size_t>(max_workers, nworkers));
+}
+
+}

include/dlaf/factorization/qr.h

@@ -27,20 +27,69 @@ namespace dlaf::factorization::internal {
 /// H0 H1 H2 ... HK-1
 /// Note: The first element of the HH reflectors is NOT implicitly assumed to be 1,
 ///       it has to be set correctly in the panel (0s as well).
-
+///
+/// It is similar to what xLARFT in LAPACK does.
+/// Given @p k elementary reflectors stored in the column of @p hh_panel together with related tau values
+/// in @p taus, in @p t will be formed the triangular factor for the H block of reflectors, such that
+///
+/// H = I - V . T . V*
+///
+/// where H = H1 . H2 . ... . Hk
+///
+/// in which Hi represents a single elementary reflector transformation.
+///
 /// A Storage-Efficient WY Representation for Products of Householder Transformations.
 /// Schreiber, Robert & VanLoan, Charles. (1989)
 /// SIAM Journal on Scientific and Statistical Computing. 10. 10.1137/0910005.
 ///
-/// @pre taus contains a vector with k elements
-/// @pre t contains a (k x k) tile
+/// @param hh_panel where the elementary reflectors are stored
+/// @param taus array of taus, associated with the related elementary reflector
+/// @param t tile where the resulting T factor will be stored in its top-left sub-matrix of size
+/// TileElementSize(k, k)
+///
+/// @pre hh_pane.getWidth() <= t.get().size().rows && hh_panel.size().getWidth() <= t.get().size().cols()
 template <Backend backend, Device device, class T>
 void computeTFactor(matrix::Panel<Coord::Col, T, device>& hh_panel,
                     pika::shared_future<common::internal::vector<T>> taus,
                     pika::future<matrix::Tile<T, device>> t) {
   QR_Tfactor<backend, device, T>::call(hh_panel, taus, std::move(t));
 }
 
+/// Forms the triangular factor T of a block of reflectors H, which is defined as a product of
+/// k := hh_panel.getWidth() elementary reflectors.
+///
+/// hh_panel should have the following form
+/// H0  0  0 ...    0
+///  . H1  0 ...    0
+///  .  . H2 ...    0
+///  .  .  . ...    0
+///  .  .  . ... HK-1
+///  .  .  . ...    .
+/// H0 H1 H2 ... HK-1
+/// Note: The first element of the HH reflectors is NOT implicitly assumed to be 1,
+///       it has to be set correctly in the panel (0s as well).
+///
+/// It is similar to what xLARFT in LAPACK does.
+/// Given @p k elementary reflectors stored in the column of @p hh_panel together with related tau values
+/// in @p taus, in @p t will be formed the triangular factor for the H block of reflectors, such that
+///
+/// H = I - V . T . V*
+///
+/// where H = H1 . H2 . ... . Hk
+///
+/// in which Hi represents a single elementary reflector transformation.
+///
+/// A Storage-Efficient WY Representation for Products of Householder Transformations.
+/// Schreiber, Robert & VanLoan, Charles. (1989)
+/// SIAM Journal on Scientific and Statistical Computing. 10. 10.1137/0910005.
+///
+/// @param hh_panel where the elementary reflectors are stored
+/// @param taus array of taus, associated with the related elementary reflector
+/// @param t tile where the resulting T factor will be stored in its top-left sub-matrix of size
+/// TileElementSize(k, k)
+/// @param mpi_col_task_chain where internal communications are issued
+///
+/// @pre hh_pane.getWidth() <= t.get().size().rows && hh_panel.size().getWidth() <= t.get().size().cols()
 template <Backend backend, Device device, class T>
 void computeTFactor(matrix::Panel<Coord::Col, T, device>& hh_panel,
                     pika::shared_future<common::internal::vector<T>> taus,

include/dlaf/factorization/qr/api.h

@@ -24,65 +24,31 @@ template <Backend backend, Device device, class T>
 struct QR {};
 
 template <Backend backend, Device device, class T>
-struct QR_Tfactor {
-  /// Forms the triangular factor T of a block of reflectors H, which is defined as a product of k
-  /// elementary reflectors.
-  ///
-  /// It is similar to what xLARFT in LAPACK does.
-  /// Given @p k elementary reflectors stored in the column of @p v starting at tile @p v_start,
-  /// together with related tau values in @p taus, in @p t will be formed the triangular factor for the H
-  /// block of reflector, such that
-  ///
-  /// H = I - V . T . V*
-  ///
-  /// where H = H1 . H2 . ... . Hk
-  ///
-  /// in which Hi represents a single elementary reflector transformation
-  ///
-  /// @param k the number of elementary reflectors to use (from the beginning of the tile)
-  /// @param v where the elementary reflectors are stored
-  /// @param v_start tile in @p v where the column of reflectors starts
-  /// @param taus array of taus, associated with the related elementary reflector
-  /// @param t tile where the resulting T factor will be stored in its top-left sub-matrix of size
-  /// TileElementSize(k, k)
-  ///
-  /// @pre k <= t.get().size().rows && k <= t.get().size().cols()
-  /// @pre k >= 0
-  /// @pre v_start.isIn(v.nrTiles())
-  static void call(matrix::Panel<Coord::Col, T, device>& panel_view,
+struct QR_Tfactor;
+
+template <class T>
+struct QR_Tfactor<Backend::MC, Device::CPU, T> {
+  static void call(matrix::Panel<Coord::Col, T, Device::CPU>& panel_view,
+                   pika::shared_future<common::internal::vector<T>> taus,
+                   pika::future<matrix::Tile<T, Device::CPU>> t);
+  static void call(matrix::Panel<Coord::Col, T, Device::CPU>& hh_panel,
                    pika::shared_future<common::internal::vector<T>> taus,
-                   pika::future<matrix::Tile<T, device>> t);
+                   pika::future<matrix::Tile<T, Device::CPU>> t,
+                   common::Pipeline<comm::Communicator>& mpi_col_task_chain);
+};
 
-  /// Forms the triangular factor T of a block of reflectors H, which is defined as a product of k
-  /// elementary reflectors.
-  ///
-  /// It is similar to what xLARFT in LAPACK does.
-  /// Given @p k elementary reflectors stored in the column of @p v starting at tile @p v_start,
-  /// together with related tau values in @p taus, in @p t will be formed the triangular factor for the H
-  /// block of reflector, such that
-  ///
-  /// H = I - V . T . V*
-  ///
-  /// where H = H1 . H2 . ... . Hk
-  ///
-  /// in which Hi represents a single elementary reflector transformation
-  ///
-  /// @param k the number of elementary reflectors to use (from the beginning of the tile)
-  /// @param v where the elementary reflectors are stored
-  /// @param v_start tile in @p v where the column of reflectors starts
-  /// @param taus array of taus, associated with the related elementary reflector
-  /// @param t tile where the resulting T factor will be stored in its top-left sub-matrix of size
-  /// TileElementSize(k, k)
-  /// @param mpi_col_task_chain where internal communications are issued
-  ///
-  /// @pre k <= t.get().size().rows && k <= t.get().size().cols()
-  /// @pre k >= 0
-  /// @pre v_start.isIn(v.nrTiles())
-  static void call(matrix::Panel<Coord::Col, T, device>& hh_panel,
+#ifdef DLAF_WITH_GPU
+template <class T>
+struct QR_Tfactor<Backend::GPU, Device::GPU, T> {
+  static void call(matrix::Panel<Coord::Col, T, Device::GPU>& panel_view,
+                   pika::shared_future<common::internal::vector<T>> taus,
+                   pika::future<matrix::Tile<T, Device::GPU>> t);
+  static void call(matrix::Panel<Coord::Col, T, Device::GPU>& hh_panel,
                    pika::shared_future<common::internal::vector<T>> taus,
-                   pika::future<matrix::Tile<T, device>> t,
+                   pika::future<matrix::Tile<T, Device::GPU>> t,
                    common::Pipeline<comm::Communicator>& mpi_col_task_chain);
 };
+#endif
 
 /// ---- ETI
 #define DLAF_FACTORIZATION_QR_TFACTOR_ETI(KWORD, BACKEND, DEVICE, DATATYPE) \