minor changes

include/dlaf/eigensolver/bt_reduction_to_band/impl.h

@@ -222,7 +222,6 @@ void BackTransformationReductionToBand<backend, device, T>::call(
     const LocalTileIndex t_index{Coord::Col, k};
 
     computeTFactor<backend>(panelV, mat_taus.read(taus_index), panelT.readwrite(t_index), panelWS);
-    panelWS.reset();
 
     // W = V T
     auto tile_t = panelT.read(t_index);
@@ -247,6 +246,7 @@ void BackTransformationReductionToBand<backend, device, T>::call(
     panelW.reset();
     panelW2.reset();
     panelT.reset();
+    panelWS.reset();
   }
 }
 
@@ -389,11 +389,11 @@ void BackTransformationReductionToBand<B, D, T>::call(comm::CommunicatorGrid& gr
                             splitTile(mat_c.readwrite(ij), mat_c_view(ij)));
     }
 
-    panelWS.reset();
     panelV.reset();
     panelW.reset();
     panelW2.reset();
     panelT.reset();
+    panelWS.reset();
   }
 }
 }

include/dlaf/factorization/qr.h

@@ -54,10 +54,10 @@ namespace dlaf::factorization::internal {
 /// @param t tile where the resulting T factor will be stored in its top-left sub-matrix of size
 /// TileElementSize(k, k)
 /// @param workspaces array of tiles used as workspace, with at least one tile per worker (see
-/// get_tfactor_num_workers), each tile should have the same size as @param tile_t
+/// get_tfactor_num_workers), each tile should be at least of size TileElementSize(k, k)
 ///
 /// @pre reflectors in hh_panel are well formed (1s on the diagonal and 0s in the upper part)
-/// @pre hh_panel.getWidth() <= t.get().size().rows && hh_panel.size().getWidth() <= t.get().size().cols()
+/// @pre hh_panel.getWidth() <= t.get().size().rows && hh_panel.getWidth() <= t.get().size().cols()
 template <Backend backend, Device device, class T>
 void computeTFactor(matrix::Panel<Coord::Col, T, device>& hh_panel,
                     matrix::ReadOnlyTileSender<T, Device::CPU> taus,
@@ -99,11 +99,11 @@ void computeTFactor(matrix::Panel<Coord::Col, T, device>& hh_panel,
 /// @param t tile where the resulting T factor will be stored in its top-left sub-matrix of size
 /// TileElementSize(k, k)
 /// @param workspaces array of tiles used as workspace, with at least one tile per worker (see
-/// get_tfactor_num_workers), each tile should have the same size as @param tile_t
+/// get_tfactor_num_workers), each tile should be at least of size TileElementSize(k, k)
 /// @param mpi_col_task_chain where internal communications are issued
 ///
 /// @pre reflectors in hh_panel are well formed (1s on the diagonal and 0s in the upper part)
-/// @pre hh_panel.getWidth() <= t.get().size().rows && hh_panel.size().getWidth() <= t.get().size().cols()
+/// @pre hh_panel.getWidth() <= t.get().size().rows && hh_panel.getWidth() <= t.get().size().cols()
 template <Backend backend, Device device, class T>
 void computeTFactor(matrix::Panel<Coord::Col, T, device>& hh_panel,
                     matrix::ReadOnlyTileSender<T, Device::CPU> taus,

include/dlaf/factorization/qr/t_factor_impl.h

@@ -158,8 +158,6 @@ struct Helpers<Backend::MC, Device::CPU, T> {
                               (worker_id == 0 ? tile_t : workspaces[worker_id - 1])
                                   .subTileReference({{0, 0}, tile_t.size()});
 
-                          DLAF_ASSERT(equal_size(ws_worker, tile_t), ws_worker.size(), tile_t.size());
-
                           tile::internal::set0<T>(ws_worker);
                           lapack::lacpy(blas::Uplo::General, 1, k, taus.get().ptr(), 1, ws_worker.ptr(),
                                         ws_worker.ld() + 1);
@@ -296,10 +294,9 @@ struct Helpers<Backend::GPU, Device::GPU, T> {
               di::Policy<Backend::GPU>(thread_priority::high),
               [k](cublasHandle_t handle, auto&& hh_tiles, auto&& taus,
                   matrix::Tile<T, Device::GPU>& tile_t_full) {
-                matrix::Tile<T, Device::GPU> tile_t = tile_t_full.subTileReference({{0, 0}, {k, k}});
-
                 DLAF_ASSERT_MODERATE(k == taus.size().rows(), k, taus.size().rows());
-                DLAF_ASSERT(tile_t.size() == TileElementSize(k, k), tile_t.size(), k);
+
+                matrix::Tile<T, Device::GPU> tile_t = tile_t_full.subTileReference({{0, 0}, {k, k}});
 
                 // Note:
                 // prepare the diagonal of taus in t and reset the rest

src/init.cpp

@@ -349,8 +349,8 @@ pika::program_options::options_description getOptionsDescription() {
 
   // Tune parameters command line options
   desc.add_options()("dlaf:tfactor-num-threads", pika::program_options::value<std::size_t>(), "The maximum number of threads to use for computing the tfactor.");
-  desc.add_options()("dlaf:tfactor-num-streams", pika::program_options::value<std::size_t>(), "The maximum number of threads to use for computing the tfactor.");
-  desc.add_options()("dlaf:tfactor-barrier-busy-wait-us", pika::program_options::value<std::size_t>(), "The duration in microseconds to busy-wait in barriers in the tfactor t algorithm.");
+  desc.add_options()("dlaf:tfactor-num-streams", pika::program_options::value<std::size_t>(), "The maximum number of GPU streams to use for computing the tfactor.");
+  desc.add_options()("dlaf:tfactor-barrier-busy-wait-us", pika::program_options::value<std::size_t>(), "The duration in microseconds to busy-wait in barriers in the tfactor algorithm.");
   desc.add_options()("dlaf:red2band-panel-nworkers", pika::program_options::value<std::size_t>(), "The maximum number of threads to use for computing the panel in the reduction to band algorithm.");
   desc.add_options()("dlaf:red2band-barrier-busy-wait-us", pika::program_options::value<std::size_t>(), "The duration in microseconds to busy-wait in barriers in the reduction to band algorithm.");
   desc.add_options()("dlaf:eigensolver-min-band", pika::program_options::value<SizeType>(), "The minimum value to start looking for a divisor of the block size. When larger than the block size, the block size will be used instead.");