Merge Pull Request #12793 from trilinos/Trilinos/master_merge_2024030…

…1_175926 Automatically Merged using Trilinos Master Merge AutoTester PR Title: b'Trilinos Master Merge PR Generator: Auto PR created to promote from master_merge_20240301_175926 branch to master' PR Author: trilinos-autotester
trilinos · Mar 2, 2024 · bee69a6 · bee69a6
2 parents a791e19 + a469209
commit bee69a6
Show file tree

Hide file tree

Showing 123 changed files with 3,242 additions and 1,482 deletions.
diff --git a/RELEASE_NOTES b/RELEASE_NOTES
@@ -1,75 +1,93 @@
 
 ###############################################################################
 #                                                                             #
-# Trilinos Release 15.1 Release Notes                               TBD, 2024 #
+# Trilinos Release 15.1.0 Release Notes                     February 26, 2024 #
 #                                                                             #
 ###############################################################################
 
 Amesos2
 
   - The interface to SuperLU_DIST now also works for the CUDA-enabled
     variant of the library.
+    https://github.com/trilinos/Trilinos/pull/12524
+
+
+Framework
+
+  - Began using semantic versioning for Trilinos with 15.1.0 release.
 
 
 Ifpack2
 
   - BlockRelaxation can now generate blocks using a Zoltan2.
+    https://github.com/trilinos/Trilinos/pull/12728
 
 
 Kokkos & Kokkos Kernels
 
   - Inclusion of version 4.2.1 of Kokkos and Kokkos Kernels
+    https://github.com/trilinos/Trilinos/pull/12707
 
 
 MueLu
 
   - The reformulated Maxwell solver (RefMaxwell) was generalized to
     also work for grad-div / Darcy flow problems.
+    https://github.com/trilinos/Trilinos/pull/12142
 
   - In an effort to consolidate the old non-Kokkos code path with the
     newer Kokkos code path, the following factories were deprecated
     and should be removed from input decks: NullspaceFactory_kokkos,
     SaPFactory_kokkos, UncoupledAggregationFactory_kokkos.
+    https://github.com/trilinos/Trilinos/pull/12720
+    https://github.com/trilinos/Trilinos/pull/12740
 
 
 Panzer
 
   - MiniEM can now also assemble and solve Darcy problems using first
     or higher order mixed finite elements.
+    https://github.com/trilinos/Trilinos/pull/12142
 
 
 PyTrilinos2
 
   - New package that auto-generates Python interfaces for Trilinos
     packages. Currently, most of Tpetra is exposed. We are planning on
     adding other packages.
+    https://github.com/trilinos/Trilinos/pull/12332
 
 
 ROL
 
   - An auto-generated Python interface was added. A standalone Python
     package can be downloaded from rol.sandia.gov
+    https://github.com/trilinos/Trilinos/pull/12770
 
 
 Teko
 
   - Block Jacobi and Gauss-Seidel methods allow now to specify
     preconditioners for the iterative solves of the diagonal blocks.
+    https://github.com/trilinos/Trilinos/pull/12675
 
 
 Tpetra
 
   - Tpetra will now assume by default that the MPI library is GPU
     aware, unless automatic detection or the user indicates otherwise.
+    https://github.com/trilinos/Trilinos/pull/12517
 
   - Reject unrecognized TPETRA_* environment variable. Misspelled or
     removed environment variables are no longer silently ignored.
+    https://github.com/trilinos/Trilinos/pull/12722
 
   - In order to allocate in shared host/device space (i.e.
     CudaUVMSpace, HIPManagedSpace or SYCLSharedUSMSpace) by default,
     please use the CMake options
     KokkosKernels_INST_MEMSPACE_CUDAUVMSPACE=ON
     Tpetra_ALLOCATE_IN_SHARED_SPACE=ON
+    https://github.com/trilinos/Trilinos/pull/12622
 
 
 ###############################################################################

diff --git a/Version.cmake b/Version.cmake
@@ -59,10 +59,10 @@
 # for release mode and set the version.
 #
 
-SET(Trilinos_VERSION 15.1)
+SET(Trilinos_VERSION 15.2.0)
 SET(Trilinos_MAJOR_VERSION 15)
-SET(Trilinos_MAJOR_MINOR_VERSION 150100)
-SET(Trilinos_VERSION_STRING "15.1 (Dev)")
+SET(Trilinos_MAJOR_MINOR_VERSION 150200)
+SET(Trilinos_VERSION_STRING "15.2.0-dev")
 SET(Trilinos_ENABLE_DEVELOPMENT_MODE_DEFAULT ON) # Change to 'OFF' for a release
 
 # Used by testing scripts and should not be used elsewhere

diff --git a/packages/framework/ini-files/config-specs.ini b/packages/framework/ini-files/config-specs.ini
@@ -2409,9 +2409,9 @@ use CUDA11-RUN-SERIAL-TESTS
 opt-set-cmake-var ROL_example_PinT_parabolic-control_AugmentedSystem_test_MPI_2_DISABLE BOOL FORCE : ON
 
 
-[rhel7_sems-cuda-11.4.2-sems-gnu-10.1.0-sems-openmpi-4.0.5_release_static_Volta70_no-asan_complex_no-fpic_mpi_pt_no-rdc_uvm_deprecated-on_all]
+[rhel7_sems-cuda-11.4.2-sems-gnu-10.1.0-sems-openmpi-4.0.5_release_static_Volta70_no-asan_complex_no-fpic_mpi_pt_no-rdc_uvm_deprecated-on_pr]
 use rhel7_sems-cuda-11.4.2-sems-gnu-10.1.0-sems-openmpi-4.0.5_release_static_Volta70_no-asan_complex_no-fpic_mpi_pt_no-rdc_uvm_deprecated-on_no-package-enables
-use PACKAGE-ENABLES|ALL
+use PACKAGE-ENABLES|PR
 opt-set-cmake-var Trilinos_ENABLE_TESTS BOOL FORCE : OFF
 
 

diff --git a/packages/ifpack2/src/Ifpack2_BlockTriDiContainer_def.hpp b/packages/ifpack2/src/Ifpack2_BlockTriDiContainer_def.hpp
@@ -187,7 +187,7 @@ namespace Ifpack2 {
     const bool useSeqMethod = false;
     const bool overlapCommAndComp = false;
     initInternal(matrix, importer, overlapCommAndComp, useSeqMethod);
-    n_subparts_per_part_ = 1;
+    n_subparts_per_part_ = -1;
     IFPACK2_BLOCKHELPER_TIMER_FENCE(typename BlockHelperDetails::ImplType<MatrixType>::execution_space)
   }
 

diff --git a/packages/ifpack2/src/Ifpack2_BlockTriDiContainer_impl.hpp b/packages/ifpack2/src/Ifpack2_BlockTriDiContainer_impl.hpp
@@ -842,22 +842,83 @@ namespace Ifpack2 {
       return Teuchos::null;
     }
 
+    template<typename local_ordinal_type>
+    local_ordinal_type costTRSM(const local_ordinal_type block_size) {
+      return block_size*block_size;
+    }
+
+    template<typename local_ordinal_type>
+    local_ordinal_type costGEMV(const local_ordinal_type block_size) {
+      return 2*block_size*block_size;
+    }
+
+    template<typename local_ordinal_type>
+    local_ordinal_type costTriDiagSolve(const local_ordinal_type subline_length, const local_ordinal_type block_size) {
+      return 2 * subline_length * costTRSM(block_size) + 2 * (subline_length-1) * costGEMV(block_size);
+    }
+
+    template<typename local_ordinal_type>
+    local_ordinal_type costSolveSchur(const local_ordinal_type num_parts,
+                                      const local_ordinal_type num_teams,
+                                      const local_ordinal_type line_length,
+                                      const local_ordinal_type block_size,
+                                      const local_ordinal_type n_subparts_per_part) {
+      const local_ordinal_type subline_length = ceil(double(line_length - (n_subparts_per_part-1) * 2) / n_subparts_per_part);
+      if (subline_length < 1) {
+        return INT_MAX;
+      }
+
+      const local_ordinal_type p_n_lines = ceil(double(num_parts)/num_teams);
+      const local_ordinal_type p_n_sublines = ceil(double(n_subparts_per_part)*num_parts/num_teams);
+      const local_ordinal_type p_n_sublines_2 = ceil(double(n_subparts_per_part-1)*num_parts/num_teams);
+
+      const local_ordinal_type p_costApplyE = p_n_sublines_2 * subline_length * 2 * costGEMV(block_size);
+      const local_ordinal_type p_costApplyS = p_n_lines * costTriDiagSolve((n_subparts_per_part-1)*2,block_size);
+      const local_ordinal_type p_costApplyAinv = p_n_sublines * costTriDiagSolve(subline_length,block_size);
+      const local_ordinal_type p_costApplyC = p_n_sublines_2 * 2 * costGEMV(block_size);
+
+      if (n_subparts_per_part == 1) {
+        return p_costApplyAinv;
+      }
+      return p_costApplyE + p_costApplyS + p_costApplyAinv + p_costApplyC;
+    }
+
+    template<typename local_ordinal_type>
+    local_ordinal_type getAutomaticNSubparts(const local_ordinal_type num_parts,
+                                             const local_ordinal_type num_teams,
+                                             const local_ordinal_type line_length,
+                                             const local_ordinal_type block_size) {
+      local_ordinal_type n_subparts_per_part_0 = 1;
+      local_ordinal_type flop_0 = costSolveSchur(num_parts, num_teams, line_length, block_size, n_subparts_per_part_0);
+      local_ordinal_type flop_1 = costSolveSchur(num_parts, num_teams, line_length, block_size, n_subparts_per_part_0+1);
+      while (flop_0 > flop_1) {
+        flop_0 = flop_1;
+        flop_1 = costSolveSchur(num_parts, num_teams, line_length, block_size, (++n_subparts_per_part_0)+1);
+      }
+      return n_subparts_per_part_0;
+    }
+
+    template<typename ArgActiveExecutionMemorySpace>
+    struct SolveTridiagsDefaultModeAndAlgo;
+
     ///
     /// setup part interface using the container partitions array
     ///
     template<typename MatrixType>
     BlockHelperDetails::PartInterface<MatrixType>
     createPartInterface(const Teuchos::RCP<const typename BlockHelperDetails::ImplType<MatrixType>::tpetra_block_crs_matrix_type> &A,
                         const Teuchos::Array<Teuchos::Array<typename BlockHelperDetails::ImplType<MatrixType>::local_ordinal_type> > &partitions,
-                        const typename BlockHelperDetails::ImplType<MatrixType>::local_ordinal_type n_subparts_per_part) {
+                        const typename BlockHelperDetails::ImplType<MatrixType>::local_ordinal_type n_subparts_per_part_in) {
       IFPACK2_BLOCKHELPER_TIMER("createPartInterface");
       using impl_type = BlockHelperDetails::ImplType<MatrixType>;
       using local_ordinal_type = typename impl_type::local_ordinal_type;
       using local_ordinal_type_1d_view = typename impl_type::local_ordinal_type_1d_view;
       using local_ordinal_type_2d_view = typename impl_type::local_ordinal_type_2d_view;
       using size_type = typename impl_type::size_type;
 
+      const auto blocksize = A->getBlockSize();
       constexpr int vector_length = impl_type::vector_length;
+      constexpr int internal_vector_length = impl_type::internal_vector_length;
 
       const auto comm = A->getRowMap()->getComm();
 
@@ -867,6 +928,40 @@ namespace Ifpack2 {
       const local_ordinal_type A_n_lclrows = A->getLocalNumRows();
       const local_ordinal_type nparts = jacobi ? A_n_lclrows : partitions.size();
 
+      typedef std::pair<local_ordinal_type,local_ordinal_type> size_idx_pair_type;
+      std::vector<size_idx_pair_type> partsz(nparts);
+
+      if (!jacobi) {
+        for (local_ordinal_type i=0;i<nparts;++i)
+          partsz[i] = size_idx_pair_type(partitions[i].size(), i);
+        std::sort(partsz.begin(), partsz.end(),
+                  [] (const size_idx_pair_type& x, const size_idx_pair_type& y) {
+                    return x.first > y.first;
+                  });
+      }
+
+      local_ordinal_type n_subparts_per_part;
+      if (n_subparts_per_part_in == -1) {
+        // If the number of subparts is set to -1, the user let the algorithm
+        // decides the value automatically
+        using execution_space = typename impl_type::execution_space;
+
+        const int line_length = partsz[0].first;
+
+        const local_ordinal_type team_size = 
+          SolveTridiagsDefaultModeAndAlgo<typename execution_space::memory_space>::
+          recommended_team_size(blocksize, vector_length, internal_vector_length);
+
+        const local_ordinal_type num_teams = execution_space().concurrency() / (team_size * vector_length);
+
+        n_subparts_per_part = getAutomaticNSubparts(nparts, num_teams, line_length, blocksize);
+
+        printf("Automatically chosen n_subparts_per_part = %d for nparts = %d, num_teams = %d, team_size = %d, line_length = %d, and blocksize = %d;\n", n_subparts_per_part, nparts, num_teams, team_size, line_length, blocksize);
+      }
+      else {
+        n_subparts_per_part = n_subparts_per_part_in;
+      }
+
       // Total number of sub lines:
       const local_ordinal_type n_sub_parts = nparts * n_subparts_per_part;
       // Total number of sub lines + the Schur complement blocks.
@@ -896,14 +991,6 @@ namespace Ifpack2 {
         // reorder parts to maximize simd packing efficiency
         p.resize(nparts);
 
-        typedef std::pair<local_ordinal_type,local_ordinal_type> size_idx_pair_type;
-        std::vector<size_idx_pair_type> partsz(nparts);
-        for (local_ordinal_type i=0;i<nparts;++i)
-          partsz[i] = size_idx_pair_type(partitions[i].size(), i);
-        std::sort(partsz.begin(), partsz.end(),
-                  [] (const size_idx_pair_type& x, const size_idx_pair_type& y) {
-                    return x.first > y.first;
-                  });
         for (local_ordinal_type i=0;i<nparts;++i)
           p[i] = partsz[i].second;
 
@@ -2074,9 +2161,6 @@ namespace Ifpack2 {
     };
 #endif
 
-    template<typename ArgActiveExecutionMemorySpace>
-    struct SolveTridiagsDefaultModeAndAlgo;
-
     template<typename impl_type, typename WWViewType>
     KOKKOS_INLINE_FUNCTION
     void
@@ -3251,7 +3335,7 @@ namespace Ifpack2 {
 
           {
 #ifdef IFPACK2_BLOCKTRIDICONTAINER_USE_PRINTF
-        printf("Star ComputeSchurTag\n");
+        printf("Start ComputeSchurTag\n");
 #endif
             IFPACK2_BLOCKHELPER_TIMER("BlockTriDi::NumericPhase::ComputeSchurTag");
             writeBTDValuesToFile(part2packrowidx0_sub.extent(0), scalar_values_schur, "before_schur.mm");
@@ -3270,7 +3354,7 @@ namespace Ifpack2 {
 
           {
 #ifdef IFPACK2_BLOCKTRIDICONTAINER_USE_PRINTF
-        printf("Star FactorizeSchurTag\n");
+        printf("Start FactorizeSchurTag\n");
 #endif
             IFPACK2_BLOCKHELPER_TIMER("BlockTriDi::NumericPhase::FactorizeSchurTag");
             Kokkos::TeamPolicy<execution_space,FactorizeSchurTag>

diff --git a/packages/ifpack2/src/Ifpack2_SparseContainer_decl.hpp b/packages/ifpack2/src/Ifpack2_SparseContainer_decl.hpp
@@ -172,7 +172,7 @@ class SparseContainer
   using inverse_mv_type = Tpetra::MultiVector<InverseScalar, InverseLocalOrdinal, InverseGlobalOrdinal, InverseNode>;
   using InverseCrs = Tpetra::CrsMatrix<InverseScalar, InverseLocalOrdinal, InverseGlobalOrdinal, InverseNode>;
   using InverseMap = typename Tpetra::Map<InverseLocalOrdinal, InverseGlobalOrdinal, InverseNode>;
-
+  using InverseGraph = typename InverseCrs::crs_graph_type;
   using typename Container<MatrixType>::HostView;
   using typename Container<MatrixType>::ConstHostView;
   using HostViewInverse = typename inverse_mv_type::dual_view_type::t_host;
@@ -287,6 +287,8 @@ class SparseContainer
 
   //! Extract the submatrices identified by the local indices set by the constructor.
   void extract ();
+  void extractGraph ();
+  void extractValues ();
 
   /// \brief Post-permutation, post-view version of apply().
   ///