First working apply for small problem

ginkgo-project · Feb 10, 2025 · 9d9a57d · 9d9a57d
1 parent 8a3296b
commit 9d9a57d
Show file tree

Hide file tree

Showing 11 changed files with 346 additions and 130 deletions.
diff --git a/common/cuda_hip/distributed/preconditioner/bddc_kernels.cpp b/common/cuda_hip/distributed/preconditioner/bddc_kernels.cpp
@@ -25,10 +25,10 @@ void classify_dofs(
     const matrix::Dense<ValueType>* labels, comm_index_type local_part,
     array<experimental::distributed::preconditioner::dof_type>& dof_types,
     array<IndexType>& permutation_array, array<IndexType>& interface_sizes,
-    array<ValueType>& owning_labels, size_type& n_inner_idxs,
-    size_type& n_face_idxs, size_type& n_edge_idxs, size_type& n_vertices,
-    size_type& n_faces, size_type& n_edges, size_type& n_constraints,
-    int& n_owning_interfaces) GKO_NOT_IMPLEMENTED;
+    array<ValueType>& unique_labels, array<ValueType>& owning_labels,
+    size_type& n_inner_idxs, size_type& n_face_idxs, size_type& n_edge_idxs,
+    size_type& n_vertices, size_type& n_faces, size_type& n_edges,
+    size_type& n_constraints, int& n_owning_interfaces) GKO_NOT_IMPLEMENTED;
 
 GKO_INSTANTIATE_FOR_EACH_NON_COMPLEX_VALUE_AND_INDEX_TYPE_BASE(
     GKO_DECLARE_CLASSIFY_DOFS);
@@ -53,16 +53,17 @@ void fill_coarse_data(std::shared_ptr<const DefaultExecutor> exec,
 GKO_INSTANTIATE_FOR_EACH_VALUE_TYPE_BASE(GKO_DECLARE_FILL_COARSE_DATA);
 
 
-template <typename ValueType>
+template <typename ValueType, typename IndexType>
 void build_coarse_contribution(
     std::shared_ptr<const DefaultExecutor> exec,
-    const matrix::Dense<remove_complex<ValueType>>* local_labels,
+    const array<remove_complex<ValueType>>& local_labels,
     const array<remove_complex<ValueType>>& global_labels,
     const matrix::Dense<ValueType>* lambda,
-    device_matrix_data<ValueType, int>& coarse_contribution)
-    GKO_NOT_IMPLEMENTED;
+    device_matrix_data<ValueType, int>& coarse_contribution,
+    array<IndexType>& permutation_array) GKO_NOT_IMPLEMENTED;
 
-GKO_INSTANTIATE_FOR_EACH_VALUE_TYPE_BASE(GKO_DECLARE_BUILD_COARSE_CONTRIBUTION);
+GKO_INSTANTIATE_FOR_EACH_VALUE_AND_INDEX_TYPE_BASE(
+    GKO_DECLARE_BUILD_COARSE_CONTRIBUTION);
 
 
 }  // namespace bddc

diff --git a/core/device_hooks/common_kernels.inc.cpp b/core/device_hooks/common_kernels.inc.cpp
@@ -387,7 +387,7 @@ GKO_STUB_NON_COMPLEX_VALUE_AND_INDEX_TYPE_BASE(GKO_DECLARE_CLASSIFY_DOFS);
 GKO_STUB_NON_COMPLEX_VALUE_AND_INDEX_TYPE_BASE(
     GKO_DECLARE_GENERATE_CONSTRAINTS);
 GKO_STUB_VALUE_TYPE_BASE(GKO_DECLARE_FILL_COARSE_DATA);
-GKO_STUB_VALUE_TYPE_BASE(GKO_DECLARE_BUILD_COARSE_CONTRIBUTION);
+GKO_STUB_VALUE_AND_INDEX_TYPE_BASE(GKO_DECLARE_BUILD_COARSE_CONTRIBUTION);
 
 
 }  // namespace bddc

diff --git a/core/distributed/dd_matrix.cpp b/core/distributed/dd_matrix.cpp
@@ -252,6 +252,7 @@ void DdMatrix<ValueType, LocalIndexType, GlobalIndexType>::read_distributed(
         exec, dim<2>{row_partition->get_size(), large_partition->get_size()},
         std::move(prolongate_row_idxs), std::move(prolongate_col_idxs),
         std::move(prolongate_values)};
+    prolongate_data.sort_row_major();
     prolongation_ =
         Matrix<ValueType, LocalIndexType, GlobalIndexType>::create(exec, comm);
     prolongation_->read_distributed(prolongate_data, row_partition,

diff --git a/core/distributed/preconditioner/bddc.cpp b/core/distributed/preconditioner/bddc.cpp
diff --git a/core/distributed/preconditioner/bddc_kernels.hpp b/core/distributed/preconditioner/bddc_kernels.hpp
@@ -1,4 +1,4 @@
-// SPDX-FileCopyrightText: 2017 - 2024 The Ginkgo authors
+// SPDX-FileCopyrightText: 2017 - 2025 The Ginkgo authors
 //
 // SPDX-License-Identifier: BSD-3-Clause
 
@@ -28,10 +28,10 @@ namespace kernels {
         const matrix::Dense<ValueType>* labels, comm_index_type local_part,    \
         array<experimental::distributed::preconditioner::dof_type>& dof_types, \
         array<IndexType>& permutation_array,                                   \
-        array<IndexType>& interface_sizes, array<ValueType>& owning_labels,    \
-        size_type& n_inner_idxs, size_type& n_face_idxs,                       \
-        size_type& n_edge_idxs, size_type& n_vertices, size_type& n_faces,     \
-        size_type& n_edges, size_type& n_constraints,                          \
+        array<IndexType>& interface_sizes, array<ValueType>& unique_labels,    \
+        array<ValueType>& owning_labels, size_type& n_inner_idxs,              \
+        size_type& n_face_idxs, size_type& n_edge_idxs, size_type& n_vertices, \
+        size_type& n_faces, size_type& n_edges, size_type& n_constraints,      \
         int& n_owning_interfaces)
 
 
@@ -49,13 +49,14 @@ namespace kernels {
                           matrix::Dense<ValueType>* lambda_rhs)
 
 
-#define GKO_DECLARE_BUILD_COARSE_CONTRIBUTION(ValueType)              \
-    void build_coarse_contribution(                                   \
-        std::shared_ptr<const DefaultExecutor> exec,                  \
-        const matrix::Dense<remove_complex<ValueType>>* local_labels, \
-        const array<remove_complex<ValueType>>& global_labels,        \
-        const matrix::Dense<ValueType>* lambda,                       \
-        device_matrix_data<ValueType, int>& coarse_contribution)
+#define GKO_DECLARE_BUILD_COARSE_CONTRIBUTION(ValueType, IndexType) \
+    void build_coarse_contribution(                                 \
+        std::shared_ptr<const DefaultExecutor> exec,                \
+        const array<remove_complex<ValueType>>& local_labels,       \
+        const array<remove_complex<ValueType>>& global_labels,      \
+        const matrix::Dense<ValueType>* lambda,                     \
+        device_matrix_data<ValueType, int>& coarse_contribution,    \
+        array<IndexType>& permutation_array)
 
 
 #define GKO_DECLARE_ALL_AS_TEMPLATES                                    \
@@ -66,8 +67,8 @@ namespace kernels {
     GKO_DECLARE_GENERATE_CONSTRAINTS(ValueType, IndexType);             \
     template <typename ValueType>                                       \
     GKO_DECLARE_FILL_COARSE_DATA(ValueType);                            \
-    template <typename ValueType>                                       \
-    GKO_DECLARE_BUILD_COARSE_CONTRIBUTION(ValueType)
+    template <typename ValueType, typename IndexType>                   \
+    GKO_DECLARE_BUILD_COARSE_CONTRIBUTION(ValueType, IndexType)
 
 
 GKO_DECLARE_FOR_ALL_EXECUTOR_NAMESPACES(bddc, GKO_DECLARE_ALL_AS_TEMPLATES);

diff --git a/dpcpp/distributed/preconditioner/bddc_kernels.dp.cpp b/dpcpp/distributed/preconditioner/bddc_kernels.dp.cpp
@@ -25,10 +25,10 @@ void classify_dofs(
     const matrix::Dense<ValueType>* labels, comm_index_type local_part,
     array<experimental::distributed::preconditioner::dof_type>& dof_types,
     array<IndexType>& permutation_array, array<IndexType>& interface_sizes,
-    array<ValueType>& owning_labels, size_type& n_inner_idxs,
-    size_type& n_face_idxs, size_type& n_edge_idxs, size_type& n_vertices,
-    size_type& n_faces, size_type& n_edges, size_type& n_constraints,
-    int& n_owning_interfaces) GKO_NOT_IMPLEMENTED;
+    array<ValueType>& unique_labels, array<ValueType>& owning_labels,
+    size_type& n_inner_idxs, size_type& n_face_idxs, size_type& n_edge_idxs,
+    size_type& n_vertices, size_type& n_faces, size_type& n_edges,
+    size_type& n_constraints, int& n_owning_interfaces) GKO_NOT_IMPLEMENTED;
 
 GKO_INSTANTIATE_FOR_EACH_NON_COMPLEX_VALUE_AND_INDEX_TYPE_BASE(
     GKO_DECLARE_CLASSIFY_DOFS);
@@ -53,16 +53,17 @@ void fill_coarse_data(std::shared_ptr<const DefaultExecutor> exec,
 GKO_INSTANTIATE_FOR_EACH_VALUE_TYP_BASE_BASE(GKO_DECLARE_FILL_COARSE_DATA);
 
 
-template <typename ValueType>
+template <typename ValueType, typename IndexType>
 void build_coarse_contribution(
     std::shared_ptr<const DefaultExecutor> exec,
-    const matrix::Dense<remove_complex<ValueType>>* local_labels,
+    const array<remove_complex<ValueType>>& local_labels,
     const array<remove_complex<ValueType>>& global_labels,
     const matrix::Dense<ValueType>* lambda,
-    device_matrix_data<ValueType, int>& coarse_contribution)
-    GKO_NOT_IMPLEMENTED;
+    device_matrix_data<ValueType, int>& coarse_contribution,
+    array<IndexType>& permutation_array) GKO_NOT_IMPLEMENTED;
 
-GKO_INSTANTIATE_FOR_EACH_VALUE_TYPE_BASE(GKO_DECLARE_BUILD_COARSE_CONTRIBUTION);
+GKO_INSTANTIATE_FOR_EACH_VALUE_AND_INDEX_TYPE_BASE(
+    GKO_DECLARE_BUILD_COARSE_CONTRIBUTION);
 
 
 }  // namespace bddc

diff --git a/include/ginkgo/core/distributed/preconditioner/bddc.hpp b/include/ginkgo/core/distributed/preconditioner/bddc.hpp
@@ -1,4 +1,4 @@
-// SPDX-FileCopyrightText: 2017 - 2024 The Ginkgo authors
+// SPDX-FileCopyrightText: 2017 - 2025 The Ginkgo authors
 //
 // SPDX-License-Identifier: BSD-3-Clause
 
@@ -102,6 +102,11 @@ class Bddc
          * Use of Face constraints.
          */
         bool GKO_FACTORY_PARAMETER_SCALAR(faces, true);
+
+        /**
+         * Use of AMD reordering to reduce fill-in for local direct solvers.
+         */
+        bool GKO_FACTORY_PARAMETER_SCALAR(amd, true);
     };
     GKO_ENABLE_LIN_OP_FACTORY(Bddc, parameters, Factory);
     GKO_ENABLE_BUILD_METHOD(Factory);
@@ -192,17 +197,35 @@ class Bddc
     std::shared_ptr<vec> coarse_buf_1_;
     std::shared_ptr<vec> coarse_buf_2_;
     std::shared_ptr<local_vec> local_buf_1_;
+    std::shared_ptr<local_vec> interior_1_;
+    std::shared_ptr<local_vec> bndry_1_;
+    std::shared_ptr<local_vec> dual_1_;
+    std::shared_ptr<local_vec> primal_1_;
     std::shared_ptr<local_vec> local_buf_2_;
+    std::shared_ptr<local_vec> interior_2_;
+    std::shared_ptr<local_vec> bndry_2_;
+    std::shared_ptr<local_vec> dual_2_;
+    std::shared_ptr<local_vec> primal_2_;
+    std::shared_ptr<local_vec> local_buf_3_;
+    std::shared_ptr<local_vec> interior_3_;
+    std::shared_ptr<local_vec> bndry_3_;
+    std::shared_ptr<local_vec> dual_3_;
+    std::shared_ptr<local_vec> primal_3_;
+    std::shared_ptr<local_vec> schur_buf_1_;
+    std::shared_ptr<local_vec> schur_buf_2_;
     std::shared_ptr<vec> broken_coarse_buf_1_;
     std::shared_ptr<vec> broken_coarse_buf_2_;
     std::shared_ptr<local_vec> local_coarse_buf_1_;
     std::shared_ptr<local_vec> local_coarse_buf_2_;
-    std::shared_ptr<const LinOp> A_LL;
-    std::shared_ptr<const LinOp> A_LP;
-    std::shared_ptr<const LinOp> A_PL;
-    std::shared_ptr<const LinOp> A_PP;
-    std::shared_ptr<const LinOp> A_IB;
-    std::shared_ptr<const LinOp> A_BI;
+    std::shared_ptr<local_mtx> A_LL;
+    std::shared_ptr<const local_mtx> A_LP;
+    std::shared_ptr<const local_mtx> A_PL;
+    std::shared_ptr<const local_mtx> A_PP;
+    std::shared_ptr<const local_mtx> A_IB;
+    std::shared_ptr<const local_mtx> A_BI;
+    std::shared_ptr<const LinOp> one_;
+    std::shared_ptr<const LinOp> neg_one_;
+    std::shared_ptr<const LinOp> zero_;
 };
 
 

diff --git a/omp/distributed/preconditioner/bddc_kernels.cpp b/omp/distributed/preconditioner/bddc_kernels.cpp
@@ -25,10 +25,10 @@ void classify_dofs(
     const matrix::Dense<ValueType>* labels, comm_index_type local_part,
     array<experimental::distributed::preconditioner::dof_type>& dof_types,
     array<IndexType>& permutation_array, array<IndexType>& interface_sizes,
-    array<ValueType>& owning_labels, size_type& n_inner_idxs,
-    size_type& n_face_idxs, size_type& n_edge_idxs, size_type& n_vertices,
-    size_type& n_faces, size_type& n_edges, size_type& n_constraints,
-    int& n_owning_interfaces) GKO_NOT_IMPLEMENTED;
+    array<ValueType>& unique_labels, array<ValueType>& owning_labels,
+    size_type& n_inner_idxs, size_type& n_face_idxs, size_type& n_edge_idxs,
+    size_type& n_vertices, size_type& n_faces, size_type& n_edges,
+    size_type& n_constraints, int& n_owning_interfaces) GKO_NOT_IMPLEMENTED;
 
 GKO_INSTANTIATE_FOR_EACH_NON_COMPLEX_VALUE_AND_INDEX_TYPE_BASE(
     GKO_DECLARE_CLASSIFY_DOFS);
@@ -53,16 +53,17 @@ void fill_coarse_data(std::shared_ptr<const DefaultExecutor> exec,
 GKO_INSTANTIATE_FOR_EACH_VALUE_TYPE_BASE(GKO_DECLARE_FILL_COARSE_DATA);
 
 
-template <typename ValueType>
+template <typename ValueType, typename IndexType>
 void build_coarse_contribution(
     std::shared_ptr<const DefaultExecutor> exec,
-    const matrix::Dense<remove_complex<ValueType>>* local_labels,
+    const array<remove_complex<ValueType>>& local_labels,
     const array<remove_complex<ValueType>>& global_labels,
     const matrix::Dense<ValueType>* lambda,
-    device_matrix_data<ValueType, int>& coarse_contribution)
-    GKO_NOT_IMPLEMENTED;
+    device_matrix_data<ValueType, int>& coarse_contribution,
+    array<IndexType>& permutation_array) GKO_NOT_IMPLEMENTED;
 
-GKO_INSTANTIATE_FOR_EACH_VALUE_TYPE_BASE(GKO_DECLARE_BUILD_COARSE_CONTRIBUTION);
+GKO_INSTANTIATE_FOR_EACH_VALUE_AND_INDEX_TYPE_BASE(
+    GKO_DECLARE_BUILD_COARSE_CONTRIBUTION);
 
 
 }  // namespace bddc

diff --git a/reference/distributed/index_map_kernels.cpp b/reference/distributed/index_map_kernels.cpp
@@ -1,4 +1,4 @@
-// SPDX-FileCopyrightText: 2017 - 2024 The Ginkgo authors
+// SPDX-FileCopyrightText: 2017 - 2025 The Ginkgo authors
 //
 // SPDX-License-Identifier: BSD-3-Clause
 
@@ -220,7 +220,7 @@ void map_to_global(
     auto map_local = [&](auto lid) {
         if (0 <= lid && lid < local_size) {
             local_range_id =
-                find_local_range(lid, rank, partition, local_range_id);
+                find_local_range(lid, rank, partition, 0);  // local_range_id);
             return map_to_global(lid, partition,
                                  local_ranges.begin[local_range_id]);
         } else {
@@ -262,6 +262,7 @@ void map_to_global(
         for (size_type i = 0; i < local_ids.get_size(); ++i) {
             auto lid = local_ids.get_const_data()[i];
             global_ids.get_data()[i] = map_combined(lid);
+            // std::cout << lid << " ==> " << map_combined(lid) << std::endl;
         }
     }
 }