Ginkgo  Generated from pipelines/2662685947 branch based on develop. Ginkgo version 2.0.0
A numerical linear algebra library targeting many-core architectures
matrix.hpp
1 // SPDX-FileCopyrightText: 2017 - 2026 The Ginkgo authors
2 //
3 // SPDX-License-Identifier: BSD-3-Clause
4 
5 #ifndef GKO_PUBLIC_CORE_DISTRIBUTED_MATRIX_HPP_
6 #define GKO_PUBLIC_CORE_DISTRIBUTED_MATRIX_HPP_
7 
8 
9 #include <ginkgo/config.hpp>
10 
11 
12 #if GINKGO_BUILD_MPI
13 
14 
15 #include <ginkgo/core/base/dense_cache.hpp>
16 #include <ginkgo/core/base/lin_op.hpp>
17 #include <ginkgo/core/base/mpi.hpp>
18 #include <ginkgo/core/base/std_extensions.hpp>
19 #include <ginkgo/core/distributed/base.hpp>
20 #include <ginkgo/core/distributed/index_map.hpp>
21 #include <ginkgo/core/distributed/row_gatherer.hpp>
22 #include <ginkgo/core/distributed/vector_cache.hpp>
23 
24 
25 namespace gko {
26 namespace matrix {
27 
28 
29 template <typename ValueType, typename IndexType>
30 class Csr;
31 
32 
33 }
34 
35 
36 namespace multigrid {
37 
38 
39 template <typename ValueType, typename IndexType>
40 class Pgm;
41 
42 
43 }
44 
45 
46 namespace detail {
47 
48 
53 template <typename Builder, typename ValueType, typename IndexType,
54  typename = void>
55 struct is_matrix_type_builder : std::false_type {};
56 
57 
58 template <typename Builder, typename ValueType, typename IndexType>
59 struct is_matrix_type_builder<
60  Builder, ValueType, IndexType,
61  xstd::void_t<
62  decltype(std::declval<Builder>().template create<ValueType, IndexType>(
63  std::declval<std::shared_ptr<const Executor>>()))>>
64  : std::true_type {};
65 
66 
67 template <template <typename, typename> class MatrixType,
68  typename... CreateArgs>
69 struct MatrixTypeBuilderFromValueAndIndex {
70  template <typename ValueType, typename IndexType, std::size_t... I>
71  auto create_impl(std::shared_ptr<const Executor> exec,
72  std::index_sequence<I...>)
73  {
74  return MatrixType<ValueType, IndexType>::create(
75  exec, std::get<I>(create_args)...);
76  }
77 
78 
79  template <typename ValueType, typename IndexType>
80  auto create(std::shared_ptr<const Executor> exec)
81  {
82  // with c++17 we could use std::apply
83  static constexpr auto size = sizeof...(CreateArgs);
84  return create_impl<ValueType, IndexType>(
85  std::move(exec), std::make_index_sequence<size>{});
86  }
87 
88  std::tuple<CreateArgs...> create_args;
89 };
90 
91 
92 } // namespace detail
93 
94 
126 template <template <typename, typename> class MatrixType, typename... Args>
127 auto with_matrix_type(Args&&... create_args)
128 {
129  return detail::MatrixTypeBuilderFromValueAndIndex<MatrixType, Args...>{
130  std::forward_as_tuple(create_args...)};
131 }
132 
133 
134 namespace experimental {
135 namespace distributed {
136 
137 
147 enum class assembly_mode { communicate, local_only };
148 
149 
150 template <typename LocalIndexType, typename GlobalIndexType>
151 class Partition;
152 template <typename ValueType>
153 class Vector;
154 
155 
260 template <typename ValueType = default_precision,
261  typename LocalIndexType = int32, typename GlobalIndexType = int64>
262 class Matrix
263  : public LinOp,
264  public EnableCloneable<
265  Matrix<ValueType, LocalIndexType, GlobalIndexType>>,
266  public ConvertibleTo<
267  Matrix<next_precision<ValueType>, LocalIndexType, GlobalIndexType>>,
268 #if GINKGO_ENABLE_HALF || GINKGO_ENABLE_BFLOAT16
269  public ConvertibleTo<Matrix<next_precision<ValueType, 2>, LocalIndexType,
270  GlobalIndexType>>,
271 #endif
272 #if GINKGO_ENABLE_HALF && GINKGO_ENABLE_BFLOAT16
273  public ConvertibleTo<Matrix<next_precision<ValueType, 3>, LocalIndexType,
274  GlobalIndexType>>,
275 #endif
276  public DistributedBase {
277  friend class EnableCloneable<Matrix>;
278  friend class Matrix<previous_precision<ValueType>, LocalIndexType,
279  GlobalIndexType>;
280 
281  friend class multigrid::Pgm<ValueType, LocalIndexType>;
282  GKO_ASSERT_SUPPORTED_VALUE_AND_DIST_INDEX_TYPE;
283 
284 public:
285  using value_type = ValueType;
286  using index_type = GlobalIndexType;
287  using local_index_type = LocalIndexType;
288  using global_index_type = GlobalIndexType;
289  using global_vector_type =
291  using local_vector_type = typename global_vector_type::local_vector_type;
292 
296  GlobalIndexType>>::convert_to;
298  GlobalIndexType>>::move_to;
299 
300  void convert_to(Matrix<next_precision<value_type>, local_index_type,
301  global_index_type>* result) const override;
302 
303  void move_to(Matrix<next_precision<value_type>, local_index_type,
304  global_index_type>* result) override;
305 
306 #if GINKGO_ENABLE_HALF || GINKGO_ENABLE_BFLOAT16
307  friend class Matrix<previous_precision<ValueType, 2>, LocalIndexType,
308  GlobalIndexType>;
310  global_index_type>>::convert_to;
312  global_index_type>>::move_to;
313 
314  void convert_to(Matrix<next_precision<value_type, 2>, local_index_type,
315  global_index_type>* result) const override;
316 
317  void move_to(Matrix<next_precision<value_type, 2>, local_index_type,
318  global_index_type>* result) override;
319 #endif
320 
321 #if GINKGO_ENABLE_HALF && GINKGO_ENABLE_BFLOAT16
322  friend class Matrix<previous_precision<ValueType, 3>, LocalIndexType,
323  GlobalIndexType>;
325  global_index_type>>::convert_to;
327  global_index_type>>::move_to;
328 
329  void convert_to(Matrix<next_precision<value_type, 3>, local_index_type,
330  global_index_type>* result) const override;
331 
332  void move_to(Matrix<next_precision<value_type, 3>, local_index_type,
333  global_index_type>* result) override;
334 #endif
335 
352  void read_distributed(
355  partition,
356  assembly_mode assembly_type = assembly_mode::local_only);
357 
367  void read_distributed(
370  partition,
371  assembly_mode assembly_type = assembly_mode::local_only);
372 
391  void read_distributed(
394  row_partition,
396  col_partition,
397  assembly_mode assembly_type = assembly_mode::local_only);
398 
408  void read_distributed(
411  row_partition,
413  col_partition,
414  assembly_mode assembly_type = assembly_mode::local_only);
415 
424  std::shared_ptr<const LinOp> get_diag_matrix() const { return diag_mtx_; }
425 
434  std::shared_ptr<const LinOp> get_off_diag_matrix() const
435  {
436  return off_diag_mtx_;
437  }
438 
442  GKO_DEPRECATED("use get_diag_matrix() instead")
443  std::shared_ptr<const LinOp> get_local_matrix() const
444  {
445  return get_diag_matrix();
446  }
447 
451  GKO_DEPRECATED("use get_off_diag_matrix() instead")
452  std::shared_ptr<const LinOp> get_non_local_matrix() const
453  {
454  return get_off_diag_matrix();
455  }
456 
462  Matrix(const Matrix& other);
463 
469  Matrix(Matrix&& other) noexcept;
470 
479  Matrix& operator=(const Matrix& other);
480 
489  Matrix& operator=(Matrix&& other);
490 
500  static std::unique_ptr<Matrix> create(std::shared_ptr<const Executor> exec,
501  mpi::communicator comm);
502 
516  static std::unique_ptr<Matrix> create(
517  std::shared_ptr<const Executor> exec,
518  std::shared_ptr<const RowGatherer<LocalIndexType>>
519  row_gatherer_template);
520 
541  template <typename MatrixType,
542  typename = std::enable_if_t<gko::detail::is_matrix_type_builder<
543  MatrixType, ValueType, LocalIndexType>::value>>
544  static std::unique_ptr<Matrix> create(std::shared_ptr<const Executor> exec,
545  mpi::communicator comm,
546  MatrixType matrix_template)
547  {
548  return create(
549  exec, comm,
550  matrix_template.template create<ValueType, LocalIndexType>(exec));
551  }
552 
581  template <typename DiagMatrixType, typename OffDiagMatrixType,
582  typename = std::enable_if_t<
583  gko::detail::is_matrix_type_builder<DiagMatrixType, ValueType,
584  LocalIndexType>::value &&
585  gko::detail::is_matrix_type_builder<
586  OffDiagMatrixType, ValueType, LocalIndexType>::value>>
587  static std::unique_ptr<Matrix> create(
588  std::shared_ptr<const Executor> exec, mpi::communicator comm,
589  DiagMatrixType diag_matrix_template,
590  OffDiagMatrixType off_diag_matrix_template)
591  {
592  return create(
593  exec, comm,
594  diag_matrix_template.template create<ValueType, LocalIndexType>(
595  exec),
596  off_diag_matrix_template.template create<ValueType, LocalIndexType>(
597  exec));
598  }
599 
614  static std::unique_ptr<Matrix> create(
615  std::shared_ptr<const Executor> exec, mpi::communicator comm,
616  ptr_param<const LinOp> matrix_template);
617 
634  static std::unique_ptr<Matrix> create(
635  std::shared_ptr<const Executor> exec, mpi::communicator comm,
636  ptr_param<const LinOp> diag_matrix_template,
637  ptr_param<const LinOp> off_diag_matrix_template);
638 
651  static std::unique_ptr<Matrix> create(std::shared_ptr<const Executor> exec,
652  mpi::communicator comm, dim<2> size,
653  std::shared_ptr<LinOp> diag_linop);
654 
673  [[deprecated(
674  "Please use the overload with an index_map instead.")]] static std::
675  unique_ptr<Matrix>
676  create(std::shared_ptr<const Executor> exec, mpi::communicator comm,
677  dim<2> size, std::shared_ptr<LinOp> diag_linop,
678  std::shared_ptr<LinOp> off_diag_linop,
679  std::vector<comm_index_type> recv_sizes,
680  std::vector<comm_index_type> recv_offsets,
681  array<local_index_type> recv_gather_idxs);
682 
696  static std::unique_ptr<Matrix> create(
697  std::shared_ptr<const Executor> exec, mpi::communicator comm,
699  std::shared_ptr<LinOp> diag_linop,
700  std::shared_ptr<LinOp> off_diag_linop);
701 
709  void col_scale(ptr_param<const global_vector_type> scaling_factors);
710 
718  void row_scale(ptr_param<const global_vector_type> scaling_factors);
719 
720 protected:
721  explicit Matrix(std::shared_ptr<const Executor> exec,
722  mpi::communicator comm);
723 
724  explicit Matrix(std::shared_ptr<const Executor> exec,
725  std::shared_ptr<const RowGatherer<LocalIndexType>>
726  row_gatherer_template,
727  ptr_param<const LinOp> diag_matrix_template,
728  ptr_param<const LinOp> off_diag_matrix_template);
729 
730  explicit Matrix(std::shared_ptr<const Executor> exec,
731  mpi::communicator comm, dim<2> size,
732  std::shared_ptr<LinOp> diag_linop);
733 
734  explicit Matrix(std::shared_ptr<const Executor> exec,
735  mpi::communicator comm,
737  std::shared_ptr<LinOp> diag_linop,
738  std::shared_ptr<LinOp> off_diag_linop);
739 
740  void apply_impl(const LinOp* b, LinOp* x) const override;
741 
742  void apply_impl(const LinOp* alpha, const LinOp* b, const LinOp* beta,
743  LinOp* x) const override;
744 
745 private:
746  std::shared_ptr<RowGatherer<LocalIndexType>> row_gatherer_;
748  gko::detail::ScalarCache one_scalar_;
749  detail::GenericVectorCache recv_buffer_;
750  detail::GenericVectorCache host_recv_buffer_;
751  std::shared_ptr<LinOp> diag_mtx_;
752  std::shared_ptr<LinOp> off_diag_mtx_;
753 };
754 
755 
756 } // namespace distributed
757 } // namespace experimental
758 } // namespace gko
759 
760 
761 #endif
762 
763 
764 #endif // GKO_PUBLIC_CORE_DISTRIBUTED_MATRIX_HPP_
gko::with_matrix_type
auto with_matrix_type(Args &&... create_args)
This function returns a type that delays a call to MatrixType::create.
Definition: matrix.hpp:127
gko::matrix::Csr
CSR is a matrix format which stores only the nonzero coefficients by compressing each row of the matr...
Definition: matrix.hpp:30
gko::experimental::distributed::assembly_mode
assembly_mode
assembly_mode defines how the read_distributed function of the distributed matrix treats non-local in...
Definition: matrix.hpp:147
gko::LinOp
Definition: lin_op.hpp:117
gko::matrix::Dense< value_type >
gko::experimental::distributed::RowGatherer
The distributed::RowGatherer gathers the rows of distributed::Vector that are located on other proces...
Definition: row_gatherer.hpp:30
gko::experimental::distributed::Matrix
The Matrix class defines a (MPI-)distributed matrix.
Definition: matrix.hpp:262
gko::experimental::distributed::Vector
Vector is a format which explicitly stores (multiple) distributed column vectors in a dense storage f...
Definition: matrix.hpp:153
gko::experimental::distributed::Matrix::get_local_matrix
std::shared_ptr< const LinOp > get_local_matrix() const
Definition: matrix.hpp:443
gko::experimental::distributed::Matrix::col_scale
void col_scale(ptr_param< const global_vector_type > scaling_factors)
Scales the columns of the matrix by the respective entries of the vector.
gko::experimental::distributed::Matrix::create
static std::unique_ptr< Matrix > create(std::shared_ptr< const Executor > exec, mpi::communicator comm)
Creates an empty distributed matrix.
gko
The Ginkgo namespace.
Definition: abstract_factory.hpp:20
gko::experimental::distributed::Matrix::get_non_local_matrix
std::shared_ptr< const LinOp > get_non_local_matrix() const
Definition: matrix.hpp:452
gko::array< local_index_type >
gko::multigrid::Pgm
Parallel graph match (Pgm) is the aggregate method introduced in the paper M.
Definition: matrix.hpp:40
gko::experimental::mpi::communicator
A thin wrapper of MPI_Comm that supports most MPI calls.
Definition: mpi.hpp:419
gko::dim< 2 >
gko::matrix_data
This structure is used as an intermediate data type to store a sparse matrix.
Definition: matrix_data.hpp:126
gko::experimental::distributed::Matrix::get_diag_matrix
std::shared_ptr< const LinOp > get_diag_matrix() const
Get read access to the stored diagonal matrix block.
Definition: matrix.hpp:424
gko::experimental::distributed::Matrix::create
static std::unique_ptr< Matrix > create(std::shared_ptr< const Executor > exec, mpi::communicator comm, DiagMatrixType diag_matrix_template, OffDiagMatrixType off_diag_matrix_template)
Creates an empty distributed matrix with specified types for the diagonal matrix and the off-diagonal...
Definition: matrix.hpp:587
gko::ptr_param
This class is used for function parameters in the place of raw pointers.
Definition: utils_helper.hpp:43
gko::next_precision
typename detail::find_precision_impl< T, step >::type next_precision
Obtains the next move type of T in the singly-linked precision corresponding bfloat16/half.
Definition: math.hpp:466
gko::experimental::distributed::Partition
Represents a partition of a range of indices [0, size) into a disjoint set of parts.
Definition: assembly.hpp:26
gko::previous_precision
typename detail::find_precision_impl< T, -step >::type previous_precision
Obtains the previous move type of T in the singly-linked precision corresponding bfloat16/half.
Definition: math.hpp:473
gko::experimental::distributed::Matrix::operator=
Matrix & operator=(const Matrix &other)
Copy assigns a Matrix.
gko::EnableCloneable
This mixin is used to enable a default Cloneable::clone() implementation and similar for objects that...
Definition: polymorphic_object.hpp:369
gko::int64
std::int64_t int64
64-bit signed integral type.
Definition: types.hpp:113
gko::default_precision
double default_precision
Precision used if no precision is explicitly specified.
Definition: types.hpp:172
gko::ConvertibleTo
ConvertibleTo interface is used to mark that the implementer can be converted to the object of Result...
Definition: polymorphic_object.hpp:140
gko::int32
std::int32_t int32
32-bit signed integral type.
Definition: types.hpp:107
gko::experimental::distributed::Matrix::create
static std::unique_ptr< Matrix > create(std::shared_ptr< const Executor > exec, mpi::communicator comm, MatrixType matrix_template)
Creates an empty distributed matrix with specified type for local matrices.
Definition: matrix.hpp:544
gko::experimental::distributed::DistributedBase
A base class for distributed objects.
Definition: base.hpp:32
gko::experimental::distributed::index_map< local_index_type, global_index_type >
gko::experimental::distributed::Matrix::read_distributed
void read_distributed(const device_matrix_data< value_type, global_index_type > &data, std::shared_ptr< const Partition< local_index_type, global_index_type >> partition, assembly_mode assembly_type=assembly_mode::local_only)
Reads a square matrix from the device_matrix_data structure and a global partition.
gko::experimental::distributed::Matrix::row_scale
void row_scale(ptr_param< const global_vector_type > scaling_factors)
Scales the rows of the matrix by the respective entries of the vector.
gko::device_matrix_data
This type is a device-side equivalent to matrix_data.
Definition: device_matrix_data.hpp:36
gko::experimental::distributed::Matrix::Matrix
Matrix(const Matrix &other)
Copy constructs a Matrix.
gko::experimental::distributed::Matrix::get_off_diag_matrix
std::shared_ptr< const LinOp > get_off_diag_matrix() const
Get read access to the stored off-diagonal matrix block.
Definition: matrix.hpp:434