Ginkgo Generated from branch based on main. Ginkgo version 1.10.0
A numerical linear algebra library targeting many-core architectures
Loading...
Searching...
No Matches
fbcsr.hpp
1// SPDX-FileCopyrightText: 2017 - 2025 The Ginkgo authors
2//
3// SPDX-License-Identifier: BSD-3-Clause
4
5#ifndef GKO_PUBLIC_CORE_MATRIX_FBCSR_HPP_
6#define GKO_PUBLIC_CORE_MATRIX_FBCSR_HPP_
7
8
9#include <ginkgo/core/base/array.hpp>
10#include <ginkgo/core/base/lin_op.hpp>
11#include <ginkgo/core/base/math.hpp>
12
13
14namespace gko {
15namespace matrix {
16
17
18template <typename ValueType>
19class Dense;
20
21template <typename ValueType, typename IndexType>
22class Csr;
23
24template <typename ValueType, typename IndexType>
25class SparsityCsr;
26
27template <typename ValueType, typename IndexType>
28class Fbcsr;
29
30template <typename ValueType, typename IndexType>
32
33
34namespace detail {
35
36
47template <typename IndexType>
48inline IndexType get_num_blocks(const int block_size, const IndexType size)
49{
50 GKO_ASSERT_BLOCK_SIZE_CONFORMANT(size, block_size);
51 return size / block_size;
52}
53
54
55} // namespace detail
56
57
98template <typename ValueType = default_precision, typename IndexType = int32>
99class Fbcsr
100 : public EnableLinOp<Fbcsr<ValueType, IndexType>>,
101 public ConvertibleTo<Fbcsr<next_precision<ValueType>, IndexType>>,
102#if GINKGO_ENABLE_HALF || GINKGO_ENABLE_BFLOAT16
103 public ConvertibleTo<Fbcsr<next_precision<ValueType, 2>, IndexType>>,
104#endif
105#if GINKGO_ENABLE_HALF && GINKGO_ENABLE_BFLOAT16
106 public ConvertibleTo<Fbcsr<next_precision<ValueType, 3>, IndexType>>,
107#endif
108 public ConvertibleTo<Dense<ValueType>>,
109 public ConvertibleTo<Csr<ValueType, IndexType>>,
110 public ConvertibleTo<SparsityCsr<ValueType, IndexType>>,
111 public DiagonalExtractable<ValueType>,
112 public ReadableFromMatrixData<ValueType, IndexType>,
113 public WritableToMatrixData<ValueType, IndexType>,
114 public Transposable,
116 remove_complex<Fbcsr<ValueType, IndexType>>> {
117 friend class EnablePolymorphicObject<Fbcsr, LinOp>;
118 friend class Csr<ValueType, IndexType>;
119 friend class Dense<ValueType>;
120 friend class SparsityCsr<ValueType, IndexType>;
121 friend class FbcsrBuilder<ValueType, IndexType>;
122 friend class Fbcsr<to_complex<ValueType>, IndexType>;
123
124public:
125 using value_type = ValueType;
126 using index_type = IndexType;
127 using transposed_type = Fbcsr<ValueType, IndexType>;
128 using mat_data = matrix_data<ValueType, IndexType>;
129 using device_mat_data = device_matrix_data<ValueType, IndexType>;
130 using absolute_type = remove_complex<Fbcsr>;
131
138
143 using EnableLinOp<Fbcsr<ValueType, IndexType>>::convert_to;
144
145 using ConvertibleTo<
146 Fbcsr<next_precision<ValueType>, IndexType>>::convert_to;
147 using ConvertibleTo<Fbcsr<next_precision<ValueType>, IndexType>>::move_to;
148 using ConvertibleTo<Dense<ValueType>>::convert_to;
149 using ConvertibleTo<Dense<ValueType>>::move_to;
150 using ConvertibleTo<Csr<ValueType, IndexType>>::convert_to;
154
155 friend class Fbcsr<previous_precision<ValueType>, IndexType>;
156
157 void convert_to(
158 Fbcsr<next_precision<ValueType>, IndexType>* result) const override;
159
160 void move_to(Fbcsr<next_precision<ValueType>, IndexType>* result) override;
161
162#if GINKGO_ENABLE_HALF || GINKGO_ENABLE_BFLOAT16
163 friend class Fbcsr<previous_precision<ValueType, 2>, IndexType>;
164 using ConvertibleTo<
165 Fbcsr<next_precision<ValueType, 2>, IndexType>>::convert_to;
166 using ConvertibleTo<
167 Fbcsr<next_precision<ValueType, 2>, IndexType>>::move_to;
168
169 void convert_to(
170 Fbcsr<next_precision<ValueType, 2>, IndexType>* result) const override;
171
172 void move_to(
173 Fbcsr<next_precision<ValueType, 2>, IndexType>* result) override;
174#endif
175
176#if GINKGO_ENABLE_HALF && GINKGO_ENABLE_BFLOAT16
177 friend class Fbcsr<previous_precision<ValueType, 3>, IndexType>;
178 using ConvertibleTo<
179 Fbcsr<next_precision<ValueType, 3>, IndexType>>::convert_to;
180 using ConvertibleTo<
181 Fbcsr<next_precision<ValueType, 3>, IndexType>>::move_to;
182
183 void convert_to(
184 Fbcsr<next_precision<ValueType, 3>, IndexType>* result) const override;
185
186 void move_to(
187 Fbcsr<next_precision<ValueType, 3>, IndexType>* result) override;
188#endif
189
190 void convert_to(Dense<ValueType>* other) const override;
191
192 void move_to(Dense<ValueType>* other) override;
193
200 void convert_to(Csr<ValueType, IndexType>* result) const override;
201
202 void move_to(Csr<ValueType, IndexType>* result) override;
203
210 void convert_to(SparsityCsr<ValueType, IndexType>* result) const override;
211
212 void move_to(SparsityCsr<ValueType, IndexType>* result) override;
213
220 void read(const mat_data& data) override;
221
222 void read(const device_mat_data& data) override;
223
224 void read(device_mat_data&& data) override;
225
226 void write(mat_data& data) const override;
227
228 std::unique_ptr<LinOp> transpose() const override;
229
230 std::unique_ptr<LinOp> conj_transpose() const override;
231
232 std::unique_ptr<Diagonal<ValueType>> extract_diagonal() const override;
233
234 std::unique_ptr<absolute_type> compute_absolute() const override;
235
237
243
251
255 value_type* get_values() noexcept { return values_.get_data(); }
256
264 const value_type* get_const_values() const noexcept
265 {
266 return values_.get_const_data();
267 }
268
272 index_type* get_col_idxs() noexcept { return col_idxs_.get_data(); }
273
281 const index_type* get_const_col_idxs() const noexcept
282 {
283 return col_idxs_.get_const_data();
284 }
285
289 index_type* get_row_ptrs() noexcept { return row_ptrs_.get_data(); }
290
298 const index_type* get_const_row_ptrs() const noexcept
299 {
300 return row_ptrs_.get_const_data();
301 }
302
307 {
308 return values_.get_size();
309 }
310
315 {
316 return col_idxs_.get_size();
317 }
318
322 int get_block_size() const noexcept { return bs_; }
323
327 index_type get_num_block_rows() const noexcept
328 {
329 return this->get_size()[0] / bs_;
330 }
331
335 index_type get_num_block_cols() const noexcept
336 {
337 return this->get_size()[1] / bs_;
338 }
339
349 static std::unique_ptr<Fbcsr> create(std::shared_ptr<const Executor> exec,
350 int block_size = 1);
351
363 static std::unique_ptr<Fbcsr> create(std::shared_ptr<const Executor> exec,
364 const dim<2>& size,
365 size_type num_nonzeros,
366 int block_size);
367
388 static std::unique_ptr<Fbcsr> create(std::shared_ptr<const Executor> exec,
389 const dim<2>& size, int block_size,
390 array<value_type> values,
391 array<index_type> col_idxs,
392 array<index_type> row_ptrs);
393
399 template <typename InputValueType, typename InputColumnIndexType,
400 typename InputRowPtrType>
401 GKO_DEPRECATED(
402 "explicitly construct the gko::array argument instead of passing "
403 "initializer lists")
404 static std::unique_ptr<Fbcsr> create(
405 std::shared_ptr<const Executor> exec, const dim<2>& size,
406 int block_size, std::initializer_list<InputValueType> values,
407 std::initializer_list<InputColumnIndexType> col_idxs,
408 std::initializer_list<InputRowPtrType> row_ptrs)
409 {
410 return create(exec, size, block_size,
411 array<value_type>{exec, std::move(values)},
412 array<index_type>{exec, std::move(col_idxs)},
413 array<index_type>{exec, std::move(row_ptrs)});
414 }
415
430 static std::unique_ptr<const Fbcsr> create_const(
431 std::shared_ptr<const Executor> exec, const dim<2>& size, int blocksize,
432 gko::detail::const_array_view<ValueType>&& values,
433 gko::detail::const_array_view<IndexType>&& col_idxs,
434 gko::detail::const_array_view<IndexType>&& row_ptrs);
435
441
448
452 Fbcsr(const Fbcsr&);
453
460
461protected:
462 Fbcsr(std::shared_ptr<const Executor> exec, int block_size = 1);
463
464 Fbcsr(std::shared_ptr<const Executor> exec, const dim<2>& size,
465 size_type num_nonzeros, int block_size);
466
467 Fbcsr(std::shared_ptr<const Executor> exec, const dim<2>& size,
468 int block_size, array<value_type> values, array<index_type> col_idxs,
469 array<index_type> row_ptrs);
470
471 void apply_impl(const LinOp* b, LinOp* x) const override;
472
473 void apply_impl(const LinOp* alpha, const LinOp* b, const LinOp* beta,
474 LinOp* x) const override;
475
476private:
477 int bs_;
478 array<value_type> values_;
479 array<index_type> col_idxs_;
480 array<index_type> row_ptrs_;
481};
482
483
484} // namespace matrix
485} // namespace gko
486
487
488#endif // GKO_PUBLIC_CORE_MATRIX_FBCSR_HPP_
ConvertibleTo interface is used to mark that the implementer can be converted to the object of Result...
Definition polymorphic_object.hpp:479
The diagonal of a LinOp implementing this interface can be extracted.
Definition lin_op.hpp:743
The EnableAbsoluteComputation mixin provides the default implementations of compute_absolute_linop an...
Definition lin_op.hpp:794
The EnableLinOp mixin can be used to provide sensible default implementations of the majority of the ...
Definition lin_op.hpp:879
This mixin inherits from (a subclass of) PolymorphicObject and provides a base implementation of a ne...
Definition polymorphic_object.hpp:668
The first step in using the Ginkgo library consists of creating an executor.
Definition executor.hpp:615
Definition lin_op.hpp:117
const dim< 2 > & get_size() const noexcept
Returns the size of the operator.
Definition lin_op.hpp:210
A LinOp implementing this interface can read its data from a matrix_data structure.
Definition lin_op.hpp:605
Linear operators which support transposition should implement the Transposable interface.
Definition lin_op.hpp:433
A LinOp implementing this interface can write its data to a matrix_data structure.
Definition lin_op.hpp:660
An array is a container which encapsulates fixed-sized arrays, stored on the Executor tied to the arr...
Definition array.hpp:166
This type is a device-side equivalent to matrix_data.
Definition device_matrix_data.hpp:36
CSR is a matrix format which stores only the nonzero coefficients by compressing each row of the matr...
Definition csr.hpp:126
Dense is a matrix format which explicitly stores all values of the matrix.
Definition dense.hpp:120
Definition fbcsr.hpp:31
Fixed-block compressed sparse row storage matrix format.
Definition fbcsr.hpp:116
size_type get_num_stored_blocks() const noexcept
Definition fbcsr.hpp:314
std::unique_ptr< LinOp > transpose() const override
Returns a LinOp representing the transpose of the Transposable object.
size_type get_num_stored_elements() const noexcept
Definition fbcsr.hpp:306
const value_type * get_const_values() const noexcept
Definition fbcsr.hpp:264
const index_type * get_const_col_idxs() const noexcept
Definition fbcsr.hpp:281
index_type get_num_block_rows() const noexcept
Definition fbcsr.hpp:327
index_type * get_row_ptrs() noexcept
Definition fbcsr.hpp:289
int get_block_size() const noexcept
Definition fbcsr.hpp:322
index_type * get_col_idxs() noexcept
Definition fbcsr.hpp:272
static std::unique_ptr< Fbcsr > create(std::shared_ptr< const Executor > exec, int block_size=1)
Creates an uninitialized FBCSR matrix with the given block size.
Fbcsr(const Fbcsr &)
Copy-constructs an Ell matrix.
static std::unique_ptr< const Fbcsr > create_const(std::shared_ptr< const Executor > exec, const dim< 2 > &size, int blocksize, gko::detail::const_array_view< ValueType > &&values, gko::detail::const_array_view< IndexType > &&col_idxs, gko::detail::const_array_view< IndexType > &&row_ptrs)
Creates a constant (immutable) Fbcsr matrix from a constant array.
void compute_absolute_inplace() override
Compute absolute inplace on each element.
static std::unique_ptr< Fbcsr > create(std::shared_ptr< const Executor > exec, const dim< 2 > &size, size_type num_nonzeros, int block_size)
Creates an uninitialized FBCSR matrix of the specified size.
void sort_by_column_index()
Sorts the values blocks and block-column indices in each row by column index.
value_type * get_values() noexcept
Definition fbcsr.hpp:255
std::unique_ptr< LinOp > conj_transpose() const override
Returns a LinOp representing the conjugate transpose of the Transposable object.
Fbcsr & operator=(Fbcsr &&)
Move-assigns an Fbcsr matrix.
std::unique_ptr< absolute_type > compute_absolute() const override
Gets the AbsoluteLinOp.
Fbcsr(Fbcsr &&)
Move-constructs an Fbcsr matrix.
std::unique_ptr< Diagonal< ValueType > > extract_diagonal() const override
Extracts the diagonal entries of the matrix into a vector.
const index_type * get_const_row_ptrs() const noexcept
Definition fbcsr.hpp:298
Fbcsr & operator=(const Fbcsr &)
Copy-assigns an Fbcsr matrix.
void convert_to(Csr< ValueType, IndexType > *result) const override
Converts the matrix to CSR format.
static std::unique_ptr< Fbcsr > create(std::shared_ptr< const Executor > exec, const dim< 2 > &size, int block_size, array< value_type > values, array< index_type > col_idxs, array< index_type > row_ptrs)
Creates a FBCSR matrix from already allocated (and initialized) row pointer, column index and value a...
void read(const mat_data &data) override
Reads a matrix_data into Fbcsr format.
bool is_sorted_by_column_index() const
Tests if all row entry pairs (value, col_idx) are sorted by column index.
index_type get_num_block_cols() const noexcept
Definition fbcsr.hpp:335
void convert_to(SparsityCsr< ValueType, IndexType > *result) const override
Get the block sparsity pattern in CSR-like format.
SparsityCsr is a matrix format which stores only the sparsity pattern of a sparse matrix by compressi...
Definition sparsity_csr.hpp:56
The matrix namespace.
Definition dense_cache.hpp:24
The Ginkgo namespace.
Definition abstract_factory.hpp:20
typename detail::remove_complex_s< T >::type remove_complex
Obtain the type which removed the complex of complex/scalar type or the template parameter of class b...
Definition math.hpp:264
typename detail::to_complex_s< T >::type to_complex
Obtain the type which adds the complex of complex/scalar type or the template parameter of class by a...
Definition math.hpp:283
std::size_t size_type
Integral type used for allocation quantities.
Definition types.hpp:90
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
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
STL namespace.
A type representing the dimensions of a multidimensional object.
Definition dim.hpp:26
This structure is used as an intermediate data type to store a sparse matrix.
Definition matrix_data.hpp:126