Ginkgo Generated from branch based on main. Ginkgo version 1.11.0
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 GKO_ASSERT_SUPPORTED_VALUE_AND_INDEX_TYPE;
124
125public:
126 using value_type = ValueType;
127 using index_type = IndexType;
128 using transposed_type = Fbcsr<ValueType, IndexType>;
129 using mat_data = matrix_data<ValueType, IndexType>;
130 using device_mat_data = device_matrix_data<ValueType, IndexType>;
131 using absolute_type = remove_complex<Fbcsr>;
132
139
144 using EnableLinOp<Fbcsr<ValueType, IndexType>>::convert_to;
145
146 using ConvertibleTo<
147 Fbcsr<next_precision<ValueType>, IndexType>>::convert_to;
148 using ConvertibleTo<Fbcsr<next_precision<ValueType>, IndexType>>::move_to;
149 using ConvertibleTo<Dense<ValueType>>::convert_to;
150 using ConvertibleTo<Dense<ValueType>>::move_to;
151 using ConvertibleTo<Csr<ValueType, IndexType>>::convert_to;
152 using ConvertibleTo<Csr<ValueType, IndexType>>::move_to;
153 using ConvertibleTo<SparsityCsr<ValueType, IndexType>>::convert_to;
154 using ConvertibleTo<SparsityCsr<ValueType, IndexType>>::move_to;
155
156 friend class Fbcsr<previous_precision<ValueType>, IndexType>;
157
158 void convert_to(
159 Fbcsr<next_precision<ValueType>, IndexType>* result) const override;
160
161 void move_to(Fbcsr<next_precision<ValueType>, IndexType>* result) override;
162
163#if GINKGO_ENABLE_HALF || GINKGO_ENABLE_BFLOAT16
164 friend class Fbcsr<previous_precision<ValueType, 2>, IndexType>;
165 using ConvertibleTo<
166 Fbcsr<next_precision<ValueType, 2>, IndexType>>::convert_to;
167 using ConvertibleTo<
168 Fbcsr<next_precision<ValueType, 2>, IndexType>>::move_to;
169
170 void convert_to(
171 Fbcsr<next_precision<ValueType, 2>, IndexType>* result) const override;
172
173 void move_to(
174 Fbcsr<next_precision<ValueType, 2>, IndexType>* result) override;
175#endif
176
177#if GINKGO_ENABLE_HALF && GINKGO_ENABLE_BFLOAT16
178 friend class Fbcsr<previous_precision<ValueType, 3>, IndexType>;
179 using ConvertibleTo<
180 Fbcsr<next_precision<ValueType, 3>, IndexType>>::convert_to;
181 using ConvertibleTo<
182 Fbcsr<next_precision<ValueType, 3>, IndexType>>::move_to;
183
184 void convert_to(
185 Fbcsr<next_precision<ValueType, 3>, IndexType>* result) const override;
186
187 void move_to(
188 Fbcsr<next_precision<ValueType, 3>, IndexType>* result) override;
189#endif
190
191 void convert_to(Dense<ValueType>* other) const override;
192
193 void move_to(Dense<ValueType>* other) override;
194
201 void convert_to(Csr<ValueType, IndexType>* result) const override;
202
203 void move_to(Csr<ValueType, IndexType>* result) override;
204
211 void convert_to(SparsityCsr<ValueType, IndexType>* result) const override;
212
213 void move_to(SparsityCsr<ValueType, IndexType>* result) override;
214
221 void read(const mat_data& data) override;
222
223 void read(const device_mat_data& data) override;
224
225 void read(device_mat_data&& data) override;
226
227 void write(mat_data& data) const override;
228
229 std::unique_ptr<LinOp> transpose() const override;
230
231 std::unique_ptr<LinOp> conj_transpose() const override;
232
233 std::unique_ptr<Diagonal<ValueType>> extract_diagonal() const override;
234
235 std::unique_ptr<absolute_type> compute_absolute() const override;
236
238
244
252
256 value_type* get_values() noexcept { return values_.get_data(); }
257
265 const value_type* get_const_values() const noexcept
266 {
267 return values_.get_const_data();
268 }
269
273 index_type* get_col_idxs() noexcept { return col_idxs_.get_data(); }
274
282 const index_type* get_const_col_idxs() const noexcept
283 {
284 return col_idxs_.get_const_data();
285 }
286
290 index_type* get_row_ptrs() noexcept { return row_ptrs_.get_data(); }
291
299 const index_type* get_const_row_ptrs() const noexcept
300 {
301 return row_ptrs_.get_const_data();
302 }
303
308 {
309 return values_.get_size();
310 }
311
316 {
317 return col_idxs_.get_size();
318 }
319
323 int get_block_size() const noexcept { return bs_; }
324
328 index_type get_num_block_rows() const noexcept
329 {
330 return this->get_size()[0] / bs_;
331 }
332
336 index_type get_num_block_cols() const noexcept
337 {
338 return this->get_size()[1] / bs_;
339 }
340
350 static std::unique_ptr<Fbcsr> create(std::shared_ptr<const Executor> exec,
351 int block_size = 1);
352
364 static std::unique_ptr<Fbcsr> create(std::shared_ptr<const Executor> exec,
365 const dim<2>& size,
366 size_type num_nonzeros,
367 int block_size);
368
389 static std::unique_ptr<Fbcsr> create(std::shared_ptr<const Executor> exec,
390 const dim<2>& size, int block_size,
391 array<value_type> values,
392 array<index_type> col_idxs,
393 array<index_type> row_ptrs);
394
400 template <typename InputValueType, typename InputColumnIndexType,
401 typename InputRowPtrType>
402 GKO_DEPRECATED(
403 "explicitly construct the gko::array argument instead of passing "
404 "initializer lists")
405 static std::unique_ptr<Fbcsr> create(
406 std::shared_ptr<const Executor> exec, const dim<2>& size,
407 int block_size, std::initializer_list<InputValueType> values,
408 std::initializer_list<InputColumnIndexType> col_idxs,
409 std::initializer_list<InputRowPtrType> row_ptrs)
410 {
411 return create(exec, size, block_size,
412 array<value_type>{exec, std::move(values)},
413 array<index_type>{exec, std::move(col_idxs)},
414 array<index_type>{exec, std::move(row_ptrs)});
415 }
416
431 static std::unique_ptr<const Fbcsr> create_const(
432 std::shared_ptr<const Executor> exec, const dim<2>& size, int blocksize,
433 gko::detail::const_array_view<ValueType>&& values,
434 gko::detail::const_array_view<IndexType>&& col_idxs,
435 gko::detail::const_array_view<IndexType>&& row_ptrs);
436
442
449
453 Fbcsr(const Fbcsr&);
454
461
462protected:
463 Fbcsr(std::shared_ptr<const Executor> exec, int block_size = 1);
464
465 Fbcsr(std::shared_ptr<const Executor> exec, const dim<2>& size,
466 size_type num_nonzeros, int block_size);
467
468 Fbcsr(std::shared_ptr<const Executor> exec, const dim<2>& size,
469 int block_size, array<value_type> values, array<index_type> col_idxs,
470 array<index_type> row_ptrs);
471
472 void apply_impl(const LinOp* b, LinOp* x) const override;
473
474 void apply_impl(const LinOp* alpha, const LinOp* b, const LinOp* beta,
475 LinOp* x) const override;
476
477private:
478 int bs_;
479 array<value_type> values_;
480 array<index_type> col_idxs_;
481 array<index_type> row_ptrs_;
482};
483
484
485} // namespace matrix
486} // namespace gko
487
488
489#endif // GKO_PUBLIC_CORE_MATRIX_FBCSR_HPP_
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
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:315
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:307
const value_type * get_const_values() const noexcept
Definition fbcsr.hpp:265
const index_type * get_const_col_idxs() const noexcept
Definition fbcsr.hpp:282
index_type get_num_block_rows() const noexcept
Definition fbcsr.hpp:328
index_type * get_row_ptrs() noexcept
Definition fbcsr.hpp:290
int get_block_size() const noexcept
Definition fbcsr.hpp:323
index_type * get_col_idxs() noexcept
Definition fbcsr.hpp:273
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:256
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:299
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:336
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