Ginkgo  Generated from pipelines/2662685947 branch based on develop. Ginkgo version 2.0.0
A numerical linear algebra library targeting many-core architectures
fbcsr.hpp
1 // SPDX-FileCopyrightText: 2017 - 2026 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 
14 namespace gko {
15 namespace matrix {
16 
17 
18 template <typename ValueType>
19 class Dense;
20 
21 template <typename ValueType, typename IndexType>
22 class Csr;
23 
24 template <typename ValueType, typename IndexType>
25 class SparsityCsr;
26 
27 template <typename ValueType, typename IndexType>
28 class Fbcsr;
29 
30 template <typename ValueType, typename IndexType>
32 
33 
34 namespace detail {
35 
36 
47 template <typename IndexType>
48 inline 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 
98 template <typename ValueType = default_precision, typename IndexType = int32>
99 class Fbcsr
100  : public LinOp,
101  public EnableCloneable<Fbcsr<ValueType, IndexType>>,
102  public ConvertibleTo<Fbcsr<next_precision<ValueType>, IndexType>>,
103 #if GINKGO_ENABLE_HALF || GINKGO_ENABLE_BFLOAT16
104  public ConvertibleTo<Fbcsr<next_precision<ValueType, 2>, IndexType>>,
105 #endif
106 #if GINKGO_ENABLE_HALF && GINKGO_ENABLE_BFLOAT16
107  public ConvertibleTo<Fbcsr<next_precision<ValueType, 3>, IndexType>>,
108 #endif
109  public ConvertibleTo<Dense<ValueType>>,
110  public ConvertibleTo<Csr<ValueType, IndexType>>,
111  public ConvertibleTo<SparsityCsr<ValueType, IndexType>>,
112  public DiagonalExtractable<ValueType>,
113  public ReadableFromMatrixData<ValueType, IndexType>,
114  public WritableToMatrixData<ValueType, IndexType>,
115  public Transposable,
117  remove_complex<Fbcsr<ValueType, IndexType>>> {
118  friend class EnableCloneable<Fbcsr>;
119  friend class Csr<ValueType, IndexType>;
120  friend class Dense<ValueType>;
121  friend class SparsityCsr<ValueType, IndexType>;
122  friend class FbcsrBuilder<ValueType, IndexType>;
123  friend class Fbcsr<to_complex<ValueType>, IndexType>;
124  GKO_ASSERT_SUPPORTED_VALUE_AND_INDEX_TYPE;
125 
126 public:
127  using value_type = ValueType;
128  using index_type = IndexType;
129  using transposed_type = Fbcsr<ValueType, IndexType>;
130  using mat_data = matrix_data<ValueType, IndexType>;
131  using device_mat_data = device_matrix_data<ValueType, IndexType>;
132  using absolute_type = remove_complex<Fbcsr>;
133 
140 
146 
147  using ConvertibleTo<
148  Fbcsr<next_precision<ValueType>, IndexType>>::convert_to;
149  using ConvertibleTo<Fbcsr<next_precision<ValueType>, IndexType>>::move_to;
150  using ConvertibleTo<Dense<ValueType>>::convert_to;
151  using ConvertibleTo<Dense<ValueType>>::move_to;
152  using ConvertibleTo<Csr<ValueType, IndexType>>::convert_to;
156 
157  friend class Fbcsr<previous_precision<ValueType>, IndexType>;
158 
159  void convert_to(
160  Fbcsr<next_precision<ValueType>, IndexType>* result) const override;
161 
162  void move_to(Fbcsr<next_precision<ValueType>, IndexType>* result) override;
163 
164 #if GINKGO_ENABLE_HALF || GINKGO_ENABLE_BFLOAT16
165  friend class Fbcsr<previous_precision<ValueType, 2>, IndexType>;
166  using ConvertibleTo<
167  Fbcsr<next_precision<ValueType, 2>, IndexType>>::convert_to;
168  using ConvertibleTo<
169  Fbcsr<next_precision<ValueType, 2>, IndexType>>::move_to;
170 
171  void convert_to(
172  Fbcsr<next_precision<ValueType, 2>, IndexType>* result) const override;
173 
174  void move_to(
175  Fbcsr<next_precision<ValueType, 2>, IndexType>* result) override;
176 #endif
177 
178 #if GINKGO_ENABLE_HALF && GINKGO_ENABLE_BFLOAT16
179  friend class Fbcsr<previous_precision<ValueType, 3>, IndexType>;
180  using ConvertibleTo<
181  Fbcsr<next_precision<ValueType, 3>, IndexType>>::convert_to;
182  using ConvertibleTo<
183  Fbcsr<next_precision<ValueType, 3>, IndexType>>::move_to;
184 
185  void convert_to(
186  Fbcsr<next_precision<ValueType, 3>, IndexType>* result) const override;
187 
188  void move_to(
189  Fbcsr<next_precision<ValueType, 3>, IndexType>* result) override;
190 #endif
191 
192  void convert_to(Dense<ValueType>* other) const override;
193 
194  void move_to(Dense<ValueType>* other) override;
195 
202  void convert_to(Csr<ValueType, IndexType>* result) const override;
203 
204  void move_to(Csr<ValueType, IndexType>* result) override;
205 
212  void convert_to(SparsityCsr<ValueType, IndexType>* result) const override;
213 
214  void move_to(SparsityCsr<ValueType, IndexType>* result) override;
215 
222  void read(const mat_data& data) override;
223 
224  void read(const device_mat_data& data) override;
225 
226  void read(device_mat_data&& data) override;
227 
228  void write(mat_data& data) const override;
229 
230  std::unique_ptr<LinOp> transpose() const override;
231 
232  std::unique_ptr<LinOp> conj_transpose() const override;
233 
234  std::unique_ptr<Diagonal<ValueType>> extract_diagonal() const override;
235 
236  std::unique_ptr<absolute_type> compute_absolute() const override;
237 
238  void compute_absolute_inplace() override;
239 
244  void sort_by_column_index();
245 
252  bool is_sorted_by_column_index() const;
253 
257  value_type* get_values() noexcept { return values_.get_data(); }
258 
266  const value_type* get_const_values() const noexcept
267  {
268  return values_.get_const_data();
269  }
270 
274  index_type* get_col_idxs() noexcept { return col_idxs_.get_data(); }
275 
283  const index_type* get_const_col_idxs() const noexcept
284  {
285  return col_idxs_.get_const_data();
286  }
287 
291  index_type* get_row_ptrs() noexcept { return row_ptrs_.get_data(); }
292 
300  const index_type* get_const_row_ptrs() const noexcept
301  {
302  return row_ptrs_.get_const_data();
303  }
304 
309  {
310  return values_.get_size();
311  }
312 
317  {
318  return col_idxs_.get_size();
319  }
320 
324  int get_block_size() const noexcept { return bs_; }
325 
329  index_type get_num_block_rows() const noexcept
330  {
331  return this->get_size()[0] / bs_;
332  }
333 
337  index_type get_num_block_cols() const noexcept
338  {
339  return this->get_size()[1] / bs_;
340  }
341 
351  static std::unique_ptr<Fbcsr> create(std::shared_ptr<const Executor> exec,
352  int block_size = 1);
353 
365  static std::unique_ptr<Fbcsr> create(std::shared_ptr<const Executor> exec,
366  const dim<2>& size,
367  size_type num_nonzeros,
368  int block_size);
369 
390  static std::unique_ptr<Fbcsr> create(std::shared_ptr<const Executor> exec,
391  const dim<2>& size, int block_size,
392  array<value_type> values,
393  array<index_type> col_idxs,
394  array<index_type> row_ptrs);
395 
401  template <typename InputValueType, typename InputColumnIndexType,
402  typename InputRowPtrType>
403  GKO_DEPRECATED(
404  "explicitly construct the gko::array argument instead of passing "
405  "initializer lists")
406  static std::unique_ptr<Fbcsr> create(
407  std::shared_ptr<const Executor> exec, const dim<2>& size,
408  int block_size, std::initializer_list<InputValueType> values,
409  std::initializer_list<InputColumnIndexType> col_idxs,
410  std::initializer_list<InputRowPtrType> row_ptrs)
411  {
412  return create(exec, size, block_size,
413  array<value_type>{exec, std::move(values)},
414  array<index_type>{exec, std::move(col_idxs)},
415  array<index_type>{exec, std::move(row_ptrs)});
416  }
417 
432  static std::unique_ptr<const Fbcsr> create_const(
433  std::shared_ptr<const Executor> exec, const dim<2>& size, int blocksize,
434  gko::detail::const_array_view<ValueType>&& values,
435  gko::detail::const_array_view<IndexType>&& col_idxs,
436  gko::detail::const_array_view<IndexType>&& row_ptrs);
437 
442  Fbcsr& operator=(const Fbcsr&);
443 
449  Fbcsr& operator=(Fbcsr&&);
450 
454  Fbcsr(const Fbcsr&);
455 
461  Fbcsr(Fbcsr&&);
462 
463 protected:
464  Fbcsr(std::shared_ptr<const Executor> exec, int block_size = 1);
465 
466  Fbcsr(std::shared_ptr<const Executor> exec, const dim<2>& size,
467  size_type num_nonzeros, int block_size);
468 
469  Fbcsr(std::shared_ptr<const Executor> exec, const dim<2>& size,
470  int block_size, array<value_type> values, array<index_type> col_idxs,
471  array<index_type> row_ptrs);
472 
473  void apply_impl(const LinOp* b, LinOp* x) const override;
474 
475  void apply_impl(const LinOp* alpha, const LinOp* b, const LinOp* beta,
476  LinOp* x) const override;
477 
478 private:
479  int bs_;
480  array<value_type> values_;
481  array<index_type> col_idxs_;
482  array<index_type> row_ptrs_;
483 };
484 
485 
486 } // namespace matrix
487 } // namespace gko
488 
489 
490 #endif // GKO_PUBLIC_CORE_MATRIX_FBCSR_HPP_
gko::matrix::Fbcsr
Fixed-block compressed sparse row storage matrix format.
Definition: csr.hpp:46
gko::matrix::Fbcsr::Fbcsr
Fbcsr(const Fbcsr &)
Copy-constructs an Ell matrix.
gko::matrix::Fbcsr::get_block_size
int get_block_size() const noexcept
Definition: fbcsr.hpp:324
gko::LinOp
Definition: lin_op.hpp:117
gko::matrix::Fbcsr::get_const_values
const value_type * get_const_values() const noexcept
Definition: fbcsr.hpp:266
gko::DiagonalExtractable
The diagonal of a LinOp implementing this interface can be extracted.
Definition: lin_op.hpp:702
gko::matrix::Fbcsr::extract_diagonal
std::unique_ptr< Diagonal< ValueType > > extract_diagonal() const override
Extracts the diagonal entries of the matrix into a vector.
gko::matrix::Fbcsr::get_num_stored_elements
size_type get_num_stored_elements() const noexcept
Definition: fbcsr.hpp:308
gko::Transposable
Linear operators which support transposition should implement the Transposable interface.
Definition: lin_op.hpp:392
gko::size_type
std::size_t size_type
Integral type used for allocation quantities.
Definition: types.hpp:90
gko::matrix::Fbcsr::is_sorted_by_column_index
bool is_sorted_by_column_index() const
Tests if all row entry pairs (value, col_idx) are sorted by column index.
gko::matrix::Fbcsr::get_values
value_type * get_values() noexcept
Definition: fbcsr.hpp:257
gko::matrix::Fbcsr::get_row_ptrs
index_type * get_row_ptrs() noexcept
Definition: fbcsr.hpp:291
gko::matrix::Fbcsr::get_num_block_cols
index_type get_num_block_cols() const noexcept
Definition: fbcsr.hpp:337
gko
The Ginkgo namespace.
Definition: abstract_factory.hpp:20
gko::matrix::Fbcsr::get_const_col_idxs
const index_type * get_const_col_idxs() const noexcept
Definition: fbcsr.hpp:283
gko::array< value_type >
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::matrix::Fbcsr::sort_by_column_index
void sort_by_column_index()
Sorts the values blocks and block-column indices in each row by column index.
gko::matrix::Fbcsr::get_const_row_ptrs
const index_type * get_const_row_ptrs() const noexcept
Definition: fbcsr.hpp:300
gko::array::get_data
value_type * get_data() noexcept
Returns a pointer to the block of memory used to store the elements of the array.
Definition: array.hpp:687
gko::ReadableFromMatrixData
A LinOp implementing this interface can read its data from a matrix_data structure.
Definition: lin_op.hpp:564
gko::matrix::Fbcsr::conj_transpose
std::unique_ptr< LinOp > conj_transpose() const override
Returns a LinOp representing the conjugate transpose of the Transposable object.
gko::WritableToMatrixData
A LinOp implementing this interface can write its data to a matrix_data structure.
Definition: lin_op.hpp:619
gko::matrix::Fbcsr::compute_absolute
std::unique_ptr< absolute_type > compute_absolute() const override
Gets the AbsoluteLinOp.
gko::matrix::Fbcsr::get_num_block_rows
index_type get_num_block_rows() const noexcept
Definition: fbcsr.hpp:329
gko::matrix::Fbcsr::create
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.
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::matrix::Fbcsr::operator=
Fbcsr & operator=(const Fbcsr &)
Copy-assigns an Fbcsr matrix.
gko::matrix::Fbcsr::get_num_stored_blocks
size_type get_num_stored_blocks() const noexcept
Definition: fbcsr.hpp:316
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::EnableCloneable
This mixin is used to enable a default Cloneable::clone() implementation and similar for objects that...
Definition: polymorphic_object.hpp:369
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::matrix::Fbcsr::get_col_idxs
index_type * get_col_idxs() noexcept
Definition: fbcsr.hpp:274
gko::Executor
The first step in using the Ginkgo library consists of creating an executor.
Definition: executor.hpp:616
gko::matrix::FbcsrBuilder
Definition: fbcsr.hpp:31
gko::array::get_const_data
const value_type * get_const_data() const noexcept
Returns a constant pointer to the block of memory used to store the elements of the array.
Definition: array.hpp:696
gko::matrix::Fbcsr::compute_absolute_inplace
void compute_absolute_inplace() override
Compute absolute inplace on each element.
gko::EnableAbsoluteComputation
The EnableAbsoluteComputation mixin provides the default implementations of compute_absolute_linop an...
Definition: lin_op.hpp:753
gko::matrix::Fbcsr::create_const
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.
gko::array::get_size
size_type get_size() const noexcept
Returns the number of elements in the array.
Definition: array.hpp:670
gko::LinOp::get_size
const dim< 2 > & get_size() const noexcept
Returns the size of the operator.
Definition: lin_op.hpp:169
gko::matrix::Fbcsr::write
void write(mat_data &data) const override
Writes a matrix to a matrix_data structure.
gko::matrix::Fbcsr::read
void read(const mat_data &data) override
Reads a matrix_data into Fbcsr format.
gko::remove_complex
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
gko::device_matrix_data
This type is a device-side equivalent to matrix_data.
Definition: device_matrix_data.hpp:36
gko::matrix::Fbcsr::transpose
std::unique_ptr< LinOp > transpose() const override
Returns a LinOp representing the transpose of the Transposable object.
gko::to_complex
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