5#ifndef GKO_PUBLIC_CORE_MATRIX_CSR_HPP_
6#define GKO_PUBLIC_CORE_MATRIX_CSR_HPP_
9#include <ginkgo/core/base/array.hpp>
10#include <ginkgo/core/base/index_set.hpp>
11#include <ginkgo/core/base/lin_op.hpp>
12#include <ginkgo/core/base/math.hpp>
13#include <ginkgo/core/matrix/permutation.hpp>
14#include <ginkgo/core/matrix/scaled_permutation.hpp>
21template <
typename ValueType>
24template <
typename ValueType>
27template <
typename ValueType,
typename IndexType>
30template <
typename ValueType,
typename IndexType>
33template <
typename ValueType,
typename IndexType>
36template <
typename ValueType,
typename IndexType>
39template <
typename ValueType,
typename IndexType>
42template <
typename ValueType,
typename IndexType>
45template <
typename ValueType,
typename IndexType>
48template <
typename ValueType,
typename IndexType>
55template <
typename ValueType = default_precision,
typename IndexType =
int32>
100template <
typename ValueType = default_precision,
typename IndexType =
int32>
102 public ConvertibleTo<Csr<next_precision<ValueType>, IndexType>>,
103#if GINKGO_ENABLE_HALF
105 Csr<next_precision<next_precision<ValueType>>, IndexType>>,
120 remove_complex<Csr<ValueType, IndexType>>>,
123 friend class Coo<ValueType, IndexType>;
124 friend class Dense<ValueType>;
126 friend class Ell<ValueType, IndexType>;
127 friend class Hybrid<ValueType, IndexType>;
128 friend class Sellp<ValueType, IndexType>;
130 friend class Fbcsr<ValueType, IndexType>;
131 friend class CsrBuilder<ValueType, IndexType>;
155 using value_type = ValueType;
156 using index_type = IndexType;
212 virtual std::shared_ptr<strategy_type>
copy() = 0;
215 void set_name(std::string name) { name_ = name; }
237 auto host_mtx_exec = mtx_row_ptrs.
get_executor()->get_master();
239 const bool is_mtx_on_host{host_mtx_exec ==
241 const index_type* row_ptrs{};
242 if (is_mtx_on_host) {
245 row_ptrs_host = mtx_row_ptrs;
248 auto num_rows = mtx_row_ptrs.
get_size() - 1;
249 max_length_per_row_ = 0;
250 for (
size_type i = 0; i < num_rows; i++) {
251 max_length_per_row_ = std::max(max_length_per_row_,
252 row_ptrs[i + 1] - row_ptrs[i]);
256 int64_t
clac_size(
const int64_t nnz)
override {
return 0; }
258 index_type get_max_length_per_row() const noexcept
260 return max_length_per_row_;
263 std::shared_ptr<strategy_type>
copy()
override
265 return std::make_shared<classical>();
269 index_type max_length_per_row_;
288 int64_t
clac_size(
const int64_t nnz)
override {
return 0; }
290 std::shared_ptr<strategy_type>
copy()
override
292 return std::make_shared<merge_path>();
313 int64_t
clac_size(
const int64_t nnz)
override {
return 0; }
315 std::shared_ptr<strategy_type>
copy()
override
317 return std::make_shared<cusparse>();
337 int64_t
clac_size(
const int64_t nnz)
override {
return 0; }
339 std::shared_ptr<strategy_type>
copy()
override
341 return std::make_shared<sparselib>();
367 :
load_balance(exec->get_num_warps(), exec->get_warp_size())
376 :
load_balance(exec->get_num_warps(), exec->get_warp_size(), false)
387 :
load_balance(exec->get_num_subgroups(), 32, false,
"intel")
402 bool cuda_strategy =
true,
403 std::string strategy_name =
"none")
406 warp_size_(warp_size),
407 cuda_strategy_(cuda_strategy),
408 strategy_name_(strategy_name)
417 auto host_srow_exec = mtx_srow->
get_executor()->get_master();
418 auto host_mtx_exec = mtx_row_ptrs.
get_executor()->get_master();
419 const bool is_srow_on_host{host_srow_exec ==
421 const bool is_mtx_on_host{host_mtx_exec ==
425 const index_type* row_ptrs{};
427 if (is_srow_on_host) {
430 srow_host = *mtx_srow;
433 if (is_mtx_on_host) {
436 row_ptrs_host = mtx_row_ptrs;
442 const auto num_rows = mtx_row_ptrs.
get_size() - 1;
443 const auto num_elems = row_ptrs[num_rows];
444 const auto bucket_divider =
445 num_elems > 0 ?
ceildiv(num_elems, warp_size_) : 1;
446 for (
size_type i = 0; i < num_rows; i++) {
450 if (bucket < nwarps) {
456 srow[i] += srow[i - 1];
458 if (!is_srow_on_host) {
459 *mtx_srow = srow_host;
466 if (warp_size_ > 0) {
468 if (nnz >=
static_cast<int64_t
>(2e8)) {
470 }
else if (nnz >=
static_cast<int64_t
>(2e7)) {
472 }
else if (nnz >=
static_cast<int64_t
>(2e6)) {
474 }
else if (nnz >=
static_cast<int64_t
>(2e5)) {
477 if (strategy_name_ ==
"intel") {
479 if (nnz >=
static_cast<int64_t
>(2e8)) {
481 }
else if (nnz >=
static_cast<int64_t
>(2e7)) {
485#if GINKGO_HIP_PLATFORM_HCC
486 if (!cuda_strategy_) {
488 if (nnz >=
static_cast<int64_t
>(1e7)) {
490 }
else if (nnz >=
static_cast<int64_t
>(1e6)) {
496 auto nwarps = nwarps_ * multiple;
503 std::shared_ptr<strategy_type>
copy()
override
505 return std::make_shared<load_balance>(
506 nwarps_, warp_size_, cuda_strategy_, strategy_name_);
513 std::string strategy_name_;
520 const index_type nvidia_row_len_limit = 1024;
523 const index_type nvidia_nnz_limit{
static_cast<index_type
>(1e6)};
526 const index_type amd_row_len_limit = 768;
529 const index_type amd_nnz_limit{
static_cast<index_type
>(1e8)};
532 const index_type intel_row_len_limit = 25600;
535 const index_type intel_nnz_limit{
static_cast<index_type
>(3e8)};
555 :
automatical(exec->get_num_warps(), exec->get_warp_size())
564 :
automatical(exec->get_num_warps(), exec->get_warp_size(), false)
575 :
automatical(exec->get_num_subgroups(), 32, false,
"intel")
590 bool cuda_strategy =
true,
591 std::string strategy_name =
"none")
594 warp_size_(warp_size),
595 cuda_strategy_(cuda_strategy),
596 strategy_name_(strategy_name),
597 max_length_per_row_(0)
606 index_type nnz_limit = nvidia_nnz_limit;
607 index_type row_len_limit = nvidia_row_len_limit;
608 if (strategy_name_ ==
"intel") {
609 nnz_limit = intel_nnz_limit;
610 row_len_limit = intel_row_len_limit;
612#if GINKGO_HIP_PLATFORM_HCC
613 if (!cuda_strategy_) {
614 nnz_limit = amd_nnz_limit;
615 row_len_limit = amd_row_len_limit;
618 auto host_mtx_exec = mtx_row_ptrs.
get_executor()->get_master();
619 const bool is_mtx_on_host{host_mtx_exec ==
622 const index_type* row_ptrs{};
623 if (is_mtx_on_host) {
626 row_ptrs_host = mtx_row_ptrs;
629 const auto num_rows = mtx_row_ptrs.
get_size() - 1;
630 if (row_ptrs[num_rows] > nnz_limit) {
632 cuda_strategy_, strategy_name_);
633 if (is_mtx_on_host) {
634 actual_strategy.
process(mtx_row_ptrs, mtx_srow);
636 actual_strategy.
process(row_ptrs_host, mtx_srow);
638 this->set_name(actual_strategy.
get_name());
640 index_type maxnum = 0;
641 for (
size_type i = 0; i < num_rows; i++) {
642 maxnum = std::max(maxnum, row_ptrs[i + 1] - row_ptrs[i]);
644 if (maxnum > row_len_limit) {
646 nwarps_, warp_size_, cuda_strategy_, strategy_name_);
647 if (is_mtx_on_host) {
648 actual_strategy.
process(mtx_row_ptrs, mtx_srow);
650 actual_strategy.
process(row_ptrs_host, mtx_srow);
652 this->set_name(actual_strategy.
get_name());
655 if (is_mtx_on_host) {
656 actual_strategy.
process(mtx_row_ptrs, mtx_srow);
657 max_length_per_row_ =
658 actual_strategy.get_max_length_per_row();
660 actual_strategy.
process(row_ptrs_host, mtx_srow);
661 max_length_per_row_ =
662 actual_strategy.get_max_length_per_row();
664 this->set_name(actual_strategy.
get_name());
671 return std::make_shared<load_balance>(
672 nwarps_, warp_size_, cuda_strategy_, strategy_name_)
676 index_type get_max_length_per_row() const noexcept
678 return max_length_per_row_;
681 std::shared_ptr<strategy_type>
copy()
override
683 return std::make_shared<automatical>(
684 nwarps_, warp_size_, cuda_strategy_, strategy_name_);
691 std::string strategy_name_;
692 index_type max_length_per_row_;
695 friend class Csr<previous_precision<ValueType>, IndexType>;
702#if GINKGO_ENABLE_HALF
703 friend class Csr<previous_precision<previous_precision<ValueType>>,
711 result)
const override;
791 bool invert =
false)
const;
823 bool invert =
false)
const;
860 bool is_sorted_by_column_index()
const;
974 strategy_ = std::move(strategy->copy());
987 GKO_ASSERT_EQUAL_DIMENSIONS(alpha,
dim<2>(1, 1));
1000 GKO_ASSERT_EQUAL_DIMENSIONS(alpha,
dim<2>(1, 1));
1012 static std::unique_ptr<Csr>
create(std::shared_ptr<const Executor> exec,
1013 std::shared_ptr<strategy_type> strategy);
1027 std::shared_ptr<const Executor> exec,
const dim<2>& size = {},
1029 std::shared_ptr<strategy_type> strategy =
nullptr);
1051 std::shared_ptr<const Executor> exec,
const dim<2>& size,
1054 std::shared_ptr<strategy_type> strategy =
nullptr);
1060 template <
typename InputValueType,
typename InputColumnIndexType,
1061 typename InputRowPtrType>
1063 "explicitly construct the gko::array argument instead of passing "
1064 "initializer lists")
1067 std::initializer_list<InputValueType> values,
1068 std::initializer_list<InputColumnIndexType> col_idxs,
1069 std::initializer_list<InputRowPtrType> row_ptrs)
1092 std::shared_ptr<const Executor> exec,
const dim<2>& size,
1093 gko::detail::const_array_view<ValueType>&& values,
1094 gko::detail::const_array_view<IndexType>&& col_idxs,
1095 gko::detail::const_array_view<IndexType>&& row_ptrs,
1096 std::shared_ptr<strategy_type> strategy =
nullptr);
1126 const span& row_span,
const span& column_span)
const;
1153 Csr(std::shared_ptr<const Executor> exec,
const dim<2>& size = {},
1155 std::shared_ptr<strategy_type> strategy =
nullptr);
1157 Csr(std::shared_ptr<const Executor> exec,
const dim<2>& size,
1160 std::shared_ptr<strategy_type> strategy =
nullptr);
1162 void apply_impl(
const LinOp* b,
LinOp* x)
const override;
1164 void apply_impl(
const LinOp* alpha,
const LinOp* b,
const LinOp* beta,
1165 LinOp* x)
const override;
1168 static std::shared_ptr<strategy_type> make_default_strategy(
1169 std::shared_ptr<const Executor> exec)
1171 auto cuda_exec = std::dynamic_pointer_cast<const CudaExecutor>(exec);
1172 auto hip_exec = std::dynamic_pointer_cast<const HipExecutor>(exec);
1173 auto dpcpp_exec = std::dynamic_pointer_cast<const DpcppExecutor>(exec);
1174 std::shared_ptr<strategy_type> new_strategy;
1176 new_strategy = std::make_shared<automatical>(cuda_exec);
1177 }
else if (hip_exec) {
1178 new_strategy = std::make_shared<automatical>(hip_exec);
1179 }
else if (dpcpp_exec) {
1180 new_strategy = std::make_shared<automatical>(dpcpp_exec);
1182 new_strategy = std::make_shared<classical>();
1184 return new_strategy;
1188 template <
typename CsrType>
1189 void convert_strategy_helper(CsrType* result)
const
1192 std::shared_ptr<typename CsrType::strategy_type> new_strat;
1193 if (
dynamic_cast<classical*
>(strat)) {
1194 new_strat = std::make_shared<typename CsrType::classical>();
1195 }
else if (
dynamic_cast<merge_path*
>(strat)) {
1196 new_strat = std::make_shared<typename CsrType::merge_path>();
1197 }
else if (
dynamic_cast<cusparse*
>(strat)) {
1198 new_strat = std::make_shared<typename CsrType::cusparse>();
1199 }
else if (
dynamic_cast<sparselib*
>(strat)) {
1200 new_strat = std::make_shared<typename CsrType::sparselib>();
1202 auto rexec = result->get_executor();
1204 std::dynamic_pointer_cast<const CudaExecutor>(rexec);
1205 auto hip_exec = std::dynamic_pointer_cast<const HipExecutor>(rexec);
1207 std::dynamic_pointer_cast<const DpcppExecutor>(rexec);
1208 auto lb =
dynamic_cast<load_balance*
>(strat);
1212 std::make_shared<typename CsrType::load_balance>(
1215 new_strat = std::make_shared<typename CsrType::automatical>(
1218 }
else if (hip_exec) {
1221 std::make_shared<typename CsrType::load_balance>(
1224 new_strat = std::make_shared<typename CsrType::automatical>(
1227 }
else if (dpcpp_exec) {
1230 std::make_shared<typename CsrType::load_balance>(
1233 new_strat = std::make_shared<typename CsrType::automatical>(
1238 auto this_cuda_exec =
1239 std::dynamic_pointer_cast<const CudaExecutor>(
1241 auto this_hip_exec =
1242 std::dynamic_pointer_cast<const HipExecutor>(
1244 auto this_dpcpp_exec =
1245 std::dynamic_pointer_cast<const DpcppExecutor>(
1247 if (this_cuda_exec) {
1250 std::make_shared<typename CsrType::load_balance>(
1254 std::make_shared<typename CsrType::automatical>(
1257 }
else if (this_hip_exec) {
1260 std::make_shared<typename CsrType::load_balance>(
1264 std::make_shared<typename CsrType::automatical>(
1267 }
else if (this_dpcpp_exec) {
1270 std::make_shared<typename CsrType::load_balance>(
1274 std::make_shared<typename CsrType::automatical>(
1282 new_strat = std::make_shared<typename CsrType::classical>();
1286 result->set_strategy(new_strat);
1295 strategy_->process(row_ptrs_, &srow_);
1304 virtual void scale_impl(
const LinOp* alpha);
1312 virtual void inv_scale_impl(
const LinOp* alpha);
1315 std::shared_ptr<strategy_type> strategy_;
1316 array<value_type> values_;
1317 array<index_type> col_idxs_;
1318 array<index_type> row_ptrs_;
1319 array<index_type> srow_;
1321 void add_scaled_identity_impl(
const LinOp* a,
const LinOp* b)
override;
1334template <
typename ValueType,
typename IndexType>
1335void strategy_rebuild_helper(Csr<ValueType, IndexType>* result)
1337 using load_balance =
typename Csr<ValueType, IndexType>::load_balance;
1338 using automatical =
typename Csr<ValueType, IndexType>::automatical;
1339 auto strategy = result->get_strategy();
1340 auto executor = result->get_executor();
1341 if (std::dynamic_pointer_cast<load_balance>(strategy)) {
1343 std::dynamic_pointer_cast<const HipExecutor>(executor)) {
1344 result->set_strategy(std::make_shared<load_balance>(exec));
1345 }
else if (
auto exec = std::dynamic_pointer_cast<const CudaExecutor>(
1347 result->set_strategy(std::make_shared<load_balance>(exec));
1349 }
else if (std::dynamic_pointer_cast<automatical>(strategy)) {
1351 std::dynamic_pointer_cast<const HipExecutor>(executor)) {
1352 result->set_strategy(std::make_shared<automatical>(exec));
1353 }
else if (
auto exec = std::dynamic_pointer_cast<const CudaExecutor>(
1355 result->set_strategy(std::make_shared<automatical>(exec));
ConvertibleTo interface is used to mark that the implementer can be converted to the object of Result...
Definition polymorphic_object.hpp:470
This is the Executor subclass which represents the CUDA device.
Definition executor.hpp:1542
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:662
The first step in using the Ginkgo library consists of creating an executor.
Definition executor.hpp:615
Definition lin_op.hpp:117
LinOp(const LinOp &)=default
Copy-constructs a LinOp.
This is the Executor subclass which represents the OpenMP device (typically CPU).
Definition executor.hpp:1387
Linear operators which support permutation should implement the Permutable interface.
Definition lin_op.hpp:484
std::shared_ptr< const Executor > get_executor() const noexcept
Returns the Executor of the object.
Definition polymorphic_object.hpp:234
A LinOp implementing this interface can read its data from a matrix_data structure.
Definition lin_op.hpp:605
Adds the operation M <- a I + b M for matrix M, identity operator I and scalars a and b,...
Definition lin_op.hpp:818
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 logger.hpp:25
void resize_and_reset(size_type size)
Resizes the array so it is able to hold the specified number of elements.
Definition array.hpp:622
value_type * get_data() noexcept
Returns a pointer to the block of memory used to store the elements of the array.
Definition array.hpp:673
std::shared_ptr< const Executor > get_executor() const noexcept
Returns the Executor associated with the array.
Definition array.hpp:689
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:682
size_type get_size() const noexcept
Returns the number of elements in the array.
Definition array.hpp:656
This type is a device-side equivalent to matrix_data.
Definition device_matrix_data.hpp:36
An index set class represents an ordered set of intervals.
Definition index_set.hpp:56
COO stores a matrix in the coordinate matrix format.
Definition ell.hpp:21
std::shared_ptr< strategy_type > copy() override
Copy a strategy.
Definition csr.hpp:681
automatical(int64_t nwarps, int warp_size=32, bool cuda_strategy=true, std::string strategy_name="none")
Creates an automatical strategy with specified parameters.
Definition csr.hpp:589
automatical()
Creates an automatical strategy.
Definition csr.hpp:544
int64_t clac_size(const int64_t nnz) override
Computes the srow size according to the number of nonzeros.
Definition csr.hpp:669
automatical(std::shared_ptr< const CudaExecutor > exec)
Creates an automatical strategy with CUDA executor.
Definition csr.hpp:554
void process(const array< index_type > &mtx_row_ptrs, array< index_type > *mtx_srow) override
Computes srow according to row pointers.
Definition csr.hpp:600
automatical(std::shared_ptr< const DpcppExecutor > exec)
Creates an automatical strategy with Dpcpp executor.
Definition csr.hpp:574
automatical(std::shared_ptr< const HipExecutor > exec)
Creates an automatical strategy with HIP executor.
Definition csr.hpp:563
classical is a strategy_type which uses the same number of threads on each row.
Definition csr.hpp:227
void process(const array< index_type > &mtx_row_ptrs, array< index_type > *mtx_srow) override
Computes srow according to row pointers.
Definition csr.hpp:234
std::shared_ptr< strategy_type > copy() override
Copy a strategy.
Definition csr.hpp:263
classical()
Creates a classical strategy.
Definition csr.hpp:232
int64_t clac_size(const int64_t nnz) override
Computes the srow size according to the number of nonzeros.
Definition csr.hpp:256
cusparse is a strategy_type which uses the sparselib csr.
Definition csr.hpp:302
int64_t clac_size(const int64_t nnz) override
Computes the srow size according to the number of nonzeros.
Definition csr.hpp:313
std::shared_ptr< strategy_type > copy() override
Copy a strategy.
Definition csr.hpp:315
cusparse()
Creates a cusparse strategy.
Definition csr.hpp:307
void process(const array< index_type > &mtx_row_ptrs, array< index_type > *mtx_srow) override
Computes srow according to row pointers.
Definition csr.hpp:309
load_balance is a strategy_type which uses the load balance algorithm.
Definition csr.hpp:348
void process(const array< index_type > &mtx_row_ptrs, array< index_type > *mtx_srow) override
Computes srow according to row pointers.
Definition csr.hpp:411
std::shared_ptr< strategy_type > copy() override
Copy a strategy.
Definition csr.hpp:503
load_balance(std::shared_ptr< const HipExecutor > exec)
Creates a load_balance strategy with HIP executor.
Definition csr.hpp:375
load_balance()
Creates a load_balance strategy.
Definition csr.hpp:356
int64_t clac_size(const int64_t nnz) override
Computes the srow size according to the number of nonzeros.
Definition csr.hpp:464
load_balance(int64_t nwarps, int warp_size=32, bool cuda_strategy=true, std::string strategy_name="none")
Creates a load_balance strategy with specified parameters.
Definition csr.hpp:401
load_balance(std::shared_ptr< const CudaExecutor > exec)
Creates a load_balance strategy with CUDA executor.
Definition csr.hpp:366
load_balance(std::shared_ptr< const DpcppExecutor > exec)
Creates a load_balance strategy with DPCPP executor.
Definition csr.hpp:386
merge_path is a strategy_type which uses the merge_path algorithm.
Definition csr.hpp:277
int64_t clac_size(const int64_t nnz) override
Computes the srow size according to the number of nonzeros.
Definition csr.hpp:288
std::shared_ptr< strategy_type > copy() override
Copy a strategy.
Definition csr.hpp:290
merge_path()
Creates a merge_path strategy.
Definition csr.hpp:282
void process(const array< index_type > &mtx_row_ptrs, array< index_type > *mtx_srow) override
Computes srow according to row pointers.
Definition csr.hpp:284
sparselib is a strategy_type which uses the sparselib csr.
Definition csr.hpp:326
int64_t clac_size(const int64_t nnz) override
Computes the srow size according to the number of nonzeros.
Definition csr.hpp:337
void process(const array< index_type > &mtx_row_ptrs, array< index_type > *mtx_srow) override
Computes srow according to row pointers.
Definition csr.hpp:333
sparselib()
Creates a sparselib strategy.
Definition csr.hpp:331
std::shared_ptr< strategy_type > copy() override
Copy a strategy.
Definition csr.hpp:339
strategy_type is to decide how to set the csr algorithm.
Definition csr.hpp:170
virtual int64_t clac_size(const int64_t nnz)=0
Computes the srow size according to the number of nonzeros.
std::string get_name()
Returns the name of strategy.
Definition csr.hpp:188
virtual std::shared_ptr< strategy_type > copy()=0
Copy a strategy.
virtual void process(const array< index_type > &mtx_row_ptrs, array< index_type > *mtx_srow)=0
Computes srow according to row pointers.
strategy_type(std::string name)
Creates a strategy_type.
Definition csr.hpp:179
CSR is a matrix format which stores only the nonzero coefficients by compressing each row of the matr...
Definition sparsity_csr.hpp:21
std::unique_ptr< LinOp > column_permute(const array< IndexType > *permutation_indices) const override
Returns a LinOp representing the column permutation of the Permutable object.
Csr & operator=(const Csr &)
Copy-assigns a Csr matrix.
std::unique_ptr< Csr > scale_permute(ptr_param< const ScaledPermutation< value_type, index_type > > permutation, permute_mode=permute_mode::symmetric) const
Creates a scaled and permuted copy of this matrix.
void write(mat_data &data) const override
Writes a matrix to a matrix_data structure.
std::unique_ptr< absolute_type > compute_absolute() const override
Gets the AbsoluteLinOp.
const index_type * get_const_row_ptrs() const noexcept
Returns the row pointers of the matrix.
Definition csr.hpp:914
std::unique_ptr< Csr< ValueType, IndexType > > create_submatrix(const span &row_span, const span &column_span) const
Creates a submatrix from this Csr matrix given row and column spans.
static std::unique_ptr< Csr > create(std::shared_ptr< const Executor > exec, const dim< 2 > &size={}, size_type num_nonzeros={}, std::shared_ptr< strategy_type > strategy=nullptr)
Creates an uninitialized CSR matrix of the specified size.
void read(device_mat_data &&data) override
Reads a matrix from a device_matrix_data structure.
const index_type * get_const_srow() const noexcept
Returns the starting rows.
Definition csr.hpp:933
void set_strategy(std::shared_ptr< strategy_type > strategy)
Set the strategy.
Definition csr.hpp:972
void inv_scale(ptr_param< const LinOp > alpha)
Scales the matrix with the inverse of a scalar.
Definition csr.hpp:997
void read(const device_mat_data &data) override
Reads a matrix from a device_matrix_data structure.
index_type * get_srow() noexcept
Returns the starting rows.
Definition csr.hpp:924
static std::unique_ptr< Csr > create(std::shared_ptr< const Executor > exec, std::shared_ptr< strategy_type > strategy)
Creates an uninitialized CSR matrix of the specified size.
size_type get_num_srow_elements() const noexcept
Returns the number of the srow stored elements (involved warps)
Definition csr.hpp:943
std::unique_ptr< LinOp > inverse_permute(const array< IndexType > *inverse_permutation_indices) const override
Returns a LinOp representing the symmetric inverse row and column permutation of the Permutable objec...
std::unique_ptr< LinOp > row_permute(const array< IndexType > *permutation_indices) const override
Returns a LinOp representing the row permutation of the Permutable object.
std::unique_ptr< Csr< ValueType, IndexType > > create_submatrix(const index_set< IndexType > &row_index_set, const index_set< IndexType > &column_index_set) const
Creates a submatrix from this Csr matrix given row and column index_set objects.
std::unique_ptr< Diagonal< ValueType > > extract_diagonal() const override
Extracts the diagonal entries of the matrix into a vector.
static std::unique_ptr< Csr > create(std::shared_ptr< const Executor > exec, const dim< 2 > &size, array< value_type > values, array< index_type > col_idxs, array< index_type > row_ptrs, std::shared_ptr< strategy_type > strategy=nullptr)
Creates a CSR matrix from already allocated (and initialized) row pointer, column index and value arr...
index_type * get_row_ptrs() noexcept
Returns the row pointers of the matrix.
Definition csr.hpp:905
std::unique_ptr< Csr > permute(ptr_param< const Permutation< index_type > > permutation, permute_mode mode=permute_mode::symmetric) const
Creates a permuted copy of this matrix with the given permutation .
static std::unique_ptr< const Csr > create_const(std::shared_ptr< const Executor > exec, const dim< 2 > &size, gko::detail::const_array_view< ValueType > &&values, gko::detail::const_array_view< IndexType > &&col_idxs, gko::detail::const_array_view< IndexType > &&row_ptrs, std::shared_ptr< strategy_type > strategy=nullptr)
Creates a constant (immutable) Csr matrix from a set of constant arrays.
Csr(const Csr &)
Copy-constructs a Csr matrix.
Csr & operator=(Csr &&)
Move-assigns a Csr matrix.
std::unique_ptr< LinOp > transpose() const override
Returns a LinOp representing the transpose of the Transposable object.
const value_type * get_const_values() const noexcept
Returns the values of the matrix.
Definition csr.hpp:876
std::unique_ptr< LinOp > inverse_column_permute(const array< IndexType > *inverse_permutation_indices) const override
Returns a LinOp representing the row permutation of the inverse permuted object.
std::unique_ptr< LinOp > inverse_row_permute(const array< IndexType > *inverse_permutation_indices) const override
Returns a LinOp representing the row permutation of the inverse permuted object.
void compute_absolute_inplace() override
Compute absolute inplace on each element.
size_type get_num_stored_elements() const noexcept
Returns the number of elements explicitly stored in the matrix.
Definition csr.hpp:953
std::shared_ptr< strategy_type > get_strategy() const noexcept
Returns the strategy.
Definition csr.hpp:962
std::unique_ptr< LinOp > permute(const array< IndexType > *permutation_indices) const override
Returns a LinOp representing the symmetric row and column permutation of the Permutable object.
const index_type * get_const_col_idxs() const noexcept
Returns the column indexes of the matrix.
Definition csr.hpp:895
void read(const mat_data &data) override
Reads a matrix from a matrix_data structure.
void sort_by_column_index()
Sorts all (value, col_idx) pairs in each row by column index.
std::unique_ptr< Csr > scale_permute(ptr_param< const ScaledPermutation< value_type, index_type > > row_permutation, ptr_param< const ScaledPermutation< value_type, index_type > > column_permutation, bool invert=false) const
Creates a scaled and permuted copy of this matrix.
void scale(ptr_param< const LinOp > alpha)
Scales the matrix with a scalar.
Definition csr.hpp:984
value_type * get_values() noexcept
Returns the values of the matrix.
Definition csr.hpp:867
index_type * get_col_idxs() noexcept
Returns the column indexes of the matrix.
Definition csr.hpp:886
Csr(Csr &&)
Move-constructs a Csr matrix.
std::unique_ptr< Csr > permute(ptr_param< const Permutation< index_type > > row_permutation, ptr_param< const Permutation< index_type > > column_permutation, bool invert=false) const
Creates a non-symmetrically permuted copy of this matrix with the given row and column permutations...
std::unique_ptr< LinOp > conj_transpose() const override
Returns a LinOp representing the conjugate transpose of the Transposable object.
Dense is a matrix format which explicitly stores all values of the matrix.
Definition sparsity_csr.hpp:25
This class is a utility which efficiently implements the diagonal matrix (a linear operator which sca...
Definition diagonal.hpp:53
ELL is a matrix format where stride with explicit zeros is used such that all rows have the same numb...
Definition ell.hpp:64
Fixed-block compressed sparse row storage matrix format.
Definition sparsity_csr.hpp:29
HYBRID is a matrix format which splits the matrix into ELLPACK and COO format.
Definition hybrid.hpp:55
Permutation is a matrix format that represents a permutation matrix, i.e.
Definition permutation.hpp:112
ScaledPermutation is a matrix combining a permutation with scaling factors.
Definition scaled_permutation.hpp:38
SELL-P is a matrix format similar to ELL format.
Definition sellp.hpp:55
SparsityCsr is a matrix format which stores only the sparsity pattern of a sparse matrix by compressi...
Definition sparsity_csr.hpp:56
This class is used for function parameters in the place of raw pointers.
Definition utils_helper.hpp:41
permute_mode
Specifies how a permutation will be applied to a matrix.
Definition permutation.hpp:42
@ symmetric
The rows and columns will be permuted.
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:260
typename detail::next_precision_impl< T >::type next_precision
Obtains the next type in the singly-linked precision list with half.
Definition math.hpp:438
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:279
constexpr int64 ceildiv(int64 num, int64 den)
Performs integer division with rounding up.
Definition math.hpp:590
std::size_t size_type
Integral type used for allocation quantities.
Definition types.hpp:89
constexpr T min(const T &x, const T &y)
Returns the smaller of the arguments.
Definition math.hpp:719
detail::temporary_clone< detail::pointee< Ptr > > make_temporary_clone(std::shared_ptr< const Executor > exec, Ptr &&ptr)
Creates a temporary_clone.
Definition temporary_clone.hpp:208
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
A span is a lightweight structure used to create sub-ranges from other ranges.
Definition range.hpp:46