The cuSparse handler class. More...

Public Member Functions
	CuSparseHandler ()
	Constructor.
	~CuSparseHandler ()
	Destructor.
cusparseHandle_t	handle ()
	Get the cuSPARSE handle.
void	clear ()
	Clear the internal workspace buffers.
cusparsePointerMode_t	setPointerMode (cusparsePointerMode_t mode)
	Set the pointer mode for cuSPARSE operations.
cusparsePointerMode_t	pointerMode ()
	Get the current pointer mode.
template<typename T_Scalar>
void	reserveSpmv (op_t opA, const T_Scalar alpha, const cusparse::SpMatBase &matA, const cusparse::DnVec< T_Scalar > &vecX, const T_Scalar beta, cusparse::DnVec< T_Scalar > &vecY, cusparseSpMVAlg_t alg=cusparseSpMVAlg_t::CUSPARSE_SPMV_CSR_ALG1)
	Reserve workspace memory for sparse matrix-vector multiplication (SpMV).
template<typename T_Scalar>
void	preprocessSpmv (op_t opA, const T_Scalar alpha, const cusparse::SpMatBase &matA, const cusparse::DnVec< T_Scalar > &vecX, const T_Scalar beta, cusparse::DnVec< T_Scalar > &vecY, cusparseSpMVAlg_t alg=cusparseSpMVAlg_t::CUSPARSE_SPMV_CSR_ALG1)
	Preprocess the sparse matrix-vector multiplication operation.
template<typename T_Scalar>
void	performSpmv (op_t opA, const T_Scalar alpha, const cusparse::SpMatBase &matA, const cusparse::DnVec< T_Scalar > &vecX, const T_Scalar beta, cusparse::DnVec< T_Scalar > &vecY, cusparseSpMVAlg_t alg=cusparseSpMVAlg_t::CUSPARSE_SPMV_CSR_ALG1)
	Perform sparse matrix-vector multiplication.
template<typename T_Scalar>
void	reserveSpmm (op_t opA, op_t opB, const T_Scalar alpha, const cusparse::SpMatBase &matA, const cusparse::DnMat< T_Scalar > &matB, const T_Scalar beta, cusparse::DnMat< T_Scalar > &matC, cusparseSpMMAlg_t alg=cusparseSpMMAlg_t::CUSPARSE_SPMM_CSR_ALG1)
	Reserve workspace memory for sparse matrix-matrix multiplication (SpMM).
template<typename T_Scalar>
void	preprocessSpmm (op_t opA, op_t opB, const T_Scalar alpha, const cusparse::SpMatBase &matA, const cusparse::DnMat< T_Scalar > &matB, const T_Scalar beta, cusparse::DnMat< T_Scalar > &matC, cusparseSpMMAlg_t alg=cusparseSpMMAlg_t::CUSPARSE_SPMM_CSR_ALG1)
	Preprocess the sparse matrix-matrix multiplication operation.
template<typename T_Scalar>
void	performSpmm (op_t opA, op_t opB, const T_Scalar alpha, const cusparse::SpMatBase &matA, const cusparse::DnMat< T_Scalar > &matB, const T_Scalar beta, cusparse::DnMat< T_Scalar > &matC, cusparseSpMMAlg_t alg=cusparseSpMMAlg_t::CUSPARSE_SPMM_CSR_ALG1)
	Perform sparse matrix-matrix multiplication.

Detailed Description

The cuSparse handler class.

This class provides a wrapper around the cuSPARSE library for performing sparse linear algebra operations on GPU devices. It manages the cuSPARSE handle and internal workspace buffers required for factorization and solve operations.

Constructor & Destructor Documentation

◆ CuSparseHandler()

culite::CuSparseHandler::CuSparseHandler ( )

Constructor.

Initializes the cuSPARSE handle and internal state.

◆ ~CuSparseHandler()

culite::CuSparseHandler::~CuSparseHandler ( )

Destructor.

Destroys the cuSPARSE handle and releases all allocated resources.

Member Function Documentation

◆ handle()

cusparseHandle_t culite::CuSparseHandler::handle ( )

Get the cuSPARSE handle.

Returns: The cuSPARSE handle.

◆ clear()

void culite::CuSparseHandler::clear ( )

Clear the internal workspace buffers.

Releases all internal workspace memory allocated for cuSPARSE operations.

◆ setPointerMode()

cusparsePointerMode_t culite::CuSparseHandler::setPointerMode ( cusparsePointerMode_t mode )

Set the pointer mode for cuSPARSE operations.

Changes the pointer mode and returns the previous mode. The pointer mode determines whether scalar values (alpha, beta) are passed by reference on the host or device.

Parameters

[in] mode The new pointer mode to set.

Returns: The pointer mode that was active before this call.

◆ pointerMode()

cusparsePointerMode_t culite::CuSparseHandler::pointerMode ( )

Get the current pointer mode.

Returns the current pointer mode setting for cuSPARSE operations.

Returns: The current pointer mode.

◆ reserveSpmv()

template<typename T_Scalar>

void culite::CuSparseHandler::reserveSpmv	(	op_t	opA,
		const T_Scalar *	alpha,
		const cusparse::SpMatBase &	matA,
		const cusparse::DnVec< T_Scalar > &	vecX,
		const T_Scalar *	beta,
		cusparse::DnVec< T_Scalar > &	vecY,
		cusparseSpMVAlg_t	alg = cusparseSpMVAlg_t::CUSPARSE_SPMV_CSR_ALG1 )

Reserve workspace memory for sparse matrix-vector multiplication (SpMV).

Queries the required buffer size for the SpMV operation and allocates the necessary workspace memory on the device.

Template Parameters

T_Scalar The scalar type (e.g., float, double, complex).

Parameters

[in]	opA	Operation to apply to matrix A (e.g., no-transpose, transpose, conjugate-transpose).
[in]	alpha	Pointer to scalar alpha.
[in]	matA	Sparse matrix.
[in]	vecX	Input dense vector X.
[in]	beta	Pointer to scalar beta.
[in,out]	vecY	Dense vector Y for buffer size computation.
[in]	alg	Algorithm to use for SpMV (default: CUSPARSE_SPMV_CSR_ALG1).

◆ preprocessSpmv()

template<typename T_Scalar>

void culite::CuSparseHandler::preprocessSpmv	(	op_t	opA,
		const T_Scalar *	alpha,
		const cusparse::SpMatBase &	matA,
		const cusparse::DnVec< T_Scalar > &	vecX,
		const T_Scalar *	beta,
		cusparse::DnVec< T_Scalar > &	vecY,
		cusparseSpMVAlg_t	alg = cusparseSpMVAlg_t::CUSPARSE_SPMV_CSR_ALG1 )

Preprocess the sparse matrix-vector multiplication operation.

Performs any necessary preprocessing steps for the SpMV operation, which may optimize subsequent repeated operations with the same sparsity pattern.

Template Parameters

T_Scalar The scalar type (e.g., float, double, complex).

Parameters

[in]	opA	Operation to apply to matrix A (e.g., no-transpose, transpose, conjugate-transpose).
[in]	alpha	Pointer to scalar alpha.
[in]	matA	Sparse matrix.
[in]	vecX	Input dense vector X.
[in]	beta	Pointer to scalar beta.
[in,out]	vecY	Dense vector Y for preprocessing.
[in]	alg	Algorithm to use for SpMV (default: CUSPARSE_SPMV_CSR_ALG1).

◆ performSpmv()

template<typename T_Scalar>

void culite::CuSparseHandler::performSpmv	(	op_t	opA,
		const T_Scalar *	alpha,
		const cusparse::SpMatBase &	matA,
		const cusparse::DnVec< T_Scalar > &	vecX,
		const T_Scalar *	beta,
		cusparse::DnVec< T_Scalar > &	vecY,
		cusparseSpMVAlg_t	alg = cusparseSpMVAlg_t::CUSPARSE_SPMV_CSR_ALG1 )

Perform sparse matrix-vector multiplication.

Computes Y = alpha * op(A) * X + beta * Y, where A is a sparse matrix, X and Y are dense vectors, and op(A) is either A, A^T, or A^H.

Template Parameters

T_Scalar The scalar type (e.g., float, double, complex).

Parameters

[in]	opA	Operation to apply to matrix A (e.g., no-transpose, transpose, conjugate-transpose).
[in]	alpha	Pointer to scalar alpha.
[in]	matA	Sparse matrix.
[in]	vecX	Input dense vector X.
[in]	beta	Pointer to scalar beta.
[in,out]	vecY	Output dense vector Y.
[in]	alg	Algorithm to use for SpMV (default: CUSPARSE_SPMV_CSR_ALG1).

◆ reserveSpmm()

template<typename T_Scalar>

void culite::CuSparseHandler::reserveSpmm	(	op_t	opA,
		op_t	opB,
		const T_Scalar *	alpha,
		const cusparse::SpMatBase &	matA,
		const cusparse::DnMat< T_Scalar > &	matB,
		const T_Scalar *	beta,
		cusparse::DnMat< T_Scalar > &	matC,
		cusparseSpMMAlg_t	alg = cusparseSpMMAlg_t::CUSPARSE_SPMM_CSR_ALG1 )

Reserve workspace memory for sparse matrix-matrix multiplication (SpMM).

Queries the required buffer size for the SpMM operation and allocates the necessary workspace memory on the device.

Template Parameters

T_Scalar The scalar type (e.g., float, double, complex).

Parameters

[in]	opA	Operation to apply to matrix A (e.g., no-transpose, transpose, conjugate-transpose).
[in]	opB	Operation to apply to matrix B (e.g., no-transpose, transpose, conjugate-transpose).
[in]	alpha	Pointer to scalar alpha.
[in]	matA	Sparse matrix A.
[in]	matB	Dense matrix B.
[in]	beta	Pointer to scalar beta.
[in,out]	matC	Dense matrix C for buffer size computation.
[in]	alg	Algorithm to use for SpMM (default: CUSPARSE_SPMM_CSR_ALG1).

◆ preprocessSpmm()

template<typename T_Scalar>

void culite::CuSparseHandler::preprocessSpmm	(	op_t	opA,
		op_t	opB,
		const T_Scalar *	alpha,
		const cusparse::SpMatBase &	matA,
		const cusparse::DnMat< T_Scalar > &	matB,
		const T_Scalar *	beta,
		cusparse::DnMat< T_Scalar > &	matC,
		cusparseSpMMAlg_t	alg = cusparseSpMMAlg_t::CUSPARSE_SPMM_CSR_ALG1 )

Preprocess the sparse matrix-matrix multiplication operation.

Performs any necessary preprocessing steps for the SpMM operation, which may optimize subsequent repeated operations with the same sparsity pattern.

Template Parameters

T_Scalar The scalar type (e.g., float, double, complex).

Parameters

[in]	opA	Operation to apply to matrix A (e.g., no-transpose, transpose, conjugate-transpose).
[in]	opB	Operation to apply to matrix B (e.g., no-transpose, transpose, conjugate-transpose).
[in]	alpha	Pointer to scalar alpha.
[in]	matA	Sparse matrix A.
[in]	matB	Dense matrix B.
[in]	beta	Pointer to scalar beta.
[in,out]	matC	Dense matrix C for preprocessing.
[in]	alg	Algorithm to use for SpMM (default: CUSPARSE_SPMM_CSR_ALG1).

◆ performSpmm()

template<typename T_Scalar>

void culite::CuSparseHandler::performSpmm	(	op_t	opA,
		op_t	opB,
		const T_Scalar *	alpha,
		const cusparse::SpMatBase &	matA,
		const cusparse::DnMat< T_Scalar > &	matB,
		const T_Scalar *	beta,
		cusparse::DnMat< T_Scalar > &	matC,
		cusparseSpMMAlg_t	alg = cusparseSpMMAlg_t::CUSPARSE_SPMM_CSR_ALG1 )

Perform sparse matrix-matrix multiplication.

Computes C = alpha * op(A) * op(B) + beta * C, where A is a sparse matrix, B and C are dense matrices, and op(X) is either X, X^T, or X^H.

Template Parameters

T_Scalar The scalar type (e.g., float, double, complex).

Parameters

[in]	opA	Operation to apply to matrix A (e.g., no-transpose, transpose, conjugate-transpose).
[in]	opB	Operation to apply to matrix B (e.g., no-transpose, transpose, conjugate-transpose).
[in]	alpha	Pointer to scalar alpha.
[in]	matA	Sparse matrix A.
[in]	matB	Dense matrix B.
[in]	beta	Pointer to scalar beta.
[in,out]	matC	Output dense matrix C.
[in]	alg	Algorithm to use for SpMM (default: CUSPARSE_SPMM_CSR_ALG1).

Public Member Functions

Detailed Description

Constructor & Destructor Documentation

◆ CuSparseHandler()

◆ ~CuSparseHandler()

Member Function Documentation

◆ handle()

◆ clear()

◆ setPointerMode()

◆ pointerMode()

◆ reserveSpmv()

◆ preprocessSpmv()

◆ performSpmv()

◆ reserveSpmm()

◆ preprocessSpmm()

◆ performSpmm()