The GpuSparseMatrixWrapper Checks CUDA/HIP version and dispatches a version either using the old or the generic CUDA API. More...

#include <GpuSparseMatrixWrapper.hpp>

Inheritance diagram for Opm::gpuistl::GpuSparseMatrixWrapper< T, ForceLegacy >:

Public Types
using	field_type = T

using	matrix_type = GpuSparseMatrix< T >

Public Member Functions
matrix_type *	operator-> ()

const matrix_type *	operator-> () const

	GpuSparseMatrixWrapper (const T nonZeroElements, const int rowIndices, const int *columnIndices, std::size_t numberOfNonzeroBlocks, std::size_t blockSize, std::size_t numberOfRows)

	GpuSparseMatrixWrapper (const GpuVector< int > &rowIndices, const GpuVector< int > &columnIndices, std::size_t blockSize)

	GpuSparseMatrixWrapper (const GpuSparseMatrixWrapper &other)

GpuSparseMatrixWrapper &	operator= (const GpuSparseMatrixWrapper &)=delete

virtual	~GpuSparseMatrixWrapper ()=default

	GpuSparseMatrixWrapper ()=default

const matrix_type &	get () const

template<class M = matrix_type, typename = std::enable_if_t<std::is_same_v<M, GpuSparseMatrix<T>>>>
void	setUpperTriangular ()

template<class M = matrix_type, typename = std::enable_if_t<std::is_same_v<M, GpuSparseMatrix<T>>>>
void	setLowerTriangular ()
	setLowerTriangular sets the CuSparse flag that this is an lower diagonal (with non-unit diagonal) matrix. More...

template<class M = matrix_type, typename = std::enable_if_t<std::is_same_v<M, GpuSparseMatrix<T>>>>
void	setUnitDiagonal ()
	setUnitDiagonal sets the CuSparse flag that this has unit diagional. More...

template<class M = matrix_type, typename = std::enable_if_t<std::is_same_v<M, GpuSparseMatrix<T>>>>
void	setNonUnitDiagonal ()
	setNonUnitDiagonal sets the CuSparse flag that this has non-unit diagional. More...

std::size_t	N () const
	N returns the number of rows (which is equal to the number of columns) More...

std::size_t	nonzeroes () const
	nonzeroes behaves as the Dune::BCRSMatrix::nonzeros() function and returns the number of non zero blocks More...

GpuVector< T > &	getNonZeroValues ()
	getNonZeroValues returns the GPU vector containing the non-zero values (ordered by block) More...

const GpuVector< T > &	getNonZeroValues () const
	getNonZeroValues returns the GPU vector containing the non-zero values (ordered by block) More...

GpuVector< int > &	getRowIndices ()
	getRowIndices returns the row indices used to represent the BSR structure. More...

const GpuVector< int > &	getRowIndices () const
	getRowIndices returns the row indices used to represent the BSR structure. More...

GpuVector< int > &	getColumnIndices ()
	getColumnIndices returns the column indices used to represent the BSR structure. More...

const GpuVector< int > &	getColumnIndices () const
	getColumnIndices returns the column indices used to represent the BSR structure. More...

std::size_t	dim () const
	dim returns the dimension of the vector space on which this matrix acts More...

std::size_t	blockSize () const
	blockSize size of the blocks More...

detail::GpuSparseMatrixDescription &	getDescription ()
	getDescription the cusparse matrix description. More...

virtual void	mv (const GpuVector< T > &x, GpuVector< T > &y) const
	mv performs matrix vector multiply y = Ax More...

virtual void	umv (const GpuVector< T > &x, GpuVector< T > &y) const
	umv computes y=Ax+y More...

virtual void	usmv (T alpha, const GpuVector< T > &x, GpuVector< T > &y) const
	umv computes y=alpha * Ax + y More...

template<class MatrixType >
void	updateNonzeroValues (const MatrixType &matrix, bool copyNonZeroElementsDirectly=false)
	updateNonzeroValues updates the non-zero values by using the non-zero values of the supplied matrix More...

template<bool OtherForceLegacy>
void	updateNonzeroValues (const GpuSparseMatrixWrapper< T, OtherForceLegacy > &matrix)
	updateNonzeroValues updates the non-zero values by using the non-zero values of the supplied matrix More...

template<class FunctionType >
auto	dispatchOnBlocksize (FunctionType function) const
	Dispatches a function based on the block size of the matrix. More...

Static Public Member Functions
template<class MatrixType >
static GpuSparseMatrixWrapper< T, ForceLegacy >	fromMatrix (const MatrixType &matrix, bool copyNonZeroElementsDirectly=false)
	fromMatrix creates a new matrix with the same block size and values as the given matrix More...

Static Public Attributes
static constexpr int	max_block_size = 6
	Maximum block size supported by this implementation. More...

Detailed Description

template<typename T, bool ForceLegacy = false>
class Opm::gpuistl::GpuSparseMatrixWrapper< T, ForceLegacy >

The GpuSparseMatrixWrapper Checks CUDA/HIP version and dispatches a version either using the old or the generic CUDA API.

Note: we currently only support simple raw primitives for T (double and float). Block size is handled through the block size parameter

Template Parameters

T	the type to store. Can be either float, double or int.

Note: we only support square matrices.; We only support Block Compressed Sparse Row Format (BSR) for now.; This class uses the legacy cuSPARSE API, to be compatible with CuSparse's ilu0 preconditioner. However, this preconditioner is deprecated and will be removed in future versions of CuSparse. So we should migrate to the new cuSPARSE generic API in the future.; To also support block size 1, we use the GpuSparseMatrixGeneric class which uses the new cuSPARSE generic API. This is a temporary solution, and we should migrate to the new API for all block sizes in the future by replacing this class with GpuSparseMatrixGeneric.

Member Typedef Documentation

◆ field_type

template<typename T , bool ForceLegacy = false>

using Opm::gpuistl::GpuSparseMatrixWrapper< T, ForceLegacy >::field_type = T

◆ matrix_type

template<typename T , bool ForceLegacy = false>

using Opm::gpuistl::GpuSparseMatrixWrapper< T, ForceLegacy >::matrix_type = GpuSparseMatrix<T>

Constructor & Destructor Documentation

◆ GpuSparseMatrixWrapper() [1/4]

template<typename T , bool ForceLegacy = false>

Opm::gpuistl::GpuSparseMatrixWrapper< T, ForceLegacy >::GpuSparseMatrixWrapper	(	const T *	nonZeroElements,
		const int *	rowIndices,
		const int *	columnIndices,
		std::size_t	numberOfNonzeroBlocks,
		std::size_t	blockSize,
		std::size_t	numberOfRows
	)

inline

Create the sparse matrix specified by the raw data.

Note: Prefer to use the constructor taking a const reference to a matrix instead.

Parameters

[in]	nonZeroElements	the non-zero values of the matrix
[in]	rowIndices	the row indices of the non-zero elements
[in]	columnIndices	the column indices of the non-zero elements
[in]	numberOfNonzeroBlocks	number of nonzero blocks
[in]	blockSize	size of each block matrix (typically 3)
[in]	numberOfRows	the number of rows

Note: We assume numberOfNonzeroBlocks, blockSize and numberOfRows all are representable as int due to restrictions in the current version of cusparse. This might change in future versions.

References Opm::gpuistl::GpuSparseMatrixWrapper< T, ForceLegacy >::blockSize().

◆ GpuSparseMatrixWrapper() [2/4]

template<typename T , bool ForceLegacy = false>

Opm::gpuistl::GpuSparseMatrixWrapper< T, ForceLegacy >::GpuSparseMatrixWrapper	(	const GpuVector< int > &	rowIndices,
		const GpuVector< int > &	columnIndices,
		std::size_t	blockSize
	)

inline

Create a sparse matrix by copying the sparsity structure of another matrix, not filling in the values

Note: Prefer to use the constructor taking a const reference to a matrix instead.

Parameters

[in]	rowIndices	the row indices of the non-zero elements
[in]	columnIndices	the column indices of the non-zero elements
[in]	blockSize	size of each block matrix (typically 3)

Note: We assume numberOfNonzeroBlocks, blockSize and numberOfRows all are representable as int due to restrictions in the current version of cusparse. This might change in future versions.

References Opm::gpuistl::GpuSparseMatrixWrapper< T, ForceLegacy >::blockSize().

◆ GpuSparseMatrixWrapper() [3/4]

template<typename T , bool ForceLegacy = false>

Opm::gpuistl::GpuSparseMatrixWrapper< T, ForceLegacy >::GpuSparseMatrixWrapper ( const GpuSparseMatrixWrapper< T, ForceLegacy > & other )

inline

◆ ~GpuSparseMatrixWrapper()

template<typename T , bool ForceLegacy = false>

virtual Opm::gpuistl::GpuSparseMatrixWrapper< T, ForceLegacy >::~GpuSparseMatrixWrapper ( )

virtualdefault

◆ GpuSparseMatrixWrapper() [4/4]

template<typename T , bool ForceLegacy = false>

Opm::gpuistl::GpuSparseMatrixWrapper< T, ForceLegacy >::GpuSparseMatrixWrapper ( )

default

Member Function Documentation

◆ blockSize()

template<typename T , bool ForceLegacy = false>

std::size_t Opm::gpuistl::GpuSparseMatrixWrapper< T, ForceLegacy >::blockSize ( ) const

inline

blockSize size of the blocks

Referenced by Opm::gpuistl::GpuSeqILU0< M, X, Y, l >::GpuSeqILU0(), Opm::gpuistl::GpuSparseMatrixWrapper< T, ForceLegacy >::GpuSparseMatrixWrapper(), and Opm::gpuistl::AmgxInterface::updateAmgxMatrixFromGpuSparseMatrix().

◆ dim()

template<typename T , bool ForceLegacy = false>

std::size_t Opm::gpuistl::GpuSparseMatrixWrapper< T, ForceLegacy >::dim ( ) const

inline

dim returns the dimension of the vector space on which this matrix acts

This is equivalent to matrix.N() * matrix.blockSize()

Returns: matrix.N() * matrix.blockSize()

Referenced by Opm::gpuistl::GpuSeqILU0< M, X, Y, l >::GpuSeqILU0().

◆ dispatchOnBlocksize()

template<typename T , bool ForceLegacy = false>

template<class FunctionType >

auto Opm::gpuistl::GpuSparseMatrixWrapper< T, ForceLegacy >::dispatchOnBlocksize ( FunctionType function ) const

inline

Dispatches a function based on the block size of the matrix.

This method allows executing different code paths depending on the block size of the matrix, up to the maximum block size specified by max_block_size.

Use this function if you need the block size to be known at compile time.

Template Parameters

FunctionType Type of the function to be dispatched

Parameters

function The function to be executed based on the block size

Returns: The result of the function execution

You can use this function as

matrix.dispatchOnBlocksize([](auto val) {
   constexpr int blockSize = decltype(val)::value;
});

◆ fromMatrix()

template<typename T , bool ForceLegacy = false>

template<class MatrixType >

static GpuSparseMatrixWrapper< T, ForceLegacy > Opm::gpuistl::GpuSparseMatrixWrapper< T, ForceLegacy >::fromMatrix	(	const MatrixType &	matrix,
		bool	copyNonZeroElementsDirectly = `false`
	)

inlinestatic

fromMatrix creates a new matrix with the same block size and values as the given matrix

Parameters

matrix	the matrix to copy from
copyNonZeroElementsDirectly	if true will do a memcpy from matrix[0][0][0][0], otherwise will build up the non-zero elements by looping over the matrix. Note that setting this to true will yield a performance increase, but might not always yield correct results depending on how the matrix has been initialized. If unsure, leave it as false.

Template Parameters

MatrixType is assumed to be a Dune::BCRSMatrix compatible matrix.

References Opm::gpuistl::GpuSparseMatrix< T >::fromMatrix().

◆ get()

template<typename T , bool ForceLegacy = false>

const matrix_type & Opm::gpuistl::GpuSparseMatrixWrapper< T, ForceLegacy >::get ( ) const

inline

Referenced by Opm::gpuistl::GpuSparseMatrixWrapper< T, ForceLegacy >::updateNonzeroValues().

◆ getColumnIndices() [1/2]

template<typename T , bool ForceLegacy = false>

GpuVector< int > & Opm::gpuistl::GpuSparseMatrixWrapper< T, ForceLegacy >::getColumnIndices ( )

inline

getColumnIndices returns the column indices used to represent the BSR structure.

Returns: Read the CuSPARSE documentation on Block Compressed Sparse Row Format (BSR) for the exact ordering.

Referenced by Opm::gpuistl::AmgxInterface::updateAmgxMatrixFromGpuSparseMatrix().

◆ getColumnIndices() [2/2]

template<typename T , bool ForceLegacy = false>

const GpuVector< int > & Opm::gpuistl::GpuSparseMatrixWrapper< T, ForceLegacy >::getColumnIndices ( ) const

inline

getColumnIndices returns the column indices used to represent the BSR structure.

Returns: Read the CuSPARSE documentation on Block Compressed Sparse Row Format (BSR) for the exact ordering.

◆ getDescription()

template<typename T , bool ForceLegacy = false>

detail::GpuSparseMatrixDescription & Opm::gpuistl::GpuSparseMatrixWrapper< T, ForceLegacy >::getDescription ( )

inline

getDescription the cusparse matrix description.

This description is needed for most calls to the CuSparse library

◆ getNonZeroValues() [1/2]

template<typename T , bool ForceLegacy = false>

GpuVector< T > & Opm::gpuistl::GpuSparseMatrixWrapper< T, ForceLegacy >::getNonZeroValues ( )

inline

getNonZeroValues returns the GPU vector containing the non-zero values (ordered by block)

Note: Read the CuSPARSE documentation on Block Compressed Sparse Row Format (BSR) for the exact ordering.

Referenced by Opm::gpuistl::AmgxInterface::updateAmgxMatrixCoefficientsFromGpuSparseMatrix(), Opm::gpuistl::AmgxInterface::updateAmgxMatrixFromGpuSparseMatrix(), and Opm::gpuistl::AmgxInterface::updateGpuSparseMatrixFromAmgxMatrix().

◆ getNonZeroValues() [2/2]

template<typename T , bool ForceLegacy = false>

const GpuVector< T > & Opm::gpuistl::GpuSparseMatrixWrapper< T, ForceLegacy >::getNonZeroValues ( ) const

inline

getNonZeroValues returns the GPU vector containing the non-zero values (ordered by block)

Note: Read the CuSPARSE documentation on Block Compressed Sparse Row Format (BSR) for the exact ordering.

◆ getRowIndices() [1/2]

template<typename T , bool ForceLegacy = false>

GpuVector< int > & Opm::gpuistl::GpuSparseMatrixWrapper< T, ForceLegacy >::getRowIndices ( )

inline

getRowIndices returns the row indices used to represent the BSR structure.

Note: Read the CuSPARSE documentation on Block Compressed Sparse Row Format (BSR) for the exact ordering.

Referenced by Opm::gpuistl::AmgxInterface::updateAmgxMatrixFromGpuSparseMatrix().

◆ getRowIndices() [2/2]

template<typename T , bool ForceLegacy = false>

const GpuVector< int > & Opm::gpuistl::GpuSparseMatrixWrapper< T, ForceLegacy >::getRowIndices ( ) const

inline

getRowIndices returns the row indices used to represent the BSR structure.

Note: Read the CuSPARSE documentation on Block Compressed Sparse Row Format (BSR) for the exact ordering.

◆ mv()

template<typename T , bool ForceLegacy = false>

virtual void Opm::gpuistl::GpuSparseMatrixWrapper< T, ForceLegacy >::mv	(	const GpuVector< T > &	x,
		GpuVector< T > &	y
	)		const

inlinevirtual

mv performs matrix vector multiply y = Ax

Parameters

[in]	x	the vector to multiply the matrix with
[out]	y	the output vector

◆ N()

template<typename T , bool ForceLegacy = false>

std::size_t Opm::gpuistl::GpuSparseMatrixWrapper< T, ForceLegacy >::N ( ) const

inline

N returns the number of rows (which is equal to the number of columns)

Referenced by Opm::gpuistl::GpuSeqILU0< M, X, Y, l >::GpuSeqILU0(), Opm::gpuistl::AmgxInterface::updateAmgxMatrixCoefficientsFromGpuSparseMatrix(), and Opm::gpuistl::AmgxInterface::updateAmgxMatrixFromGpuSparseMatrix().

◆ nonzeroes()

template<typename T , bool ForceLegacy = false>

std::size_t Opm::gpuistl::GpuSparseMatrixWrapper< T, ForceLegacy >::nonzeroes ( ) const

inline

nonzeroes behaves as the Dune::BCRSMatrix::nonzeros() function and returns the number of non zero blocks

Returns: number of non zero blocks.

Referenced by Opm::gpuistl::GpuSeqILU0< M, X, Y, l >::GpuSeqILU0(), Opm::gpuistl::AmgxInterface::updateAmgxMatrixCoefficientsFromGpuSparseMatrix(), and Opm::gpuistl::AmgxInterface::updateAmgxMatrixFromGpuSparseMatrix().

◆ operator->() [1/2]

template<typename T , bool ForceLegacy = false>

matrix_type * Opm::gpuistl::GpuSparseMatrixWrapper< T, ForceLegacy >::operator-> ( )

inline

◆ operator->() [2/2]

template<typename T , bool ForceLegacy = false>

const matrix_type * Opm::gpuistl::GpuSparseMatrixWrapper< T, ForceLegacy >::operator-> ( ) const

inline

◆ operator=()

template<typename T , bool ForceLegacy = false>

GpuSparseMatrixWrapper & Opm::gpuistl::GpuSparseMatrixWrapper< T, ForceLegacy >::operator= ( const GpuSparseMatrixWrapper< T, ForceLegacy > & )

delete

◆ setLowerTriangular()

template<typename T , bool ForceLegacy = false>

template<class M = matrix_type, typename = std::enable_if_t<std::is_same_v<M, GpuSparseMatrix<T>>>>

void Opm::gpuistl::GpuSparseMatrixWrapper< T, ForceLegacy >::setLowerTriangular ( )

inline

setLowerTriangular sets the CuSparse flag that this is an lower diagonal (with non-unit diagonal) matrix.

◆ setNonUnitDiagonal()

template<typename T , bool ForceLegacy = false>

template<class M = matrix_type, typename = std::enable_if_t<std::is_same_v<M, GpuSparseMatrix<T>>>>

void Opm::gpuistl::GpuSparseMatrixWrapper< T, ForceLegacy >::setNonUnitDiagonal ( )

inline

setNonUnitDiagonal sets the CuSparse flag that this has non-unit diagional.

◆ setUnitDiagonal()

template<typename T , bool ForceLegacy = false>

template<class M = matrix_type, typename = std::enable_if_t<std::is_same_v<M, GpuSparseMatrix<T>>>>

void Opm::gpuistl::GpuSparseMatrixWrapper< T, ForceLegacy >::setUnitDiagonal ( )

inline

setUnitDiagonal sets the CuSparse flag that this has unit diagional.

◆ setUpperTriangular()

template<typename T , bool ForceLegacy = false>

template<class M = matrix_type, typename = std::enable_if_t<std::is_same_v<M, GpuSparseMatrix<T>>>>

void Opm::gpuistl::GpuSparseMatrixWrapper< T, ForceLegacy >::setUpperTriangular ( )

inline

◆ umv()

template<typename T , bool ForceLegacy = false>

virtual void Opm::gpuistl::GpuSparseMatrixWrapper< T, ForceLegacy >::umv	(	const GpuVector< T > &	x,
		GpuVector< T > &	y
	)		const

inlinevirtual

umv computes y=Ax+y

Parameters

[in]	x	the vector to multiply with A
[in,out]	y	the vector to add and store the output in

◆ updateNonzeroValues() [1/2]

template<typename T , bool ForceLegacy = false>

template<bool OtherForceLegacy>

void Opm::gpuistl::GpuSparseMatrixWrapper< T, ForceLegacy >::updateNonzeroValues ( const GpuSparseMatrixWrapper< T, OtherForceLegacy > & matrix )

inline

updateNonzeroValues updates the non-zero values by using the non-zero values of the supplied matrix

Parameters

matrix the matrix to extract the non-zero values from

Note: This assumes the given matrix has the same sparsity pattern.

References Opm::gpuistl::GpuSparseMatrixWrapper< T, ForceLegacy >::get().

◆ updateNonzeroValues() [2/2]

template<typename T , bool ForceLegacy = false>

template<class MatrixType >

void Opm::gpuistl::GpuSparseMatrixWrapper< T, ForceLegacy >::updateNonzeroValues	(	const MatrixType &	matrix,
		bool	copyNonZeroElementsDirectly = `false`
	)

inline

updateNonzeroValues updates the non-zero values by using the non-zero values of the supplied matrix

Parameters

matrix	the matrix to extract the non-zero values from
copyNonZeroElementsDirectly	if true will do a memcpy from matrix[0][0][0][0], otherwise will build up the non-zero elements by looping over the matrix. Note that setting this to true will yield a performance increase, but might not always yield correct results depending on how the matrix matrix has been initialized. If unsure, leave it as false.

Note: This assumes the given matrix has the same sparsity pattern.

Template Parameters

MatrixType is assumed to be a Dune::BCRSMatrix compatible matrix.

Referenced by Opm::gpuistl::SolverAdapter< Operator, UnderlyingSolver, X >::apply(), and Opm::gpuistl::PreconditionerCPUMatrixToGPUMatrix< X, Y, CudaPreconditionerType, CPUMatrixType >::update().

◆ usmv()

template<typename T , bool ForceLegacy = false>

virtual void Opm::gpuistl::GpuSparseMatrixWrapper< T, ForceLegacy >::usmv	(	T	alpha,
		const GpuVector< T > &	x,
		GpuVector< T > &	y
	)		const

inlinevirtual

umv computes y=alpha * Ax + y

Parameters

[in]	alpha	Scaling factor
[in]	x	the vector to multiply with A
[in,out]	y	the vector to add and store the output in

Member Data Documentation

◆ max_block_size

template<typename T , bool ForceLegacy = false>

constexpr int Opm::gpuistl::GpuSparseMatrixWrapper< T, ForceLegacy >::max_block_size = 6

staticconstexpr

Maximum block size supported by this implementation.

This constant defines an upper bound on the block size to ensure reasonable compilation times. While this class itself could support larger values, functions that call dispatchOnBlocksize() might have limitations. This value can be increased if needed, but will increase compilation time due to template instantiations.

The documentation for this class was generated from the following files:

Public Types

Public Member Functions

Static Public Member Functions

Static Public Attributes

Detailed Description

Member Typedef Documentation

◆ field_type

◆ matrix_type

Constructor & Destructor Documentation

◆ GpuSparseMatrixWrapper() [1/4]

◆ GpuSparseMatrixWrapper() [2/4]

◆ GpuSparseMatrixWrapper() [3/4]

◆ ~GpuSparseMatrixWrapper()

◆ GpuSparseMatrixWrapper() [4/4]

Member Function Documentation

◆ blockSize()

◆ dim()

◆ dispatchOnBlocksize()

◆ fromMatrix()

◆ get()

◆ getColumnIndices() [1/2]

◆ getColumnIndices() [2/2]

◆ getDescription()

◆ getNonZeroValues() [1/2]

◆ getNonZeroValues() [2/2]

◆ getRowIndices() [1/2]

◆ getRowIndices() [2/2]

◆ mv()

◆ N()

◆ nonzeroes()

◆ operator->() [1/2]

◆ operator->() [2/2]

◆ operator=()

◆ setLowerTriangular()

◆ setNonUnitDiagonal()

◆ setUnitDiagonal()

◆ setUpperTriangular()

◆ umv()

◆ updateNonzeroValues() [1/2]

◆ updateNonzeroValues() [2/2]

◆ usmv()

Member Data Documentation

◆ max_block_size