nl_cuda.h File Reference

Internal OpenNL functions that interface with CUDA. More...

#include "nl_private.h"
#include "nl_matrix.h"
#include "nl_blas.h"

Go to the source code of this file.

Functions
NLboolean	nlInitExtension_CUDA (void)
	Initializes the CUDA extension.
NLboolean	nlExtensionIsInitialized_CUDA (void)
	Tests whether the CUDA extension is initialized.
NLMatrix	nlCUDAMatrixNewFromCRSMatrix (NLMatrix M)
	Creates a CUDA on-GPU matrix from an OpenNL CRS matrix.
NLMatrix	nlCUDAJacobiPreconditionerNewFromCRSMatrix (NLMatrix M)
	Creates a CUDA on-GPU Jacobi preconditioner from an OpenNL CRS matrix.
void	nlCUDAMatrixSpMV (NLMatrix M, const double *x, double *y, double alpha, double beta)
	Computes a sparse matrix vector product.
NLBlas_t	nlCUDABlas (void)
	Gets a pointer to the BLAS abstraction layer for BLAS operation on the GPU using CUDA.

Detailed Description

Internal OpenNL functions that interface with CUDA.

Definition in file nl_cuda.h.

Function Documentation

nlCUDABlas()

NLBlas_t nlCUDABlas

(

void

)

Gets a pointer to the BLAS abstraction layer for BLAS operation on the GPU using CUDA.

Returns

a pointer to the BLAS abstraction layer.

nlCUDAJacobiPreconditionerNewFromCRSMatrix()

NLMatrix nlCUDAJacobiPreconditionerNewFromCRSMatrix

(

NLMatrix

M

)

Creates a CUDA on-GPU Jacobi preconditioner from an OpenNL CRS matrix.

Calling nlMultMatrixVector() with the created matrix only works with vectors that reside on the GPU.

Parameters

[in]

M

the OpenNL CRS matrix.

Returns

a handle to the created CUDA matrix.

nlCUDAMatrixNewFromCRSMatrix()

NLMatrix nlCUDAMatrixNewFromCRSMatrix

(

NLMatrix

M

)

Creates a CUDA on-GPU matrix from an OpenNL CRS matrix.

Calling nlMultMatrixVector() with the created matrix only works with vectors that reside on the GPU.

Parameters

[in]

M

the OpenNL CRS matrix to be copied.

Returns

a handle to the CUDA matrix.

nlCUDAMatrixSpMV()

void nlCUDAMatrixSpMV	(	NLMatrix	M,
		const double *	x,
		double *	y,
		double	alpha,
		double	beta
	)

Computes a sparse matrix vector product.

Computes \( y \leftarrow alpha M x + beta y \) As compared to NL abstract matrix API, it has the alpha and beta parameters.

Parameters

[in]	M	a matrix created from nlCUDAMAtrixNewFromCRSMatrix()
[in]	x	device pointer
[in,out]	y	device pointer
[in]	alpha,beta	two scalars

nlExtensionIsInitialized_CUDA()

NLboolean nlExtensionIsInitialized_CUDA

(

void

)

Tests whether the CUDA extension is initialized.

Return values

NL_TRUE	if the CUDA extension is initialized.
NL_FALSE	otherwise.

nlInitExtension_CUDA()

NLboolean nlInitExtension_CUDA

(

void

)

Initializes the CUDA extension.

This dynamically loads the CUDA libraries available in the system (if available) and retrieves the symbols in there.

Return values

NL_TRUE	if CUDA could be successfully dynamically loaded and all functions could be found in it.
NL_FALSE	otherwise.

Note

For now, only implemented under Linux in dynamic libraries mode.