Functions
void	init ()
	Create the BLAS context. More...

void	destroy ()
	Destroy the BLAS context. More...

long long	BatchInvertMatrix (void Ainv, void A, const int n, const uint64_t batch, QudaPrecision precision, QudaFieldLocation location)
	Batch inversion the matrix field using an LU decomposition method. More...

long long	stridedBatchGEMM (void A, void B, void *C, QudaBLASParam blas_param, QudaFieldLocation location)
	Strided Batch GEMM. This function performs N GEMM type operations in a strided batched fashion. If the user passes. More...

Detailed Description

The native namespace is where we can deploy target specific blas/lapack operations, using vendor-specific libraries. In the case of CUDA, this corresponds to the use of cuBLAS.

Function Documentation

◆ BatchInvertMatrix()

long long quda::blas_lapack::native::BatchInvertMatrix	(	void *	Ainv,
		void *	A,
		const int	n,
		const uint64_t	batch,
		QudaPrecision	precision,
		QudaFieldLocation	location
	)

Batch inversion the matrix field using an LU decomposition method.

Parameters

[out]	Ainv	Matrix field containing the inverse matrices
[in]	A	Matrix field containing the input matrices
[in]	n	Dimension each matrix
[in]	batch	Problem batch size
[in]	precision	Precision of the input/output data
[in]	Location	of the input/output data

Returns: Number of flops done in this computation

Definition at line 76 of file blas_lapack_cublas.cpp.

◆ destroy()

void quda::blas_lapack::native::destroy ( )

Destroy the BLAS context.

Definition at line 42 of file blas_lapack_cublas.cpp.

◆ init()

void quda::blas_lapack::native::init ( )

Create the BLAS context.

Definition at line 28 of file blas_lapack_cublas.cpp.

◆ stridedBatchGEMM()

long long quda::blas_lapack::native::stridedBatchGEMM	(	void *	A,
		void *	B,
		void *	C,
		QudaBLASParam	blas_param,
		QudaFieldLocation	location
	)

Strided Batch GEMM. This function performs N GEMM type operations in a strided batched fashion. If the user passes.

stride<A,B,C> = -1

it deduces the strides for the A, B, and C arrays from the matrix dimensions, leading dims, etc, and will behave identically to the batched GEMM. If any of the stride<A,B,C> values passed in the parameter structure are greater than or equal to 0, the routine accepts the user's values instead.

Example: If the user passes

a_stride = 0

the routine will use only the first matrix in the A array and compute

C_{n} <- a * A_{0} * B_{n} + b * C_{n}

where n is the batch index.