#include <stdio.h>
#include <stdlib.h>
#include <quda_internal.h>
#include <color_spinor_field.h>
#include <blas_quda.h>
#include <host_utils.h>
#include <command_line_params.h>
#include <dslash_reference.h>
#include <gtest/gtest.h>

Classes
class	BlasTest

Macros
#define	ERROR(a) fabs(blas::norm2(a##D) - blas::norm2(a##H)) / blas::norm2(*a##H)

Typedefs
using	prec_pair_t = std::pair< QudaPrecision, QudaPrecision >

Enumerations
enum class	Kernel { copyHS , copyLS , axpbyz , ax , caxpy , caxpby , cxpaypbz , axpyBzpcx , axpyZpbx , caxpbypzYmbw , cabxpyAx , caxpyXmaz , norm2 , reDotProduct , axpbyzNorm , axpyCGNorm , caxpyNorm , caxpyXmazNormX , cabxpyzAxNorm , cDotProduct , caxpyDotzy , cDotProductNormA , caxpbypzYmbwcDotProductUYNormY , HeavyQuarkResidualNorm , xpyHeavyQuarkResidualNorm , tripleCGReduction , tripleCGUpdate , axpyReDot , caxpyBxpz , caxpyBzpx , axpy_block , caxpy_block , axpyBzpcx_block , reDotProductNorm_block , reDotProduct_block , cDotProductNorm_block , cDotProduct_block , caxpyXmazMR }

Functions
void	setPrec (ColorSpinorParam &param, QudaPrecision precision)

void	display_test_info ()

bool	is_multi (Kernel kernel)

bool	is_copy (Kernel kernel)

bool	is_site_unroll (Kernel kernel)

bool	skip_kernel (prec_pair_t pair, Kernel kernel)

void	initFields (prec_pair_t prec_pair)

void	freeFields ()

double	benchmark (Kernel kernel, const int niter)

double	test (Kernel kernel)

int	main (int argc, char **argv)

prec_pair_t	prec_idx_map (int idx)

	TEST_P (BlasTest, verify)

	TEST_P (BlasTest, benchmark)

std::string	getblasname (testing::TestParamInfo<::testing::tuple< int, int >> param)

	INSTANTIATE_TEST_SUITE_P (QUDA, BlasTest, Combine(Range(0,(Nprec *(Nprec+1))/2), Range(0, Nkernels)), getblasname)

Variables
ColorSpinorField *	xH

ColorSpinorField *	yH

ColorSpinorField *	zH

ColorSpinorField *	wH

ColorSpinorField *	vH

ColorSpinorField *	xoH

ColorSpinorField *	yoH

ColorSpinorField *	zoH

ColorSpinorField *	xD

ColorSpinorField *	yD

ColorSpinorField *	zD

ColorSpinorField *	wD

ColorSpinorField *	vD

ColorSpinorField *	xmD

ColorSpinorField *	ymD

ColorSpinorField *	zmD

ColorSpinorField *	xoD

ColorSpinorField *	yoD

ColorSpinorField *	zoD

ColorSpinorField *	woD

ColorSpinorField *	voD

ColorSpinorField *	xmoD

ColorSpinorField *	ymoD

ColorSpinorField *	zmoD

std::vector< cpuColorSpinorField * >	xmH

std::vector< cpuColorSpinorField * >	ymH

std::vector< cpuColorSpinorField * >	zmH

int	Nspin

int	Ncolor

const std::map< QudaPrecision, std::string >	prec_map

const int	Nprec = prec_map.size()

const std::map< Kernel, std::string >	kernel_map

const int	Nkernels = kernel_map.size()

Macro Definition Documentation

◆ ERROR

#define ERROR ( a ) fabs(blas::norm2(*a##D) - blas::norm2(*a##H)) / blas::norm2(*a##H)

Definition at line 541 of file blas_test.cpp.

Typedef Documentation

◆ prec_pair_t

using prec_pair_t = std::pair<QudaPrecision, QudaPrecision>

Definition at line 65 of file blas_test.cpp.

Enumeration Type Documentation

◆ Kernel

enum Kernel

strong

Enumerator
copyHS
copyLS
axpbyz
ax
caxpy
caxpby
cxpaypbz
axpyBzpcx
axpyZpbx
caxpbypzYmbw
cabxpyAx
caxpyXmaz
norm2
reDotProduct
axpbyzNorm
axpyCGNorm
caxpyNorm
caxpyXmazNormX
cabxpyzAxNorm
cDotProduct
caxpyDotzy
cDotProductNormA
caxpbypzYmbwcDotProductUYNormY
HeavyQuarkResidualNorm
xpyHeavyQuarkResidualNorm
tripleCGReduction
tripleCGUpdate
axpyReDot
caxpyBxpz
caxpyBzpx
axpy_block
caxpy_block
axpyBzpcx_block
reDotProductNorm_block
reDotProduct_block
cDotProductNorm_block
cDotProduct_block
caxpyXmazMR

Definition at line 74 of file blas_test.cpp.

Function Documentation

◆ benchmark()

double benchmark	(	Kernel	kernel,
		const int	niter
	)

Definition at line 349 of file blas_test.cpp.

◆ display_test_info()

void display_test_info ( )

Definition at line 55 of file blas_test.cpp.

◆ freeFields()

void freeFields ( )

Definition at line 315 of file blas_test.cpp.

◆ getblasname()

std::string getblasname ( testing::TestParamInfo<::testing::tuple< int, int >> param )

Definition at line 1144 of file blas_test.cpp.

◆ initFields()

void initFields ( prec_pair_t prec_pair )

Definition at line 201 of file blas_test.cpp.

◆ INSTANTIATE_TEST_SUITE_P()

INSTANTIATE_TEST_SUITE_P	(	QUDA	,
		BlasTest	,
		Combine(Range(0,(Nprec *(Nprec+1))/2), Range(0, Nkernels))	,
		getblasname
	)

◆ is_copy()

bool is_copy ( Kernel kernel )

Definition at line 165 of file blas_test.cpp.

◆ is_multi()

bool is_multi ( Kernel kernel )

Definition at line 160 of file blas_test.cpp.

◆ is_site_unroll()

bool is_site_unroll ( Kernel kernel )

Definition at line 168 of file blas_test.cpp.

◆ main()

int main	(	int	argc,
		char **	argv
	)

Definition at line 1001 of file blas_test.cpp.

◆ prec_idx_map()

prec_pair_t prec_idx_map ( int idx )

Definition at line 1060 of file blas_test.cpp.

◆ setPrec()

void setPrec	(	ColorSpinorParam &	param,
		QudaPrecision	precision
	)

Definition at line 53 of file blas_test.cpp.

◆ skip_kernel()

bool skip_kernel	(	prec_pair_t	pair,
		Kernel	kernel
	)

Definition at line 173 of file blas_test.cpp.

◆ test()

double test ( Kernel kernel )

Definition at line 543 of file blas_test.cpp.

◆ TEST_P() [1/2]

TEST_P	(	BlasTest	,
		benchmark
	)

Definition at line 1121 of file blas_test.cpp.

◆ TEST_P() [2/2]

TEST_P	(	BlasTest	,
		verify
	)

Definition at line 1097 of file blas_test.cpp.

Variable Documentation

◆ kernel_map

const std::map<Kernel, std::string> kernel_map

Definition at line 117 of file blas_test.cpp.

◆ Ncolor

int Ncolor

Definition at line 51 of file blas_test.cpp.

◆ Nkernels

const int Nkernels = kernel_map.size()

Definition at line 157 of file blas_test.cpp.

◆ Nprec

const int Nprec = prec_map.size()

Definition at line 72 of file blas_test.cpp.

◆ Nspin

int Nspin

Definition at line 50 of file blas_test.cpp.

◆ prec_map

const std::map<QudaPrecision, std::string> prec_map

Initial value:

= {{QUDA_QUARTER_PRECISION, "quarter"},
                                                       {QUDA_HALF_PRECISION, "half"},
                                                       {QUDA_SINGLE_PRECISION, "single"},
                                                       {QUDA_DOUBLE_PRECISION, "double"}}

Definition at line 67 of file blas_test.cpp.

◆ vD

ColorSpinorField * vD

Definition at line 35 of file blas_test.cpp.

◆ vH

ColorSpinorField * vH

Definition at line 32 of file blas_test.cpp.

◆ voD

ColorSpinorField * voD

Definition at line 41 of file blas_test.cpp.

◆ wD

ColorSpinorField * wD

Definition at line 35 of file blas_test.cpp.

◆ wH

ColorSpinorField * wH

Definition at line 32 of file blas_test.cpp.

◆ woD

ColorSpinorField * woD

Definition at line 41 of file blas_test.cpp.

◆ xD

ColorSpinorField* xD

Definition at line 35 of file blas_test.cpp.

◆ xH

ColorSpinorField* xH

This is the blas_test for checking correctness of the blas and reduction functions used by the lienar solvers. Any blas kernels that are added to QUDA should have a test added here.

For kernels that have mixed-precision support, we can test this functionality as well. The number of precision cominations we test are Nprec * (Nprec + 1)/2, which corresponds to all uni-precision combinations ("this precision") and all combinations with the "other precision" > "this precision".

Definition at line 32 of file blas_test.cpp.