inline void trmv( const Order, const UpLo, const Trans, const Diag,
const int n, const float* a, const int lda, float* x,
const int incx ) {
BOOST_STATIC_ASSERT( (is_same<Order, tag::column_major>::value) );
cublasStrmv( blas_option< UpLo >::value, blas_option< Trans >::value,
blas_option< Diag >::value, n, a, lda, x, incx );
}
void magma_strmv(
magma_uplo_t uplo, magma_trans_t trans, magma_diag_t diag,
magma_int_t n,
const float *dA, magma_int_t ldda,
float *dx, magma_int_t incx )
{
cublasStrmv(
cublas_uplo_const( uplo ),
cublas_trans_const( trans ),
cublas_diag_const( diag ),
n,
dA, ldda,
dx, incx );
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing strmv
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gflops, cublas_perf, cublas_time, cpu_perf, cpu_time;
float cublas_error, Cnorm, work[1];
magma_int_t N;
magma_int_t Ak;
magma_int_t sizeA;
magma_int_t lda, ldda;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
float *h_A, *h_x, *h_xcublas;
magmaFloat_ptr d_A, d_x;
float c_neg_one = MAGMA_S_NEG_ONE;
magma_int_t status = 0;
magma_opts opts;
opts.parse_opts( argc, argv );
opts.lapack |= opts.check; // check (-c) implies lapack (-l)
float tol = opts.tolerance * lapackf77_slamch("E");
printf("%% If running lapack (option --lapack), CUBLAS error is computed\n"
"%% relative to CPU BLAS result.\n\n");
printf("%% uplo = %s, transA = %s, diag = %s \n",
lapack_uplo_const(opts.uplo), lapack_trans_const(opts.transA),
lapack_diag_const(opts.diag) );
printf("%% N CUBLAS Gflop/s (ms) CPU Gflop/s (ms) CUBLAS error\n");
printf("%%=================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
N = opts.nsize[itest];
gflops = FLOPS_STRMM(opts.side, N, 1) / 1e9;
lda = N;
Ak = N;
ldda = magma_roundup( lda, opts.align ); // multiple of 32 by default
sizeA = lda*Ak;
TESTING_MALLOC_CPU( h_A, float, lda*Ak );
TESTING_MALLOC_CPU( h_x, float, N );
TESTING_MALLOC_CPU( h_xcublas, float, N );
TESTING_MALLOC_DEV( d_A, float, ldda*Ak );
TESTING_MALLOC_DEV( d_x, float, N );
/* Initialize the matrices */
lapackf77_slarnv( &ione, ISEED, &sizeA, h_A );
lapackf77_slarnv( &ione, ISEED, &N, h_x );
/* =====================================================================
Performs operation using CUBLAS
=================================================================== */
magma_ssetmatrix( Ak, Ak, h_A, lda, d_A, ldda, opts.queue );
magma_ssetvector( N, h_x, 1, d_x, 1, opts.queue );
cublas_time = magma_sync_wtime( opts.queue );
#ifdef HAVE_CUBLAS
cublasStrmv( opts.handle, cublas_uplo_const(opts.uplo), cublas_trans_const(opts.transA),
cublas_diag_const(opts.diag),
N,
d_A, ldda,
d_x, 1 );
#else
magma_strmv( opts.uplo, opts.transA, opts.diag,
N,
d_A, 0, ldda,
d_x, 0, 1, opts.queue );
#endif
cublas_time = magma_sync_wtime( opts.queue ) - cublas_time;
cublas_perf = gflops / cublas_time;
magma_sgetvector( N, d_x, 1, h_xcublas, 1, opts.queue );
/* =====================================================================
Performs operation using CPU BLAS
=================================================================== */
if ( opts.lapack ) {
cpu_time = magma_wtime();
blasf77_strmv( lapack_uplo_const(opts.uplo), lapack_trans_const(opts.transA), lapack_diag_const(opts.diag),
&N,
h_A, &lda,
h_x, &ione );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
}
/* =====================================================================
Check the result
=================================================================== */
if ( opts.lapack ) {
// compute relative error for both magma & cublas, relative to lapack,
// |C_magma - C_lapack| / |C_lapack|
Cnorm = lapackf77_slange( "M", &N, &ione, h_x, &N, work );
blasf77_saxpy( &N, &c_neg_one, h_x, &ione, h_xcublas, &ione );
cublas_error = lapackf77_slange( "M", &N, &ione, h_xcublas, &N, work ) / Cnorm;
printf("%5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %s\n",
(int) N,
cublas_perf, 1000.*cublas_time,
cpu_perf, 1000.*cpu_time,
cublas_error, (cublas_error < tol ? "ok" : "failed"));
status += ! (cublas_error < tol);
}
else {
printf("%5d %7.2f (%7.2f) --- ( --- ) --- ---\n",
(int) N,
cublas_perf, 1000.*cublas_time);
}
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_x );
TESTING_FREE_CPU( h_xcublas );
TESTING_FREE_DEV( d_A );
TESTING_FREE_DEV( d_x );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
opts.cleanup();
TESTING_FINALIZE();
return status;
}
