void magma_chemv(
magma_uplo_t uplo,
magma_int_t n,
magmaFloatComplex alpha,
const magmaFloatComplex *dA, magma_int_t ldda,
const magmaFloatComplex *dx, magma_int_t incx,
magmaFloatComplex beta,
magmaFloatComplex *dy, magma_int_t incy )
{
cublasChemv(
cublas_uplo_const( uplo ),
n,
alpha, dA, ldda,
dx, incx,
beta, dy, incy );
}
int main(int argc, char **argv)
{
TESTING_INIT();
real_Double_t gflops, magma_perf, magma_time, cublas_perf, cublas_time, cpu_perf, cpu_time;
float magma_error, cublas_error, work[1];
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t N, lda, sizeA, sizeX, sizeY, blocks, ldwork;
magma_int_t incx = 1;
magma_int_t incy = 1;
magma_int_t nb = 64;
magmaFloatComplex c_neg_one = MAGMA_C_NEG_ONE;
magmaFloatComplex alpha = MAGMA_C_MAKE( 1.5, -2.3 );
magmaFloatComplex beta = MAGMA_C_MAKE( -0.6, 0.8 );
magmaFloatComplex *A, *X, *Y, *Ycublas, *Ymagma;
magmaFloatComplex *dA, *dX, *dY, *dC_work;
magma_opts opts;
parse_opts( argc, argv, &opts );
printf(" N MAGMA Gflop/s (ms) CUBLAS Gflop/s (ms) CPU Gflop/s (ms) MAGMA error CUBLAS error\n");
printf("=============================================================================================\n");
for( int i = 0; i < opts.ntest; ++i ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
N = opts.nsize[i];
lda = ((N + 31)/32)*32;
sizeA = N*lda;
sizeX = N*incx;
sizeY = N*incy;
gflops = FLOPS_CHEMV( N ) / 1e9;
TESTING_MALLOC( A, magmaFloatComplex, sizeA );
TESTING_MALLOC( X, magmaFloatComplex, sizeX );
TESTING_MALLOC( Y, magmaFloatComplex, sizeY );
TESTING_MALLOC( Ycublas, magmaFloatComplex, sizeY );
TESTING_MALLOC( Ymagma, magmaFloatComplex, sizeY );
TESTING_DEVALLOC( dA, magmaFloatComplex, sizeA );
TESTING_DEVALLOC( dX, magmaFloatComplex, sizeX );
TESTING_DEVALLOC( dY, magmaFloatComplex, sizeY );
blocks = (N + nb - 1) / nb;
ldwork = lda * (blocks + 1);
TESTING_DEVALLOC( dC_work, magmaFloatComplex, ldwork );
/* Initialize the matrix */
lapackf77_clarnv( &ione, ISEED, &sizeA, A );
magma_cmake_hermitian( N, A, lda );
lapackf77_clarnv( &ione, ISEED, &sizeX, X );
lapackf77_clarnv( &ione, ISEED, &sizeY, Y );
/* =====================================================================
Performs operation using CUBLAS
=================================================================== */
magma_csetmatrix( N, N, A, lda, dA, lda );
magma_csetvector( N, X, incx, dX, incx );
magma_csetvector( N, Y, incy, dY, incy );
cublas_time = magma_sync_wtime( 0 );
cublasChemv( opts.uplo, N, alpha, dA, lda, dX, incx, beta, dY, incy );
cublas_time = magma_sync_wtime( 0 ) - cublas_time;
cublas_perf = gflops / cublas_time;
magma_cgetvector( N, dY, incy, Ycublas, incy );
/* =====================================================================
Performs operation using MAGMA BLAS
=================================================================== */
magma_csetvector( N, Y, incy, dY, incy );
magma_time = magma_sync_wtime( 0 );
#if (GPUSHMEM >= 200)
magmablas_chemv2( opts.uplo, N, alpha, dA, lda, dX, incx, beta, dY, incy, dC_work, ldwork );
#else
magmablas_chemv( opts.uplo, N, alpha, dA, lda, dX, incx, beta, dY, incy );
#endif
magma_time = magma_sync_wtime( 0 ) - magma_time;
magma_perf = gflops / magma_time;
magma_cgetvector( N, dY, incy, Ymagma, incy );
/* =====================================================================
Performs operation using CPU BLAS
=================================================================== */
cpu_time = magma_wtime();
blasf77_chemv( &opts.uplo, &N, &alpha, A, &lda, X, &incx, &beta, Y, &incy );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
/* =====================================================================
Check the result
=================================================================== */
blasf77_caxpy( &N, &c_neg_one, Y, &incy, Ymagma, &incy );
magma_error = lapackf77_clange( "M", &N, &ione, Ymagma, &N, work ) / N;
blasf77_caxpy( &N, &c_neg_one, Y, &incy, Ycublas, &incy );
cublas_error = lapackf77_clange( "M", &N, &ione, Ycublas, &N, work ) / N;
printf("%5d %7.2f (%7.2f) %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %8.2e\n",
(int) N,
magma_perf, 1000.*magma_time,
cublas_perf, 1000.*cublas_time,
cpu_perf, 1000.*cpu_time,
magma_error, cublas_error );
TESTING_FREE( A );
TESTING_FREE( X );
TESTING_FREE( Y );
TESTING_FREE( Ycublas );
TESTING_FREE( Ymagma );
TESTING_DEVFREE( dA );
TESTING_DEVFREE( dX );
TESTING_DEVFREE( dY );
TESTING_DEVFREE( dC_work );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return 0;
}
int main(int argc, char **argv)
{
TESTING_INIT();
magma_setdevice(0);
magma_timestr_t start, end;
float flops, magma_perf, cuda_perf, error, work[1];
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
magmaFloatComplex c_neg_one = MAGMA_C_NEG_ONE;
magma_int_t n_local[4];
FILE *fp ;
magma_int_t N, m, i, j, lda, LDA, M;
magma_int_t matsize;
magma_int_t vecsize;
magma_int_t istart = 64;
magma_int_t incx = 1;
char uplo = MagmaLower;
magmaFloatComplex alpha = MAGMA_C_MAKE(1., 0.); // MAGMA_C_MAKE( 1.5, -2.3 );
magmaFloatComplex beta = MAGMA_C_MAKE(0., 0.); // MAGMA_C_MAKE( -0.6, 0.8 );
magmaFloatComplex *A, *X, *Y[4], *Ycublas, *Ymagma;
magmaFloatComplex *dA, *dX[4], *dY[4], *d_lA[4], *dYcublas ;
magma_queue_t stream[4][10];
magmaFloatComplex *C_work;
magmaFloatComplex *dC_work[4];
int max_num_gpus;
magma_int_t num_gpus = 1, nb;
magma_int_t blocks, lwork;
magma_int_t offset = 0;
M = 0;
N = 0;
if (argc != 1){
for(i = 1; i<argc; i++){
if (strcmp("-N", argv[i])==0)
{
N = atoi(argv[++i]);
istart = N;
}
else if (strcmp("-M", argv[i])==0)
M = atoi(argv[++i]);
else if (strcmp("-NGPU", argv[i])==0)
num_gpus = atoi(argv[++i]);
else if (strcmp("-offset", argv[i])==0)
offset = atoi(argv[++i]);
}
if ( M == 0 ) {
M = N;
}
if ( N == 0 ) {
N = M;
}
if (M>0 && N>0)
{ printf(" testing_chemv_mgpu -M %d -N %d -NGPU %d\n\n", (int) M, (int) N, (int) num_gpus);
printf(" in %c side \n", uplo);
}
else
{
printf("\nUsage: \n");
printf(" testing_chemv_mgpu -M %d -N %d -NGPU %d\n\n",
1024, 1024, 1);
exit(1);
}
}
else {
#if defined(PRECISION_z)
M = N = 8000;
#else
M = N = 12480;
#endif
num_gpus = 2;
offset = 0;
printf("\nUsage: \n");
printf(" testing_chemv_mgpu -M %d -N %d -NGPU %d\n\n", (int) M, (int) N, (int) num_gpus);
}
//////////////////////////////////////////////////////////////////////////
cudaGetDeviceCount(&max_num_gpus);
if (num_gpus > max_num_gpus){
printf("More GPUs requested than available. Have to change it.\n");
num_gpus = max_num_gpus;
}
printf("Number of GPUs to be used = %d\n", (int) num_gpus);
for(int i=0; i< num_gpus; i++)
{
magma_queue_create(&stream[i][0]);
}
LDA = ((N+31)/32)*32;
matsize = N*LDA;
vecsize = N*incx;
nb = 32;
//nb = 64;
printf("block size = %d\n", (int) nb);
TESTING_MALLOC_CPU( A, magmaFloatComplex, matsize );
TESTING_MALLOC_CPU( X, magmaFloatComplex, vecsize );
TESTING_MALLOC_CPU( Ycublas, magmaFloatComplex, vecsize );
TESTING_MALLOC_CPU( Ymagma, magmaFloatComplex, vecsize );
for(i=0; i<num_gpus; i++)
{
TESTING_MALLOC_CPU( Y[i], magmaFloatComplex, vecsize );
}
magma_setdevice(0);
TESTING_MALLOC_DEV( dA, magmaFloatComplex, matsize );
TESTING_MALLOC_DEV( dYcublas, magmaFloatComplex, vecsize );
for(i=0; i<num_gpus; i++)
{
n_local[i] = ((N/nb)/num_gpus)*nb;
if (i < (N/nb)%num_gpus)
n_local[i] += nb;
else if (i == (N/nb)%num_gpus)
n_local[i] += N%nb;
magma_setdevice(i);
TESTING_MALLOC_DEV( d_lA[i], magmaFloatComplex, LDA*n_local[i] );// potentially bugged
TESTING_MALLOC_DEV( dX[i], magmaFloatComplex, vecsize );
TESTING_MALLOC_DEV( dY[i], magmaFloatComplex, vecsize );
printf("device %2d n_local = %4d\n", (int) i, (int) n_local[i]);
}
magma_setdevice(0);
//////////////////////////////////////////////////////////////////////////
/* Initialize the matrix */
lapackf77_clarnv( &ione, ISEED, &matsize, A );
magma_cmake_hermitian( N, A, LDA );
blocks = N / nb + (N % nb != 0);
lwork = LDA * (blocks + 1);
TESTING_MALLOC_CPU( C_work, magmaFloatComplex, lwork );
for(i=0; i<num_gpus; i++){
magma_setdevice(i);
TESTING_MALLOC_DEV( dC_work[i], magmaFloatComplex, lwork );
//fillZero(dC_work[i], lwork);
}
magma_setdevice(0);
//////////////////////////////////////////////////////////////////////////
fp = fopen ("results_chemv_mgpu.csv", "w") ;
if( fp == NULL ){ printf("Couldn't open output file\n"); exit(1);}
printf("CHEMV magmaFloatComplex precision\n\n");
printf( " n CUBLAS,Gflop/s MAGMABLAS,Gflop/s \"error\"\n"
"==============================================================\n");
fprintf(fp, " n CUBLAS,Gflop/s MAGMABLAS,Gflop/s \"error\"\n"
"==============================================================\n");
// for( offset = 0; offset< N; offset ++ )
for(int size = istart ; size <= N ; size += 128)
{
// printf("offset = %d ", offset);
m = size ;
// m = N;
// lda = ((m+31)/32)*32;//
lda = LDA;
flops = FLOPS( (float)m ) / 1e6;
printf( "N %5d ", (int) m );
fprintf( fp, "%5d, ", (int) m );
vecsize = m * incx;
lapackf77_clarnv( &ione, ISEED, &vecsize, X );
lapackf77_clarnv( &ione, ISEED, &vecsize, Y[0] );
/* =====================================================================
Performs operation using CUDA-BLAS
=================================================================== */
magma_setdevice(0);
magma_csetmatrix_1D_col_bcyclic(m, m, A, LDA, d_lA, lda, num_gpus, nb);
magma_setdevice(0);
magma_csetmatrix( m, m, A, LDA, dA, lda );
magma_csetvector( m, Y[0], incx, dYcublas, incx );
for(i=0; i<num_gpus; i++){
magma_setdevice(i);
magma_csetvector( m, X, incx, dX[i], incx );
magma_csetvector( m, Y[0], incx, dY[i], incx );
blocks = m / nb + (m % nb != 0);
magma_csetmatrix( lda, blocks, C_work, LDA, dC_work[i], lda );
}
magma_setdevice(0);
start = get_current_time();
cublasChemv( uplo, m-offset, alpha, dA + offset + offset * lda, lda, dX[0] + offset, incx, beta, dYcublas + offset, incx );
end = get_current_time();
magma_cgetvector( m, dYcublas, incx, Ycublas, incx );
cuda_perf = flops / GetTimerValue(start,end);
printf( "%11.2f", cuda_perf );
fprintf(fp, "%11.2f,", cuda_perf );
magma_setdevice(0);
start = get_current_time();
if(nb == 32)
{
magmablas_chemv2_mgpu_32_offset( uplo, m, alpha, d_lA, lda, dX, incx, beta, dY, incx,
dC_work, lwork, num_gpus, nb, offset);
}
else // nb = 64
{
magmablas_chemv2_mgpu_offset( uplo, m, alpha, d_lA, lda, dX, incx, beta, dY, incx,
dC_work, lwork, num_gpus, nb, offset);
}
for(i=1; i<num_gpus; i++)
{
magma_setdevice(i);
cudaDeviceSynchronize();
}
end = get_current_time();
magma_perf = flops / GetTimerValue(start,end);
printf( "%11.2f", magma_perf );
fprintf(fp, "%11.2f,", magma_perf );
for(i=0; i<num_gpus; i++)
{
magma_setdevice(i);
magma_cgetvector( m, dY[i], incx, Y[i], incx );
}
magma_setdevice(0);
#ifdef validate
for( j= offset;j<m;j++)
{
for(i=1; i<num_gpus; i++)
{
// printf("Y[%d][%d] = %15.14f\n", i, j, Y[i][j].x);
#if defined(PRECISION_z) || defined(PRECISION_c)
Y[0][j].x = Y[0][j].x + Y[i][j].x;
Y[0][j].y = Y[0][j].y + Y[i][j].y;
#else
Y[0][j] = Y[0][j] + Y[i][j];
#endif
}
}
/*
#if defined(PRECISION_z) || defined(PRECISION_c)
for( j=offset;j<m;j++)
{
if(Y[0][j].x != Ycublas[j].x)
{
printf("Y-multi[%d] = %f, %f\n", j, Y[0][j].x, Y[0][j].y );
printf("Ycublas[%d] = %f, %f\n", j, Ycublas[j].x, Ycublas[j].y);
}
}
#else
for( j=offset;j<m;j++)
{
if(Y[0][j] != Ycublas[j])
{
printf("Y-multi[%d] = %f\n", j, Y[0][j] );
printf("Ycublas[%d] = %f\n", j, Ycublas[j]);
}
}
#endif
*/
/* =====================================================================
Computing the Difference Cublas VS Magma
=================================================================== */
magma_int_t nw = m - offset ;
blasf77_caxpy( &nw, &c_neg_one, Y[0] + offset, &incx, Ycublas + offset, &incx);
error = lapackf77_clange( "M", &nw, &ione, Ycublas + offset, &nw, work );
#if 0
printf( "\t\t %8.6e", error / m );
fprintf( fp, "\t\t %8.6e", error / m );
/*
* Extra check with cblas vs magma
*/
cblas_ccopy( m, Y, incx, Ycublas, incx );
cblas_chemv( CblasColMajor, CblasLower, m,
CBLAS_SADDR(alpha), A, LDA, X, incx,
CBLAS_SADDR(beta), Ycublas, incx );
blasf77_caxpy( &m, &c_neg_one, Ymagma, &incx, Ycublas, &incx);
error = lapackf77_clange( "M", &m, &ione, Ycublas, &m, work );
#endif
printf( "\t\t %8.6e", error / m );
fprintf( fp, "\t\t %8.6e", error / m );
#endif
printf("\n");
fprintf(fp, "\n");
}
fclose( fp ) ;
/* Free Memory */
TESTING_FREE_CPU( A );
TESTING_FREE_CPU( X );
TESTING_FREE_CPU( Ycublas );
TESTING_FREE_CPU( Ymagma );
TESTING_FREE_CPU( C_work );
magma_setdevice(0);
TESTING_FREE_DEV( dA );
TESTING_FREE_DEV( dYcublas );
for(i=0; i<num_gpus; i++)
{
TESTING_FREE_CPU( Y[i] );
magma_setdevice(i);
TESTING_FREE_DEV( d_lA[i] );
TESTING_FREE_DEV( dX[i] );
TESTING_FREE_DEV( dY[i] );
TESTING_FREE_DEV( dC_work[i] );
}
magma_setdevice(0);
///////////////////////////////////////////////////////////
/* Free device */
TESTING_FINALIZE();
return 0;
}
int main( int argc, char** argv )
{
TESTING_INIT();
real_Double_t gflops, t1, t2;
magmaFloatComplex c_neg_one = MAGMA_C_NEG_ONE;
magma_int_t ione = 1;
magma_trans_t trans[] = { MagmaNoTrans, MagmaConjTrans, MagmaTrans };
magma_uplo_t uplo [] = { MagmaLower, MagmaUpper };
magma_diag_t diag [] = { MagmaUnit, MagmaNonUnit };
magma_side_t side [] = { MagmaLeft, MagmaRight };
magmaFloatComplex *A, *B, *C, *C2, *LU;
magmaFloatComplex_ptr dA, dB, dC1, dC2;
magmaFloatComplex alpha = MAGMA_C_MAKE( 0.5, 0.1 );
magmaFloatComplex beta = MAGMA_C_MAKE( 0.7, 0.2 );
float dalpha = 0.6;
float dbeta = 0.8;
float work[1], error, total_error;
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t m, n, k, size, maxn, ld, info;
magma_int_t *piv;
magma_int_t err;
magma_opts opts;
parse_opts( argc, argv, &opts );
printf( "Compares magma wrapper function to cublas function; all diffs should be exactly 0.\n\n" );
total_error = 0.;
for( int itest = 0; itest < opts.ntest; ++itest ) {
m = opts.msize[itest];
n = opts.nsize[itest];
k = opts.ksize[itest];
printf("=========================================================================\n");
printf( "m=%d, n=%d, k=%d\n", (int) m, (int) n, (int) k );
// allocate matrices
// over-allocate so they can be any combination of {m,n,k} x {m,n,k}.
maxn = max( max( m, n ), k );
ld = max( 1, maxn );
size = ld*maxn;
err = magma_malloc_cpu( (void**) &piv, maxn*sizeof(magma_int_t) ); assert( err == 0 );
err = magma_cmalloc_pinned( &A, size ); assert( err == 0 );
err = magma_cmalloc_pinned( &B, size ); assert( err == 0 );
err = magma_cmalloc_pinned( &C, size ); assert( err == 0 );
err = magma_cmalloc_pinned( &C2, size ); assert( err == 0 );
err = magma_cmalloc_pinned( &LU, size ); assert( err == 0 );
err = magma_cmalloc( &dA, size ); assert( err == 0 );
err = magma_cmalloc( &dB, size ); assert( err == 0 );
err = magma_cmalloc( &dC1, size ); assert( err == 0 );
err = magma_cmalloc( &dC2, size ); assert( err == 0 );
// initialize matrices
size = maxn*maxn;
lapackf77_clarnv( &ione, ISEED, &size, A );
lapackf77_clarnv( &ione, ISEED, &size, B );
lapackf77_clarnv( &ione, ISEED, &size, C );
printf( "========== Level 1 BLAS ==========\n" );
// ----- test CSWAP
// swap columns 2 and 3 of dA, then copy to C2 and compare with A
if ( n >= 3 ) {
magma_csetmatrix( m, n, A, ld, dA, ld );
magma_csetmatrix( m, n, A, ld, dB, ld );
magma_cswap( m, dA(0,1), 1, dA(0,2), 1 );
magma_cswap( m, dB(0,1), 1, dB(0,2), 1 );
// check results, storing diff between magma and cuda calls in C2
cublasCaxpy( opts.handle, ld*n, &c_neg_one, dA, 1, dB, 1 );
magma_cgetmatrix( m, n, dB, ld, C2, ld );
error = lapackf77_clange( "F", &m, &k, C2, &ld, work );
total_error += error;
printf( "cswap diff %.2g\n", error );
}
else {
printf( "cswap skipped for n < 3\n" );
}
// ----- test ICAMAX
// get argmax of column of A
magma_csetmatrix( m, k, A, ld, dA, ld );
error = 0;
for( int j = 0; j < k; ++j ) {
magma_int_t i1 = magma_icamax( m, dA(0,j), 1 );
int i2; // NOT magma_int_t, for cublas
cublasIcamax( opts.handle, m, dA(0,j), 1, &i2 );
// todo need sync here?
assert( i1 == i2 );
error += abs( i1 - i2 );
}
total_error += error;
gflops = (float)m * k / 1e9;
printf( "icamax diff %.2g\n", error );
printf( "\n" );
printf( "========== Level 2 BLAS ==========\n" );
// ----- test CGEMV
// c = alpha*A*b + beta*c, with A m*n; b,c m or n-vectors
// try no-trans/trans
for( int ia = 0; ia < 3; ++ia ) {
magma_csetmatrix( m, n, A, ld, dA, ld );
magma_csetvector( maxn, B, 1, dB, 1 );
magma_csetvector( maxn, C, 1, dC1, 1 );
magma_csetvector( maxn, C, 1, dC2, 1 );
t1 = magma_sync_wtime( 0 );
magma_cgemv( trans[ia], m, n, alpha, dA, ld, dB, 1, beta, dC1, 1 );
t1 = magma_sync_wtime( 0 ) - t1;
t2 = magma_sync_wtime( 0 );
cublasCgemv( opts.handle, cublas_trans_const(trans[ia]),
m, n, &alpha, dA, ld, dB, 1, &beta, dC2, 1 );
t2 = magma_sync_wtime( 0 ) - t2;
// check results, storing diff between magma and cuda call in C2
size = (trans[ia] == MagmaNoTrans ? m : n);
cublasCaxpy( opts.handle, size, &c_neg_one, dC1, 1, dC2, 1 );
magma_cgetvector( size, dC2, 1, C2, 1 );
error = lapackf77_clange( "F", &size, &ione, C2, &ld, work );
total_error += error;
gflops = FLOPS_CGEMV( m, n ) / 1e9;
printf( "cgemv( %c ) diff %.2g, Gflop/s %7.2f, %7.2f\n",
lapacke_trans_const(trans[ia]), error, gflops/t1, gflops/t2 );
}
printf( "\n" );
// ----- test CHEMV
// c = alpha*A*b + beta*c, with A m*m symmetric; b,c m-vectors
// try upper/lower
for( int iu = 0; iu < 2; ++iu ) {
magma_csetmatrix( m, m, A, ld, dA, ld );
magma_csetvector( m, B, 1, dB, 1 );
magma_csetvector( m, C, 1, dC1, 1 );
magma_csetvector( m, C, 1, dC2, 1 );
t1 = magma_sync_wtime( 0 );
magma_chemv( uplo[iu], m, alpha, dA, ld, dB, 1, beta, dC1, 1 );
t1 = magma_sync_wtime( 0 ) - t1;
t2 = magma_sync_wtime( 0 );
cublasChemv( opts.handle, cublas_uplo_const(uplo[iu]),
m, &alpha, dA, ld, dB, 1, &beta, dC2, 1 );
t2 = magma_sync_wtime( 0 ) - t2;
// check results, storing diff between magma and cuda call in C2
cublasCaxpy( opts.handle, m, &c_neg_one, dC1, 1, dC2, 1 );
magma_cgetvector( m, dC2, 1, C2, 1 );
error = lapackf77_clange( "F", &m, &ione, C2, &ld, work );
total_error += error;
gflops = FLOPS_CHEMV( m ) / 1e9;
printf( "chemv( %c ) diff %.2g, Gflop/s %7.2f, %7.2f\n",
lapacke_uplo_const(uplo[iu]), error, gflops/t1, gflops/t2 );
}
printf( "\n" );
// ----- test CTRSV
// solve A*c = c, with A m*m triangular; c m-vector
// try upper/lower, no-trans/trans, unit/non-unit diag
// Factor A into LU to get well-conditioned triangles, else solve yields garbage.
// Still can give garbage if solves aren't consistent with LU factors,
// e.g., using unit diag for U, so copy lower triangle to upper triangle.
// Also used for trsm later.
lapackf77_clacpy( "Full", &maxn, &maxn, A, &ld, LU, &ld );
lapackf77_cgetrf( &maxn, &maxn, LU, &ld, piv, &info );
for( int j = 0; j < maxn; ++j ) {
for( int i = 0; i < j; ++i ) {
*LU(i,j) = *LU(j,i);
}
}
for( int iu = 0; iu < 2; ++iu ) {
for( int it = 0; it < 3; ++it ) {
for( int id = 0; id < 2; ++id ) {
magma_csetmatrix( m, m, LU, ld, dA, ld );
magma_csetvector( m, C, 1, dC1, 1 );
magma_csetvector( m, C, 1, dC2, 1 );
t1 = magma_sync_wtime( 0 );
magma_ctrsv( uplo[iu], trans[it], diag[id], m, dA, ld, dC1, 1 );
t1 = magma_sync_wtime( 0 ) - t1;
t2 = magma_sync_wtime( 0 );
cublasCtrsv( opts.handle, cublas_uplo_const(uplo[iu]), cublas_trans_const(trans[it]),
cublas_diag_const(diag[id]), m, dA, ld, dC2, 1 );
t2 = magma_sync_wtime( 0 ) - t2;
// check results, storing diff between magma and cuda call in C2
cublasCaxpy( opts.handle, m, &c_neg_one, dC1, 1, dC2, 1 );
magma_cgetvector( m, dC2, 1, C2, 1 );
error = lapackf77_clange( "F", &m, &ione, C2, &ld, work );
total_error += error;
gflops = FLOPS_CTRSM( MagmaLeft, m, 1 ) / 1e9;
printf( "ctrsv( %c, %c, %c ) diff %.2g, Gflop/s %7.2f, %7.2f\n",
lapacke_uplo_const(uplo[iu]), lapacke_trans_const(trans[it]), lapacke_diag_const(diag[id]),
error, gflops/t1, gflops/t2 );
}}}
printf( "\n" );
printf( "========== Level 3 BLAS ==========\n" );
// ----- test CGEMM
// C = alpha*A*B + beta*C, with A m*k or k*m; B k*n or n*k; C m*n
// try combinations of no-trans/trans
for( int ia = 0; ia < 3; ++ia ) {
for( int ib = 0; ib < 3; ++ib ) {
bool nta = (trans[ia] == MagmaNoTrans);
bool ntb = (trans[ib] == MagmaNoTrans);
magma_csetmatrix( (nta ? m : k), (nta ? m : k), A, ld, dA, ld );
magma_csetmatrix( (ntb ? k : n), (ntb ? n : k), B, ld, dB, ld );
magma_csetmatrix( m, n, C, ld, dC1, ld );
magma_csetmatrix( m, n, C, ld, dC2, ld );
t1 = magma_sync_wtime( 0 );
magma_cgemm( trans[ia], trans[ib], m, n, k, alpha, dA, ld, dB, ld, beta, dC1, ld );
t1 = magma_sync_wtime( 0 ) - t1;
t2 = magma_sync_wtime( 0 );
cublasCgemm( opts.handle, cublas_trans_const(trans[ia]), cublas_trans_const(trans[ib]),
m, n, k, &alpha, dA, ld, dB, ld, &beta, dC2, ld );
t2 = magma_sync_wtime( 0 ) - t2;
// check results, storing diff between magma and cuda call in C2
cublasCaxpy( opts.handle, ld*n, &c_neg_one, dC1, 1, dC2, 1 );
magma_cgetmatrix( m, n, dC2, ld, C2, ld );
error = lapackf77_clange( "F", &m, &n, C2, &ld, work );
total_error += error;
gflops = FLOPS_CGEMM( m, n, k ) / 1e9;
printf( "cgemm( %c, %c ) diff %.2g, Gflop/s %7.2f, %7.2f\n",
lapacke_trans_const(trans[ia]), lapacke_trans_const(trans[ib]),
error, gflops/t1, gflops/t2 );
}}
printf( "\n" );
// ----- test CHEMM
// C = alpha*A*B + beta*C (left) with A m*m symmetric; B,C m*n; or
// C = alpha*B*A + beta*C (right) with A n*n symmetric; B,C m*n
// try left/right, upper/lower
for( int is = 0; is < 2; ++is ) {
for( int iu = 0; iu < 2; ++iu ) {
magma_csetmatrix( m, m, A, ld, dA, ld );
magma_csetmatrix( m, n, B, ld, dB, ld );
magma_csetmatrix( m, n, C, ld, dC1, ld );
magma_csetmatrix( m, n, C, ld, dC2, ld );
t1 = magma_sync_wtime( 0 );
magma_chemm( side[is], uplo[iu], m, n, alpha, dA, ld, dB, ld, beta, dC1, ld );
t1 = magma_sync_wtime( 0 ) - t1;
t2 = magma_sync_wtime( 0 );
cublasChemm( opts.handle, cublas_side_const(side[is]), cublas_uplo_const(uplo[iu]),
m, n, &alpha, dA, ld, dB, ld, &beta, dC2, ld );
t2 = magma_sync_wtime( 0 ) - t2;
// check results, storing diff between magma and cuda call in C2
cublasCaxpy( opts.handle, ld*n, &c_neg_one, dC1, 1, dC2, 1 );
magma_cgetmatrix( m, n, dC2, ld, C2, ld );
error = lapackf77_clange( "F", &m, &n, C2, &ld, work );
total_error += error;
gflops = FLOPS_CHEMM( side[is], m, n ) / 1e9;
printf( "chemm( %c, %c ) diff %.2g, Gflop/s %7.2f, %7.2f\n",
lapacke_side_const(side[is]), lapacke_uplo_const(uplo[iu]),
error, gflops/t1, gflops/t2 );
}}
printf( "\n" );
// ----- test CHERK
// C = alpha*A*A^H + beta*C (no-trans) with A m*k and C m*m symmetric; or
// C = alpha*A^H*A + beta*C (trans) with A k*m and C m*m symmetric
// try upper/lower, no-trans/trans
for( int iu = 0; iu < 2; ++iu ) {
for( int it = 0; it < 3; ++it ) {
magma_csetmatrix( n, k, A, ld, dA, ld );
magma_csetmatrix( n, n, C, ld, dC1, ld );
magma_csetmatrix( n, n, C, ld, dC2, ld );
t1 = magma_sync_wtime( 0 );
magma_cherk( uplo[iu], trans[it], n, k, dalpha, dA, ld, dbeta, dC1, ld );
t1 = magma_sync_wtime( 0 ) - t1;
t2 = magma_sync_wtime( 0 );
cublasCherk( opts.handle, cublas_uplo_const(uplo[iu]), cublas_trans_const(trans[it]),
n, k, &dalpha, dA, ld, &dbeta, dC2, ld );
t2 = magma_sync_wtime( 0 ) - t2;
// check results, storing diff between magma and cuda call in C2
cublasCaxpy( opts.handle, ld*n, &c_neg_one, dC1, 1, dC2, 1 );
magma_cgetmatrix( n, n, dC2, ld, C2, ld );
error = lapackf77_clange( "F", &n, &n, C2, &ld, work );
total_error += error;
gflops = FLOPS_CHERK( k, n ) / 1e9;
printf( "cherk( %c, %c ) diff %.2g, Gflop/s %7.2f, %7.2f\n",
lapacke_uplo_const(uplo[iu]), lapacke_trans_const(trans[it]),
error, gflops/t1, gflops/t2 );
}}
printf( "\n" );
// ----- test CHER2K
// C = alpha*A*B^H + ^alpha*B*A^H + beta*C (no-trans) with A,B n*k; C n*n symmetric; or
// C = alpha*A^H*B + ^alpha*B^H*A + beta*C (trans) with A,B k*n; C n*n symmetric
// try upper/lower, no-trans/trans
for( int iu = 0; iu < 2; ++iu ) {
for( int it = 0; it < 3; ++it ) {
bool nt = (trans[it] == MagmaNoTrans);
magma_csetmatrix( (nt ? n : k), (nt ? n : k), A, ld, dA, ld );
magma_csetmatrix( n, n, C, ld, dC1, ld );
magma_csetmatrix( n, n, C, ld, dC2, ld );
t1 = magma_sync_wtime( 0 );
magma_cher2k( uplo[iu], trans[it], n, k, alpha, dA, ld, dB, ld, dbeta, dC1, ld );
t1 = magma_sync_wtime( 0 ) - t1;
t2 = magma_sync_wtime( 0 );
cublasCher2k( opts.handle, cublas_uplo_const(uplo[iu]), cublas_trans_const(trans[it]),
n, k, &alpha, dA, ld, dB, ld, &dbeta, dC2, ld );
t2 = magma_sync_wtime( 0 ) - t2;
// check results, storing diff between magma and cuda call in C2
cublasCaxpy( opts.handle, ld*n, &c_neg_one, dC1, 1, dC2, 1 );
magma_cgetmatrix( n, n, dC2, ld, C2, ld );
error = lapackf77_clange( "F", &n, &n, C2, &ld, work );
total_error += error;
gflops = FLOPS_CHER2K( k, n ) / 1e9;
printf( "cher2k( %c, %c ) diff %.2g, Gflop/s %7.2f, %7.2f\n",
lapacke_uplo_const(uplo[iu]), lapacke_trans_const(trans[it]),
error, gflops/t1, gflops/t2 );
}}
printf( "\n" );
// ----- test CTRMM
// C = alpha*A*C (left) with A m*m triangular; C m*n; or
// C = alpha*C*A (right) with A n*n triangular; C m*n
// try left/right, upper/lower, no-trans/trans, unit/non-unit
for( int is = 0; is < 2; ++is ) {
for( int iu = 0; iu < 2; ++iu ) {
for( int it = 0; it < 3; ++it ) {
for( int id = 0; id < 2; ++id ) {
bool left = (side[is] == MagmaLeft);
magma_csetmatrix( (left ? m : n), (left ? m : n), A, ld, dA, ld );
magma_csetmatrix( m, n, C, ld, dC1, ld );
magma_csetmatrix( m, n, C, ld, dC2, ld );
t1 = magma_sync_wtime( 0 );
magma_ctrmm( side[is], uplo[iu], trans[it], diag[id], m, n, alpha, dA, ld, dC1, ld );
t1 = magma_sync_wtime( 0 ) - t1;
// note cublas does trmm out-of-place (i.e., adds output matrix C),
// but allows C=B to do in-place.
t2 = magma_sync_wtime( 0 );
cublasCtrmm( opts.handle, cublas_side_const(side[is]), cublas_uplo_const(uplo[iu]),
cublas_trans_const(trans[it]), cublas_diag_const(diag[id]),
m, n, &alpha, dA, ld, dC2, ld, dC2, ld );
t2 = magma_sync_wtime( 0 ) - t2;
// check results, storing diff between magma and cuda call in C2
cublasCaxpy( opts.handle, ld*n, &c_neg_one, dC1, 1, dC2, 1 );
magma_cgetmatrix( m, n, dC2, ld, C2, ld );
error = lapackf77_clange( "F", &n, &n, C2, &ld, work );
total_error += error;
gflops = FLOPS_CTRMM( side[is], m, n ) / 1e9;
printf( "ctrmm( %c, %c ) diff %.2g, Gflop/s %7.2f, %7.2f\n",
lapacke_uplo_const(uplo[iu]), lapacke_trans_const(trans[it]),
error, gflops/t1, gflops/t2 );
}}}}
printf( "\n" );
// ----- test CTRSM
// solve A*X = alpha*B (left) with A m*m triangular; B m*n; or
// solve X*A = alpha*B (right) with A n*n triangular; B m*n
// try left/right, upper/lower, no-trans/trans, unit/non-unit
for( int is = 0; is < 2; ++is ) {
for( int iu = 0; iu < 2; ++iu ) {
for( int it = 0; it < 3; ++it ) {
for( int id = 0; id < 2; ++id ) {
bool left = (side[is] == MagmaLeft);
magma_csetmatrix( (left ? m : n), (left ? m : n), LU, ld, dA, ld );
magma_csetmatrix( m, n, C, ld, dC1, ld );
magma_csetmatrix( m, n, C, ld, dC2, ld );
t1 = magma_sync_wtime( 0 );
magma_ctrsm( side[is], uplo[iu], trans[it], diag[id], m, n, alpha, dA, ld, dC1, ld );
t1 = magma_sync_wtime( 0 ) - t1;
t2 = magma_sync_wtime( 0 );
cublasCtrsm( opts.handle, cublas_side_const(side[is]), cublas_uplo_const(uplo[iu]),
cublas_trans_const(trans[it]), cublas_diag_const(diag[id]),
m, n, &alpha, dA, ld, dC2, ld );
t2 = magma_sync_wtime( 0 ) - t2;
// check results, storing diff between magma and cuda call in C2
cublasCaxpy( opts.handle, ld*n, &c_neg_one, dC1, 1, dC2, 1 );
magma_cgetmatrix( m, n, dC2, ld, C2, ld );
error = lapackf77_clange( "F", &n, &n, C2, &ld, work );
total_error += error;
gflops = FLOPS_CTRSM( side[is], m, n ) / 1e9;
printf( "ctrsm( %c, %c ) diff %.2g, Gflop/s %7.2f, %7.2f\n",
lapacke_uplo_const(uplo[iu]), lapacke_trans_const(trans[it]),
error, gflops/t1, gflops/t2 );
}}}}
printf( "\n" );
// cleanup
magma_free_cpu( piv );
magma_free_pinned( A );
magma_free_pinned( B );
magma_free_pinned( C );
magma_free_pinned( C2 );
magma_free_pinned( LU );
magma_free( dA );
magma_free( dB );
magma_free( dC1 );
magma_free( dC2 );
fflush( stdout );
}
if ( total_error != 0. ) {
printf( "total error %.2g -- ought to be 0 -- some test failed (see above).\n",
total_error );
}
else {
printf( "all tests passed\n" );
}
TESTING_FINALIZE();
int status = (total_error != 0.);
return status;
}
