int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gflops, gpu_perf, gpu_time, cpu_perf, cpu_time;
magmaDoubleComplex *h_A, *h_R;
magmaDoubleComplex *d_A;
magma_int_t N, n2, lda, ldda, info;
magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
double work[1], error;
magma_int_t status = 0;
magmaDoubleComplex **d_A_array = NULL;
magma_int_t *dinfo_magma;
magma_int_t batchCount;
magma_queue_t queue = magma_stream;
magma_opts opts;
parse_opts( argc, argv, &opts );
opts.lapack |= opts.check; // check (-c) implies lapack (-l)
batchCount = opts.batchcount;
double tol = opts.tolerance * lapackf77_dlamch("E");
printf("BatchCount N CPU GFlop/s (ms) GPU GFlop/s (ms) ||R_magma - R_lapack||_F / ||R_lapack||_F\n");
printf("========================================================\n");
for( int i = 0; i < opts.ntest; ++i ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
N = opts.nsize[i];
ldda = lda = ((N+31)/32)*32;
n2 = lda* N * batchCount;
gflops = batchCount * FLOPS_ZPOTRF( N ) / 1e9 ;
TESTING_MALLOC_CPU( h_A, magmaDoubleComplex, n2);
TESTING_MALLOC_PIN( h_R, magmaDoubleComplex, n2);
TESTING_MALLOC_DEV( d_A, magmaDoubleComplex, ldda * N * batchCount);
TESTING_MALLOC_DEV( dinfo_magma, magma_int_t, batchCount);
magma_malloc((void**)&d_A_array, batchCount * sizeof(*d_A_array));
/* Initialize the matrix */
lapackf77_zlarnv( &ione, ISEED, &n2, h_A );
for(int i=0; i<batchCount; i++)
{
magma_zmake_hpd( N, h_A + i * lda * N, lda );// need modification
}
magma_int_t columns = N * batchCount;
lapackf77_zlacpy( MagmaUpperLowerStr, &N, &(columns), h_A, &lda, h_R, &lda );
magma_zsetmatrix( N, columns, h_A, lda, d_A, ldda );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
zset_pointer(d_A_array, d_A, ldda, 0, 0, ldda * N, batchCount, queue);
gpu_time = magma_sync_wtime(NULL);
info = magma_zpotrf_batched( opts.uplo, N, d_A_array, ldda, dinfo_magma, batchCount, queue);
gpu_time = magma_sync_wtime(NULL) - gpu_time;
gpu_perf = gflops / gpu_time;
magma_int_t *cpu_info = (magma_int_t*) malloc(batchCount*sizeof(magma_int_t));
magma_getvector( batchCount, sizeof(magma_int_t), dinfo_magma, 1, cpu_info, 1);
for(int i=0; i<batchCount; i++)
{
if(cpu_info[i] != 0 ){
printf("magma_zpotrf_batched matrix %d returned internal error %d\n",i, (int)cpu_info[i] );
}
}
if (info != 0)
printf("magma_zpotrf_batched returned argument error %d: %s.\n", (int) info, magma_strerror( info ));
if ( opts.lapack ) {
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
cpu_time = magma_wtime();
for(int i=0; i<batchCount; i++)
{
lapackf77_zpotrf( lapack_uplo_const(opts.uplo), &N, h_A + i * lda * N, &lda, &info );
}
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
if (info != 0)
printf("lapackf77_zpotrf returned error %d: %s.\n",
(int) info, magma_strerror( info ));
/* =====================================================================
Check the result compared to LAPACK
=================================================================== */
magma_zgetmatrix( N, columns, d_A, ldda, h_R, lda );
magma_int_t NN = lda*N;
char const uplo = 'l'; // lapack_uplo_const(opts.uplo)
double err = 0.0;
for(int i=0; i<batchCount; i++)
{
error = lapackf77_zlanhe("f", &uplo, &N, h_A + i * lda*N, &lda, work);
blasf77_zaxpy(&NN, &c_neg_one, h_A + i * lda*N, &ione, h_R + i * lda*N, &ione);
error = lapackf77_zlanhe("f", &uplo, &N, h_R + i * lda*N, &lda, work) / error;
if ( isnan(error) || isinf(error) ) {
err = error;
break;
}
err = max(fabs(error),err);
}
printf("%5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %s\n",
(int)batchCount, (int) N, cpu_perf, cpu_time*1000., gpu_perf, gpu_time*1000., err, (error < tol ? "ok" : "failed"));
status += ! (err < tol);
}
else {
printf("%5d %5d --- ( --- ) %7.2f (%7.2f) --- \n",
(int)batchCount, (int) N, gpu_perf, gpu_time*1000. );
}
TESTING_FREE_CPU( h_A );
TESTING_FREE_PIN( h_R );
TESTING_FREE_DEV( d_A );
TESTING_FREE_DEV( d_A_array );
TESTING_FREE_DEV( dinfo_magma );
free(cpu_info);
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing dswap, dswapblk, dpermute, dlaswp, dlaswpx
*/
int main( int argc, char** argv)
{
TESTING_INIT();
double *h_A1, *h_A2;
double *d_A1, *d_A2;
double *h_R1, *h_R2;
// row-major and column-major performance
real_Double_t row_perf0, col_perf0;
real_Double_t row_perf1, col_perf1;
real_Double_t row_perf2, col_perf2;
real_Double_t row_perf3;
real_Double_t row_perf4;
real_Double_t row_perf5, col_perf5;
real_Double_t row_perf6, col_perf6;
real_Double_t row_perf7;
real_Double_t cpu_perf;
real_Double_t time, gbytes;
magma_int_t N, lda, ldda, nb, j;
magma_int_t ione = 1;
magma_int_t *ipiv, *ipiv2;
magma_int_t *d_ipiv;
magma_int_t status = 0;
magma_opts opts;
parse_opts( argc, argv, &opts );
magma_queue_t queue = 0;
printf(" cublasDswap dswap dswapblk dlaswp dpermute dlaswp2 dlaswpx dcopymatrix CPU (all in )\n");
printf(" N nb row-maj/col-maj row-maj/col-maj row-maj/col-maj row-maj row-maj row-maj row-maj/col-maj row-blk/col-blk dlaswp (GByte/s)\n");
printf("==================================================================================================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
// For an N x N matrix, swap nb rows or nb columns using various methods.
// Each test is assigned one bit in the 'check' bitmask; bit=1 indicates failure.
// The variable 'shift' keeps track of which bit is for current test
int shift = 1;
int check = 0;
N = opts.nsize[itest];
lda = N;
ldda = ((N+31)/32)*32;
nb = (opts.nb > 0 ? opts.nb : magma_get_dgetrf_nb( N ));
nb = min( N, nb );
// each swap does 2N loads and 2N stores, for nb swaps
gbytes = sizeof(double) * 4.*N*nb / 1e9;
TESTING_MALLOC_PIN( h_A1, double, lda*N );
TESTING_MALLOC_PIN( h_A2, double, lda*N );
TESTING_MALLOC_PIN( h_R1, double, lda*N );
TESTING_MALLOC_PIN( h_R2, double, lda*N );
TESTING_MALLOC_CPU( ipiv, magma_int_t, nb );
TESTING_MALLOC_CPU( ipiv2, magma_int_t, nb );
TESTING_MALLOC_DEV( d_ipiv, magma_int_t, nb );
TESTING_MALLOC_DEV( d_A1, double, ldda*N );
TESTING_MALLOC_DEV( d_A2, double, ldda*N );
for( j=0; j < nb; j++ ) {
ipiv[j] = (magma_int_t) ((rand()*1.*N) / (RAND_MAX * 1.)) + 1;
}
/* =====================================================================
* cublasDswap, row-by-row (2 matrices)
*/
/* Row Major */
init_matrix( N, N, h_A1, lda, 0 );
init_matrix( N, N, h_A2, lda, 100 );
magma_dsetmatrix( N, N, h_A1, lda, d_A1, ldda );
magma_dsetmatrix( N, N, h_A2, lda, d_A2, ldda );
time = magma_sync_wtime( queue );
for( j=0; j < nb; j++) {
if ( j != (ipiv[j]-1)) {
cublasDswap( N, d_A1+ldda*j, 1, d_A2+ldda*(ipiv[j]-1), 1);
}
}
time = magma_sync_wtime( queue ) - time;
row_perf0 = gbytes / time;
for( j=0; j < nb; j++) {
if ( j != (ipiv[j]-1)) {
blasf77_dswap( &N, h_A1+lda*j, &ione, h_A2+lda*(ipiv[j]-1), &ione);
}
}
magma_dgetmatrix( N, N, d_A1, ldda, h_R1, lda );
magma_dgetmatrix( N, N, d_A2, ldda, h_R2, lda );
check += (diff_matrix( N, N, h_A1, lda, h_R1, lda ) ||
diff_matrix( N, N, h_A2, lda, h_R2, lda ))*shift;
shift *= 2;
/* Column Major */
init_matrix( N, N, h_A1, lda, 0 );
init_matrix( N, N, h_A2, lda, 100 );
magma_dsetmatrix( N, N, h_A1, lda, d_A1, ldda );
magma_dsetmatrix( N, N, h_A2, lda, d_A2, ldda );
time = magma_sync_wtime( queue );
for( j=0; j < nb; j++) {
if ( j != (ipiv[j]-1)) {
cublasDswap( N, d_A1+j, ldda, d_A2+ipiv[j]-1, ldda);
}
}
time = magma_sync_wtime( queue ) - time;
col_perf0 = gbytes / time;
for( j=0; j < nb; j++) {
if ( j != (ipiv[j]-1)) {
blasf77_dswap( &N, h_A1+j, &lda, h_A2+(ipiv[j]-1), &lda);
}
}
magma_dgetmatrix( N, N, d_A1, ldda, h_R1, lda );
magma_dgetmatrix( N, N, d_A2, ldda, h_R2, lda );
check += (diff_matrix( N, N, h_A1, lda, h_R1, lda ) ||
diff_matrix( N, N, h_A2, lda, h_R2, lda ))*shift;
shift *= 2;
/* =====================================================================
* dswap, row-by-row (2 matrices)
*/
/* Row Major */
init_matrix( N, N, h_A1, lda, 0 );
init_matrix( N, N, h_A2, lda, 100 );
magma_dsetmatrix( N, N, h_A1, lda, d_A1, ldda );
magma_dsetmatrix( N, N, h_A2, lda, d_A2, ldda );
time = magma_sync_wtime( queue );
for( j=0; j < nb; j++) {
if ( j != (ipiv[j]-1)) {
magmablas_dswap( N, d_A1+ldda*j, 1, d_A2+ldda*(ipiv[j]-1), 1);
}
}
time = magma_sync_wtime( queue ) - time;
row_perf1 = gbytes / time;
for( j=0; j < nb; j++) {
if ( j != (ipiv[j]-1)) {
blasf77_dswap( &N, h_A1+lda*j, &ione, h_A2+lda*(ipiv[j]-1), &ione);
}
}
magma_dgetmatrix( N, N, d_A1, ldda, h_R1, lda );
magma_dgetmatrix( N, N, d_A2, ldda, h_R2, lda );
check += (diff_matrix( N, N, h_A1, lda, h_R1, lda ) ||
diff_matrix( N, N, h_A2, lda, h_R2, lda ))*shift;
shift *= 2;
/* Column Major */
init_matrix( N, N, h_A1, lda, 0 );
init_matrix( N, N, h_A2, lda, 100 );
magma_dsetmatrix( N, N, h_A1, lda, d_A1, ldda );
magma_dsetmatrix( N, N, h_A2, lda, d_A2, ldda );
time = magma_sync_wtime( queue );
for( j=0; j < nb; j++) {
if ( j != (ipiv[j]-1)) {
magmablas_dswap( N, d_A1+j, ldda, d_A2+ipiv[j]-1, ldda );
}
}
time = magma_sync_wtime( queue ) - time;
col_perf1 = gbytes / time;
for( j=0; j < nb; j++) {
if ( j != (ipiv[j]-1)) {
blasf77_dswap( &N, h_A1+j, &lda, h_A2+(ipiv[j]-1), &lda);
}
}
magma_dgetmatrix( N, N, d_A1, ldda, h_R1, lda );
magma_dgetmatrix( N, N, d_A2, ldda, h_R2, lda );
check += (diff_matrix( N, N, h_A1, lda, h_R1, lda ) ||
diff_matrix( N, N, h_A2, lda, h_R2, lda ))*shift;
shift *= 2;
/* =====================================================================
* dswapblk, blocked version (2 matrices)
*/
/* Row Major */
init_matrix( N, N, h_A1, lda, 0 );
init_matrix( N, N, h_A2, lda, 100 );
magma_dsetmatrix( N, N, h_A1, lda, d_A1, ldda );
magma_dsetmatrix( N, N, h_A2, lda, d_A2, ldda );
time = magma_sync_wtime( queue );
magmablas_dswapblk( MagmaRowMajor, N, d_A1, ldda, d_A2, ldda, 1, nb, ipiv, 1, 0);
time = magma_sync_wtime( queue ) - time;
row_perf2 = gbytes / time;
for( j=0; j < nb; j++) {
if ( j != (ipiv[j]-1)) {
blasf77_dswap( &N, h_A1+lda*j, &ione, h_A2+lda*(ipiv[j]-1), &ione);
}
}
magma_dgetmatrix( N, N, d_A1, ldda, h_R1, lda );
magma_dgetmatrix( N, N, d_A2, ldda, h_R2, lda );
check += (diff_matrix( N, N, h_A1, lda, h_R1, lda ) ||
diff_matrix( N, N, h_A2, lda, h_R2, lda ))*shift;
shift *= 2;
/* Column Major */
init_matrix( N, N, h_A1, lda, 0 );
init_matrix( N, N, h_A2, lda, 100 );
magma_dsetmatrix( N, N, h_A1, lda, d_A1, ldda );
magma_dsetmatrix( N, N, h_A2, lda, d_A2, ldda );
time = magma_sync_wtime( queue );
magmablas_dswapblk( MagmaColMajor, N, d_A1, ldda, d_A2, ldda, 1, nb, ipiv, 1, 0);
time = magma_sync_wtime( queue ) - time;
col_perf2 = gbytes / time;
for( j=0; j < nb; j++) {
if ( j != (ipiv[j]-1)) {
blasf77_dswap( &N, h_A1+j, &lda, h_A2+(ipiv[j]-1), &lda);
}
}
magma_dgetmatrix( N, N, d_A1, ldda, h_R1, lda );
magma_dgetmatrix( N, N, d_A2, ldda, h_R2, lda );
check += (diff_matrix( N, N, h_A1, lda, h_R1, lda ) ||
diff_matrix( N, N, h_A2, lda, h_R2, lda ))*shift;
shift *= 2;
/* =====================================================================
* dpermute_long (1 matrix)
*/
/* Row Major */
memcpy( ipiv2, ipiv, nb*sizeof(magma_int_t) ); // dpermute updates ipiv2
init_matrix( N, N, h_A1, lda, 0 );
magma_dsetmatrix( N, N, h_A1, lda, d_A1, ldda );
time = magma_sync_wtime( queue );
magmablas_dpermute_long2( N, d_A1, ldda, ipiv2, nb, 0 );
time = magma_sync_wtime( queue ) - time;
row_perf3 = gbytes / time;
for( j=0; j < nb; j++) {
if ( j != (ipiv[j]-1)) {
blasf77_dswap( &N, h_A1+lda*j, &ione, h_A1+lda*(ipiv[j]-1), &ione);
}
}
magma_dgetmatrix( N, N, d_A1, ldda, h_R1, lda );
check += diff_matrix( N, N, h_A1, lda, h_R1, lda )*shift;
shift *= 2;
/* =====================================================================
* LAPACK-style dlaswp (1 matrix)
*/
/* Row Major */
init_matrix( N, N, h_A1, lda, 0 );
magma_dsetmatrix( N, N, h_A1, lda, d_A1, ldda );
time = magma_sync_wtime( queue );
magmablas_dlaswp( N, d_A1, ldda, 1, nb, ipiv, 1);
time = magma_sync_wtime( queue ) - time;
row_perf4 = gbytes / time;
for( j=0; j < nb; j++) {
if ( j != (ipiv[j]-1)) {
blasf77_dswap( &N, h_A1+lda*j, &ione, h_A1+lda*(ipiv[j]-1), &ione);
}
}
magma_dgetmatrix( N, N, d_A1, ldda, h_R1, lda );
check += diff_matrix( N, N, h_A1, lda, h_R1, lda )*shift;
shift *= 2;
/* =====================================================================
* LAPACK-style dlaswp (1 matrix) - d_ipiv on GPU
*/
/* Row Major */
init_matrix( N, N, h_A1, lda, 0 );
magma_dsetmatrix( N, N, h_A1, lda, d_A1, ldda );
time = magma_sync_wtime( queue );
magma_setvector( nb, sizeof(magma_int_t), ipiv, 1, d_ipiv, 1 );
magmablas_dlaswp2( N, d_A1, ldda, 1, nb, d_ipiv, 1 );
time = magma_sync_wtime( queue ) - time;
row_perf7 = gbytes / time;
for( j=0; j < nb; j++) {
if ( j != (ipiv[j]-1)) {
blasf77_dswap( &N, h_A1+lda*j, &ione, h_A1+lda*(ipiv[j]-1), &ione);
}
}
magma_dgetmatrix( N, N, d_A1, ldda, h_R1, lda );
check += diff_matrix( N, N, h_A1, lda, h_R1, lda )*shift;
shift *= 2;
/* =====================================================================
* LAPACK-style dlaswpx (extended for row- and col-major) (1 matrix)
*/
/* Row Major */
init_matrix( N, N, h_A1, lda, 0 );
magma_dsetmatrix( N, N, h_A1, lda, d_A1, ldda );
time = magma_sync_wtime( queue );
magmablas_dlaswpx( N, d_A1, ldda, 1, 1, nb, ipiv, 1);
time = magma_sync_wtime( queue ) - time;
row_perf5 = gbytes / time;
for( j=0; j < nb; j++) {
if ( j != (ipiv[j]-1)) {
blasf77_dswap( &N, h_A1+lda*j, &ione, h_A1+lda*(ipiv[j]-1), &ione);
}
}
magma_dgetmatrix( N, N, d_A1, ldda, h_R1, lda );
check += diff_matrix( N, N, h_A1, lda, h_R1, lda )*shift;
shift *= 2;
/* Col Major */
init_matrix( N, N, h_A1, lda, 0 );
magma_dsetmatrix( N, N, h_A1, lda, d_A1, ldda );
time = magma_sync_wtime( queue );
magmablas_dlaswpx( N, d_A1, 1, ldda, 1, nb, ipiv, 1);
time = magma_sync_wtime( queue ) - time;
col_perf5 = gbytes / time;
time = magma_wtime();
lapackf77_dlaswp( &N, h_A1, &lda, &ione, &nb, ipiv, &ione);
time = magma_wtime() - time;
cpu_perf = gbytes / time;
magma_dgetmatrix( N, N, d_A1, ldda, h_R1, lda );
check += diff_matrix( N, N, h_A1, lda, h_R1, lda )*shift;
shift *= 2;
/* =====================================================================
* Copy matrix.
*/
time = magma_sync_wtime( queue );
magma_dcopymatrix( N, nb, d_A1, ldda, d_A2, ldda );
time = magma_sync_wtime( queue ) - time;
// copy reads 1 matrix and writes 1 matrix, so has half gbytes of swap
col_perf6 = 0.5 * gbytes / time;
time = magma_sync_wtime( queue );
magma_dcopymatrix( nb, N, d_A1, ldda, d_A2, ldda );
time = magma_sync_wtime( queue ) - time;
// copy reads 1 matrix and writes 1 matrix, so has half gbytes of swap
row_perf6 = 0.5 * gbytes / time;
printf("%5d %3d %6.2f%c/ %6.2f%c %6.2f%c/ %6.2f%c %6.2f%c/ %6.2f%c %6.2f%c %6.2f%c %6.2f%c %6.2f%c/ %6.2f%c %6.2f / %6.2f %6.2f %10s\n",
(int) N, (int) nb,
row_perf0, ((check & 0x001) != 0 ? '*' : ' '),
col_perf0, ((check & 0x002) != 0 ? '*' : ' '),
row_perf1, ((check & 0x004) != 0 ? '*' : ' '),
col_perf1, ((check & 0x008) != 0 ? '*' : ' '),
row_perf2, ((check & 0x010) != 0 ? '*' : ' '),
col_perf2, ((check & 0x020) != 0 ? '*' : ' '),
row_perf3, ((check & 0x040) != 0 ? '*' : ' '),
row_perf4, ((check & 0x080) != 0 ? '*' : ' '),
row_perf7, ((check & 0x100) != 0 ? '*' : ' '),
row_perf5, ((check & 0x200) != 0 ? '*' : ' '),
col_perf5, ((check & 0x400) != 0 ? '*' : ' '),
row_perf6,
col_perf6,
cpu_perf,
(check == 0 ? "ok" : "* failed") );
status += ! (check == 0);
TESTING_FREE_PIN( h_A1 );
TESTING_FREE_PIN( h_A2 );
TESTING_FREE_PIN( h_R1 );
TESTING_FREE_PIN( h_R2 );
TESTING_FREE_CPU( ipiv );
TESTING_FREE_CPU( ipiv2 );
TESTING_FREE_DEV( d_ipiv );
TESTING_FREE_DEV( d_A1 );
TESTING_FREE_DEV( d_A2 );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return status;
}
int main( int argc, char** argv)
{
real_Double_t gflops, gpu_perf, cpu_perf, gpu_time, cpu_time, error;
float matnorm, work[1];
magmaFloatComplex c_neg_one = MAGMA_C_NEG_ONE;
magmaFloatComplex *h_A, *h_R, *tau, *h_work, tmp[1];
magmaFloatComplex_ptr d_lA[MagmaMaxGPUs];
/* Matrix size */
magma_int_t M = 0, N = 0, n2, n_local[4], lda, ldda, lhwork;
magma_int_t size[10] = {1000,2000,3000,4000,5000,6000,7000,8000,9000,10000};
magma_int_t i, k, nk, info, min_mn;
int max_num_gpus = 2, num_gpus = 2;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
if (argc != 1){
for(i = 1; i<argc; i++){
if (strcmp("-N", argv[i])==0)
N = atoi(argv[++i]);
else if (strcmp("-M", argv[i])==0)
M = atoi(argv[++i]);
else if (strcmp("-NGPU", argv[i])==0)
num_gpus = atoi(argv[++i]);
}
if ( M == 0 ) {
M = N;
}
if ( N == 0 ) {
N = M;
}
if (M>0 && N>0)
printf(" testing_cgeqrf_gpu -M %d -N %d -NGPU %d\n\n", (int) M, (int) N, (int) num_gpus);
else
{
printf("\nUsage: \n");
printf(" testing_cgeqrf_gpu -M %d -N %d -NGPU %d\n\n",
1024, 1024, 1);
exit(1);
}
}
else {
printf("\nUsage: \n");
printf(" testing_cgeqrf_gpu -M %d -N %d -NGPU %d\n\n", 1024, 1024, 1);
M = N = size[9];
}
ldda = ((M+31)/32)*32;
n2 = M * N;
min_mn = min(M, N);
magma_int_t nb = magma_get_cgeqrf_nb(M);
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);
/* Initialize */
magma_queue_t queues[MagmaMaxGPUs * 2];
magma_device_t devices[ MagmaMaxGPUs ];
int num = 0;
magma_err_t err;
magma_init();
err = magma_get_devices( devices, MagmaMaxGPUs, &num );
if ( err != 0 || num < 1 ) {
fprintf( stderr, "magma_get_devices failed: %d\n", err );
exit(-1);
}
for(i=0;i<num_gpus;i++){
err = magma_queue_create( devices[i], &queues[2*i] );
if ( err != 0 ) {
fprintf( stderr, "magma_queue_create failed: %d\n", err );
exit(-1);
}
err = magma_queue_create( devices[i], &queues[2*i+1] );
if ( err != 0 ) {
fprintf( stderr, "magma_queue_create failed: %d\n", err );
exit(-1);
}
}
/* Allocate host memory for the matrix */
TESTING_MALLOC_CPU( tau, magmaFloatComplex, min_mn );
TESTING_MALLOC_CPU( h_A, magmaFloatComplex, n2 );
TESTING_MALLOC_CPU( h_R, magmaFloatComplex, n2 );
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;
TESTING_MALLOC_DEV( d_lA[i], magmaFloatComplex, ldda*n_local[i] );
printf("device %2d n_local = %4d\n", (int) i, (int) n_local[i]);
}
lhwork = -1;
lapackf77_cgeqrf(&M, &N, h_A, &M, tau, tmp, &lhwork, &info);
lhwork = (magma_int_t)MAGMA_C_REAL( tmp[0] );
TESTING_MALLOC_CPU( h_work, magmaFloatComplex, lhwork );
printf(" M N CPU GFlop/s (sec) GPU GFlop/s (sec) ||R||_F / ||A||_F\n");
printf("======================================================================\n");
for(i=0; i<10; i++){
if (argc == 1){
M = N = size[i];
}
min_mn= min(M, N);
lda = M;
n2 = lda*N;
ldda = ((M+31)/32)*32;
gflops = FLOPS( (float)M, (float)N ) * 1e-9;
/* Initialize the matrix */
lapackf77_clarnv( &ione, ISEED, &n2, h_A );
lapackf77_clacpy( MagmaUpperLowerStr, &M, &N, h_A, &lda, h_R, &lda );
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
cpu_time = magma_wtime();
lapackf77_cgeqrf(&M, &N, h_A, &M, tau, h_work, &lhwork, &info);
cpu_time = magma_wtime() - cpu_time;
if (info < 0)
printf("Argument %d of lapack_cgeqrf had an illegal value.\n", (int) -info);
cpu_perf = gflops / cpu_time;
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
int j;
magma_queue_t *trans_queues = (magma_queue_t*)malloc(num_gpus*sizeof(magma_queue_t));
for(j=0;j<num_gpus;j++){
trans_queues[j] = queues[2*j];
}
// warm-up
magmablas_csetmatrix_1D_bcyclic(M, N, h_R, lda, d_lA, ldda, num_gpus, nb, trans_queues);
magma_cgeqrf2_mgpu( num_gpus, M, N, d_lA, ldda, tau, &info, queues);
magmablas_csetmatrix_1D_bcyclic(M, N, h_R, lda, d_lA, ldda, num_gpus, nb, trans_queues);
gpu_time = magma_wtime();
magma_cgeqrf2_mgpu( num_gpus, M, N, d_lA, ldda, tau, &info, queues);
gpu_time = magma_wtime() - gpu_time;
if (info < 0)
printf("Argument %d of magma_cgeqrf2 had an illegal value.\n", (int) -info);
gpu_perf = gflops / gpu_time;
/* =====================================================================
Check the result compared to LAPACK
=================================================================== */
magmablas_cgetmatrix_1D_bcyclic(M, N, d_lA, ldda, h_R, lda, num_gpus, nb, trans_queues);
matnorm = lapackf77_clange("f", &M, &N, h_A, &M, work);
blasf77_caxpy(&n2, &c_neg_one, h_A, &ione, h_R, &ione);
printf("%5d %5d %6.2f (%6.2f) %6.2f (%6.2f) %e\n",
(int) M, (int) N, cpu_perf, cpu_time, gpu_perf, gpu_time,
lapackf77_clange("f", &M, &N, h_R, &M, work) / matnorm);
if (argc != 1)
break;
}
/* Memory clean up */
TESTING_FREE_PIN( tau );
TESTING_FREE_PIN( h_A );
TESTING_FREE_PIN( h_work );
TESTING_FREE_PIN( h_R );
for(i=0; i<num_gpus; i++){
TESTING_FREE_DEV( d_lA[i] );
magma_queue_destroy(queues[2*i]);
magma_queue_destroy(queues[2*i+1]);
}
/* Shutdown */
magma_finalize();
}
int main(int argc, char **argv)
{
TESTING_INIT();
const magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
const magma_int_t ione = 1;
real_Double_t atomics_perf=0, atomics_time=0;
real_Double_t gflops, magma_perf=0, magma_time=0, cublas_perf, cublas_time, cpu_perf, cpu_time;
double magma_error=0, atomics_error=0, cublas_error, work[1];
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t N, lda, ldda, sizeA, sizeX, sizeY, blocks, ldwork;
magma_int_t incx = 1;
magma_int_t incy = 1;
magma_int_t nb = 64;
magmaDoubleComplex alpha = MAGMA_Z_MAKE( 1.5, -2.3 );
magmaDoubleComplex beta = MAGMA_Z_MAKE( -0.6, 0.8 );
magmaDoubleComplex *A, *X, *Y, *Yatomics, *Ycublas, *Ymagma;
magmaDoubleComplex_ptr dA, dX, dY, dwork;
magma_int_t status = 0;
magma_opts opts;
opts.parse_opts( argc, argv );
double tol = opts.tolerance * lapackf77_dlamch("E");
printf("%% uplo = %s\n", lapack_uplo_const(opts.uplo) );
printf("%% N MAGMA Gflop/s (ms) Atomics Gflop/s CUBLAS Gflop/s CPU Gflop/s MAGMA error Atomics CUBLAS\n");
printf("%%=====================================================================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
N = opts.nsize[itest];
lda = N;
ldda = magma_roundup( N, opts.align ); // multiple of 32 by default
sizeA = N*lda;
sizeX = N*incx;
sizeY = N*incy;
gflops = FLOPS_ZHEMV( N ) / 1e9;
TESTING_MALLOC_CPU( A, magmaDoubleComplex, sizeA );
TESTING_MALLOC_CPU( X, magmaDoubleComplex, sizeX );
TESTING_MALLOC_CPU( Y, magmaDoubleComplex, sizeY );
TESTING_MALLOC_CPU( Yatomics, magmaDoubleComplex, sizeY );
TESTING_MALLOC_CPU( Ycublas, magmaDoubleComplex, sizeY );
TESTING_MALLOC_CPU( Ymagma, magmaDoubleComplex, sizeY );
TESTING_MALLOC_DEV( dA, magmaDoubleComplex, ldda*N );
TESTING_MALLOC_DEV( dX, magmaDoubleComplex, sizeX );
TESTING_MALLOC_DEV( dY, magmaDoubleComplex, sizeY );
blocks = magma_ceildiv( N, nb );
ldwork = ldda*blocks;
TESTING_MALLOC_DEV( dwork, magmaDoubleComplex, ldwork );
magmablas_zlaset( MagmaFull, ldwork, 1, MAGMA_Z_NAN, MAGMA_Z_NAN, dwork, ldwork );
magmablas_zlaset( MagmaFull, ldda, N, MAGMA_Z_NAN, MAGMA_Z_NAN, dA, ldda );
/* Initialize the matrix */
lapackf77_zlarnv( &ione, ISEED, &sizeA, A );
magma_zmake_hermitian( N, A, lda );
// should not use data from the opposite triangle -- fill with NAN to check
magma_int_t N1 = N-1;
if ( opts.uplo == MagmaUpper ) {
lapackf77_zlaset( "Lower", &N1, &N1, &MAGMA_Z_NAN, &MAGMA_Z_NAN, &A[1], &lda );
}
else {
lapackf77_zlaset( "Upper", &N1, &N1, &MAGMA_Z_NAN, &MAGMA_Z_NAN, &A[lda], &lda );
}
lapackf77_zlarnv( &ione, ISEED, &sizeX, X );
lapackf77_zlarnv( &ione, ISEED, &sizeY, Y );
/* =====================================================================
Performs operation using CUBLAS
=================================================================== */
magma_zsetmatrix( N, N, A, lda, dA, ldda );
magma_zsetvector( N, X, incx, dX, incx );
magma_zsetvector( N, Y, incy, dY, incy );
magmablasSetKernelStream( opts.queue ); // opts.handle also uses opts.queue
cublas_time = magma_sync_wtime( opts.queue );
#ifdef HAVE_CUBLAS
cublasZhemv( opts.handle, cublas_uplo_const(opts.uplo),
N, &alpha, dA, ldda, dX, incx, &beta, dY, incy );
#else
magma_zhemv( opts.uplo, N, alpha, dA, 0, ldda, dX, 0, incx, beta, dY, 0, incy, opts.queue );
#endif
cublas_time = magma_sync_wtime( opts.queue ) - cublas_time;
cublas_perf = gflops / cublas_time;
magma_zgetvector( N, dY, incy, Ycublas, incy );
/* =====================================================================
Performs operation using CUBLAS - using atomics
=================================================================== */
#ifdef HAVE_CUBLAS
cublasSetAtomicsMode( opts.handle, CUBLAS_ATOMICS_ALLOWED );
magma_zsetvector( N, Y, incy, dY, incy );
// sync on queue doesn't work -- need device sync or use NULL stream -- bug in CUBLAS?
atomics_time = magma_sync_wtime( NULL /*opts.queue*/ );
cublasZhemv( opts.handle, cublas_uplo_const(opts.uplo),
N, &alpha, dA, ldda, dX, incx, &beta, dY, incy );
atomics_time = magma_sync_wtime( NULL /*opts.queue*/ ) - atomics_time;
atomics_perf = gflops / atomics_time;
magma_zgetvector( N, dY, incy, Yatomics, incy );
cublasSetAtomicsMode( opts.handle, CUBLAS_ATOMICS_NOT_ALLOWED );
#endif
/* =====================================================================
Performs operation using MAGMABLAS
=================================================================== */
#ifdef HAVE_CUBLAS
magma_zsetvector( N, Y, incy, dY, incy );
magma_time = magma_sync_wtime( opts.queue );
if ( opts.version == 1 ) {
magmablas_zhemv_work( opts.uplo, N, alpha, dA, ldda, dX, incx, beta, dY, incy, dwork, ldwork, opts.queue );
}
else {
// non-work interface (has added overhead)
magmablas_zhemv( opts.uplo, N, alpha, dA, ldda, dX, incx, beta, dY, incy );
}
magma_time = magma_sync_wtime( opts.queue ) - magma_time;
magma_perf = gflops / magma_time;
magma_zgetvector( N, dY, incy, Ymagma, incy );
#endif
/* =====================================================================
Performs operation using CPU BLAS
=================================================================== */
cpu_time = magma_wtime();
blasf77_zhemv( lapack_uplo_const(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_zaxpy( &N, &c_neg_one, Y, &incy, Ycublas, &incy );
cublas_error = lapackf77_zlange( "M", &N, &ione, Ycublas, &N, work ) / N;
#ifdef HAVE_CUBLAS
blasf77_zaxpy( &N, &c_neg_one, Y, &incy, Yatomics, &incy );
atomics_error = lapackf77_zlange( "M", &N, &ione, Yatomics, &N, work ) / N;
blasf77_zaxpy( &N, &c_neg_one, Y, &incy, Ymagma, &incy );
magma_error = lapackf77_zlange( "M", &N, &ione, Ymagma, &N, work ) / N;
#endif
bool okay = (magma_error < tol && cublas_error < tol && atomics_error < tol);
status += ! okay;
printf("%5d %7.2f (%7.2f) %7.2f (%7.2f) %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %8.2e %8.2e %s\n",
(int) N,
magma_perf, 1000.*magma_time,
atomics_perf, 1000.*atomics_time,
cublas_perf, 1000.*cublas_time,
cpu_perf, 1000.*cpu_time,
magma_error, cublas_error, atomics_error,
(okay ? "ok" : "failed"));
TESTING_FREE_CPU( A );
TESTING_FREE_CPU( X );
TESTING_FREE_CPU( Y );
TESTING_FREE_CPU( Ycublas );
TESTING_FREE_CPU( Yatomics );
TESTING_FREE_CPU( Ymagma );
TESTING_FREE_DEV( dA );
TESTING_FREE_DEV( dX );
TESTING_FREE_DEV( dY );
TESTING_FREE_DEV( dwork );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
opts.cleanup();
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing zpotrf
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gflops, gpu_perf, gpu_time, cpu_perf, cpu_time;
magmaDoubleComplex *h_A, *h_R;
magmaDoubleComplex *d_A;
magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
magma_int_t N, n2, lda, ldda, info;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
double work[1], error;
magma_int_t status = 0;
magma_opts opts;
parse_opts( argc, argv, &opts );
opts.lapack |= opts.check; // check (-c) implies lapack (-l)
double tol = opts.tolerance * lapackf77_dlamch("E");
printf("uplo = %s\n", lapack_uplo_const(opts.uplo) );
printf(" N CPU GFlop/s (sec) GPU GFlop/s (sec) ||R||_F / ||A||_F\n");
printf("=================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
N = opts.nsize[itest];
lda = N;
n2 = lda*N;
ldda = ((N+31)/32)*32;
gflops = FLOPS_ZPOTRI( N ) / 1e9;
TESTING_MALLOC_CPU( h_A, magmaDoubleComplex, n2 );
TESTING_MALLOC_PIN( h_R, magmaDoubleComplex, n2 );
TESTING_MALLOC_DEV( d_A, magmaDoubleComplex, ldda*N );
/* Initialize the matrix */
lapackf77_zlarnv( &ione, ISEED, &n2, h_A );
magma_zmake_hpd( N, h_A, lda );
lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_A, &lda, h_R, &lda );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
/* factorize matrix */
magma_zsetmatrix( N, N, h_A, lda, d_A, ldda );
magma_zpotrf_gpu( opts.uplo, N, d_A, ldda, &info );
// check for exact singularity
//magma_zgetmatrix( N, N, d_A, ldda, h_R, lda );
//h_R[ 10 + 10*lda ] = MAGMA_Z_MAKE( 0.0, 0.0 );
//magma_zsetmatrix( N, N, h_R, lda, d_A, ldda );
gpu_time = magma_wtime();
magma_zpotri_gpu( opts.uplo, N, d_A, ldda, &info );
gpu_time = magma_wtime() - gpu_time;
gpu_perf = gflops / gpu_time;
if (info != 0)
printf("magma_zpotri_gpu returned error %d: %s.\n",
(int) info, magma_strerror( info ));
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
if ( opts.lapack ) {
lapackf77_zpotrf( lapack_uplo_const(opts.uplo), &N, h_A, &lda, &info );
cpu_time = magma_wtime();
lapackf77_zpotri( lapack_uplo_const(opts.uplo), &N, h_A, &lda, &info );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
if (info != 0)
printf("lapackf77_zpotri returned error %d: %s.\n",
(int) info, magma_strerror( info ));
/* =====================================================================
Check the result compared to LAPACK
=================================================================== */
magma_zgetmatrix( N, N, d_A, ldda, h_R, lda );
error = lapackf77_zlange("f", &N, &N, h_A, &lda, work);
blasf77_zaxpy(&n2, &c_neg_one, h_A, &ione, h_R, &ione);
error = lapackf77_zlange("f", &N, &N, h_R, &lda, work) / error;
printf("%5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %s\n",
(int) N, cpu_perf, cpu_time, gpu_perf, gpu_time,
error, (error < tol ? "ok" : "failed") );
status += ! (error < tol);
}
else {
printf("%5d --- ( --- ) %7.2f (%7.2f) ---\n",
(int) N, gpu_perf, gpu_time );
}
TESTING_FREE_CPU( h_A );
TESTING_FREE_PIN( h_R );
TESTING_FREE_DEV( d_A );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing zlarfb_gpu
*/
int main( int argc, char** argv )
{
TESTING_INIT();
magmaDoubleComplex c_zero = MAGMA_Z_ZERO;
magmaDoubleComplex c_one = MAGMA_Z_ONE;
magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
magma_int_t M, N, K, size, ldc, ldv, ldt, ldw, nv;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
double error, work[1];
magma_int_t status = 0;
// test all combinations of input parameters
magma_side_t side [] = { MagmaLeft, MagmaRight };
magma_trans_t trans [] = { MagmaConjTrans, MagmaNoTrans };
magma_direct_t direct[] = { MagmaForward, MagmaBackward };
magma_storev_t storev[] = { MagmaColumnwise, MagmaRowwise };
magma_opts opts;
parse_opts( argc, argv, &opts );
double tol = opts.tolerance * lapackf77_dlamch("E");
printf(" M N K storev side direct trans ||R||_F / ||HC||_F\n");
printf("========================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
M = opts.msize[itest];
N = opts.nsize[itest];
K = opts.ksize[itest];
if ( M < K || N < K || K <= 0 ) {
printf( "%5d %5d %5d skipping because zlarfb requires M >= K, N >= K, K >= 0\n",
(int) M, (int) N, (int) K );
continue;
}
for( int istor = 0; istor < 2; ++istor ) {
for( int iside = 0; iside < 2; ++iside ) {
for( int idir = 0; idir < 2; ++idir ) {
for( int itran = 0; itran < 2; ++itran ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
ldc = ((M+31)/32)*32;
ldt = ((K+31)/32)*32;
ldw = (side[iside] == MagmaLeft ? N : M);
// (ldv, nv) get swapped later if rowwise
ldv = (side[iside] == MagmaLeft ? M : N);
nv = K;
// Allocate memory for matrices
magmaDoubleComplex *C, *R, *V, *T, *W;
TESTING_MALLOC_CPU( C, magmaDoubleComplex, ldc*N );
TESTING_MALLOC_CPU( R, magmaDoubleComplex, ldc*N );
TESTING_MALLOC_CPU( V, magmaDoubleComplex, ldv*K );
TESTING_MALLOC_CPU( T, magmaDoubleComplex, ldt*K );
TESTING_MALLOC_CPU( W, magmaDoubleComplex, ldw*K );
magmaDoubleComplex *dC, *dV, *dT, *dW;
TESTING_MALLOC_DEV( dC, magmaDoubleComplex, ldc*N );
TESTING_MALLOC_DEV( dV, magmaDoubleComplex, ldv*K );
TESTING_MALLOC_DEV( dT, magmaDoubleComplex, ldt*K );
TESTING_MALLOC_DEV( dW, magmaDoubleComplex, ldw*K );
// C is M x N.
size = ldc*N;
lapackf77_zlarnv( &ione, ISEED, &size, C );
//printf( "C=" ); magma_zprint( M, N, C, ldc );
// V is ldv x nv. See larfb docs for description.
// if column-wise and left, M x K
// if column-wise and right, N x K
// if row-wise and left, K x M
// if row-wise and right, K x N
size = ldv*nv;
lapackf77_zlarnv( &ione, ISEED, &size, V );
if ( storev[istor] == MagmaColumnwise ) {
if ( direct[idir] == MagmaForward ) {
lapackf77_zlaset( MagmaUpperStr, &K, &K, &c_zero, &c_one, V, &ldv );
}
else {
lapackf77_zlaset( MagmaLowerStr, &K, &K, &c_zero, &c_one, &V[(ldv-K)], &ldv );
}
}
else {
// rowwise, swap V's dimensions
std::swap( ldv, nv );
if ( direct[idir] == MagmaForward ) {
lapackf77_zlaset( MagmaLowerStr, &K, &K, &c_zero, &c_one, V, &ldv );
}
else {
lapackf77_zlaset( MagmaUpperStr, &K, &K, &c_zero, &c_one, &V[(nv-K)*ldv], &ldv );
}
}
//printf( "# ldv %d, nv %d\n", ldv, nv );
//printf( "V=" ); magma_zprint( ldv, nv, V, ldv );
// T is K x K, upper triangular for forward, and lower triangular for backward
magma_int_t k1 = K-1;
size = ldt*K;
lapackf77_zlarnv( &ione, ISEED, &size, T );
if ( direct[idir] == MagmaForward ) {
lapackf77_zlaset( MagmaLowerStr, &k1, &k1, &c_zero, &c_zero, &T[1], &ldt );
}
else {
lapackf77_zlaset( MagmaUpperStr, &k1, &k1, &c_zero, &c_zero, &T[1*ldt], &ldt );
}
//printf( "T=" ); magma_zprint( K, K, T, ldt );
magma_zsetmatrix( M, N, C, ldc, dC, ldc );
magma_zsetmatrix( ldv, nv, V, ldv, dV, ldv );
magma_zsetmatrix( K, K, T, ldt, dT, ldt );
lapackf77_zlarfb( lapack_side_const( side[iside] ), lapack_trans_const( trans[itran] ),
lapack_direct_const( direct[idir] ), lapack_storev_const( storev[istor] ),
&M, &N, &K,
V, &ldv, T, &ldt, C, &ldc, W, &ldw );
//printf( "HC=" ); magma_zprint( M, N, C, ldc );
magma_zlarfb_gpu( side[iside], trans[itran], direct[idir], storev[istor],
M, N, K,
dV, ldv, dT, ldt, dC, ldc, dW, ldw );
magma_zgetmatrix( M, N, dC, ldc, R, ldc );
//printf( "dHC=" ); magma_zprint( M, N, R, ldc );
// compute relative error |HC_magma - HC_lapack| / |HC_lapack|
error = lapackf77_zlange( "Fro", &M, &N, C, &ldc, work );
size = ldc*N;
blasf77_zaxpy( &size, &c_neg_one, C, &ione, R, &ione );
error = lapackf77_zlange( "Fro", &M, &N, R, &ldc, work ) / error;
printf( "%5d %5d %5d %c %c %c %c %8.2e %s\n",
(int) M, (int) N, (int) K,
lapacke_storev_const(storev[istor]), lapacke_side_const(side[iside]),
lapacke_direct_const(direct[idir]), lapacke_trans_const(trans[itran]),
error, (error < tol ? "ok" : "failed") );
status += ! (error < tol);
TESTING_FREE_CPU( C );
TESTING_FREE_CPU( R );
TESTING_FREE_CPU( V );
TESTING_FREE_CPU( T );
TESTING_FREE_CPU( W );
TESTING_FREE_DEV( dC );
TESTING_FREE_DEV( dV );
TESTING_FREE_DEV( dT );
TESTING_FREE_DEV( dW );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}}}}
printf( "\n" );
}
TESTING_FINALIZE();
return status;
}
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 M, N, Xm, Ym, lda, sizeA, sizeX, sizeY;
magma_int_t incx = 1;
magma_int_t incy = 1;
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;
magma_int_t status = 0;
magma_opts opts;
parse_opts( argc, argv, &opts );
float tol = opts.tolerance * lapackf77_slamch("E");
printf("trans = %s\n", lapack_trans_const(opts.transA) );
printf(" M N MAGMA Gflop/s (ms) CUBLAS Gflop/s (ms) CPU Gflop/s (ms) MAGMA error CUBLAS error\n");
printf("===================================================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
M = opts.msize[itest];
N = opts.nsize[itest];
lda = ((M+31)/32)*32;
gflops = FLOPS_CGEMV( M, N ) / 1e9;
if ( opts.transA == MagmaNoTrans ) {
Xm = N;
Ym = M;
} else {
Xm = M;
Ym = N;
}
sizeA = lda*N;
sizeX = incx*Xm;
sizeY = incy*Ym;
TESTING_MALLOC_CPU( A, magmaFloatComplex, sizeA );
TESTING_MALLOC_CPU( X, magmaFloatComplex, sizeX );
TESTING_MALLOC_CPU( Y, magmaFloatComplex, sizeY );
TESTING_MALLOC_CPU( Ycublas, magmaFloatComplex, sizeY );
TESTING_MALLOC_CPU( Ymagma, magmaFloatComplex, sizeY );
TESTING_MALLOC_DEV( dA, magmaFloatComplex, sizeA );
TESTING_MALLOC_DEV( dX, magmaFloatComplex, sizeX );
TESTING_MALLOC_DEV( dY, magmaFloatComplex, sizeY );
/* Initialize the matrix */
lapackf77_clarnv( &ione, ISEED, &sizeA, A );
lapackf77_clarnv( &ione, ISEED, &sizeX, X );
lapackf77_clarnv( &ione, ISEED, &sizeY, Y );
/* =====================================================================
Performs operation using CUBLAS
=================================================================== */
magma_csetmatrix( M, N, A, lda, dA, lda );
magma_csetvector( Xm, X, incx, dX, incx );
magma_csetvector( Ym, Y, incy, dY, incy );
cublas_time = magma_sync_wtime( 0 );
cublasCgemv( handle, cublas_trans_const(opts.transA),
M, N, &alpha, dA, lda, dX, incx, &beta, dY, incy );
cublas_time = magma_sync_wtime( 0 ) - cublas_time;
cublas_perf = gflops / cublas_time;
magma_cgetvector( Ym, dY, incy, Ycublas, incy );
/* =====================================================================
Performs operation using MAGMABLAS
=================================================================== */
magma_csetvector( Ym, Y, incy, dY, incy );
magma_time = magma_sync_wtime( 0 );
magmablas_cgemv( opts.transA, M, N, alpha, dA, lda, dX, incx, beta, dY, incy );
magma_time = magma_sync_wtime( 0 ) - magma_time;
magma_perf = gflops / magma_time;
magma_cgetvector( Ym, dY, incx, Ymagma, incx );
/* =====================================================================
Performs operation using CPU BLAS
=================================================================== */
cpu_time = magma_wtime();
blasf77_cgemv( lapack_trans_const(opts.transA), &M, &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( &Ym, &c_neg_one, Y, &incy, Ymagma, &incy );
magma_error = lapackf77_clange( "M", &Ym, &ione, Ymagma, &Ym, work ) / Ym;
blasf77_caxpy( &Ym, &c_neg_one, Y, &incy, Ycublas, &incy );
cublas_error = lapackf77_clange( "M", &Ym, &ione, Ycublas, &Ym, work ) / Ym;
printf("%5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %8.2e %s\n",
(int) M, (int) N,
magma_perf, 1000.*magma_time,
cublas_perf, 1000.*cublas_time,
cpu_perf, 1000.*cpu_time,
magma_error, cublas_error,
(magma_error < tol && cublas_error < tol ? "ok" : "failed"));
status += ! (magma_error < tol && cublas_error < tol);
TESTING_FREE_CPU( A );
TESTING_FREE_CPU( X );
TESTING_FREE_CPU( Y );
TESTING_FREE_CPU( Ycublas );
TESTING_FREE_CPU( Ymagma );
TESTING_FREE_DEV( dA );
TESTING_FREE_DEV( dX );
TESTING_FREE_DEV( dY );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing csymmetrize
Code is very similar to testing_ctranspose.cpp
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gbytes, gpu_perf, gpu_time, cpu_perf, cpu_time;
float error, work[1];
magmaFloatComplex c_neg_one = MAGMA_C_NEG_ONE;
magmaFloatComplex *h_A, *h_R;
magmaFloatComplex_ptr d_A;
magma_int_t i, j, N, nb, size, lda, ldda, mstride, nstride, ntile, tile, offset;
magma_int_t ione = 1;
magma_int_t status = 0;
magma_opts opts;
opts.parse_opts( argc, argv );
nb = (opts.nb == 0 ? 64 : opts.nb);
mstride = 2*nb;
nstride = 3*nb;
printf("%% uplo = %s, nb = %d, mstride = %d, nstride = %d\n",
lapack_uplo_const(opts.uplo), (int) nb, (int) mstride, (int) nstride );
printf("%% N ntile CPU GByte/s (ms) GPU GByte/s (ms) check\n");
printf("%%==========================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
N = opts.nsize[itest];
lda = N;
ldda = magma_roundup( N, opts.align ); // multiple of 32 by default
size = lda*N;
if ( N < nb ) {
ntile = 0;
} else {
ntile = min( (N - nb)/mstride + 1,
(N - nb)/nstride + 1 );
}
// load each tile, save each tile
gbytes = sizeof(magmaFloatComplex) * 2.*nb*nb*ntile / 1e9;
TESTING_MALLOC_CPU( h_A, magmaFloatComplex, size );
TESTING_MALLOC_CPU( h_R, magmaFloatComplex, size );
TESTING_MALLOC_DEV( d_A, magmaFloatComplex, ldda*N );
/* Initialize the matrix */
for( j = 0; j < N; ++j ) {
for( i = 0; i < N; ++i ) {
h_A[i + j*lda] = MAGMA_C_MAKE( i + j/10000., j );
}
}
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
magma_csetmatrix( N, N, h_A, lda, d_A, ldda, opts.queue );
gpu_time = magma_sync_wtime( opts.queue );
magmablas_csymmetrize_tiles( opts.uplo, nb, d_A, ldda, ntile, mstride, nstride, opts.queue );
gpu_time = magma_sync_wtime( opts.queue ) - gpu_time;
gpu_perf = gbytes / gpu_time;
/* =====================================================================
Performs operation using naive in-place algorithm
(LAPACK doesn't implement symmetrize)
=================================================================== */
cpu_time = magma_wtime();
for( tile = 0; tile < ntile; ++tile ) {
offset = tile*mstride + tile*nstride*lda;
for( j = 0; j < nb; ++j ) {
for( i = 0; i < j; ++i ) {
if ( opts.uplo == MagmaLower ) {
h_A[offset + i + j*lda] = MAGMA_C_CONJ( h_A[offset + j + i*lda] );
}
else {
h_A[offset + j + i*lda] = MAGMA_C_CONJ( h_A[offset + i + j*lda] );
}
}
}
}
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gbytes / cpu_time;
/* =====================================================================
Check the result
=================================================================== */
magma_cgetmatrix( N, N, d_A, ldda, h_R, lda, opts.queue );
blasf77_caxpy(&size, &c_neg_one, h_A, &ione, h_R, &ione);
error = lapackf77_clange("f", &N, &N, h_R, &lda, work);
printf("%5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %s\n",
(int) N, (int) ntile,
cpu_perf, cpu_time*1000., gpu_perf, gpu_time*1000.,
(error == 0. ? "ok" : "failed") );
status += ! (error == 0.);
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_R );
TESTING_FREE_DEV( d_A );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
opts.cleanup();
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing chesv_nopiv_gpu
*/
int main(int argc, char **argv)
{
TESTING_INIT();
real_Double_t gflops, cpu_perf, cpu_time, gpu_perf, gpu_time;
float error, Rnorm, Anorm, Xnorm, *work;
magmaFloatComplex c_one = MAGMA_C_ONE;
magmaFloatComplex c_neg_one = MAGMA_C_NEG_ONE;
magmaFloatComplex *h_A, *h_B, *h_X, temp, *hwork;
magmaFloatComplex_ptr d_A, d_B;
magma_int_t *ipiv;
magma_int_t N, nrhs, lda, ldb, ldda, lddb, info, sizeA, sizeB, lwork;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t status = 0;
magma_opts opts;
parse_opts( argc, argv, &opts );
float tol = opts.tolerance * lapackf77_slamch("E");
nrhs = opts.nrhs;
printf(" N NRHS CPU GFlop/s (sec) GPU GFlop/s (sec) ||B - AX|| / N*||A||*||X||\n");
printf("================================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
N = opts.nsize[itest];
lda = N;
ldb = lda;
ldda = ((N+31)/32)*32;
lddb = ldda;
gflops = ( FLOPS_CGETRF( N, N ) + FLOPS_CGETRS( N, nrhs ) ) / 1e9;
TESTING_MALLOC_CPU( h_A, magmaFloatComplex, lda*N );
TESTING_MALLOC_CPU( h_B, magmaFloatComplex, ldb*nrhs );
TESTING_MALLOC_CPU( h_X, magmaFloatComplex, ldb*nrhs );
TESTING_MALLOC_CPU( work, float, N );
TESTING_MALLOC_CPU( ipiv, magma_int_t, N );
TESTING_MALLOC_DEV( d_A, magmaFloatComplex, ldda*N );
TESTING_MALLOC_DEV( d_B, magmaFloatComplex, lddb*nrhs );
/* Initialize the matrices */
sizeA = lda*N;
sizeB = ldb*nrhs;
lapackf77_clarnv( &ione, ISEED, &sizeA, h_A );
lapackf77_clarnv( &ione, ISEED, &sizeB, h_B );
bool nopiv = true;
if ( nopiv ) {
magma_cmake_hpd( N, h_A, lda ); // SPD / HPD does not require pivoting
}
else {
magma_cmake_hermitian( N, h_A, lda ); // symmetric/Hermitian generally requires pivoting
}
magma_csetmatrix( N, N, h_A, lda, d_A, ldda );
magma_csetmatrix( N, nrhs, h_B, ldb, d_B, lddb );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
gpu_time = magma_wtime();
magma_chesv_nopiv_gpu( opts.uplo, N, nrhs, d_A, ldda, d_B, lddb, &info );
gpu_time = magma_wtime() - gpu_time;
gpu_perf = gflops / gpu_time;
if (info != 0)
printf("magma_cgesv_gpu returned error %d: %s.\n",
(int) info, magma_strerror( info ));
//=====================================================================
// Residual
//=====================================================================
magma_cgetmatrix( N, nrhs, d_B, lddb, h_X, ldb );
Anorm = lapackf77_clange("I", &N, &N, h_A, &lda, work);
Xnorm = lapackf77_clange("I", &N, &nrhs, h_X, &ldb, work);
blasf77_cgemm( MagmaNoTransStr, MagmaNoTransStr, &N, &nrhs, &N,
&c_one, h_A, &lda,
h_X, &ldb,
&c_neg_one, h_B, &ldb);
Rnorm = lapackf77_clange("I", &N, &nrhs, h_B, &ldb, work);
error = Rnorm/(N*Anorm*Xnorm);
status += ! (error < tol);
/* ====================================================================
Performs operation using LAPACK
=================================================================== */
if ( opts.lapack ) {
lwork = -1;
lapackf77_chesv( lapack_uplo_const(opts.uplo), &N,&nrhs,
h_A, &lda, ipiv, h_B, &ldb, &temp, &lwork, &info );
lwork = (magma_int_t) MAGMA_C_REAL( temp );
TESTING_MALLOC_PIN( hwork, magmaFloatComplex, lwork );
cpu_time = magma_wtime();
lapackf77_chesv( lapack_uplo_const(opts.uplo), &N, &nrhs,
h_A, &lda, ipiv, h_B, &ldb, hwork, &lwork, &info );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
if (info != 0)
printf("lapackf77_chesv returned error %d: %s.\n",
(int) info, magma_strerror( info ));
printf( "%5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %s\n",
(int) N, (int) nrhs, cpu_perf, cpu_time, gpu_perf, gpu_time,
error, (error < tol ? "ok" : "failed"));
TESTING_FREE_CPU( hwork );
}
else {
printf( "%5d %5d --- ( --- ) %7.2f (%7.2f) %8.2e %s\n",
(int) N, (int) nrhs, gpu_perf, gpu_time,
error, (error < tol ? "ok" : "failed"));
}
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_B );
TESTING_FREE_CPU( h_X );
TESTING_FREE_CPU( work );
TESTING_FREE_CPU( ipiv );
TESTING_FREE_DEV( d_A );
TESTING_FREE_DEV( d_B );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return status;
}
int main( int argc, char** argv)
{
real_Double_t gflops, gpu_perf, cpu_perf, gpu_time, cpu_time;
magmaDoubleComplex *hA, *hR;
magmaDoubleComplex_ptr dA;
magma_int_t N = 0, n2, lda, ldda;
magma_int_t size[10] =
{ 1000, 2000, 3000, 4000, 5000, 6000, 7000, 8000, 8160, 8192 };
magma_int_t i, info;
magmaDoubleComplex mz_one = MAGMA_Z_NEG_ONE;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
double work[1], matnorm, diffnorm;
if (argc != 1){
for(i = 1; i<argc; i++){
if (strcmp("-N", argv[i])==0)
N = atoi(argv[++i]);
}
if (N>0) size[0] = size[9] = N;
else exit(1);
}
else {
printf("\nUsage: \n");
printf(" testing_zpotrf2_gpu -N %d\n\n", 1024);
}
/* Initialize */
magma_queue_t queue1, queue2;
magma_device_t device;
magma_int_t num = 0;
magma_int_t err;
magma_init();
err = magma_getdevices( &device, 2, &num );
if ( err != 0 or num < 1 ) {
fprintf( stderr, "magma_getdevices failed: %d\n", (int) err );
exit(-1);
}
err = magma_queue_create( device, &queue1 );
if ( err != 0 ) {
fprintf( stderr, "magma_queue_create failed: %d\n", (int) err );
exit(-1);
}
err = magma_queue_create( device, &queue2 );
if ( err != 0 ) {
fprintf( stderr, "magma_queue_create failed: %d\n", (int) err );
exit(-1);
}
magma_queue_t queues[2] = {queue1, queue2};
/* Allocate memory for the largest matrix */
N = size[9];
n2 = N * N;
ldda = ((N+31)/32) * 32;
TESTING_MALLOC_CPU( hA, magmaDoubleComplex, n2 );
TESTING_MALLOC_PIN( hR, magmaDoubleComplex, n2 );
TESTING_MALLOC_DEV( dA, magmaDoubleComplex, ldda*N );
printf("\n\n");
printf(" N CPU GFlop/s (sec) GPU GFlop/s (sec) ||R_magma-R_lapack||_F / ||R_lapack||_F\n");
printf("========================================================================================\n");
for(i=0; i<10; i++){
N = size[i];
lda = N;
n2 = lda*N;
ldda = ((N+31)/32)*32;
gflops = FLOPS( (double)N ) * 1e-9;
/* Initialize the matrix */
lapackf77_zlarnv( &ione, ISEED, &n2, hA );
/* Symmetrize and increase the diagonal */
for( int i = 0; i < N; ++i ) {
hA(i,i) = MAGMA_Z_MAKE( MAGMA_Z_REAL(hA(i,i)) + N, 0 );
for( int j = 0; j < i; ++j ) {
hA(i, j) = MAGMA_Z_CNJG( hA(j,i) );
}
}
lapackf77_zlacpy( MagmaFullStr, &N, &N, hA, &lda, hR, &lda );
/* Warm up to measure the performance */
magma_zsetmatrix( N, N, hA, lda, dA, 0, ldda, queue1);
clFinish(queue1);
magma_zpotrf2_gpu( MagmaLower, N, dA, 0, ldda, queues, &info );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
magma_zsetmatrix( N, N, hA, lda, dA, 0, ldda, queue1 );
clFinish(queue1);
gpu_time = magma_wtime();
magma_zpotrf2_gpu( MagmaLower, N, dA, 0, ldda, queues, &info );
gpu_time = magma_wtime() - gpu_time;
if (info != 0)
printf( "magma_zpotrf2 had error %d.\n", info );
gpu_perf = gflops / gpu_time;
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
cpu_time = magma_wtime();
lapackf77_zpotrf( MagmaLowerStr, &N, hA, &lda, &info );
cpu_time = magma_wtime() - cpu_time;
if (info != 0)
printf( "lapackf77_zpotrf had error %d.\n", info );
cpu_perf = gflops / cpu_time;
/* =====================================================================
Check the result compared to LAPACK
|R_magma - R_lapack| / |R_lapack|
=================================================================== */
magma_zgetmatrix( N, N, dA, 0, ldda, hR, lda, queue1 );
matnorm = lapackf77_zlange("f", &N, &N, hA, &lda, work);
blasf77_zaxpy(&n2, &mz_one, hA, &ione, hR, &ione);
diffnorm = lapackf77_zlange("f", &N, &N, hR, &lda, work);
printf( "%5d %6.2f (%6.2f) %6.2f (%6.2f) %e\n",
N, cpu_perf, cpu_time, gpu_perf, gpu_time, diffnorm / matnorm );
if (argc != 1)
break;
}
/* clean up */
TESTING_FREE_CPU( hA );
TESTING_FREE_PIN( hR );
TESTING_FREE_DEV( dA );
magma_queue_destroy( queue1 );
magma_queue_destroy( queue2 );
magma_finalize();
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing zgetrf_mgpu
*/
int main( int argc, char** argv )
{
TESTING_INIT();
real_Double_t gflops, gpu_perf, gpu_time, cpu_perf=0, cpu_time=0;
double error;
magmaDoubleComplex *h_A;
magmaDoubleComplex_ptr d_lA[ MagmaMaxGPUs ];
magma_int_t *ipiv;
magma_int_t M, N, n2, lda, ldda, n_local, ngpu;
magma_int_t info, min_mn, nb, ldn_local;
magma_int_t status = 0;
magma_opts opts;
opts.parse_opts( argc, argv );
opts.ngpu = abs( opts.ngpu ); // always uses multi-GPU code
double tol = opts.tolerance * lapackf77_dlamch("E");
magma_queue_t queues[MagmaMaxGPUs];
for( int dev = 0; dev < opts.ngpu; ++dev ) {
magma_queue_create( dev, &queues[dev] );
}
printf("%% ngpu %d\n", (int) opts.ngpu );
if ( opts.check == 2 ) {
printf("%% M N CPU Gflop/s (sec) GPU Gflop/s (sec) |Ax-b|/(N*|A|*|x|)\n");
}
else {
printf("%% M N CPU Gflop/s (sec) GPU Gflop/s (sec) |PA-LU|/(N*|A|)\n");
}
printf("%%========================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
M = opts.msize[itest];
N = opts.nsize[itest];
min_mn = min(M, N);
lda = M;
n2 = lda*N;
ldda = magma_roundup( M, opts.align ); // multiple of 32 by default
nb = magma_get_zgetrf_nb( M, N );
gflops = FLOPS_ZGETRF( M, N ) / 1e9;
// ngpu must be at least the number of blocks
ngpu = min( opts.ngpu, magma_ceildiv(N,nb) );
if ( ngpu < opts.ngpu ) {
printf( " * too many GPUs for the matrix size, using %d GPUs\n", (int) ngpu );
}
// Allocate host memory for the matrix
TESTING_MALLOC_CPU( ipiv, magma_int_t, min_mn );
TESTING_MALLOC_CPU( h_A, magmaDoubleComplex, n2 );
// Allocate device memory
for( int dev=0; dev < ngpu; dev++ ) {
n_local = ((N/nb)/ngpu)*nb;
if (dev < (N/nb) % ngpu)
n_local += nb;
else if (dev == (N/nb) % ngpu)
n_local += N % nb;
ldn_local = magma_roundup( n_local, opts.align ); // multiple of 32 by default // TODO why?
magma_setdevice( dev );
TESTING_MALLOC_DEV( d_lA[dev], magmaDoubleComplex, ldda*ldn_local );
}
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
if ( opts.lapack ) {
init_matrix( M, N, h_A, lda );
cpu_time = magma_wtime();
lapackf77_zgetrf( &M, &N, h_A, &lda, ipiv, &info );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
if (info != 0) {
printf("lapackf77_zgetrf returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
}
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
init_matrix( M, N, h_A, lda );
magma_zsetmatrix_1D_col_bcyclic( M, N, h_A, lda, d_lA, ldda, ngpu, nb, queues );
gpu_time = magma_wtime();
magma_zgetrf_mgpu( ngpu, M, N, d_lA, ldda, ipiv, &info );
gpu_time = magma_wtime() - gpu_time;
gpu_perf = gflops / gpu_time;
if (info != 0) {
printf("magma_zgetrf_mgpu returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
magma_zgetmatrix_1D_col_bcyclic( M, N, d_lA, ldda, h_A, lda, ngpu, nb, queues );
/* =====================================================================
Check the factorization
=================================================================== */
if ( opts.lapack ) {
printf("%5d %5d %7.2f (%7.2f) %7.2f (%7.2f)",
(int) M, (int) N, cpu_perf, cpu_time, gpu_perf, gpu_time );
}
else {
printf("%5d %5d --- ( --- ) %7.2f (%7.2f)",
(int) M, (int) N, gpu_perf, gpu_time );
}
if ( opts.check == 2 ) {
error = get_residual( M, N, h_A, lda, ipiv );
printf(" %8.2e %s\n", error, (error < tol ? "ok" : "failed"));
status += ! (error < tol);
}
else if ( opts.check ) {
error = get_LU_error( M, N, h_A, lda, ipiv );
printf(" %8.2e %s\n", error, (error < tol ? "ok" : "failed"));
status += ! (error < tol);
}
else {
printf( " ---\n" );
}
TESTING_FREE_CPU( ipiv );
TESTING_FREE_CPU( h_A );
for( int dev=0; dev < ngpu; dev++ ) {
magma_setdevice( dev );
TESTING_FREE_DEV( d_lA[dev] );
}
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
for( int dev = 0; dev < opts.ngpu; ++dev ) {
magma_queue_destroy( queues[dev] );
}
opts.cleanup();
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing cgeqrf
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gflops, gpu_perf, gpu_time, cpu_perf, cpu_time;
float error, work[1];
magmaFloatComplex c_neg_one = MAGMA_C_NEG_ONE;
magmaFloatComplex *h_A, *h_T, *h_R, *tau, *h_work, tmp[1];
magmaFloatComplex *d_A, *d_T, *ddA, *dtau;
magmaFloatComplex *d_A2, *d_T2, *ddA2, *dtau2;
float *dwork, *dwork2;
magma_int_t M, N, lda, ldda, lwork, n2, info, min_mn;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t status = 0;
#define BLOCK_SIZE 64
magma_opts opts;
parse_opts( argc, argv, &opts );
float tol = 10. * opts.tolerance * lapackf77_slamch("E");
magma_queue_t stream[2];
magma_queue_create( &stream[0] );
magma_queue_create( &stream[1] );
printf("version %d\n", (int) opts.version );
printf(" M N CPU GFlop/s (ms) GPU GFlop/s (ms) ||R||_F/||A||_F ||R_T||\n");
printf("=============================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
M = opts.msize[itest];
N = opts.nsize[itest];
if (N > 128) {
printf("%5d %5d skipping because cgeqr2x requires N <= 128\n",
(int) M, (int) N);
continue;
}
if (M < N) {
printf("%5d %5d skipping because cgeqr2x requires M >= N\n",
(int) M, (int) N);
continue;
}
min_mn = min(M, N);
lda = M;
n2 = lda*N;
ldda = ((M+31)/32)*32;
gflops = (FLOPS_CGEQRF( M, N ) + FLOPS_CGEQRT( M, N )) / 1e9;
/* Allocate memory for the matrix */
TESTING_MALLOC_CPU( tau, magmaFloatComplex, min_mn );
TESTING_MALLOC_CPU( h_A, magmaFloatComplex, n2 );
TESTING_MALLOC_CPU( h_T, magmaFloatComplex, N*N );
TESTING_MALLOC_PIN( h_R, magmaFloatComplex, n2 );
TESTING_MALLOC_DEV( d_A, magmaFloatComplex, ldda*N );
TESTING_MALLOC_DEV( d_T, magmaFloatComplex, N*N );
TESTING_MALLOC_DEV( ddA, magmaFloatComplex, N*N );
TESTING_MALLOC_DEV( dtau, magmaFloatComplex, min_mn );
TESTING_MALLOC_DEV( d_A2, magmaFloatComplex, ldda*N );
TESTING_MALLOC_DEV( d_T2, magmaFloatComplex, N*N );
TESTING_MALLOC_DEV( ddA2, magmaFloatComplex, N*N );
TESTING_MALLOC_DEV( dtau2, magmaFloatComplex, min_mn );
TESTING_MALLOC_DEV( dwork, float, max(5*min_mn, (BLOCK_SIZE*2+2)*min_mn) );
TESTING_MALLOC_DEV( dwork2, float, max(5*min_mn, (BLOCK_SIZE*2+2)*min_mn) );
// todo replace with magma_claset
cudaMemset(ddA, 0, N*N*sizeof(magmaFloatComplex));
cudaMemset(d_T, 0, N*N*sizeof(magmaFloatComplex));
cudaMemset(ddA2, 0, N*N*sizeof(magmaFloatComplex));
cudaMemset(d_T2, 0, N*N*sizeof(magmaFloatComplex));
lwork = -1;
lapackf77_cgeqrf(&M, &N, NULL, &M, NULL, tmp, &lwork, &info);
lwork = (magma_int_t)MAGMA_C_REAL( tmp[0] );
lwork = max(lwork, N*N);
TESTING_MALLOC_CPU( h_work, magmaFloatComplex, lwork );
/* Initialize the matrix */
lapackf77_clarnv( &ione, ISEED, &n2, h_A );
lapackf77_clacpy( MagmaUpperLowerStr, &M, &N, h_A, &lda, h_R, &lda );
magma_csetmatrix( M, N, h_R, lda, d_A, ldda );
magma_csetmatrix( M, N, h_R, lda, d_A2, ldda );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
gpu_time = magma_sync_wtime(0);
if (opts.version == 1)
magma_cgeqr2x_gpu(M, N, d_A, ldda, dtau, d_T, ddA, dwork, &info);
else if (opts.version == 2)
magma_cgeqr2x2_gpu(M, N, d_A, ldda, dtau, d_T, ddA, dwork, &info);
else if (opts.version == 3)
magma_cgeqr2x3_gpu(M, N, d_A, ldda, dtau, d_T, ddA, dwork, &info);
else {
printf( "call magma_cgeqr2x4_gpu\n" );
/*
Going through NULL stream is faster
Going through any stream is slower
Doing two streams in parallel is slower than doing them sequentially
Queuing happens on the NULL stream - user defined buffers are smaller?
*/
magma_cgeqr2x4_gpu(M, N, d_A, ldda, dtau, d_T, ddA, dwork, &info, NULL);
//magma_cgeqr2x4_gpu(M, N, d_A, ldda, dtau, d_T, ddA, dwork, &info, stream[1]);
//magma_cgeqr2x4_gpu(M, N, d_A2, ldda, dtau2, d_T2, ddA2, dwork2, &info, stream[0]);
//magma_cgeqr2x4_gpu(M, N, d_A2, ldda, dtau2, d_T2, ddA2, dwork2, &info, NULL);
//gflops *= 2;
}
gpu_time = magma_sync_wtime(0) - gpu_time;
gpu_perf = gflops / gpu_time;
if (info != 0) {
printf("magma_cgeqr2x_gpu version %d returned error %d: %s.\n",
(int) opts.version, (int) info, magma_strerror( info ));
}
else {
if ( opts.check ) {
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
cpu_time = magma_wtime();
lapackf77_cgeqrf(&M, &N, h_A, &lda, tau, h_work, &lwork, &info);
lapackf77_clarft( MagmaForwardStr, MagmaColumnwiseStr,
&M, &N, h_A, &lda, tau, h_work, &N);
//magma_cgeqr2(&M, &N, h_A, &lda, tau, h_work, &info);
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
if (info != 0)
printf("lapackf77_cgeqrf returned error %d: %s.\n",
(int) info, magma_strerror( info ));
/* =====================================================================
Check the result compared to LAPACK
=================================================================== */
magma_cgetmatrix( M, N, d_A, ldda, h_R, M );
magma_cgetmatrix( N, N, ddA, N, h_T, N );
// Restore the upper triangular part of A before the check
for(int col=0; col < N; col++){
for(int row=0; row <= col; row++)
h_R[row + col*M] = h_T[row + col*N];
}
error = lapackf77_clange("M", &M, &N, h_A, &lda, work);
blasf77_caxpy(&n2, &c_neg_one, h_A, &ione, h_R, &ione);
error = lapackf77_clange("M", &M, &N, h_R, &lda, work) / (N * error);
// Check if T is the same
magma_cgetmatrix( N, N, d_T, N, h_T, N );
float terr = 0.;
for(int col=0; col < N; col++)
for(int row=0; row <= col; row++)
terr += ( MAGMA_C_ABS(h_work[row + col*N] - h_T[row + col*N])*
MAGMA_C_ABS(h_work[row + col*N] - h_T[row + col*N]) );
terr = magma_ssqrt(terr);
printf("%5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %8.2e %s\n",
(int) M, (int) N, cpu_perf, 1000.*cpu_time, gpu_perf, 1000.*gpu_time,
error, terr, (error < tol ? "ok" : "failed") );
status += ! (error < tol);
}
else {
printf("%5d %5d --- ( --- ) %7.2f (%7.2f) --- \n",
(int) M, (int) N, gpu_perf, 1000.*gpu_time);
}
}
TESTING_FREE_CPU( tau );
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_T );
TESTING_FREE_CPU( h_work );
TESTING_FREE_PIN( h_R );
TESTING_FREE_DEV( d_A );
TESTING_FREE_DEV( d_T );
TESTING_FREE_DEV( ddA );
TESTING_FREE_DEV( dtau );
TESTING_FREE_DEV( dwork );
TESTING_FREE_DEV( d_A2 );
TESTING_FREE_DEV( d_T2 );
TESTING_FREE_DEV( ddA2 );
TESTING_FREE_DEV( dtau2 );
TESTING_FREE_DEV( dwork2 );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
magma_queue_destroy( stream[0] );
magma_queue_destroy( stream[1] );
TESTING_FINALIZE();
return status;
}
int main(int argc, char **argv)
{
TESTING_INIT();
real_Double_t gflops, magma_perf, magma_time, cublas_perf, cublas_time, cpu_perf, cpu_time;
double 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, ldda, sizeA, sizeX, sizeY, blocks, ldwork;
magma_int_t incx = 1;
magma_int_t incy = 1;
magma_int_t nb = 64;
magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
magmaDoubleComplex alpha = MAGMA_Z_MAKE( 1.5, -2.3 );
magmaDoubleComplex beta = MAGMA_Z_MAKE( -0.6, 0.8 );
magmaDoubleComplex *A, *X, *Y, *Ycublas, *Ymagma;
magmaDoubleComplex *dA, *dX, *dY, *dwork;
magma_int_t status = 0;
magma_opts opts;
parse_opts( argc, argv, &opts );
double tol = opts.tolerance * lapackf77_dlamch("E");
printf("uplo = %s\n", lapack_uplo_const(opts.uplo) );
if ( opts.uplo == MagmaUpper ) {
printf("** for uplo=MagmaUpper, magmablas_zhemv simply calls cublas_zhemv.\n");
}
printf(" N MAGMA Gflop/s (ms) CUBLAS Gflop/s (ms) CPU Gflop/s (ms) MAGMA error 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];
lda = N;
ldda = ((N + 31)/32)*32;
sizeA = N*lda;
sizeX = N*incx;
sizeY = N*incy;
gflops = FLOPS_ZHEMV( N ) / 1e9;
TESTING_MALLOC_CPU( A, magmaDoubleComplex, sizeA );
TESTING_MALLOC_CPU( X, magmaDoubleComplex, sizeX );
TESTING_MALLOC_CPU( Y, magmaDoubleComplex, sizeY );
TESTING_MALLOC_CPU( Ycublas, magmaDoubleComplex, sizeY );
TESTING_MALLOC_CPU( Ymagma, magmaDoubleComplex, sizeY );
TESTING_MALLOC_DEV( dA, magmaDoubleComplex, ldda*N );
TESTING_MALLOC_DEV( dX, magmaDoubleComplex, sizeX );
TESTING_MALLOC_DEV( dY, magmaDoubleComplex, sizeY );
blocks = (N + nb - 1) / nb;
ldwork = ldda*blocks;
TESTING_MALLOC_DEV( dwork, magmaDoubleComplex, ldwork );
magmablas_zlaset( MagmaFull, ldwork, 1, MAGMA_Z_NAN, MAGMA_Z_NAN, dwork, ldwork );
magmablas_zlaset( MagmaFull, ldda, N, MAGMA_Z_NAN, MAGMA_Z_NAN, dA, ldda );
/* Initialize the matrix */
lapackf77_zlarnv( &ione, ISEED, &sizeA, A );
magma_zmake_hermitian( N, A, lda );
lapackf77_zlarnv( &ione, ISEED, &sizeX, X );
lapackf77_zlarnv( &ione, ISEED, &sizeY, Y );
/* =====================================================================
Performs operation using CUBLAS
=================================================================== */
magma_zsetmatrix( N, N, A, lda, dA, ldda );
magma_zsetvector( N, X, incx, dX, incx );
magma_zsetvector( N, Y, incy, dY, incy );
cublas_time = magma_sync_wtime( 0 );
cublasZhemv( handle, cublas_uplo_const(opts.uplo),
N, &alpha, dA, ldda, dX, incx, &beta, dY, incy );
cublas_time = magma_sync_wtime( 0 ) - cublas_time;
cublas_perf = gflops / cublas_time;
magma_zgetvector( N, dY, incy, Ycublas, incy );
/* =====================================================================
Performs operation using MAGMABLAS
=================================================================== */
magma_zsetvector( N, Y, incy, dY, incy );
//magma_zprint_gpu( ldda, blocks, dwork, ldda );
magma_time = magma_sync_wtime( 0 );
magmablas_zhemv_work( opts.uplo, N, alpha, dA, ldda, dX, incx, beta, dY, incy, dwork, ldwork );
// TODO provide option to test non-work interface
//magmablas_zhemv( opts.uplo, N, alpha, dA, ldda, dX, incx, beta, dY, incy );
magma_time = magma_sync_wtime( 0 ) - magma_time;
magma_perf = gflops / magma_time;
magma_zgetvector( N, dY, incy, Ymagma, incy );
//magma_zprint_gpu( ldda, blocks, dwork, ldda );
/* =====================================================================
Performs operation using CPU BLAS
=================================================================== */
cpu_time = magma_wtime();
blasf77_zhemv( lapack_uplo_const(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_zaxpy( &N, &c_neg_one, Y, &incy, Ymagma, &incy );
magma_error = lapackf77_zlange( "M", &N, &ione, Ymagma, &N, work ) / N;
blasf77_zaxpy( &N, &c_neg_one, Y, &incy, Ycublas, &incy );
cublas_error = lapackf77_zlange( "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 %s\n",
(int) N,
magma_perf, 1000.*magma_time,
cublas_perf, 1000.*cublas_time,
cpu_perf, 1000.*cpu_time,
magma_error, cublas_error,
(magma_error < tol && cublas_error < tol ? "ok" : "failed"));
status += ! (magma_error < tol && cublas_error < tol);
TESTING_FREE_CPU( A );
TESTING_FREE_CPU( X );
TESTING_FREE_CPU( Y );
TESTING_FREE_CPU( Ycublas );
TESTING_FREE_CPU( Ymagma );
TESTING_FREE_DEV( dA );
TESTING_FREE_DEV( dX );
TESTING_FREE_DEV( dY );
TESTING_FREE_DEV( dwork );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing ctrsm
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gflops, magma_perf=0, magma_time=0, cublas_perf, cublas_time, cpu_perf=0, cpu_time=0;
float magma_error=0, cublas_error, lapack_error, work[1];
magma_int_t M, N, info;
magma_int_t Ak;
magma_int_t sizeA, sizeB;
magma_int_t lda, ldb, ldda, lddb;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t *ipiv;
magmaFloatComplex *h_A, *h_B, *h_Bcublas, *h_Bmagma, *h_Blapack, *h_X;
magmaFloatComplex_ptr d_A, d_B;
magmaFloatComplex c_neg_one = MAGMA_C_NEG_ONE;
magmaFloatComplex c_one = MAGMA_C_ONE;
magmaFloatComplex alpha = MAGMA_C_MAKE( 0.29, -0.86 );
magma_int_t status = 0;
magma_opts opts;
opts.parse_opts( argc, argv );
float tol = opts.tolerance * lapackf77_slamch("E");
// pass ngpu = -1 to test multi-GPU code using 1 gpu
magma_int_t abs_ngpu = abs( opts.ngpu );
printf("%% side = %s, uplo = %s, transA = %s, diag = %s, ngpu = %d\n",
lapack_side_const(opts.side), lapack_uplo_const(opts.uplo),
lapack_trans_const(opts.transA), lapack_diag_const(opts.diag), int(abs_ngpu) );
printf("%% M N MAGMA Gflop/s (ms) CUBLAS Gflop/s (ms) CPU Gflop/s (ms) MAGMA CUBLAS LAPACK error\n");
printf("%%============================================================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
M = opts.msize[itest];
N = opts.nsize[itest];
gflops = FLOPS_CTRSM(opts.side, M, N) / 1e9;
if ( opts.side == MagmaLeft ) {
lda = M;
Ak = M;
} else {
lda = N;
Ak = N;
}
ldb = M;
ldda = magma_roundup( lda, opts.align ); // multiple of 32 by default
lddb = magma_roundup( ldb, opts.align ); // multiple of 32 by default
sizeA = lda*Ak;
sizeB = ldb*N;
TESTING_MALLOC_CPU( h_A, magmaFloatComplex, lda*Ak );
TESTING_MALLOC_CPU( h_B, magmaFloatComplex, ldb*N );
TESTING_MALLOC_CPU( h_X, magmaFloatComplex, ldb*N );
TESTING_MALLOC_CPU( h_Blapack, magmaFloatComplex, ldb*N );
TESTING_MALLOC_CPU( h_Bcublas, magmaFloatComplex, ldb*N );
TESTING_MALLOC_CPU( h_Bmagma, magmaFloatComplex, ldb*N );
TESTING_MALLOC_CPU( ipiv, magma_int_t, Ak );
TESTING_MALLOC_DEV( d_A, magmaFloatComplex, ldda*Ak );
TESTING_MALLOC_DEV( d_B, magmaFloatComplex, lddb*N );
/* Initialize the matrices */
/* Factor A into LU to get well-conditioned triangular matrix.
* Copy L to U, since L seems okay when used with non-unit diagonal
* (i.e., from U), while U fails when used with unit diagonal. */
lapackf77_clarnv( &ione, ISEED, &sizeA, h_A );
lapackf77_cgetrf( &Ak, &Ak, h_A, &lda, ipiv, &info );
for( int j = 0; j < Ak; ++j ) {
for( int i = 0; i < j; ++i ) {
*h_A(i,j) = *h_A(j,i);
}
}
lapackf77_clarnv( &ione, ISEED, &sizeB, h_B );
memcpy( h_Blapack, h_B, sizeB*sizeof(magmaFloatComplex) );
magma_csetmatrix( Ak, Ak, h_A, lda, d_A, ldda, opts.queue );
/* =====================================================================
Performs operation using MAGMABLAS
=================================================================== */
#if defined(HAVE_CUBLAS)
magma_csetmatrix( M, N, h_B, ldb, d_B, lddb, opts.queue );
magma_time = magma_sync_wtime( opts.queue );
if (opts.ngpu == 1) {
magmablas_ctrsm( opts.side, opts.uplo, opts.transA, opts.diag,
M, N,
alpha, d_A, ldda,
d_B, lddb, opts.queue );
}
else {
magma_ctrsm_m( abs_ngpu, opts.side, opts.uplo, opts.transA, opts.diag,
M, N,
alpha, d_A, ldda,
d_B, lddb );
}
magma_time = magma_sync_wtime( opts.queue ) - magma_time;
magma_perf = gflops / magma_time;
magma_cgetmatrix( M, N, d_B, lddb, h_Bmagma, ldb, opts.queue );
#endif
/* =====================================================================
Performs operation using CUBLAS
=================================================================== */
magma_csetmatrix( M, N, h_B, ldb, d_B, lddb, opts.queue );
cublas_time = magma_sync_wtime( opts.queue );
#if defined(HAVE_CUBLAS)
// opts.handle also uses opts.queue
cublasCtrsm( opts.handle,
cublas_side_const(opts.side), cublas_uplo_const(opts.uplo),
cublas_trans_const(opts.transA), cublas_diag_const(opts.diag),
M, N,
&alpha, d_A, ldda,
d_B, lddb );
#elif defined(HAVE_clBLAS)
clblasCtrsm( clblasColumnMajor,
clblas_side_const(opts.side), clblas_uplo_const(opts.uplo),
clblas_trans_const(opts.transA), clblas_diag_const(opts.diag),
M, N,
alpha, d_A, 0, ldda,
d_B, 0, lddb,
1, &opts.queue, 0, NULL, NULL );
#endif
cublas_time = magma_sync_wtime( opts.queue ) - cublas_time;
cublas_perf = gflops / cublas_time;
magma_cgetmatrix( M, N, d_B, lddb, h_Bcublas, ldb, opts.queue );
/* =====================================================================
Performs operation using CPU BLAS
=================================================================== */
if ( opts.lapack ) {
cpu_time = magma_wtime();
blasf77_ctrsm( lapack_side_const(opts.side), lapack_uplo_const(opts.uplo),
lapack_trans_const(opts.transA), lapack_diag_const(opts.diag),
&M, &N,
&alpha, h_A, &lda,
h_Blapack, &ldb );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
}
/* =====================================================================
Check the result
=================================================================== */
// ||b - 1/alpha*A*x|| / (||A||*||x||)
magmaFloatComplex inv_alpha = MAGMA_C_DIV( c_one, alpha );
float normR, normX, normA;
normA = lapackf77_clange( "M", &Ak, &Ak, h_A, &lda, work );
#if defined(HAVE_CUBLAS)
// check magma
memcpy( h_X, h_Bmagma, sizeB*sizeof(magmaFloatComplex) );
blasf77_ctrmm( lapack_side_const(opts.side), lapack_uplo_const(opts.uplo),
lapack_trans_const(opts.transA), lapack_diag_const(opts.diag),
&M, &N,
&inv_alpha, h_A, &lda,
h_X, &ldb );
blasf77_caxpy( &sizeB, &c_neg_one, h_B, &ione, h_X, &ione );
normR = lapackf77_clange( "M", &M, &N, h_X, &ldb, work );
normX = lapackf77_clange( "M", &M, &N, h_Bmagma, &ldb, work );
magma_error = normR/(normX*normA);
#endif
// check cublas
memcpy( h_X, h_Bcublas, sizeB*sizeof(magmaFloatComplex) );
blasf77_ctrmm( lapack_side_const(opts.side), lapack_uplo_const(opts.uplo),
lapack_trans_const(opts.transA), lapack_diag_const(opts.diag),
&M, &N,
&inv_alpha, h_A, &lda,
h_X, &ldb );
blasf77_caxpy( &sizeB, &c_neg_one, h_B, &ione, h_X, &ione );
normR = lapackf77_clange( "M", &M, &N, h_X, &ldb, work );
normX = lapackf77_clange( "M", &M, &N, h_Bcublas, &ldb, work );
cublas_error = normR/(normX*normA);
if ( opts.lapack ) {
// check lapack
// this verifies that the matrix wasn't so bad that it couldn't be solved accurately.
memcpy( h_X, h_Blapack, sizeB*sizeof(magmaFloatComplex) );
blasf77_ctrmm( lapack_side_const(opts.side), lapack_uplo_const(opts.uplo),
lapack_trans_const(opts.transA), lapack_diag_const(opts.diag),
&M, &N,
&inv_alpha, h_A, &lda,
h_X, &ldb );
blasf77_caxpy( &sizeB, &c_neg_one, h_B, &ione, h_X, &ione );
normR = lapackf77_clange( "M", &M, &N, h_X, &ldb, work );
normX = lapackf77_clange( "M", &M, &N, h_Blapack, &ldb, work );
lapack_error = normR/(normX*normA);
printf("%5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %8.2e %8.2e %s\n",
(int) M, (int) N,
magma_perf, 1000.*magma_time,
cublas_perf, 1000.*cublas_time,
cpu_perf, 1000.*cpu_time,
magma_error, cublas_error, lapack_error,
(magma_error < tol && cublas_error < tol? "ok" : "failed"));
status += ! (magma_error < tol && cublas_error < tol);
}
else {
printf("%5d %5d %7.2f (%7.2f) %7.2f (%7.2f) --- ( --- ) %8.2e %8.2e --- %s\n",
(int) M, (int) N,
magma_perf, 1000.*magma_time,
cublas_perf, 1000.*cublas_time,
magma_error, cublas_error,
(magma_error < tol && cublas_error < tol ? "ok" : "failed"));
status += ! (magma_error < tol && cublas_error < tol);
}
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_B );
TESTING_FREE_CPU( h_X );
TESTING_FREE_CPU( h_Blapack );
TESTING_FREE_CPU( h_Bcublas );
TESTING_FREE_CPU( h_Bmagma );
TESTING_FREE_CPU( ipiv );
TESTING_FREE_DEV( d_A );
TESTING_FREE_DEV( d_B );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
opts.cleanup();
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing cgegqr
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gflops, gpu_perf, gpu_time, cpu_perf, cpu_time;
float e1, e2, work[1];
magmaFloatComplex *h_A, *h_R, *tau, *dtau, *h_work, tmp[1];
magmaFloatComplex *d_A, *dwork, *ddA, *d_T;
magma_int_t M, N, n2, lda, ldda, lwork, info, min_mn;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
magma_opts opts;
parse_opts( argc, argv, &opts );
opts.lapack |= opts.check; // check (-c) implies lapack (-l)
printf(" M N CPU GFlop/s (ms) GPU GFlop/s (ms) ||I - Q'Q||_F \n");
printf("=======================================================================\n");
for( int i = 0; i < opts.ntest; ++i ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
M = opts.msize[i];
N = opts.nsize[i];
min_mn = min(M, N);
lda = M;
n2 = lda*N;
ldda = ((M+31)/32)*32;
gflops = FLOPS_CGEQRF( M, N ) / 1e9 + FLOPS_CUNGQR( M, N, N ) / 1e9;
// query for workspace size
lwork = -1;
lapackf77_cgeqrf(&M, &N, NULL, &M, NULL, tmp, &lwork, &info);
lwork = (magma_int_t)MAGMA_C_REAL( tmp[0] );
lwork = max(lwork, 3*N*N);
TESTING_MALLOC_PIN( tau, magmaFloatComplex, min_mn );
TESTING_MALLOC_PIN( h_work, magmaFloatComplex, lwork );
TESTING_MALLOC_CPU( h_A, magmaFloatComplex, n2 );
TESTING_MALLOC_CPU( h_R, magmaFloatComplex, n2 );
TESTING_MALLOC_DEV( d_A, magmaFloatComplex, ldda*N );
TESTING_MALLOC_DEV( dtau, magmaFloatComplex, min_mn );
TESTING_MALLOC_DEV( dwork, magmaFloatComplex, N*N );
TESTING_MALLOC_DEV( ddA, magmaFloatComplex, N*N );
TESTING_MALLOC_DEV( d_T, magmaFloatComplex, N*N );
cudaMemset( ddA, 0, N*N*sizeof(magmaFloatComplex) );
cudaMemset( d_T, 0, N*N*sizeof(magmaFloatComplex) );
/* Initialize the matrix */
lapackf77_clarnv( &ione, ISEED, &n2, h_A );
lapackf77_clacpy( MagmaUpperLowerStr, &M, &N, h_A, &lda, h_R, &lda );
magma_csetmatrix( M, N, h_R, lda, d_A, ldda );
// warmup
magma_cgegqr_gpu( M, N, d_A, ldda, dwork, h_work, &info );
magma_csetmatrix( M, N, h_R, lda, d_A, ldda );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
gpu_time = magma_sync_wtime( 0 );
if (opts.version == 2) {
int min_mn = min(M, N);
int nb = N;
cuFloatComplex *dtau = dwork;
magma_cgeqr2x3_gpu(&M, &N, d_A, &ldda, dtau, d_T, ddA,
(float *)(dwork+min_mn), &info);
magma_cgetmatrix( min_mn, 1, dtau, min_mn, tau, min_mn);
magma_cungqr_gpu( M, N, N, d_A, ldda, tau, d_T, nb, &info );
}
else
magma_cgegqr_gpu( M, N, d_A, ldda, dwork, h_work, &info );
gpu_time = magma_sync_wtime( 0 ) - gpu_time;
gpu_perf = gflops / gpu_time;
if (info != 0)
printf("magma_cgegqr returned error %d: %s.\n",
(int) info, magma_strerror( info ));
if ( opts.lapack ) {
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
cpu_time = magma_wtime();
/* Orthogonalize on the CPU */
lapackf77_cgeqrf(&M, &N, h_A, &lda, tau, h_work, &lwork, &info);
lapackf77_cungqr(&M, &N, &N, h_A, &lda, tau, h_work, &lwork, &info );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
if (info != 0)
printf("lapackf77_cungqr returned error %d: %s.\n",
(int) info, magma_strerror( info ));
/* =====================================================================
Check the result compared to LAPACK
=================================================================== */
magma_cgetmatrix( M, N, d_A, ldda, h_R, M );
magmaFloatComplex one = MAGMA_C_ONE, zero = MAGMA_C_ZERO;
blasf77_cgemm("t", "n", &N, &N, &M, &one, h_R, &M, h_R, &M, &zero, h_work, &N);
for(int ii=0; ii<N*N; ii+=(N+1)) h_work[ii] = MAGMA_C_SUB(h_work[ii], one);
e1 = lapackf77_clange("f", &N, &N, h_work, &N, work);
blasf77_cgemm("t", "n", &N, &N, &M, &one, h_A, &M, h_A, &M, &zero, h_work, &N);
for(int ii=0; ii<N*N; ii+=(N+1)) h_work[ii] = MAGMA_C_SUB(h_work[ii], one);
e2 = lapackf77_clange("f", &N, &N, h_work, &N, work);
printf("%5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %8.2e\n",
(int) M, (int) N, cpu_perf, 1000.*cpu_time, gpu_perf, 1000.*gpu_time, e1, e2 );
}
else {
printf("%5d %5d --- ( --- ) %7.2f (%7.2f) --- \n",
(int) M, (int) N, gpu_perf, 1000.*gpu_time );
}
TESTING_FREE_PIN( tau );
TESTING_FREE_PIN( h_work );
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_R );
TESTING_FREE_DEV( d_A );
TESTING_FREE_DEV( dtau );
TESTING_FREE_DEV( dwork );
TESTING_FREE_DEV( ddA );
TESTING_FREE_DEV( d_T );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return 0;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing cgels
*/
int main( int argc, char** argv )
{
TESTING_INIT();
real_Double_t gflops, gpu_perf, gpu_time, cpu_perf, cpu_time;
float gpu_error, cpu_error, error, Anorm, work[1];
magmaFloatComplex c_one = MAGMA_C_ONE;
magmaFloatComplex c_neg_one = MAGMA_C_NEG_ONE;
magmaFloatComplex *h_A, *h_A2, *h_B, *h_X, *h_R, *tau, *h_work, tmp[1];
magmaFloatComplex *d_A, *d_B;
magma_int_t M, N, size, nrhs, lda, ldb, ldda, lddb, min_mn, max_mn, nb, info;
magma_int_t lworkgpu, lhwork;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
magma_opts opts;
parse_opts( argc, argv, &opts );
magma_int_t status = 0;
float tol = opts.tolerance * lapackf77_slamch("E");
nrhs = opts.nrhs;
printf(" ||b-Ax|| / (N||A||) ||dx-x||/(N||A||)\n");
printf(" M N NRHS CPU GFlop/s (sec) GPU GFlop/s (sec) CPU GPU \n");
printf("===================================================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
M = opts.msize[itest];
N = opts.nsize[itest];
if ( M < N ) {
printf( "%5d %5d %5d skipping because M < N is not yet supported.\n", (int) M, (int) N, (int) nrhs );
continue;
}
min_mn = min(M, N);
max_mn = max(M, N);
lda = M;
ldb = max_mn;
ldda = ((M+31)/32)*32;
lddb = ((max_mn+31)/32)*32;
nb = magma_get_cgeqrf_nb(M);
gflops = (FLOPS_CGEQRF( M, N ) + FLOPS_CGEQRS( M, N, nrhs )) / 1e9;
lworkgpu = (M - N + nb)*(nrhs + nb) + nrhs*nb;
// query for workspace size
lhwork = -1;
lapackf77_cgels( MagmaNoTransStr, &M, &N, &nrhs,
NULL, &lda, NULL, &ldb, tmp, &lhwork, &info );
lhwork = (magma_int_t) MAGMA_C_REAL( tmp[0] );
lhwork = max( lhwork, lworkgpu );
TESTING_MALLOC_CPU( tau, magmaFloatComplex, min_mn );
TESTING_MALLOC_CPU( h_A, magmaFloatComplex, lda*N );
TESTING_MALLOC_CPU( h_A2, magmaFloatComplex, lda*N );
TESTING_MALLOC_CPU( h_B, magmaFloatComplex, ldb*nrhs );
TESTING_MALLOC_CPU( h_X, magmaFloatComplex, ldb*nrhs );
TESTING_MALLOC_CPU( h_R, magmaFloatComplex, ldb*nrhs );
TESTING_MALLOC_CPU( h_work, magmaFloatComplex, lhwork );
TESTING_MALLOC_DEV( d_A, magmaFloatComplex, ldda*N );
TESTING_MALLOC_DEV( d_B, magmaFloatComplex, lddb*nrhs );
/* Initialize the matrices */
size = lda*N;
lapackf77_clarnv( &ione, ISEED, &size, h_A );
lapackf77_clacpy( MagmaUpperLowerStr, &M, &N, h_A, &lda, h_A2, &lda );
// make random RHS
size = ldb*nrhs;
lapackf77_clarnv( &ione, ISEED, &size, h_B );
lapackf77_clacpy( MagmaUpperLowerStr, &M, &nrhs, h_B, &ldb, h_R, &ldb );
// make consistent RHS
//size = N*nrhs;
//lapackf77_clarnv( &ione, ISEED, &size, h_X );
//blasf77_cgemm( MagmaNoTransStr, MagmaNoTransStr, &M, &nrhs, &N,
// &c_one, h_A, &lda,
// h_X, &ldb,
// &c_zero, h_B, &ldb );
//lapackf77_clacpy( MagmaUpperLowerStr, &M, &nrhs, h_B, &ldb, h_R, &ldb );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
magma_csetmatrix( M, N, h_A, lda, d_A, ldda );
magma_csetmatrix( M, nrhs, h_B, ldb, d_B, lddb );
gpu_time = magma_wtime();
magma_cgels_gpu( MagmaNoTrans, M, N, nrhs, d_A, ldda,
d_B, lddb, h_work, lworkgpu, &info );
gpu_time = magma_wtime() - gpu_time;
gpu_perf = gflops / gpu_time;
if (info != 0)
printf("magma_cgels_gpu returned error %d: %s.\n",
(int) info, magma_strerror( info ));
// compute the residual
magma_cgetmatrix( N, nrhs, d_B, lddb, h_X, ldb );
blasf77_cgemm( MagmaNoTransStr, MagmaNoTransStr, &M, &nrhs, &N,
&c_neg_one, h_A, &lda,
h_X, &ldb,
&c_one, h_R, &ldb );
Anorm = lapackf77_clange("f", &M, &N, h_A, &lda, work);
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
lapackf77_clacpy( MagmaUpperLowerStr, &M, &nrhs, h_B, &ldb, h_X, &ldb );
cpu_time = magma_wtime();
lapackf77_cgels( MagmaNoTransStr, &M, &N, &nrhs,
h_A, &lda, h_X, &ldb, h_work, &lhwork, &info );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
if (info != 0)
printf("lapackf77_cgels returned error %d: %s.\n",
(int) info, magma_strerror( info ));
blasf77_cgemm( MagmaNoTransStr, MagmaNoTransStr, &M, &nrhs, &N,
&c_neg_one, h_A2, &lda,
h_X, &ldb,
&c_one, h_B, &ldb );
cpu_error = lapackf77_clange("f", &M, &nrhs, h_B, &ldb, work) / (min_mn*Anorm);
gpu_error = lapackf77_clange("f", &M, &nrhs, h_R, &ldb, work) / (min_mn*Anorm);
// error relative to LAPACK
size = M*nrhs;
blasf77_caxpy( &size, &c_neg_one, h_B, &ione, h_R, &ione );
error = lapackf77_clange("f", &M, &nrhs, h_R, &ldb, work) / (min_mn*Anorm);
printf("%5d %5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %8.2e %8.2e",
(int) M, (int) N, (int) nrhs,
cpu_perf, cpu_time, gpu_perf, gpu_time, cpu_error, gpu_error, error );
if ( M == N ) {
printf( " %s\n", (gpu_error < tol && error < tol ? "ok" : "failed"));
status += ! (gpu_error < tol && error < tol);
}
else {
printf( " %s\n", (error < tol ? "ok" : "failed"));
status += ! (error < tol);
}
TESTING_FREE_CPU( tau );
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_A2 );
TESTING_FREE_CPU( h_B );
TESTING_FREE_CPU( h_X );
TESTING_FREE_CPU( h_R );
TESTING_FREE_CPU( h_work );
TESTING_FREE_DEV( d_A );
TESTING_FREE_DEV( d_B );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return status;
}
int main(int argc, char **argv)
{
TESTING_INIT();
real_Double_t gflopsF, gflopsS, gpu_perf, gpu_time /*cpu_perf, cpu_time*/;
real_Double_t gpu_perfdf, gpu_perfds;
real_Double_t gpu_perfsf, gpu_perfss;
double error, Rnorm, Anorm;
magmaDoubleComplex c_one = MAGMA_Z_ONE;
magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
magmaDoubleComplex *h_A, *h_B, *h_X;
magmaDoubleComplex_ptr d_A, d_B, d_X, d_workd;
magmaFloatComplex_ptr d_As, d_Bs, d_works;
double *h_workd;
magma_int_t lda, ldb, ldx;
magma_int_t N, nrhs, posv_iter, info, size;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
printf("%% Epsilon(double): %8.6e\n"
"%% Epsilon(single): %8.6e\n\n",
lapackf77_dlamch("Epsilon"), lapackf77_slamch("Epsilon") );
magma_int_t status = 0;
magma_opts opts;
opts.parse_opts( argc, argv );
double tol = opts.tolerance * lapackf77_dlamch("E");
nrhs = opts.nrhs;
printf("%% uplo = %s\n",
lapack_uplo_const(opts.uplo));
printf("%% N NRHS DP-Factor DP-Solve SP-Factor SP-Solve MP-Solve Iter |b-Ax|/|A|\n");
printf("%%====================================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
N = opts.nsize[itest];
ldb = ldx = lda = N;
gflopsF = FLOPS_ZPOTRF( N ) / 1e9;
gflopsS = gflopsF + FLOPS_ZPOTRS( N, nrhs ) / 1e9;
TESTING_MALLOC_CPU( h_A, magmaDoubleComplex, lda*N );
TESTING_MALLOC_CPU( h_B, magmaDoubleComplex, ldb*nrhs );
TESTING_MALLOC_CPU( h_X, magmaDoubleComplex, ldx*nrhs );
TESTING_MALLOC_CPU( h_workd, double, N );
TESTING_MALLOC_DEV( d_A, magmaDoubleComplex, lda*N );
TESTING_MALLOC_DEV( d_B, magmaDoubleComplex, ldb*nrhs );
TESTING_MALLOC_DEV( d_X, magmaDoubleComplex, ldx*nrhs );
TESTING_MALLOC_DEV( d_works, magmaFloatComplex, lda*(N+nrhs) );
TESTING_MALLOC_DEV( d_workd, magmaDoubleComplex, N*nrhs );
/* Initialize the matrix */
size = lda * N;
lapackf77_zlarnv( &ione, ISEED, &size, h_A );
magma_zmake_hpd( N, h_A, lda );
size = ldb * nrhs;
lapackf77_zlarnv( &ione, ISEED, &size, h_B );
magma_zsetmatrix( N, N, h_A, lda, d_A, lda );
magma_zsetmatrix( N, nrhs, h_B, ldb, d_B, ldb );
//=====================================================================
// Mixed Precision Iterative Refinement - GPU
//=====================================================================
gpu_time = magma_wtime();
magma_zcposv_gpu(opts.uplo, N, nrhs, d_A, lda, d_B, ldb, d_X, ldx,
d_workd, d_works, &posv_iter, &info);
gpu_time = magma_wtime() - gpu_time;
gpu_perf = gflopsS / gpu_time;
if (info != 0) {
printf("magma_zcposv returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
//=====================================================================
// Error Computation
//=====================================================================
magma_zgetmatrix( N, nrhs, d_X, ldx, h_X, ldx );
Anorm = safe_lapackf77_zlanhe( "I", lapack_uplo_const(opts.uplo), &N, h_A, &lda, h_workd);
blasf77_zhemm( "L", lapack_uplo_const(opts.uplo), &N, &nrhs,
&c_one, h_A, &lda,
h_X, &ldx,
&c_neg_one, h_B, &ldb);
Rnorm = lapackf77_zlange( "I", &N, &nrhs, h_B, &ldb, h_workd);
error = Rnorm / Anorm;
//=====================================================================
// Double Precision Factor
//=====================================================================
magma_zsetmatrix( N, N, h_A, lda, d_A, lda );
gpu_time = magma_wtime();
magma_zpotrf_gpu(opts.uplo, N, d_A, lda, &info);
gpu_time = magma_wtime() - gpu_time;
gpu_perfdf = gflopsF / gpu_time;
if (info != 0) {
printf("magma_zpotrf returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
//=====================================================================
// Double Precision Solve
//=====================================================================
magma_zsetmatrix( N, N, h_A, lda, d_A, lda );
magma_zsetmatrix( N, nrhs, h_B, ldb, d_B, ldb );
gpu_time = magma_wtime();
magma_zpotrf_gpu(opts.uplo, N, d_A, lda, &info);
magma_zpotrs_gpu(opts.uplo, N, nrhs, d_A, lda, d_B, ldb, &info);
gpu_time = magma_wtime() - gpu_time;
gpu_perfds = gflopsS / gpu_time;
if (info != 0) {
printf("magma_zpotrs returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
//=====================================================================
// Single Precision Factor
//=====================================================================
d_As = d_works;
d_Bs = d_works + lda*N;
magma_zsetmatrix( N, N, h_A, lda, d_A, lda );
magma_zsetmatrix( N, nrhs, h_B, ldb, d_B, ldb );
magmablas_zlag2c( N, N, d_A, lda, d_As, N, &info );
magmablas_zlag2c( N, nrhs, d_B, ldb, d_Bs, N, &info );
gpu_time = magma_wtime();
magma_cpotrf_gpu(opts.uplo, N, d_As, N, &info);
gpu_time = magma_wtime() - gpu_time;
gpu_perfsf = gflopsF / gpu_time;
if (info != 0) {
printf("magma_cpotrf returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
//=====================================================================
// Single Precision Solve
//=====================================================================
magmablas_zlag2c(N, N, d_A, lda, d_As, N, &info );
magmablas_zlag2c(N, nrhs, d_B, ldb, d_Bs, N, &info );
gpu_time = magma_wtime();
magma_cpotrf_gpu(opts.uplo, N, d_As, lda, &info);
magma_cpotrs_gpu(opts.uplo, N, nrhs, d_As, N, d_Bs, N, &info);
gpu_time = magma_wtime() - gpu_time;
gpu_perfss = gflopsS / gpu_time;
if (info != 0) {
printf("magma_cpotrs returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
printf("%5d %5d %7.2f %7.2f %7.2f %7.2f %7.2f %4d %8.2e %s\n",
(int) N, (int) nrhs,
gpu_perfdf, gpu_perfds, gpu_perfsf, gpu_perfss, gpu_perf,
(int) posv_iter, error, (error < tol ? "ok" : "failed"));
status += ! (error < tol);
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_B );
TESTING_FREE_CPU( h_X );
TESTING_FREE_CPU( h_workd );
TESTING_FREE_DEV( d_A );
TESTING_FREE_DEV( d_B );
TESTING_FREE_DEV( d_X );
TESTING_FREE_DEV( d_works );
TESTING_FREE_DEV( d_workd );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
opts.cleanup();
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing zlaset
Code is very similar to testing_zlacpy.cpp
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gbytes, gpu_perf, gpu_time, cpu_perf, cpu_time;
double error, work[1];
magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
magmaDoubleComplex *h_A, *h_R;
magmaDoubleComplex *d_A;
magmaDoubleComplex offdiag = MAGMA_Z_MAKE( 1.2000, 6.7000 );
magmaDoubleComplex diag = MAGMA_Z_MAKE( 3.1415, 2.7183 );
magma_int_t M, N, size, lda, ldb, ldda;
magma_int_t ione = 1;
magma_int_t status = 0;
magma_opts opts;
parse_opts( argc, argv, &opts );
magma_uplo_t uplo[] = { MagmaLower, MagmaUpper, MagmaFull };
printf("uplo M N CPU GByte/s (ms) GPU GByte/s (ms) check\n");
printf("==================================================================\n");
for( int iuplo = 0; iuplo < 3; ++iuplo ) {
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
M = opts.msize[itest];
N = opts.nsize[itest];
//M += 2; // space for insets
//N += 2;
lda = M;
ldb = lda;
ldda = ((M+31)/32)*32;
size = lda*N;
if ( uplo[iuplo] == MagmaLower || uplo[iuplo] == MagmaUpper ) {
// save triangle (with diagonal)
// TODO wrong for trapezoid
gbytes = sizeof(magmaDoubleComplex) * 0.5*N*(N+1) / 1e9;
}
else {
// save entire matrix
gbytes = sizeof(magmaDoubleComplex) * 1.*M*N / 1e9;
}
TESTING_MALLOC_CPU( h_A, magmaDoubleComplex, size );
TESTING_MALLOC_CPU( h_R, magmaDoubleComplex, size );
TESTING_MALLOC_DEV( d_A, magmaDoubleComplex, ldda*N );
/* Initialize the matrix */
for( int j = 0; j < N; ++j ) {
for( int i = 0; i < M; ++i ) {
h_A[i + j*lda] = MAGMA_Z_MAKE( i + j/10000., j );
}
}
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
magma_zsetmatrix( M, N, h_A, lda, d_A, ldda );
gpu_time = magma_sync_wtime( 0 );
//magmablas_zlaset( uplo[iuplo], M-2, N-2, offdiag, diag, d_A+1+ldda, ldda ); // inset by 1 row & col
magmablas_zlaset( uplo[iuplo], M, N, offdiag, diag, d_A, ldda );
gpu_time = magma_sync_wtime( 0 ) - gpu_time;
gpu_perf = gbytes / gpu_time;
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
cpu_time = magma_wtime();
//magma_int_t M2 = M-2; // inset by 1 row & col
//magma_int_t N2 = N-2;
//lapackf77_zlaset( lapack_uplo_const( uplo[iuplo] ), &M2, &N2, &offdiag, &diag, h_A+1+lda, &lda );
lapackf77_zlaset( lapack_uplo_const( uplo[iuplo] ), &M, &N, &offdiag, &diag, h_A, &lda );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gbytes / cpu_time;
/* =====================================================================
Check the result
=================================================================== */
magma_zgetmatrix( M, N, d_A, ldda, h_R, lda );
blasf77_zaxpy(&size, &c_neg_one, h_A, &ione, h_R, &ione);
error = lapackf77_zlange("f", &M, &N, h_R, &lda, work);
printf("%4c %5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %s\n",
lapacke_uplo_const( uplo[iuplo] ), (int) M, (int) N,
cpu_perf, cpu_time*1000., gpu_perf, gpu_time*1000.,
(error == 0. ? "ok" : "failed") );
status += ! (error == 0.);
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_R );
TESTING_FREE_DEV( d_A );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
printf( "\n" );
}
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing cgegqr
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gflops, gpu_perf, gpu_time, cpu_perf, cpu_time;
float error, e1, e2, e3, e4, e5, *work;
magmaFloatComplex c_neg_one = MAGMA_C_NEG_ONE;
magmaFloatComplex c_one = MAGMA_C_ONE;
magmaFloatComplex c_zero = MAGMA_C_ZERO;
magmaFloatComplex *h_A, *h_R, *tau, *dtau, *h_work, *h_rwork, tmp[1];
magmaFloatComplex_ptr d_A, dwork;
magma_int_t M, N, n2, lda, ldda, lwork, info, min_mn;
magma_int_t ione = 1, ldwork;
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t status = 0;
magma_opts opts;
opts.parse_opts( argc, argv );
opts.lapack |= opts.check; // check (-c) implies lapack (-l)
// versions 1...4 are valid
if (opts.version < 1 || opts.version > 4) {
printf("Unknown version %d; exiting\n", (int) opts.version );
return -1;
}
float tol = 10. * opts.tolerance * lapackf77_slamch("E");
printf("%% version %d\n", (int) opts.version );
printf("%% M N CPU Gflop/s (ms) GPU Gflop/s (ms) ||I-Q'Q||_F / M ||I-Q'Q||_I / M ||A-Q R||_I\n");
printf("%% MAGMA / LAPACK MAGMA / LAPACK\n");
printf("%%=========================================================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
M = opts.msize[itest];
N = opts.nsize[itest];
if (N > 128) {
printf("%5d %5d skipping because cgegqr requires N <= 128\n",
(int) M, (int) N);
continue;
}
if (M < N) {
printf("%5d %5d skipping because cgegqr requires M >= N\n",
(int) M, (int) N);
continue;
}
min_mn = min(M, N);
lda = M;
n2 = lda*N;
ldda = magma_roundup( M, opts.align ); // multiple of 32 by default
gflops = FLOPS_CGEQRF( M, N ) / 1e9 + FLOPS_CUNGQR( M, N, N ) / 1e9;
// query for workspace size
lwork = -1;
lapackf77_cgeqrf(&M, &N, NULL, &M, NULL, tmp, &lwork, &info);
lwork = (magma_int_t)MAGMA_C_REAL( tmp[0] );
lwork = max(lwork, 3*N*N);
ldwork = N*N;
if (opts.version == 2) {
ldwork = 3*N*N + min_mn + 2;
}
TESTING_MALLOC_PIN( tau, magmaFloatComplex, min_mn );
TESTING_MALLOC_PIN( h_work, magmaFloatComplex, lwork );
TESTING_MALLOC_PIN(h_rwork, magmaFloatComplex, lwork );
TESTING_MALLOC_CPU( h_A, magmaFloatComplex, n2 );
TESTING_MALLOC_CPU( h_R, magmaFloatComplex, n2 );
TESTING_MALLOC_CPU( work, float, M );
TESTING_MALLOC_DEV( d_A, magmaFloatComplex, ldda*N );
TESTING_MALLOC_DEV( dtau, magmaFloatComplex, min_mn );
TESTING_MALLOC_DEV( dwork, magmaFloatComplex, ldwork );
/* Initialize the matrix */
lapackf77_clarnv( &ione, ISEED, &n2, h_A );
lapackf77_clacpy( MagmaFullStr, &M, &N, h_A, &lda, h_R, &lda );
magma_csetmatrix( M, N, h_R, lda, d_A, ldda );
// warmup
if ( opts.warmup ) {
magma_cgegqr_gpu( 1, M, N, d_A, ldda, dwork, h_work, &info );
magma_csetmatrix( M, N, h_R, lda, d_A, ldda );
}
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
magmablasSetKernelStream( opts.queue );
gpu_time = magma_sync_wtime( opts.queue );
magma_cgegqr_gpu( opts.version, M, N, d_A, ldda, dwork, h_rwork, &info );
gpu_time = magma_sync_wtime( opts.queue ) - gpu_time;
gpu_perf = gflops / gpu_time;
if (info != 0) {
printf("magma_cgegqr returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
magma_cgetmatrix( M, N, d_A, ldda, h_R, lda );
// Regenerate R
// blasf77_cgemm("t", "n", &N, &N, &M, &c_one, h_R, &lda, h_A, &lda, &c_zero, h_rwork, &N);
// magma_cprint(N, N, h_work, N);
blasf77_ctrmm("r", "u", "n", "n", &M, &N, &c_one, h_rwork, &N, h_R, &lda);
blasf77_caxpy( &n2, &c_neg_one, h_A, &ione, h_R, &ione );
e5 = lapackf77_clange("i", &M, &N, h_R, &lda, work) /
lapackf77_clange("i", &M, &N, h_A, &lda, work);
magma_cgetmatrix( M, N, d_A, ldda, h_R, lda );
if ( opts.lapack ) {
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
cpu_time = magma_wtime();
/* Orthogonalize on the CPU */
lapackf77_cgeqrf(&M, &N, h_A, &lda, tau, h_work, &lwork, &info);
lapackf77_cungqr(&M, &N, &N, h_A, &lda, tau, h_work, &lwork, &info );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
if (info != 0) {
printf("lapackf77_cungqr returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
/* =====================================================================
Check the result compared to LAPACK
=================================================================== */
blasf77_cgemm("c", "n", &N, &N, &M, &c_one, h_R, &lda, h_R, &lda, &c_zero, h_work, &N);
for (int ii = 0; ii < N*N; ii += N+1 ) {
h_work[ii] = MAGMA_C_SUB(h_work[ii], c_one);
}
e1 = lapackf77_clange("f", &N, &N, h_work, &N, work) / N;
e3 = lapackf77_clange("i", &N, &N, h_work, &N, work) / N;
blasf77_cgemm("c", "n", &N, &N, &M, &c_one, h_A, &lda, h_A, &lda, &c_zero, h_work, &N);
for (int ii = 0; ii < N*N; ii += N+1 ) {
h_work[ii] = MAGMA_C_SUB(h_work[ii], c_one);
}
e2 = lapackf77_clange("f", &N, &N, h_work, &N, work) / N;
e4 = lapackf77_clange("i", &N, &N, h_work, &N, work) / N;
if (opts.version != 4)
error = e1;
else
error = e1 / (10.*max(M,N));
printf("%5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e / %8.2e %8.2e / %8.2e %8.2e %s\n",
(int) M, (int) N, cpu_perf, 1000.*cpu_time, gpu_perf, 1000.*gpu_time,
e1, e2, e3, e4, e5,
(error < tol ? "ok" : "failed"));
status += ! (error < tol);
}
else {
printf("%5d %5d --- ( --- ) %7.2f (%7.2f) --- \n",
(int) M, (int) N, gpu_perf, 1000.*gpu_time );
}
TESTING_FREE_PIN( tau );
TESTING_FREE_PIN( h_work );
TESTING_FREE_PIN( h_rwork );
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_R );
TESTING_FREE_CPU( work );
TESTING_FREE_DEV( d_A );
TESTING_FREE_DEV( dtau );
TESTING_FREE_DEV( dwork );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
opts.cleanup();
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing cgeqrf_mgpu
*/
int main( int argc, char** argv )
{
TESTING_INIT();
real_Double_t gflops, gpu_perf, gpu_time, cpu_perf=0, cpu_time=0;
float error, work[1];
magmaFloatComplex c_neg_one = MAGMA_C_NEG_ONE;
magmaFloatComplex *h_A, *h_R, *tau, *h_work, tmp[1];
magmaFloatComplex *d_lA[ MagmaMaxGPUs ];
magma_int_t M, N, n2, lda, ldda, n_local, ngpu;
magma_int_t info, min_mn, nb, lhwork;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1}, ISEED2[4];
magma_opts opts;
parse_opts( argc, argv, &opts );
opts.lapack |= (opts.check == 2); // check (-c2) implies lapack (-l)
magma_int_t status = 0;
float tol, eps = lapackf77_slamch("E");
tol = opts.tolerance * eps;
printf("ngpu %d\n", (int) opts.ngpu );
if ( opts.check == 1 ) {
printf(" M N CPU GFlop/s (sec) GPU GFlop/s (sec) ||R-Q'A||_1 / (M*||A||_1) ||I-Q'Q||_1 / M\n");
printf("================================================================================================\n");
} else {
printf(" M N CPU GFlop/s (sec) GPU GFlop/s (sec) ||R||_F /(M*||A||_F)\n");
printf("==========================================================================\n");
}
for( int i = 0; i < opts.ntest; ++i ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
M = opts.msize[i];
N = opts.nsize[i];
min_mn = min(M, N);
lda = M;
n2 = lda*N;
ldda = ((M+31)/32)*32;
nb = magma_get_cgeqrf_nb( M );
gflops = FLOPS_CGEQRF( M, N ) / 1e9;
// ngpu must be at least the number of blocks
ngpu = min( opts.ngpu, int((N+nb-1)/nb) );
if ( ngpu < opts.ngpu ) {
printf( " * too many GPUs for the matrix size, using %d GPUs\n", (int) ngpu );
}
// query for workspace size
lhwork = -1;
lapackf77_cgeqrf( &M, &N, NULL, &M, NULL, tmp, &lhwork, &info );
lhwork = (magma_int_t) MAGMA_C_REAL( tmp[0] );
// Allocate host memory for the matrix
TESTING_MALLOC_CPU( tau, magmaFloatComplex, min_mn );
TESTING_MALLOC_CPU( h_A, magmaFloatComplex, n2 );
TESTING_MALLOC_CPU( h_work, magmaFloatComplex, lhwork );
TESTING_MALLOC_PIN( h_R, magmaFloatComplex, n2 );
// Allocate device memory
for( int dev = 0; dev < ngpu; dev++ ) {
n_local = ((N/nb)/ngpu)*nb;
if (dev < (N/nb) % ngpu)
n_local += nb;
else if (dev == (N/nb) % ngpu)
n_local += N % nb;
magma_setdevice( dev );
TESTING_MALLOC_DEV( d_lA[dev], magmaFloatComplex, ldda*n_local );
}
/* Initialize the matrix */
for ( int j=0; j<4; j++ ) ISEED2[j] = ISEED[j]; // saving seeds
lapackf77_clarnv( &ione, ISEED, &n2, h_A );
lapackf77_clacpy( MagmaUpperLowerStr, &M, &N, h_A, &lda, h_R, &lda );
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
if ( opts.lapack ) {
magmaFloatComplex *tau2;
TESTING_MALLOC_CPU( tau2, magmaFloatComplex, min_mn );
cpu_time = magma_wtime();
lapackf77_cgeqrf( &M, &N, h_A, &M, tau2, h_work, &lhwork, &info );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
if (info != 0)
printf("lapack_cgeqrf returned error %d: %s.\n",
(int) info, magma_strerror( info ));
TESTING_FREE_CPU( tau2 );
}
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
magma_csetmatrix_1D_col_bcyclic( M, N, h_R, lda, d_lA, ldda, ngpu, nb );
gpu_time = magma_wtime();
magma_cgeqrf2_mgpu( ngpu, M, N, d_lA, ldda, tau, &info );
gpu_time = magma_wtime() - gpu_time;
gpu_perf = gflops / gpu_time;
if (info != 0)
printf("magma_cgeqrf2 returned error %d: %s.\n",
(int) info, magma_strerror( info ));
magma_cgetmatrix_1D_col_bcyclic( M, N, d_lA, ldda, h_R, lda, ngpu, nb );
magma_queue_sync( NULL );
if ( opts.check == 1 ) {
/* =====================================================================
Check the result
=================================================================== */
magma_int_t lwork = n2+N;
magmaFloatComplex *h_W1, *h_W2, *h_W3;
float *h_RW, results[2];
TESTING_MALLOC_CPU( h_W1, magmaFloatComplex, n2 ); // Q
TESTING_MALLOC_CPU( h_W2, magmaFloatComplex, n2 ); // R
TESTING_MALLOC_CPU( h_W3, magmaFloatComplex, lwork ); // WORK
TESTING_MALLOC_CPU( h_RW, float, M ); // RWORK
lapackf77_clarnv( &ione, ISEED2, &n2, h_A );
lapackf77_cqrt02( &M, &N, &min_mn, h_A, h_R, h_W1, h_W2, &lda, tau, h_W3, &lwork,
h_RW, results );
results[0] *= eps;
results[1] *= eps;
if ( opts.lapack ) {
printf("%5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %8.2e",
(int) M, (int) N, cpu_perf, cpu_time, gpu_perf, gpu_time, results[0],results[1] );
printf("%s\n", (results[0] < tol ? "" : " failed"));
} else {
printf("%5d %5d --- ( --- ) %7.2f (%7.2f) %8.2e %8.2e",
(int) M, (int) N, gpu_perf, gpu_time, results[0],results[1] );
printf("%s\n", (results[0] < tol ? "" : " failed"));
}
status |= ! (results[0] < tol);
TESTING_FREE_CPU( h_W1 );
TESTING_FREE_CPU( h_W2 );
TESTING_FREE_CPU( h_W3 );
TESTING_FREE_CPU( h_RW );
}
else if ( opts.check == 2 ) {
/* =====================================================================
Check the result compared to LAPACK
=================================================================== */
error = lapackf77_clange("f", &M, &N, h_A, &lda, work );
blasf77_caxpy( &n2, &c_neg_one, h_A, &ione, h_R, &ione );
error = lapackf77_clange("f", &M, &N, h_R, &lda, work ) / (min_mn*error);
printf("%5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e",
(int) M, (int) N, cpu_perf, cpu_time, gpu_perf, gpu_time, error );
printf("%s\n", (error < tol ? "" : " failed"));
status |= ! (error < tol);
}
else {
if ( opts.lapack ) {
printf("%5d %5d %7.2f (%7.2f) %7.2f (%7.2f) ---\n",
(int) M, (int) N, cpu_perf, cpu_time, gpu_perf, gpu_time );
} else {
printf("%5d %5d --- ( --- ) %7.2f (%7.2f) --- \n",
(int) M, (int) N, gpu_perf, gpu_time);
}
}
TESTING_FREE_CPU( tau );
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_work );
TESTING_FREE_PIN( h_R );
for( int dev=0; dev < ngpu; dev++ ){
magma_setdevice( dev );
TESTING_FREE_DEV( d_lA[dev] );
}
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing ztrmv
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gflops, cublas_perf, cublas_time, cpu_perf, cpu_time;
double 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};
magmaDoubleComplex *h_A, *h_x, *h_xcublas;
magmaDoubleComplex_ptr d_A, d_x;
magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
magma_int_t status = 0;
magma_opts opts;
parse_opts( argc, argv, &opts );
opts.lapack |= opts.check; // check (-c) implies lapack (-l)
double tol = opts.tolerance * lapackf77_dlamch("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_ZTRMM(opts.side, N, 1) / 1e9;
lda = N;
Ak = N;
ldda = ((lda+31)/32)*32;
sizeA = lda*Ak;
TESTING_MALLOC_CPU( h_A, magmaDoubleComplex, lda*Ak );
TESTING_MALLOC_CPU( h_x, magmaDoubleComplex, N );
TESTING_MALLOC_CPU( h_xcublas, magmaDoubleComplex, N );
TESTING_MALLOC_DEV( d_A, magmaDoubleComplex, ldda*Ak );
TESTING_MALLOC_DEV( d_x, magmaDoubleComplex, N );
/* Initialize the matrices */
lapackf77_zlarnv( &ione, ISEED, &sizeA, h_A );
lapackf77_zlarnv( &ione, ISEED, &N, h_x );
/* =====================================================================
Performs operation using CUBLAS
=================================================================== */
magma_zsetmatrix( Ak, Ak, h_A, lda, d_A, ldda );
magma_zsetvector( N, h_x, 1, d_x, 1 );
cublas_time = magma_sync_wtime( NULL );
cublasZtrmv( opts.handle, cublas_uplo_const(opts.uplo), cublas_trans_const(opts.transA),
cublas_diag_const(opts.diag),
N,
d_A, ldda,
d_x, 1 );
cublas_time = magma_sync_wtime( NULL ) - cublas_time;
cublas_perf = gflops / cublas_time;
magma_zgetvector( N, d_x, 1, h_xcublas, 1 );
/* =====================================================================
Performs operation using CPU BLAS
=================================================================== */
if ( opts.lapack ) {
cpu_time = magma_wtime();
blasf77_ztrmv( 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_zlange( "M", &N, &ione, h_x, &N, work );
blasf77_zaxpy( &N, &c_neg_one, h_x, &ione, h_xcublas, &ione );
cublas_error = lapackf77_zlange( "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" );
}
}
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing zgemm_batched
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gflops, magma_perf, magma_time, cpu_perf, cpu_time;
double magma_error, magma_err, Ynorm, work[1];
magma_int_t M, N, Xm, Ym, lda, ldda;
magma_int_t sizeA, sizeX, sizeY;
magma_int_t incx = 1;
magma_int_t incy = 1;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t status = 0;
magma_int_t batchCount;
magmaDoubleComplex *h_A, *h_X, *h_Y, *h_Ymagma;
magmaDoubleComplex *d_A, *d_X, *d_Y;
magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
magmaDoubleComplex alpha = MAGMA_Z_MAKE( 0.29, -0.86 );
magmaDoubleComplex beta = MAGMA_Z_MAKE( -0.48, 0.38 );
magmaDoubleComplex **A_array = NULL;
magmaDoubleComplex **X_array = NULL;
magmaDoubleComplex **Y_array = NULL;
magma_opts opts;
parse_opts( argc, argv, &opts );
batchCount = opts.batchcount;
opts.lapack |= opts.check;
//double tol = opts.tolerance * lapackf77_dlamch("E");
printf("trans = %s\n", lapack_trans_const(opts.transA) );
printf("BatchCount M N MAGMA Gflop/s (ms) CPU Gflop/s (ms) MAGMA error\n");
printf("===================================================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
M = opts.msize[itest];
N = opts.nsize[itest];
lda = ((M+31)/32)*32;
gflops = FLOPS_ZGEMV( M, N ) / 1e9 * batchCount;
if ( opts.transA == MagmaNoTrans ) {
Xm = N;
Ym = M;
} else {
Xm = M;
Ym = N;
}
sizeA = lda*N*batchCount;
sizeX = incx*Xm*batchCount;
sizeY = incy*Ym*batchCount;
ldda = ((lda+31)/32)*32;
TESTING_MALLOC_CPU( h_A, magmaDoubleComplex, sizeA );
TESTING_MALLOC_CPU( h_X, magmaDoubleComplex, sizeX );
TESTING_MALLOC_CPU( h_Y, magmaDoubleComplex, sizeY );
TESTING_MALLOC_CPU( h_Ymagma, magmaDoubleComplex, sizeY );
TESTING_MALLOC_DEV( d_A, magmaDoubleComplex, ldda*N*batchCount );
TESTING_MALLOC_DEV( d_X, magmaDoubleComplex, sizeX );
TESTING_MALLOC_DEV( d_Y, magmaDoubleComplex, sizeY );
magma_malloc((void**)&A_array, batchCount * sizeof(*A_array));
magma_malloc((void**)&X_array, batchCount * sizeof(*X_array));
magma_malloc((void**)&Y_array, batchCount * sizeof(*Y_array));
/* Initialize the matrices */
lapackf77_zlarnv( &ione, ISEED, &sizeA, h_A );
lapackf77_zlarnv( &ione, ISEED, &sizeX, h_X );
lapackf77_zlarnv( &ione, ISEED, &sizeY, h_Y );
/* =====================================================================
Performs operation using MAGMABLAS
=================================================================== */
magma_zsetmatrix( M, N*batchCount, h_A, lda, d_A, ldda );
magma_zsetvector( Xm*batchCount, h_X, incx, d_X, incx );
magma_zsetvector( Ym*batchCount, h_Y, incy, d_Y, incy );
zset_pointer(A_array, d_A, ldda, 0, 0, ldda*N, batchCount, magma_stream);
zset_pointer(X_array, d_X, 1, 0, 0, incx*Xm, batchCount, magma_stream);
zset_pointer(Y_array, d_Y, 1, 0, 0, incy*Ym, batchCount, magma_stream);
magma_time = magma_sync_wtime( NULL );
magmablas_zgemv_batched(opts.transA, M, N,
alpha, A_array, ldda,
X_array, incx,
beta, Y_array, incy, batchCount, magma_stream);
magma_time = magma_sync_wtime( NULL ) - magma_time;
magma_perf = gflops / magma_time;
magma_zgetvector( Ym*batchCount, d_Y, incy, h_Ymagma, incy );
/* =====================================================================
Performs operation using CPU BLAS
=================================================================== */
if ( opts.lapack ) {
cpu_time = magma_wtime();
for(int i=0; i<batchCount; i++)
{
blasf77_zgemv(
lapack_trans_const(opts.transA),
&M, &N,
&alpha, h_A + i*lda*N, &lda,
h_X + i*Xm, &incx,
&beta, h_Y + i*Ym, &incy );
}
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
}
/* =====================================================================
Check the result
=================================================================== */
if ( opts.lapack ) {
// compute relative error for both magma relative to lapack,
// |C_magma - C_lapack| / |C_lapack|
magma_error = 0.0;
for(int s=0; s<batchCount; s++)
{
Ynorm = lapackf77_zlange( "M", &M, &ione, h_Y + s*Ym, &incy, work );
blasf77_zaxpy( &Ym, &c_neg_one, h_Y + s*Ym, &ione, h_Ymagma + s*Ym, &ione );
magma_err = lapackf77_zlange( "M", &M, &ione, h_Ymagma + s*Ym, &incy, work ) / Ynorm;
if ( isnan(magma_err) || isinf(magma_err) ) {
magma_error = magma_err;
break;
}
magma_error = max(fabs(magma_err), magma_error);
}
printf("%10d %5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e \n",
(int) batchCount, (int) M, (int) N,
magma_perf, 1000.*magma_time,
cpu_perf, 1000.*cpu_time,
magma_error);
}
else {
printf("%10d %5d %5d %7.2f (%7.2f) --- ( --- ) ---\n",
(int) batchCount, (int) M, (int) N,
magma_perf, 1000.*magma_time);
}
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_X );
TESTING_FREE_CPU( h_Y );
TESTING_FREE_CPU( h_Ymagma );
TESTING_FREE_DEV( d_A );
TESTING_FREE_DEV( d_X );
TESTING_FREE_DEV( d_Y );
TESTING_FREE_DEV( A_array );
TESTING_FREE_DEV( X_array );
TESTING_FREE_DEV( Y_array );
fflush( stdout);
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing zgesv_gpu
*/
int main(int argc, char **argv)
{
TESTING_INIT();
real_Double_t gflops, cpu_perf, cpu_time, gpu_perf, gpu_time;
double error, Rnorm, Anorm, Xnorm, *work;
magmaDoubleComplex c_one = MAGMA_Z_ONE;
magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
magmaDoubleComplex *h_A, *h_B, *h_X;
magmaDoubleComplex *d_A, *d_B;
magma_int_t *ipiv;
magma_int_t N, nrhs, lda, ldb, ldda, lddb, info, sizeA, sizeB;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t status = 0;
magma_opts opts;
parse_opts( argc, argv, &opts );
double tol = opts.tolerance * lapackf77_dlamch("E");
nrhs = opts.nrhs;
printf(" N NRHS CPU GFlop/s (sec) GPU GFlop/s (sec) ||B - AX|| / N*||A||*||X||\n");
printf("================================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
N = opts.nsize[itest];
lda = N;
ldb = lda;
ldda = ((N+31)/32)*32;
lddb = ldda;
gflops = ( FLOPS_ZGETRF( N, N ) + FLOPS_ZGETRS( N, nrhs ) ) / 1e9;
TESTING_MALLOC_CPU( h_A, magmaDoubleComplex, lda*N );
TESTING_MALLOC_CPU( h_B, magmaDoubleComplex, ldb*nrhs );
TESTING_MALLOC_CPU( h_X, magmaDoubleComplex, ldb*nrhs );
TESTING_MALLOC_CPU( work, double, N );
TESTING_MALLOC_CPU( ipiv, magma_int_t, N );
TESTING_MALLOC_DEV( d_A, magmaDoubleComplex, ldda*N );
TESTING_MALLOC_DEV( d_B, magmaDoubleComplex, lddb*nrhs );
/* Initialize the matrices */
sizeA = lda*N;
sizeB = ldb*nrhs;
lapackf77_zlarnv( &ione, ISEED, &sizeA, h_A );
lapackf77_zlarnv( &ione, ISEED, &sizeB, h_B );
magma_zsetmatrix( N, N, h_A, lda, d_A, ldda );
magma_zsetmatrix( N, nrhs, h_B, ldb, d_B, lddb );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
gpu_time = magma_wtime();
magma_zgesv_gpu( N, nrhs, d_A, ldda, ipiv, d_B, lddb, &info );
gpu_time = magma_wtime() - gpu_time;
gpu_perf = gflops / gpu_time;
if (info != 0)
printf("magma_zgesv_gpu returned error %d: %s.\n",
(int) info, magma_strerror( info ));
//=====================================================================
// Residual
//=====================================================================
magma_zgetmatrix( N, nrhs, d_B, lddb, h_X, ldb );
Anorm = lapackf77_zlange("I", &N, &N, h_A, &lda, work);
Xnorm = lapackf77_zlange("I", &N, &nrhs, h_X, &ldb, work);
blasf77_zgemm( MagmaNoTransStr, MagmaNoTransStr, &N, &nrhs, &N,
&c_one, h_A, &lda,
h_X, &ldb,
&c_neg_one, h_B, &ldb);
Rnorm = lapackf77_zlange("I", &N, &nrhs, h_B, &ldb, work);
error = Rnorm/(N*Anorm*Xnorm);
status += ! (error < tol);
/* ====================================================================
Performs operation using LAPACK
=================================================================== */
if ( opts.lapack ) {
cpu_time = magma_wtime();
lapackf77_zgesv( &N, &nrhs, h_A, &lda, ipiv, h_B, &ldb, &info );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
if (info != 0)
printf("lapackf77_zgesv returned error %d: %s.\n",
(int) info, magma_strerror( info ));
printf( "%5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %s\n",
(int) N, (int) nrhs, cpu_perf, cpu_time, gpu_perf, gpu_time,
error, (error < tol ? "ok" : "failed"));
}
else {
printf( "%5d %5d --- ( --- ) %7.2f (%7.2f) %8.2e %s\n",
(int) N, (int) nrhs, gpu_perf, gpu_time,
error, (error < tol ? "ok" : "failed"));
}
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_B );
TESTING_FREE_CPU( h_X );
TESTING_FREE_CPU( work );
TESTING_FREE_CPU( ipiv );
TESTING_FREE_DEV( d_A );
TESTING_FREE_DEV( d_B );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing zhetrd_he2hb
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gpu_time, gpu_perf, gflops;
magmaDoubleComplex *h_A, *h_R, *h_work, *dT1;
magmaDoubleComplex *tau;
double *D, *E;
/* Matrix size */
magma_int_t N, n2, lda, lwork, lwork0; //ldt
magma_int_t info;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
#if defined(CHECKEIG)
#if defined(PRECISION_z) || defined(PRECISION_d)
magma_int_t WANTZ=0;
magma_int_t THREADS=1;
#endif
#endif
magma_int_t NE = 0;
magma_int_t NB = 0;
magma_int_t ngpu = 1;
magma_opts opts;
opts.parse_opts( argc, argv );
NB = opts.nb;
if (NB < 1)
NB = 64; //64; //magma_get_zhetrd_he2hb_nb(N);
// what is NE ?
if (NE < 1)
NE = 64; //N; //magma_get_zhetrd_he2hb_nb(N); // N not yet initialized
printf("%% N GPU GFlop/s \n");
printf("%%====================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
N = opts.nsize[itest];
lda = N;
//ldt = N;
n2 = lda*N;
gflops = FLOPS_ZHETRD( N ) / 1e9;
/* We suppose the magma NB is bigger than lapack NB */
lwork0 = N*NB;
/* Allocate host memory for the matrix */
TESTING_MALLOC_CPU( h_A, magmaDoubleComplex, lda*N );
TESTING_MALLOC_CPU( tau, magmaDoubleComplex, N-1 );
TESTING_MALLOC_PIN( h_R, magmaDoubleComplex, lda*N );
TESTING_MALLOC_PIN( h_work, magmaDoubleComplex, lwork0 );
TESTING_MALLOC_PIN( D, double, N );
TESTING_MALLOC_PIN( E, double, N );
//TESTING_MALLOC_DEV( dT1, magmaDoubleComplex, (2*min(N,N) + roundup( N, 32 ))*NB );
TESTING_MALLOC_DEV( dT1, magmaDoubleComplex, (N*NB) );
// if (WANTZ) gflops = 2.0*gflops;
/* ====================================================================
Initialize the matrix
=================================================================== */
lapackf77_zlarnv( &ione, ISEED, &n2, h_A );
magma_zmake_hermitian( N, h_A, lda );
lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_A, &lda, h_R, &lda );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
magma_device_t cdev;
magma_getdevice( &cdev );
gpu_time = magma_wtime();
/*
magma_zhetrd_he2hb( opts.uplo, N, NB, h_R, lda, tau, h_work, lwork0, dT1, THREADS, &info);
tband = magma_wtime - gpu_time();
printf(" Finish BAND N %d NB %d ngpu %d timing= %f\n", N, NB, ngpu, tband);
magma_zhetrd_bhe2trc_v5(THREADS, WANTZ, opts.uplo, NE, N, NB, h_R, lda, D, E, dT1, ldt);
*/
/*
magma_zhetrd_he2hb( opts.uplo, N, NB, h_R, lda, tau, h_work, lwork, dT1, THREADS, &info);
tband = magma_wtime - gpu_time();
printf(" Finish BAND N %d NB %d ngpu %d timing= %f\n", N, NB, ngpu, tband);
magma_zhetrd_bhe2trc(THREADS, WANTZ, opts.uplo, NE, N, NB, h_R, lda, D, E, dT1, ldt);
*/
magma_range_t range = MagmaRangeAll;
magma_int_t m1 = 0;
double vl = 0;
double vu = 0;
magma_int_t il = 0;
magma_int_t iu = 0;
if (opts.fraction == 0) {
il = max( 1, magma_int_t(0.1*N) );
iu = max( 1, magma_int_t(0.3*N) );
}
else {
il = 1;
iu = max( 1, magma_int_t(opts.fraction*N) );
}
magmaDoubleComplex *hh_work;
magma_int_t *iwork;
magma_int_t /*nb,*/ /*lwork,*/ liwork;
magma_int_t threads = magma_get_parallel_numthreads();
#ifdef COMPLEX
double *rwork;
magma_int_t lrwork;
#endif
magma_zheevdx_getworksize(N, threads, (opts.jobz == MagmaVec),
&lwork,
#ifdef COMPLEX
&lrwork,
#endif
&liwork);
TESTING_MALLOC_PIN( hh_work, magmaDoubleComplex, lwork );
TESTING_MALLOC_CPU( iwork, magma_int_t, liwork );
#ifdef COMPLEX
TESTING_MALLOC_PIN( rwork, double, lrwork );
#endif
if (ngpu == 1) {
printf("calling zheevdx_2stage 1 GPU\n");
magma_zheevdx_2stage( opts.jobz, range, opts.uplo, N,
h_R, lda,
vl, vu, il, iu,
&m1, D,
hh_work, lwork,
#ifdef COMPLEX
rwork, lrwork,
#endif
iwork, liwork,
&info);
} else {
printf("calling zheevdx_2stage_m %d GPU\n", (int) ngpu);
magma_zheevdx_2stage_m(ngpu, opts.jobz, range, opts.uplo, N,
h_R, lda,
vl, vu, il, iu,
&m1, D,
hh_work, lwork,
#ifdef COMPLEX
rwork, lrwork,
#endif
iwork, liwork,
&info);
}
magma_setdevice( cdev );
gpu_time = magma_wtime() - gpu_time;
gpu_perf = gflops / gpu_time;
/* =====================================================================
Check the factorization
=================================================================== */
/*
if ( opts.check ) {
FILE *fp;
printf("Writing input matrix in matlab_i_mat.txt ...\n");
fp = fopen ("matlab_i_mat.txt", "w");
if ( fp == NULL ) {
printf("Couldn't open output file\n");
return -1;
}
for (j=0; j < N; j++) {
for (k=0; k < N; k++) {
#ifdef COMPLEX
fprintf(fp, "%5d %5d %11.8f %11.8f\n", k+1, j+1,
h_A[k+j*lda].x, h_A[k+j*lda].y);
#else
fprintf(fp, "%5d %5d %11.8f\n", k+1, j+1, h_A[k+j*lda]);
#endif
}
}
fclose( fp );
printf("Writing output matrix in matlab_o_mat.txt ...\n");
fp = fopen ("matlab_o_mat.txt", "w");
if ( fp == NULL ) {
printf("Couldn't open output file\n");
return -1;
}
for (j=0; j < N; j++) {
for (k=0; k < N; k++) {
#ifdef COMPLEX
fprintf(fp, "%5d %5d %11.8f %11.8f\n", k+1, j+1,
h_R[k+j*lda].x, h_R[k+j*lda].y);
#else
fprintf(fp, "%5d %5d %11.8f\n", k+1, j+1, h_R[k+j*lda]);
#endif
}
}
fclose( fp );
}
*/
/* =====================================================================
Print performance and error.
=================================================================== */
#if defined(CHECKEIG)
#if defined(PRECISION_z) || defined(PRECISION_d)
if ( opts.check ) {
printf(" Total N %5d gflops %6.2f timing %6.2f seconds\n", (int) N, gpu_perf, gpu_time );
const char* JOBZ;
if (WANTZ == 0)
JOBZ = MagmaNoVecStr;
else
JOBZ = MagmaVecStr;
double nrmI=0.0, nrm1=0.0, nrm2=0.0;
int lwork2 = 256*N;
magmaDoubleComplex *work2, *AINIT;
double *rwork2, *D2;
// TODO free this memory !
magma_zmalloc_cpu( &work2, lwork2 );
magma_dmalloc_cpu( &rwork2, N );
magma_dmalloc_cpu( &D2, N );
magma_zmalloc_cpu( &AINIT, N*lda );
memcpy(AINIT, h_A, N*lda*sizeof(magmaDoubleComplex));
/* compute the eigenvalues using lapack routine to be able to compare to it and used as ref */
cpu_time = magma_wtime();
i= min(12, THREADS);
magma_set_lapack_numthreads( i );
lapackf77_zheev( "N", "L", &N, h_A, &lda, D2, work2, &lwork2,
#ifdef COMPLEX
rwork2,
#endif
&info );
///* call eigensolver for our resulting tridiag [D E] and for Q */
//dstedc_withZ('V', N, D, E, h_R, lda);
////dsterf_( &N, D, E, &info);
////
cpu_time = magma_wtime() - cpu_time;
printf(" Finish CHECK - EIGEN timing= %f threads %d\n", cpu_time, i);
/*
for (i=0; i < 10; i++)
printf(" voici lpk D[%d] %8.2e\n", i, D2[i]);
*/
//magmaDoubleComplex mydz=0.0, mydo=1.0;
//magmaDoubleComplex *Z;
// magma_zmalloc_cpu( &Z, N*lda );
// dgemm_("N", "N", &N, &N, &N, &mydo, h_R, &lda, h_A, &lda, &mydz, Z, &lda);
/* compare result */
cmp_vals(N, D2, D, &nrmI, &nrm1, &nrm2);
magmaDoubleComplex *WORKAJETER;
double *RWORKAJETER, *RESU;
// TODO free this memory !
magma_zmalloc_cpu( &WORKAJETER, (2* N * N + N) );
magma_dmalloc_cpu( &RWORKAJETER, N );
magma_dmalloc_cpu( &RESU, 10 );
int MATYPE;
memset(RESU, 0, 10*sizeof(double));
MATYPE=3;
double NOTHING=0.0;
cpu_time = magma_wtime();
// check results
zcheck_eig_( JOBZ, &MATYPE, &N, &NB, AINIT, &lda, &NOTHING, &NOTHING, D2, D, h_R, &lda, WORKAJETER, RWORKAJETER, RESU );
cpu_time = magma_wtime() - cpu_time;
printf(" Finish CHECK - results timing= %f\n", cpu_time);
magma_set_lapack_numthreads( 1 );
printf("\n");
printf(" ================================================================================================================\n");
printf(" ==> INFO voici threads=%d N=%d NB=%d WANTZ=%d\n", (int) THREADS, (int) N, (int) NB, (int) WANTZ);
printf(" ================================================================================================================\n");
printf(" DSBTRD : %15s \n", "STATblgv9withQ ");
printf(" ================================================================================================================\n");
if (WANTZ > 0)
printf(" | A - U S U' | / ( |A| n ulp ) : %15.3E \n", RESU[0]);
if (WANTZ > 0)
printf(" | I - U U' | / ( n ulp ) : %15.3E \n", RESU[1]);
printf(" | D1 - EVEIGS | / (|D| ulp) : %15.3E \n", RESU[2]);
printf(" max | D1 - EVEIGS | : %15.3E \n", RESU[6]);
printf(" ================================================================================================================\n\n\n");
printf(" ****************************************************************************************************************\n");
printf(" * Hello here are the norm Infinite (max)=%8.2e norm one (sum)=%8.2e norm2(sqrt)=%8.2e *\n", nrmI, nrm1, nrm2);
printf(" ****************************************************************************************************************\n\n");
}
#endif
#endif
printf(" Total N %5d gflops %6.2f timing %6.2f seconds\n", (int) N, gpu_perf, gpu_time );
printf("%%===========================================================================\n\n\n");
/* Memory clean up */
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( tau );
TESTING_FREE_PIN( h_R );
TESTING_FREE_PIN( h_work );
TESTING_FREE_PIN( D );
TESTING_FREE_PIN( E );
TESTING_FREE_DEV( dT1 );
/* TODO - not all memory has been freed inside loop */
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return EXIT_SUCCESS;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing zpotf2_gpu
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gflops, gpu_perf, gpu_time, cpu_perf, cpu_time;
magmaDoubleComplex *h_A, *h_R;
magmaDoubleComplex_ptr d_A;
magma_int_t N, n2, lda, ldda, info;
magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
double Anorm, error, work[1];
magma_int_t status = 0;
magma_opts opts;
opts.parse_opts( argc, argv );
double tol = opts.tolerance * lapackf77_dlamch("E");
opts.lapack |= opts.check; // check (-c) implies lapack (-l)
printf("%% uplo = %s\n", lapack_uplo_const(opts.uplo) );
printf("%% N CPU Gflop/s (ms) GPU Gflop/s (ms) ||R_magma - R_lapack||_F / ||R_lapack||_F\n");
printf("%%=======================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
N = opts.nsize[itest];
lda = N;
n2 = lda*N;
ldda = magma_roundup( N, opts.align ); // multiple of 32 by default
gflops = FLOPS_ZPOTRF( N ) / 1e9;
if ( N > 512 ) {
printf( "%5d skipping because zpotf2 does not support N > 512\n", (int) N );
continue;
}
TESTING_MALLOC_CPU( h_A, magmaDoubleComplex, n2 );
TESTING_MALLOC_PIN( h_R, magmaDoubleComplex, n2 );
TESTING_MALLOC_DEV( d_A, magmaDoubleComplex, ldda*N );
/* Initialize the matrix */
lapackf77_zlarnv( &ione, ISEED, &n2, h_A );
magma_zmake_hpd( N, h_A, lda );
lapackf77_zlacpy( MagmaFullStr, &N, &N, h_A, &lda, h_R, &lda );
magma_zsetmatrix( N, N, h_A, lda, d_A, ldda, opts.queue );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
gpu_time = magma_sync_wtime( opts.queue );
magma_zpotf2_gpu( opts.uplo, N, d_A, ldda, opts.queue, &info );
gpu_time = magma_sync_wtime( opts.queue ) - gpu_time;
gpu_perf = gflops / gpu_time;
if (info != 0) {
printf("magma_zpotf2_gpu returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
if ( opts.lapack ) {
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
cpu_time = magma_wtime();
lapackf77_zpotrf( lapack_uplo_const(opts.uplo), &N, h_A, &lda, &info );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
if (info != 0) {
printf("lapackf77_zpotrf returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
/* =====================================================================
Check the result compared to LAPACK
=================================================================== */
magma_zgetmatrix( N, N, d_A, ldda, h_R, lda, opts.queue );
blasf77_zaxpy(&n2, &c_neg_one, h_A, &ione, h_R, &ione);
Anorm = lapackf77_zlange("f", &N, &N, h_A, &lda, work);
error = lapackf77_zlange("f", &N, &N, h_R, &lda, work) / Anorm;
printf("%5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %s\n",
(int) N, cpu_perf, cpu_time*1000., gpu_perf, gpu_time*1000.,
error, (error < tol ? "ok" : "failed"));
status += ! (error < tol);
}
else {
printf("%5d --- ( --- ) %7.2f (%7.2f) --- \n",
(int) N, gpu_perf, gpu_time*1000. );
}
TESTING_FREE_CPU( h_A );
TESTING_FREE_PIN( h_R );
TESTING_FREE_DEV( d_A );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
opts.cleanup();
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing ztrsm
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gflops, magma_perf, magma_time=0, cublas_perf, cublas_time, cpu_perf=0, cpu_time=0;
double magma_error, cublas_error, work[1];
magma_int_t M, N, info;
magma_int_t Ak;
magma_int_t sizeA, sizeB;
magma_int_t lda, ldb, ldda, lddb;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t *ipiv;
magmaDoubleComplex *h_A, *h_B, *h_Bcublas, *h_Bmagma, *h_B1, *h_X1, *h_X2;
magmaDoubleComplex *d_A, *d_B;
magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
magmaDoubleComplex c_one = MAGMA_Z_ONE;
magmaDoubleComplex alpha = MAGMA_Z_MAKE( 0.29, -0.86 );
magma_int_t status = 0;
magma_opts opts;
parse_opts( argc, argv, &opts );
double tol = opts.tolerance * lapackf77_dlamch("E");
printf("side = %s, uplo = %s, transA = %s, diag = %s \n",
lapack_side_const(opts.side), lapack_uplo_const(opts.uplo),
lapack_trans_const(opts.transA), lapack_diag_const(opts.diag) );
printf(" M N MAGMA Gflop/s (ms) CUBLAS Gflop/s (ms) CPU Gflop/s (ms) MAGMA error CUBLAS error\n");
printf("==================================================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
M = opts.msize[itest];
N = opts.nsize[itest];
gflops = FLOPS_ZTRSM(opts.side, M, N) / 1e9;
if ( opts.side == MagmaLeft ) {
lda = M;
Ak = M;
} else {
lda = N;
Ak = N;
}
ldb = M;
ldda = ((lda+31)/32)*32;
lddb = ((ldb+31)/32)*32;
sizeA = lda*Ak;
sizeB = ldb*N;
TESTING_MALLOC_CPU( h_A, magmaDoubleComplex, lda*Ak );
TESTING_MALLOC_CPU( h_B, magmaDoubleComplex, ldb*N );
TESTING_MALLOC_CPU( h_B1, magmaDoubleComplex, ldb*N );
TESTING_MALLOC_CPU( h_X1, magmaDoubleComplex, ldb*N );
TESTING_MALLOC_CPU( h_X2, magmaDoubleComplex, ldb*N );
TESTING_MALLOC_CPU( h_Bcublas, magmaDoubleComplex, ldb*N );
TESTING_MALLOC_CPU( h_Bmagma, magmaDoubleComplex, ldb*N );
TESTING_MALLOC_CPU( ipiv, magma_int_t, Ak );
TESTING_MALLOC_DEV( d_A, magmaDoubleComplex, ldda*Ak );
TESTING_MALLOC_DEV( d_B, magmaDoubleComplex, lddb*N );
/* Initialize the matrices */
/* Factor A into LU to get well-conditioned triangular matrix.
* Copy L to U, since L seems okay when used with non-unit diagonal
* (i.e., from U), while U fails when used with unit diagonal. */
lapackf77_zlarnv( &ione, ISEED, &sizeA, h_A );
lapackf77_zgetrf( &Ak, &Ak, h_A, &lda, ipiv, &info );
for( int j = 0; j < Ak; ++j ) {
for( int i = 0; i < j; ++i ) {
*h_A(i,j) = *h_A(j,i);
}
}
lapackf77_zlarnv( &ione, ISEED, &sizeB, h_B );
memcpy(h_B1, h_B, sizeB*sizeof(magmaDoubleComplex));
/* =====================================================================
Performs operation using MAGMABLAS
=================================================================== */
magma_zsetmatrix( Ak, Ak, h_A, lda, d_A, ldda );
magma_zsetmatrix( M, N, h_B, ldb, d_B, lddb );
magma_time = magma_sync_wtime( NULL );
magmablas_ztrsm( opts.side, opts.uplo, opts.transA, opts.diag,
M, N,
alpha, d_A, ldda,
d_B, lddb );
magma_time = magma_sync_wtime( NULL ) - magma_time;
magma_perf = gflops / magma_time;
magma_zgetmatrix( M, N, d_B, lddb, h_Bmagma, ldb );
/* =====================================================================
Performs operation using CUBLAS
=================================================================== */
magma_zsetmatrix( M, N, h_B, ldb, d_B, lddb );
cublas_time = magma_sync_wtime( NULL );
cublasZtrsm( handle, cublas_side_const(opts.side), cublas_uplo_const(opts.uplo),
cublas_trans_const(opts.transA), cublas_diag_const(opts.diag),
M, N,
&alpha, d_A, ldda,
d_B, lddb );
cublas_time = magma_sync_wtime( NULL ) - cublas_time;
cublas_perf = gflops / cublas_time;
magma_zgetmatrix( M, N, d_B, lddb, h_Bcublas, ldb );
/* =====================================================================
Performs operation using CPU BLAS
=================================================================== */
if ( opts.lapack ) {
cpu_time = magma_wtime();
blasf77_ztrsm( lapack_side_const(opts.side), lapack_uplo_const(opts.uplo), lapack_trans_const(opts.transA), lapack_diag_const(opts.diag),
&M, &N,
&alpha, h_A, &lda,
h_B, &ldb );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
}
/* =====================================================================
Check the result
=================================================================== */
// ||b - Ax|| / (||A||*||x||)
memcpy(h_X1, h_Bmagma, sizeB*sizeof(magmaDoubleComplex));
magmaDoubleComplex alpha2 = MAGMA_Z_DIV( c_one, alpha );
blasf77_ztrmm( lapack_side_const(opts.side), lapack_uplo_const(opts.uplo), lapack_trans_const(opts.transA), lapack_diag_const(opts.diag),
&M, &N,
&alpha2, h_A, &lda,
h_X1, &ldb );
blasf77_zaxpy( &sizeB, &c_neg_one, h_B1, &ione, h_X1, &ione );
double norm1 = lapackf77_zlange( "M", &M, &N, h_X1, &ldb, work );
double normx = lapackf77_zlange( "M", &M, &N, h_Bmagma, &ldb, work );
double normA = lapackf77_zlange( "M", &Ak, &Ak, h_A, &lda, work );
magma_error = norm1/(normx*normA);
memcpy(h_X2, h_Bcublas, sizeB*sizeof(magmaDoubleComplex));
blasf77_ztrmm( lapack_side_const(opts.side), lapack_uplo_const(opts.uplo), lapack_trans_const(opts.transA), lapack_diag_const(opts.diag),
&M, &N,
&alpha2, h_A, &lda,
h_X2, &ldb );
blasf77_zaxpy( &sizeB, &c_neg_one, h_B1, &ione, h_X2, &ione );
norm1 = lapackf77_zlange( "M", &M, &N, h_X2, &ldb, work );
normx = lapackf77_zlange( "M", &M, &N, h_Bcublas, &ldb, work );
normA = lapackf77_zlange( "M", &Ak, &Ak, h_A, &lda, work );
cublas_error = norm1/(normx*normA);
if ( opts.lapack ) {
printf("%5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %8.2e %s\n",
(int) M, (int) N,
magma_perf, 1000.*magma_time,
cublas_perf, 1000.*cublas_time,
cpu_perf, 1000.*cpu_time,
magma_error, cublas_error,
(magma_error < tol && cublas_error < tol? "ok" : "failed"));
status += ! (magma_error < tol && cublas_error < tol);
}
else {
printf("%5d %5d %7.2f (%7.2f) %7.2f (%7.2f) --- ( --- ) %8.2e %8.2e %s\n",
(int) M, (int) N,
magma_perf, 1000.*magma_time,
cublas_perf, 1000.*cublas_time,
magma_error, cublas_error,
(magma_error < tol && cublas_error < tol? "ok" : "failed"));
status += ! (magma_error < tol && cublas_error < tol);
}
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_B );
TESTING_FREE_CPU( h_B1 );
TESTING_FREE_CPU( h_X1 );
TESTING_FREE_CPU( h_X2 );
TESTING_FREE_CPU( h_Bcublas );
TESTING_FREE_CPU( h_Bmagma );
TESTING_FREE_DEV( d_A );
TESTING_FREE_DEV( d_B );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return status;
}
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gflops, gpu_perf, gpu_time, cpu_perf, cpu_time;
magmaFloatComplex *h_x, *h_x2, *h_tau, *h_tau2;
magmaFloatComplex *d_x, *d_tau;
magmaFloatComplex c_neg_one = MAGMA_C_NEG_ONE;
float error, error2, work[1];
magma_int_t N, nb, lda, ldda, size;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1}; magma_int_t status = 0;
magma_opts opts;
parse_opts( argc, argv, &opts );
float tol = opts.tolerance * lapackf77_slamch("E");
// does larfg on nb columns, one after another
nb = (opts.nb > 0 ? opts.nb : 64);
magma_queue_t queue = 0;
printf(" N nb CPU GFLop/s (ms) GPU GFlop/s (ms) error tau error\n");
printf("==========================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
N = opts.nsize[itest];
lda = N;
ldda = ((N+31)/32)*32;
gflops = FLOPS_CLARFG( N ) / 1e9 * nb;
TESTING_MALLOC_CPU( h_x, magmaFloatComplex, N*nb );
TESTING_MALLOC_CPU( h_x2, magmaFloatComplex, N*nb );
TESTING_MALLOC_CPU( h_tau, magmaFloatComplex, nb );
TESTING_MALLOC_CPU( h_tau2, magmaFloatComplex, nb );
TESTING_MALLOC_DEV( d_x, magmaFloatComplex, ldda*nb );
TESTING_MALLOC_DEV( d_tau, magmaFloatComplex, nb );
/* Initialize the vectors */
size = N*nb;
lapackf77_clarnv( &ione, ISEED, &size, h_x );
/* =====================================================================
Performs operation using MAGMABLAS
=================================================================== */
magma_csetmatrix( N, nb, h_x, N, d_x, ldda );
gpu_time = magma_sync_wtime( queue );
for( int j = 0; j < nb; ++j ) {
magmablas_clarfg( N, &d_x[0+j*ldda], &d_x[1+j*ldda], ione, &d_tau[j] );
}
gpu_time = magma_sync_wtime( queue ) - gpu_time;
gpu_perf = gflops / gpu_time;
magma_cgetmatrix( N, nb, d_x, ldda, h_x2, N );
magma_cgetvector( nb, d_tau, 1, h_tau2, 1 );
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
cpu_time = magma_wtime();
for( int j = 0; j < nb; ++j ) {
lapackf77_clarfg( &N, &h_x[0+j*lda], &h_x[1+j*lda], &ione, &h_tau[j] );
}
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
/* =====================================================================
Error Computation and Performance Comparison
=================================================================== */
blasf77_caxpy( &size, &c_neg_one, h_x, &ione, h_x2, &ione );
error = lapackf77_clange( "F", &N, &nb, h_x2, &N, work )
/ lapackf77_clange( "F", &N, &nb, h_x, &N, work );
// tau can be 0
blasf77_caxpy( &nb, &c_neg_one, h_tau, &ione, h_tau2, &ione );
error2 = lapackf77_clange( "F", &nb, &ione, h_tau, &nb, work );
if ( error2 != 0 ) {
error2 = lapackf77_clange( "F", &nb, &ione, h_tau2, &nb, work ) / error2;
}
printf("%5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %8.2e %s\n",
(int) N, (int) nb, cpu_perf, 1000.*cpu_time, gpu_perf, 1000.*gpu_time,
error, error2,
(error < tol && error2 < tol ? "ok" : "failed") );
status += ! (error < tol && error2 < tol);
TESTING_FREE_CPU( h_x );
TESTING_FREE_CPU( h_x2 );
TESTING_FREE_CPU( h_tau );
TESTING_FREE_DEV( d_x );
TESTING_FREE_DEV( d_tau );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing clacpy_batched
Code is very similar to testing_cgeadd_batched.cpp
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gbytes, gpu_perf, gpu_time, cpu_perf, cpu_time;
float error, work[1];
magmaFloatComplex c_neg_one = MAGMA_C_NEG_ONE;
magmaFloatComplex *h_A, *h_B;
magmaFloatComplex *d_A, *d_B;
magmaFloatComplex **hAarray, **hBarray, **dAarray, **dBarray;
magma_int_t M, N, mb, nb, size, lda, ldda, mstride, nstride, ntile;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t status = 0;
magma_opts opts;
parse_opts( argc, argv, &opts );
mb = (opts.nb == 0 ? 32 : opts.nb);
nb = (opts.nb == 0 ? 64 : opts.nb);
mstride = 2*mb;
nstride = 3*nb;
printf("mb=%d, nb=%d, mstride=%d, nstride=%d\n", (int) mb, (int) nb, (int) mstride, (int) nstride );
printf(" M N ntile CPU GFlop/s (ms) GPU GFlop/s (ms) check\n");
printf("=================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
M = opts.msize[itest];
N = opts.nsize[itest];
lda = M;
ldda = ((M+31)/32)*32;
size = lda*N;
if ( N < nb || M < nb ) {
ntile = 0;
} else {
ntile = min( (M - nb)/mstride + 1,
(N - nb)/nstride + 1 );
}
gbytes = 2.*mb*nb*ntile / 1e9;
TESTING_MALLOC_CPU( h_A, magmaFloatComplex, lda *N );
TESTING_MALLOC_CPU( h_B, magmaFloatComplex, lda *N );
TESTING_MALLOC_DEV( d_A, magmaFloatComplex, ldda*N );
TESTING_MALLOC_DEV( d_B, magmaFloatComplex, ldda*N );
TESTING_MALLOC_CPU( hAarray, magmaFloatComplex*, ntile );
TESTING_MALLOC_CPU( hBarray, magmaFloatComplex*, ntile );
TESTING_MALLOC_DEV( dAarray, magmaFloatComplex*, ntile );
TESTING_MALLOC_DEV( dBarray, magmaFloatComplex*, ntile );
lapackf77_clarnv( &ione, ISEED, &size, h_A );
lapackf77_clarnv( &ione, ISEED, &size, h_B );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
magma_csetmatrix( M, N, h_A, lda, d_A, ldda );
magma_csetmatrix( M, N, h_B, lda, d_B, ldda );
// setup pointers
for( int tile = 0; tile < ntile; ++tile ) {
int offset = tile*mstride + tile*nstride*ldda;
hAarray[tile] = &d_A[offset];
hBarray[tile] = &d_B[offset];
}
magma_setvector( ntile, sizeof(magmaFloatComplex*), hAarray, 1, dAarray, 1 );
magma_setvector( ntile, sizeof(magmaFloatComplex*), hBarray, 1, dBarray, 1 );
gpu_time = magma_sync_wtime( 0 );
magmablas_clacpy_batched( MagmaUpperLower, mb, nb, dAarray, ldda, dBarray, ldda, ntile );
gpu_time = magma_sync_wtime( 0 ) - gpu_time;
gpu_perf = gbytes / gpu_time;
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
cpu_time = magma_wtime();
for( int tile = 0; tile < ntile; ++tile ) {
int offset = tile*mstride + tile*nstride*lda;
lapackf77_clacpy( MagmaUpperLowerStr, &mb, &nb,
&h_A[offset], &lda,
&h_B[offset], &lda );
}
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gbytes / cpu_time;
/* =====================================================================
Check the result
=================================================================== */
magma_cgetmatrix( M, N, d_B, ldda, h_A, lda );
blasf77_caxpy(&size, &c_neg_one, h_A, &ione, h_B, &ione);
error = lapackf77_clange("f", &M, &N, h_B, &lda, work);
printf("%5d %5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %s\n",
(int) M, (int) N, (int) ntile,
cpu_perf, cpu_time*1000., gpu_perf, gpu_time*1000.,
(error == 0. ? "ok" : "failed") );
status += ! (error == 0.);
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_B );
TESTING_FREE_DEV( d_A );
TESTING_FREE_DEV( d_B );
TESTING_FREE_CPU( hAarray );
TESTING_FREE_CPU( hBarray );
TESTING_FREE_DEV( dAarray );
TESTING_FREE_DEV( dBarray );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing cposv_gpu
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gflops, cpu_perf, cpu_time, gpu_perf, gpu_time;
float error, Rnorm, Anorm, Xnorm, *work;
magmaFloatComplex c_one = MAGMA_C_ONE;
magmaFloatComplex c_neg_one = MAGMA_C_NEG_ONE;
magmaFloatComplex *h_A, *h_B, *h_X;
magmaFloatComplex_ptr d_A, d_B;
magma_int_t N, lda, ldb, ldda, lddb, info, sizeA, sizeB;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t status = 0;
magma_opts opts;
opts.parse_opts( argc, argv );
float tol = opts.tolerance * lapackf77_slamch("E");
printf("%% uplo = %s\n", lapack_uplo_const(opts.uplo) );
printf("%% N NRHS CPU Gflop/s (sec) GPU Gflop/s (sec) ||B - AX|| / N*||A||*||X||\n");
printf("%%===============================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
N = opts.nsize[itest];
lda = ldb = N;
ldda = magma_roundup( N, opts.align ); // multiple of 32 by default
lddb = ldda;
gflops = ( FLOPS_CPOTRF( N ) + FLOPS_CPOTRS( N, opts.nrhs ) ) / 1e9;
TESTING_MALLOC_CPU( h_A, magmaFloatComplex, lda*N );
TESTING_MALLOC_CPU( h_B, magmaFloatComplex, ldb*opts.nrhs );
TESTING_MALLOC_CPU( h_X, magmaFloatComplex, ldb*opts.nrhs );
TESTING_MALLOC_CPU( work, float, N );
TESTING_MALLOC_DEV( d_A, magmaFloatComplex, ldda*N );
TESTING_MALLOC_DEV( d_B, magmaFloatComplex, lddb*opts.nrhs );
/* ====================================================================
Initialize the matrix
=================================================================== */
sizeA = lda*N;
sizeB = ldb*opts.nrhs;
lapackf77_clarnv( &ione, ISEED, &sizeA, h_A );
lapackf77_clarnv( &ione, ISEED, &sizeB, h_B );
magma_cmake_hpd( N, h_A, lda );
magma_csetmatrix( N, N, h_A, lda, d_A, ldda );
magma_csetmatrix( N, opts.nrhs, h_B, lda, d_B, lddb );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
gpu_time = magma_wtime();
magma_cposv_gpu( opts.uplo, N, opts.nrhs, d_A, ldda, d_B, lddb, &info );
gpu_time = magma_wtime() - gpu_time;
gpu_perf = gflops / gpu_time;
if (info != 0) {
printf("magma_cpotrf_gpu returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
/* =====================================================================
Residual
=================================================================== */
magma_cgetmatrix( N, opts.nrhs, d_B, lddb, h_X, ldb );
Anorm = lapackf77_clange("I", &N, &N, h_A, &lda, work);
Xnorm = lapackf77_clange("I", &N, &opts.nrhs, h_X, &ldb, work);
blasf77_cgemm( MagmaNoTransStr, MagmaNoTransStr, &N, &opts.nrhs, &N,
&c_one, h_A, &lda,
h_X, &ldb,
&c_neg_one, h_B, &ldb );
Rnorm = lapackf77_clange("I", &N, &opts.nrhs, h_B, &ldb, work);
error = Rnorm/(N*Anorm*Xnorm);
status += ! (error < tol);
/* ====================================================================
Performs operation using LAPACK
=================================================================== */
if ( opts.lapack ) {
cpu_time = magma_wtime();
lapackf77_cposv( lapack_uplo_const(opts.uplo), &N, &opts.nrhs, h_A, &lda, h_B, &ldb, &info );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
if (info != 0) {
printf("lapackf77_cposv returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
printf( "%5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %s\n",
(int) N, (int) opts.nrhs, cpu_perf, cpu_time, gpu_perf, gpu_time,
error, (error < tol ? "ok" : "failed"));
}
else {
printf( "%5d %5d --- ( --- ) %7.2f (%7.2f) %8.2e %s\n",
(int) N, (int) opts.nrhs, gpu_perf, gpu_time,
error, (error < tol ? "ok" : "failed"));
}
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_B );
TESTING_FREE_CPU( h_X );
TESTING_FREE_CPU( work );
TESTING_FREE_DEV( d_A );
TESTING_FREE_DEV( d_B );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
opts.cleanup();
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing zhetrd_he2hb
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gflops, gpu_time, gpu_perf;
magmaDoubleComplex *h_A, *h_R, *h_work;
magmaDoubleComplex *tau;
double *D, *E;
magma_int_t N, n2, lda, ldda, lwork, ldt, info, nstream;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t status = 0;
// TODO add these options to parse_opts
magma_int_t NE = 0;
magma_int_t distblk = 0;
magma_opts opts;
parse_opts( argc, argv, &opts );
magma_int_t WANTZ = (opts.jobz == MagmaVec);
double tol = opts.tolerance * lapackf77_dlamch("E");
if (opts.nb == 0)
opts.nb = 64; //magma_get_zhetrd_he2hb_nb(N);
if (NE < 1)
NE = N; //64; //magma_get_zhetrd_he2hb_nb(N);
nstream = max(3, opts.ngpu+2);
magma_queue_t streams[MagmaMaxGPUs][20];
magmaDoubleComplex_ptr da[MagmaMaxGPUs], dT1[MagmaMaxGPUs];
if ((distblk == 0) || (distblk < opts.nb))
distblk = max(256, opts.nb);
printf("voici ngpu %d distblk %d NB %d nstream %d\n ",
(int) opts.ngpu, (int) distblk, (int) opts.nb, (int) nstream);
for( magma_int_t dev = 0; dev < opts.ngpu; ++dev ) {
magma_setdevice( dev );
for( int i = 0; i < nstream; ++i ) {
magma_queue_create( &streams[dev][i] );
}
}
magma_setdevice( 0 );
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
N = opts.nsize[itest];
lda = N;
ldt = N;
ldda = ((N+31)/32)*32;
n2 = N*lda;
/* We suppose the magma NB is bigger than lapack NB */
lwork = N*opts.nb;
//gflops = ....?
/* Allocate host memory for the matrix */
TESTING_MALLOC_CPU( tau, magmaDoubleComplex, N-1 );
TESTING_MALLOC_PIN( h_A, magmaDoubleComplex, lda*N );
TESTING_MALLOC_PIN( h_R, magmaDoubleComplex, lda*N );
TESTING_MALLOC_PIN( h_work, magmaDoubleComplex, lwork );
TESTING_MALLOC_PIN( D, double, N );
TESTING_MALLOC_PIN( E, double, N );
for( magma_int_t dev = 0; dev < opts.ngpu; ++dev ) {
magma_int_t mlocal = ((N / distblk) / opts.ngpu + 1) * distblk;
magma_setdevice( dev );
TESTING_MALLOC_DEV( da[dev], magmaDoubleComplex, ldda*mlocal );
TESTING_MALLOC_DEV( dT1[dev], magmaDoubleComplex, N*opts.nb );
}
/* ====================================================================
Initialize the matrix
=================================================================== */
lapackf77_zlarnv( &ione, ISEED, &n2, h_A );
magma_zmake_hermitian( N, h_A, lda );
lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_A, &lda, h_R, &lda );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
/* Copy the matrix to the GPU */
magma_zsetmatrix_1D_col_bcyclic( N, N, h_R, lda, da, ldda, opts.ngpu, distblk);
//magmaDoubleComplex_ptr dabis;
//TESTING_MALLOC_DEV( dabis, magmaDoubleComplex, ldda*N );
//magma_zsetmatrix(N, N, h_R, lda, dabis, ldda);
for (int count=0; count < 1; ++count) {
magma_setdevice(0);
gpu_time = magma_wtime();
if (opts.version == 30) {
magma_zhetrd_he2hb_mgpu_spec(
opts.uplo, N, opts.nb, h_R, lda, tau, h_work, lwork,
da, ldda, dT1, opts.nb, opts.ngpu, distblk,
streams, nstream, opts.nthread, &info);
} else {
nstream = 3;
magma_zhetrd_he2hb_mgpu(
opts.uplo, N, opts.nb, h_R, lda, tau, h_work, lwork,
da, ldda, dT1, opts.nb, opts.ngpu, distblk,
streams, nstream, opts.nthread, &info);
}
// magma_zhetrd_he2hb(opts.uplo, N, opts.nb, h_R, lda, tau, h_work, lwork, dT1[0], &info);
gpu_time = magma_wtime() - gpu_time;
printf(" Finish BAND N %d NB %d dist %d ngpu %d version %d timing= %f\n",
N, opts.nb, distblk, opts.ngpu, opts.version, gpu_time);
}
magma_setdevice(0);
for( magma_int_t dev = 0; dev < opts.ngpu; ++dev ) {
magma_setdevice(dev);
magma_device_sync();
}
magma_setdevice(0);
magmablasSetKernelStream( NULL );
// todo neither of these is declared in headers
// magma_zhetrd_bhe2trc_v5(opts.nthread, WANTZ, opts.uplo, NE, N, opts.nb, h_R, lda, D, E, dT1[0], ldt);
// magma_zhetrd_bhe2trc(opts.nthread, WANTZ, opts.uplo, NE, N, opts.nb, h_R, lda, D, E, dT1[0], ldt);
// todo where is this timer started?
// gpu_time = magma_wtime() - gpu_time;
// todo what are the gflops?
gpu_perf = gflops / gpu_time;
if (info != 0)
printf("magma_zhetrd_he2hb returned error %d: %s.\n",
(int) info, magma_strerror( info ));
/* =====================================================================
Print performance and error.
=================================================================== */
#if defined(CHECKEIG)
#if defined(PRECISION_z) || defined(PRECISION_d)
if ( opts.check ) {
printf(" Total N %5d flops %6.2f timing %6.2f seconds\n", (int) N, gpu_perf, gpu_time );
char JOBZ;
if (WANTZ == 0)
JOBZ = 'N';
else
JOBZ = 'V';
double nrmI=0.0, nrm1=0.0, nrm2=0.0;
int lwork2 = 256*N;
magmaDoubleComplex *work2, *AINIT;
double *rwork2, *D2;
// TODO free this memory !
magma_zmalloc_cpu( &work2, lwork2 );
magma_dmalloc_cpu( &rwork2, N );
magma_dmalloc_cpu( &D2, N );
magma_zmalloc_cpu( &AINIT, N*lda );
memcpy(AINIT, h_A, N*lda*sizeof(magmaDoubleComplex));
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
cpu_time = magma_wtime();
int nt = min(12, opts.nthread);
#if defined(USEMKL)
mkl_set_num_threads(nt);
#endif
#if defined(USEACML)
omp_set_num_threads(nt);
#endif
#if defined(PRECISION_z) || defined (PRECISION_c)
lapackf77_zheev( "N", "L", &N, h_A, &lda, D2, work2, &lwork2, rwork2, &info );
#else
lapackf77_dsyev( "N", "L", &N, h_A, &lda, D2, work2, &lwork2, &info );
#endif
///* call eigensolver for our resulting tridiag [D E] and for Q */
//dstedc_withZ('V', N, D, E, h_R, lda);
////dsterf_( &N, D, E, &info);
cpu_time = magma_wtime() - cpu_time;
printf(" Finish CHECK - EIGEN timing= %f threads %d\n", cpu_time, nt);
/* compare result */
cmp_vals(N, D2, D, &nrmI, &nrm1, &nrm2);
magmaDoubleComplex *WORKAJETER;
double *RWORKAJETER, *RESU;
// TODO free this memory !
magma_zmalloc_cpu( &WORKAJETER, (2* N * N + N) );
magma_dmalloc_cpu( &RWORKAJETER, N );
magma_dmalloc_cpu( &RESU, 10 );
int MATYPE;
memset(RESU, 0, 10*sizeof(double));
MATYPE=3;
double NOTHING=0.0;
cpu_time = magma_wtime();
// check results
zcheck_eig_( lapack_vec_const(opts.jobz), &MATYPE, &N, &opts.nb,
AINIT, &lda, &NOTHING, &NOTHING, D2, D,
h_R, &lda, WORKAJETER, RWORKAJETER, RESU );
cpu_time = magma_wtime() - cpu_time;
printf(" Finish CHECK - results timing= %f\n", cpu_time);
#if defined(USEMKL)
mkl_set_num_threads(1);
#endif
#if defined(USEACML)
omp_set_num_threads(1);
#endif
printf("\n");
printf(" ================================================================================================================\n");
printf(" ==> INFO voici threads=%d N=%d NB=%d WANTZ=%d\n", (int) opts.nthread, (int) N, (int) opts.nb, (int) WANTZ);
printf(" ================================================================================================================\n");
printf(" DSBTRD : %15s \n", "STATblgv9withQ ");
printf(" ================================================================================================================\n");
if (WANTZ > 0)
printf(" | A - U S U' | / ( |A| n ulp ) : %15.3E \n", RESU[0]);
if (WANTZ > 0)
printf(" | I - U U' | / ( n ulp ) : %15.3E \n", RESU[1]);
printf(" | D1 - EVEIGS | / (|D| ulp) : %15.3E \n", RESU[2]);
printf(" max | D1 - EVEIGS | : %15.3E \n", RESU[6]);
printf(" ================================================================================================================\n\n\n");
printf(" ****************************************************************************************************************\n");
printf(" * Hello here are the norm Infinite (max)=%8.2e norm one (sum)=%8.2e norm2(sqrt)=%8.2e *\n", nrmI, nrm1, nrm2);
printf(" ****************************************************************************************************************\n\n");
}
#endif // PRECISION_z || PRECISION_d
#endif // CHECKEIG
printf(" Total N %5d flops %6.2f timing %6.2f seconds\n", (int) N, 0.0, gpu_time );
printf("============================================================================\n\n\n");
TESTING_FREE_CPU( tau );
TESTING_FREE_PIN( h_A );
TESTING_FREE_PIN( h_R );
TESTING_FREE_PIN( h_work );
TESTING_FREE_PIN( D );
TESTING_FREE_PIN( E );
for( magma_int_t dev = 0; dev < opts.ngpu; ++dev ) {
magma_setdevice( dev );
TESTING_FREE_DEV( da[dev] );
TESTING_FREE_DEV( dT1[dev] );
}
magma_setdevice( 0 );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
for( magma_int_t dev = 0; dev < opts.ngpu; ++dev ) {
for( int i = 0; i < nstream; ++i ) {
magma_queue_destroy( streams[dev][i] );
}
}
TESTING_FINALIZE();
return status;
}