/* ////////////////////////////////////////////////////////////////////////////
-- Testing sormbr
*/
int main( int argc, char** argv )
{
TESTING_INIT();
real_Double_t gflops, gpu_perf, gpu_time, cpu_perf, cpu_time;
float Cnorm, error, dwork[1];
float c_neg_one = MAGMA_S_NEG_ONE;
magma_int_t ione = 1;
magma_int_t m, n, k, mi, ni, mm, nn, nq, size, info;
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t nb, ldc, lda, lwork, lwork_max;
float *C, *R, *A, *work, *tau, *tauq, *taup;
float *d, *e;
magma_int_t status = 0;
magma_opts opts;
opts.parse_opts( argc, argv );
// need slightly looser bound (60*eps instead of 30*eps) for some tests
opts.tolerance = max( 60., opts.tolerance );
float tol = opts.tolerance * lapackf77_slamch("E");
// test all combinations of input parameters
magma_vect_t vect [] = { MagmaQ, MagmaP };
magma_side_t side [] = { MagmaLeft, MagmaRight };
magma_trans_t trans[] = { MagmaTrans, MagmaNoTrans };
printf("%% M N K vect side trans CPU Gflop/s (sec) GPU Gflop/s (sec) ||R||_F / ||QC||_F\n");
printf("%%==============================================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int ivect = 0; ivect < 2; ++ivect ) {
for( int iside = 0; iside < 2; ++iside ) {
for( int itran = 0; itran < 2; ++itran ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
m = opts.msize[itest];
n = opts.nsize[itest];
k = opts.ksize[itest];
nb = magma_get_sgebrd_nb( m, n );
ldc = m;
// A is nq x k (vect=Q) or k x nq (vect=P)
// where nq=m (left) or nq=n (right)
nq = (side[iside] == MagmaLeft ? m : n );
mm = (vect[ivect] == MagmaQ ? nq : k );
nn = (vect[ivect] == MagmaQ ? k : nq);
lda = mm;
// MBR calls either MQR or MLQ in various ways
if ( vect[ivect] == MagmaQ ) {
if ( nq >= k ) {
gflops = FLOPS_SORMQR( m, n, k, side[iside] ) / 1e9;
}
else {
if ( side[iside] == MagmaLeft ) {
mi = m - 1;
ni = n;
}
else {
mi = m;
ni = n - 1;
}
gflops = FLOPS_SORMQR( mi, ni, nq-1, side[iside] ) / 1e9;
}
}
else {
if ( nq > k ) {
gflops = FLOPS_SORMLQ( m, n, k, side[iside] ) / 1e9;
}
else {
if ( side[iside] == MagmaLeft ) {
mi = m - 1;
ni = n;
}
else {
mi = m;
ni = n - 1;
}
gflops = FLOPS_SORMLQ( mi, ni, nq-1, side[iside] ) / 1e9;
}
}
// workspace for gebrd is (mm + nn)*nb
// workspace for unmbr is m*nb or n*nb, depending on side
lwork_max = max( (mm + nn)*nb, max( m*nb, n*nb ));
// this rounds it up slightly if needed to agree with lwork query below
lwork_max = int( real( magma_smake_lwork( lwork_max )));
TESTING_MALLOC_CPU( C, float, ldc*n );
TESTING_MALLOC_CPU( R, float, ldc*n );
TESTING_MALLOC_CPU( A, float, lda*nn );
TESTING_MALLOC_CPU( work, float, lwork_max );
TESTING_MALLOC_CPU( d, float, min(mm,nn) );
TESTING_MALLOC_CPU( e, float, min(mm,nn) );
TESTING_MALLOC_CPU( tauq, float, min(mm,nn) );
TESTING_MALLOC_CPU( taup, float, min(mm,nn) );
// C is full, m x n
size = ldc*n;
lapackf77_slarnv( &ione, ISEED, &size, C );
lapackf77_slacpy( "Full", &m, &n, C, &ldc, R, &ldc );
size = lda*nn;
lapackf77_slarnv( &ione, ISEED, &size, A );
// compute BRD factorization to get Householder vectors in A, tauq, taup
//lapackf77_sgebrd( &mm, &nn, A, &lda, d, e, tauq, taup, work, &lwork_max, &info );
magma_sgebrd( mm, nn, A, lda, d, e, tauq, taup, work, lwork_max, &info );
if (info != 0) {
printf("magma_sgebrd returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
if ( vect[ivect] == MagmaQ ) {
tau = tauq;
} else {
tau = taup;
}
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
cpu_time = magma_wtime();
lapackf77_sormbr( lapack_vect_const( vect[ivect] ),
lapack_side_const( side[iside] ),
lapack_trans_const( trans[itran] ),
&m, &n, &k,
A, &lda, tau, C, &ldc, work, &lwork_max, &info );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
if (info != 0) {
printf("lapackf77_sormbr returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
// query for workspace size
lwork = -1;
magma_sormbr( vect[ivect], side[iside], trans[itran],
m, n, k,
A, lda, tau, R, ldc, work, lwork, &info );
if (info != 0) {
printf("magma_sormbr (lwork query) returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
lwork = (magma_int_t) MAGMA_S_REAL( work[0] );
if ( lwork < 0 || lwork > lwork_max ) {
printf("Warning: optimal lwork %d > allocated lwork_max %d\n", (int) lwork, (int) lwork_max );
lwork = lwork_max;
}
gpu_time = magma_wtime();
magma_sormbr( vect[ivect], side[iside], trans[itran],
m, n, k,
A, lda, tau, R, ldc, work, lwork, &info );
gpu_time = magma_wtime() - gpu_time;
gpu_perf = gflops / gpu_time;
if (info != 0) {
printf("magma_sormbr returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
/* =====================================================================
compute relative error |QC_magma - QC_lapack| / |QC_lapack|
=================================================================== */
size = ldc*n;
blasf77_saxpy( &size, &c_neg_one, C, &ione, R, &ione );
Cnorm = lapackf77_slange( "Fro", &m, &n, C, &ldc, dwork );
error = lapackf77_slange( "Fro", &m, &n, R, &ldc, dwork ) / (magma_ssqrt(m*n) * Cnorm);
printf( "%5d %5d %5d %c %4c %5c %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %s\n",
(int) m, (int) n, (int) k,
lapacke_vect_const( vect[ivect] ),
lapacke_side_const( side[iside] ),
lapacke_trans_const( trans[itran] ),
cpu_perf, cpu_time, gpu_perf, gpu_time,
error, (error < tol ? "ok" : "failed") );
status += ! (error < tol);
TESTING_FREE_CPU( C );
TESTING_FREE_CPU( R );
TESTING_FREE_CPU( A );
TESTING_FREE_CPU( work );
TESTING_FREE_CPU( d );
TESTING_FREE_CPU( e );
TESTING_FREE_CPU( taup );
TESTING_FREE_CPU( tauq );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}}} // end ivect, iside, itran
printf( "\n" );
}
opts.cleanup();
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing cunmlq
*/
int main( int argc, char** argv )
{
TESTING_INIT();
real_Double_t gflops, gpu_perf, gpu_time, cpu_perf, cpu_time;
float Cnorm, error, work[1];
magmaFloatComplex c_neg_one = MAGMA_C_NEG_ONE;
magma_int_t ione = 1;
magma_int_t mm, m, n, k, size, info;
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t nb, ldc, lda, lwork, lwork_max;
magmaFloatComplex *C, *R, *A, *W, *tau;
magma_int_t status = 0;
magma_opts opts;
opts.parse_opts( argc, argv );
// need slightly looser bound (60*eps instead of 30*eps) for some tests
opts.tolerance = max( 60., opts.tolerance );
float tol = opts.tolerance * lapackf77_slamch("E");
// test all combinations of input parameters
magma_side_t side [] = { MagmaLeft, MagmaRight };
magma_trans_t trans[] = { Magma_ConjTrans, MagmaNoTrans };
printf("%% M N K side trans CPU Gflop/s (sec) GPU Gflop/s (sec) ||R||_F / ||QC||_F\n");
printf("%%==============================================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iside = 0; iside < 2; ++iside ) {
for( int itran = 0; itran < 2; ++itran ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
m = opts.msize[itest];
n = opts.nsize[itest];
k = opts.ksize[itest];
nb = magma_get_cgelqf_nb( m, n );
ldc = m;
// A is k x m (left) or k x n (right)
mm = (side[iside] == MagmaLeft ? m : n);
lda = k;
gflops = FLOPS_CUNMLQ( m, n, k, side[iside] ) / 1e9;
if ( side[iside] == MagmaLeft && m < k ) {
printf( "%5d %5d %5d %4c %5c skipping because side=left and m < k\n",
(int) m, (int) n, (int) k,
lapacke_side_const( side[iside] ),
lapacke_trans_const( trans[itran] ) );
continue;
}
if ( side[iside] == MagmaRight && n < k ) {
printf( "%5d %5d %5d %4c %5c skipping because side=right and n < k\n",
(int) m, (int) n, (int) k,
lapacke_side_const( side[iside] ),
lapacke_trans_const( trans[itran] ) );
continue;
}
// need at least 2*nb*nb for gelqf
lwork_max = max( max( m*nb, n*nb ), 2*nb*nb );
// this rounds it up slightly if needed to agree with lwork query
lwork_max = int( real( magma_cmake_lwork( lwork_max )));
TESTING_MALLOC_CPU( C, magmaFloatComplex, ldc*n );
TESTING_MALLOC_CPU( R, magmaFloatComplex, ldc*n );
TESTING_MALLOC_CPU( A, magmaFloatComplex, lda*mm );
TESTING_MALLOC_CPU( W, magmaFloatComplex, lwork_max );
TESTING_MALLOC_CPU( tau, magmaFloatComplex, k );
// C is full, m x n
size = ldc*n;
lapackf77_clarnv( &ione, ISEED, &size, C );
lapackf77_clacpy( "Full", &m, &n, C, &ldc, R, &ldc );
size = lda*mm;
lapackf77_clarnv( &ione, ISEED, &size, A );
// compute LQ factorization to get Householder vectors in A, tau
magma_cgelqf( k, mm, A, lda, tau, W, lwork_max, &info );
if (info != 0) {
printf("magma_cgelqf returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
cpu_time = magma_wtime();
lapackf77_cunmlq( lapack_side_const( side[iside] ), lapack_trans_const( trans[itran] ),
&m, &n, &k,
A, &lda, tau, C, &ldc, W, &lwork_max, &info );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
if (info != 0) {
printf("lapackf77_cunmlq returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
// query for workspace size
lwork = -1;
magma_cunmlq( side[iside], trans[itran],
m, n, k,
A, lda, tau, R, ldc, W, lwork, &info );
if (info != 0) {
printf("magma_cunmlq (lwork query) returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
lwork = (magma_int_t) MAGMA_C_REAL( W[0] );
if ( lwork < 0 || lwork > lwork_max ) {
printf("Warning: optimal lwork %d > allocated lwork_max %d\n", (int) lwork, (int) lwork_max );
lwork = lwork_max;
}
gpu_time = magma_wtime();
magma_cunmlq( side[iside], trans[itran],
m, n, k,
A, lda, tau, R, ldc, W, lwork, &info );
gpu_time = magma_wtime() - gpu_time;
gpu_perf = gflops / gpu_time;
if (info != 0) {
printf("magma_cunmlq returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
/* =====================================================================
compute relative error |QC_magma - QC_lapack| / |QC_lapack|
=================================================================== */
size = ldc*n;
blasf77_caxpy( &size, &c_neg_one, C, &ione, R, &ione );
Cnorm = lapackf77_clange( "Fro", &m, &n, C, &ldc, work );
error = lapackf77_clange( "Fro", &m, &n, R, &ldc, work ) / (magma_ssqrt(m*n) * Cnorm);
printf( "%5d %5d %5d %4c %5c %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %s\n",
(int) m, (int) n, (int) k,
lapacke_side_const( side[iside] ),
lapacke_trans_const( trans[itran] ),
cpu_perf, cpu_time, gpu_perf, gpu_time,
error, (error < tol ? "ok" : "failed") );
status += ! (error < tol);
TESTING_FREE_CPU( C );
TESTING_FREE_CPU( R );
TESTING_FREE_CPU( A );
TESTING_FREE_CPU( W );
TESTING_FREE_CPU( tau );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}} // end iside, itran
printf( "\n" );
}
opts.cleanup();
TESTING_FINALIZE();
return status;
}
int main( int argc, char** argv )
{
TESTING_INIT();
real_Double_t gflops, t1, t2;
magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
magma_int_t ione = 1;
magma_trans_t trans[] = { MagmaNoTrans, MagmaConjTrans, MagmaTrans };
magma_uplo_t uplo [] = { MagmaLower, MagmaUpper };
magma_diag_t diag [] = { MagmaUnit, MagmaNonUnit };
magma_side_t side [] = { MagmaLeft, MagmaRight };
magmaDoubleComplex *A, *B, *C, *C2, *LU;
magmaDoubleComplex *dA, *dB, *dC1, *dC2;
magmaDoubleComplex alpha = MAGMA_Z_MAKE( 0.5, 0.1 );
magmaDoubleComplex beta = MAGMA_Z_MAKE( 0.7, 0.2 );
double dalpha = 0.6;
double dbeta = 0.8;
double work[1], error, total_error;
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t m, n, k, size, maxn, ld, info;
magma_int_t *piv;
magma_int_t err;
magma_opts opts;
parse_opts( argc, argv, &opts );
printf( "Compares magma wrapper function to cublas function; all diffs should be exactly 0.\n\n" );
total_error = 0.;
for( int itest = 0; itest < opts.ntest; ++itest ) {
m = opts.msize[itest];
n = opts.nsize[itest];
k = opts.ksize[itest];
printf("=========================================================================\n");
printf( "m=%d, n=%d, k=%d\n", (int) m, (int) n, (int) k );
// allocate matrices
// over-allocate so they can be any combination of {m,n,k} x {m,n,k}.
maxn = max( max( m, n ), k );
ld = max( 1, maxn );
size = ld*maxn;
err = magma_malloc_cpu( (void**) &piv, maxn*sizeof(magma_int_t) ); assert( err == 0 );
err = magma_zmalloc_pinned( &A, size ); assert( err == 0 );
err = magma_zmalloc_pinned( &B, size ); assert( err == 0 );
err = magma_zmalloc_pinned( &C, size ); assert( err == 0 );
err = magma_zmalloc_pinned( &C2, size ); assert( err == 0 );
err = magma_zmalloc_pinned( &LU, size ); assert( err == 0 );
err = magma_zmalloc( &dA, size ); assert( err == 0 );
err = magma_zmalloc( &dB, size ); assert( err == 0 );
err = magma_zmalloc( &dC1, size ); assert( err == 0 );
err = magma_zmalloc( &dC2, size ); assert( err == 0 );
// initialize matrices
size = maxn*maxn;
lapackf77_zlarnv( &ione, ISEED, &size, A );
lapackf77_zlarnv( &ione, ISEED, &size, B );
lapackf77_zlarnv( &ione, ISEED, &size, C );
printf( "========== Level 1 BLAS ==========\n" );
// ----- test ZSWAP
// swap columns 2 and 3 of dA, then copy to C2 and compare with A
if ( n >= 3 ) {
magma_zsetmatrix( m, n, A, ld, dA, ld );
magma_zsetmatrix( m, n, A, ld, dB, ld );
magma_zswap( m, dA(0,1), 1, dA(0,2), 1 );
magma_zswap( m, dB(0,1), 1, dB(0,2), 1 );
// check results, storing diff between magma and cuda calls in C2
cublasZaxpy( handle, ld*n, &c_neg_one, dA, 1, dB, 1 );
magma_zgetmatrix( m, n, dB, ld, C2, ld );
error = lapackf77_zlange( "F", &m, &k, C2, &ld, work );
total_error += error;
printf( "zswap diff %.2g\n", error );
}
else {
printf( "zswap skipped for n < 3\n" );
}
// ----- test IZAMAX
// get argmax of column of A
magma_zsetmatrix( m, k, A, ld, dA, ld );
error = 0;
for( int j = 0; j < k; ++j ) {
magma_int_t i1 = magma_izamax( m, dA(0,j), 1 );
int i2; // NOT magma_int_t, for cublas
cublasIzamax( handle, m, dA(0,j), 1, &i2 );
// todo need sync here?
assert( i1 == i2 );
error += abs( i1 - i2 );
}
total_error += error;
gflops = (double)m * k / 1e9;
printf( "izamax diff %.2g\n", error );
printf( "\n" );
printf( "========== Level 2 BLAS ==========\n" );
// ----- test ZGEMV
// c = alpha*A*b + beta*c, with A m*n; b,c m or n-vectors
// try no-trans/trans
for( int ia = 0; ia < 3; ++ia ) {
magma_zsetmatrix( m, n, A, ld, dA, ld );
magma_zsetvector( maxn, B, 1, dB, 1 );
magma_zsetvector( maxn, C, 1, dC1, 1 );
magma_zsetvector( maxn, C, 1, dC2, 1 );
t1 = magma_sync_wtime( 0 );
magma_zgemv( trans[ia], m, n, alpha, dA, ld, dB, 1, beta, dC1, 1 );
t1 = magma_sync_wtime( 0 ) - t1;
t2 = magma_sync_wtime( 0 );
cublasZgemv( handle, cublas_trans_const(trans[ia]),
m, n, &alpha, dA, ld, dB, 1, &beta, dC2, 1 );
t2 = magma_sync_wtime( 0 ) - t2;
// check results, storing diff between magma and cuda call in C2
size = (trans[ia] == MagmaNoTrans ? m : n);
cublasZaxpy( handle, size, &c_neg_one, dC1, 1, dC2, 1 );
magma_zgetvector( size, dC2, 1, C2, 1 );
error = lapackf77_zlange( "F", &size, &ione, C2, &ld, work );
total_error += error;
gflops = FLOPS_ZGEMV( m, n ) / 1e9;
printf( "zgemv( %c ) diff %.2g, Gflop/s %7.2f, %7.2f\n",
lapacke_trans_const(trans[ia]), error, gflops/t1, gflops/t2 );
}
printf( "\n" );
// ----- test ZHEMV
// c = alpha*A*b + beta*c, with A m*m symmetric; b,c m-vectors
// try upper/lower
for( int iu = 0; iu < 2; ++iu ) {
magma_zsetmatrix( m, m, A, ld, dA, ld );
magma_zsetvector( m, B, 1, dB, 1 );
magma_zsetvector( m, C, 1, dC1, 1 );
magma_zsetvector( m, C, 1, dC2, 1 );
t1 = magma_sync_wtime( 0 );
magma_zhemv( uplo[iu], m, alpha, dA, ld, dB, 1, beta, dC1, 1 );
t1 = magma_sync_wtime( 0 ) - t1;
t2 = magma_sync_wtime( 0 );
cublasZhemv( handle, cublas_uplo_const(uplo[iu]),
m, &alpha, dA, ld, dB, 1, &beta, dC2, 1 );
t2 = magma_sync_wtime( 0 ) - t2;
// check results, storing diff between magma and cuda call in C2
cublasZaxpy( handle, m, &c_neg_one, dC1, 1, dC2, 1 );
magma_zgetvector( m, dC2, 1, C2, 1 );
error = lapackf77_zlange( "F", &m, &ione, C2, &ld, work );
total_error += error;
gflops = FLOPS_ZHEMV( m ) / 1e9;
printf( "zhemv( %c ) diff %.2g, Gflop/s %7.2f, %7.2f\n",
lapacke_uplo_const(uplo[iu]), error, gflops/t1, gflops/t2 );
}
printf( "\n" );
// ----- test ZTRSV
// solve A*c = c, with A m*m triangular; c m-vector
// try upper/lower, no-trans/trans, unit/non-unit diag
// Factor A into LU to get well-conditioned triangles, else solve yields garbage.
// Still can give garbage if solves aren't consistent with LU factors,
// e.g., using unit diag for U, so copy lower triangle to upper triangle.
// Also used for trsm later.
lapackf77_zlacpy( "Full", &maxn, &maxn, A, &ld, LU, &ld );
lapackf77_zgetrf( &maxn, &maxn, LU, &ld, piv, &info );
for( int j = 0; j < maxn; ++j ) {
for( int i = 0; i < j; ++i ) {
*LU(i,j) = *LU(j,i);
}
}
for( int iu = 0; iu < 2; ++iu ) {
for( int it = 0; it < 3; ++it ) {
for( int id = 0; id < 2; ++id ) {
magma_zsetmatrix( m, m, LU, ld, dA, ld );
magma_zsetvector( m, C, 1, dC1, 1 );
magma_zsetvector( m, C, 1, dC2, 1 );
t1 = magma_sync_wtime( 0 );
magma_ztrsv( uplo[iu], trans[it], diag[id], m, dA, ld, dC1, 1 );
t1 = magma_sync_wtime( 0 ) - t1;
t2 = magma_sync_wtime( 0 );
cublasZtrsv( handle, cublas_uplo_const(uplo[iu]), cublas_trans_const(trans[it]),
cublas_diag_const(diag[id]), m, dA, ld, dC2, 1 );
t2 = magma_sync_wtime( 0 ) - t2;
// check results, storing diff between magma and cuda call in C2
cublasZaxpy( handle, m, &c_neg_one, dC1, 1, dC2, 1 );
magma_zgetvector( m, dC2, 1, C2, 1 );
error = lapackf77_zlange( "F", &m, &ione, C2, &ld, work );
total_error += error;
gflops = FLOPS_ZTRSM( MagmaLeft, m, 1 ) / 1e9;
printf( "ztrsv( %c, %c, %c ) diff %.2g, Gflop/s %7.2f, %7.2f\n",
lapacke_uplo_const(uplo[iu]), lapacke_trans_const(trans[it]), lapacke_diag_const(diag[id]),
error, gflops/t1, gflops/t2 );
}}}
printf( "\n" );
printf( "========== Level 3 BLAS ==========\n" );
// ----- test ZGEMM
// C = alpha*A*B + beta*C, with A m*k or k*m; B k*n or n*k; C m*n
// try combinations of no-trans/trans
for( int ia = 0; ia < 3; ++ia ) {
for( int ib = 0; ib < 3; ++ib ) {
bool nta = (trans[ia] == MagmaNoTrans);
bool ntb = (trans[ib] == MagmaNoTrans);
magma_zsetmatrix( (nta ? m : k), (nta ? m : k), A, ld, dA, ld );
magma_zsetmatrix( (ntb ? k : n), (ntb ? n : k), B, ld, dB, ld );
magma_zsetmatrix( m, n, C, ld, dC1, ld );
magma_zsetmatrix( m, n, C, ld, dC2, ld );
t1 = magma_sync_wtime( 0 );
magma_zgemm( trans[ia], trans[ib], m, n, k, alpha, dA, ld, dB, ld, beta, dC1, ld );
t1 = magma_sync_wtime( 0 ) - t1;
t2 = magma_sync_wtime( 0 );
cublasZgemm( handle, cublas_trans_const(trans[ia]), cublas_trans_const(trans[ib]),
m, n, k, &alpha, dA, ld, dB, ld, &beta, dC2, ld );
t2 = magma_sync_wtime( 0 ) - t2;
// check results, storing diff between magma and cuda call in C2
cublasZaxpy( handle, ld*n, &c_neg_one, dC1, 1, dC2, 1 );
magma_zgetmatrix( m, n, dC2, ld, C2, ld );
error = lapackf77_zlange( "F", &m, &n, C2, &ld, work );
total_error += error;
gflops = FLOPS_ZGEMM( m, n, k ) / 1e9;
printf( "zgemm( %c, %c ) diff %.2g, Gflop/s %7.2f, %7.2f\n",
lapacke_trans_const(trans[ia]), lapacke_trans_const(trans[ib]),
error, gflops/t1, gflops/t2 );
}}
printf( "\n" );
// ----- test ZHEMM
// C = alpha*A*B + beta*C (left) with A m*m symmetric; B,C m*n; or
// C = alpha*B*A + beta*C (right) with A n*n symmetric; B,C m*n
// try left/right, upper/lower
for( int is = 0; is < 2; ++is ) {
for( int iu = 0; iu < 2; ++iu ) {
magma_zsetmatrix( m, m, A, ld, dA, ld );
magma_zsetmatrix( m, n, B, ld, dB, ld );
magma_zsetmatrix( m, n, C, ld, dC1, ld );
magma_zsetmatrix( m, n, C, ld, dC2, ld );
t1 = magma_sync_wtime( 0 );
magma_zhemm( side[is], uplo[iu], m, n, alpha, dA, ld, dB, ld, beta, dC1, ld );
t1 = magma_sync_wtime( 0 ) - t1;
t2 = magma_sync_wtime( 0 );
cublasZhemm( handle, cublas_side_const(side[is]), cublas_uplo_const(uplo[iu]),
m, n, &alpha, dA, ld, dB, ld, &beta, dC2, ld );
t2 = magma_sync_wtime( 0 ) - t2;
// check results, storing diff between magma and cuda call in C2
cublasZaxpy( handle, ld*n, &c_neg_one, dC1, 1, dC2, 1 );
magma_zgetmatrix( m, n, dC2, ld, C2, ld );
error = lapackf77_zlange( "F", &m, &n, C2, &ld, work );
total_error += error;
gflops = FLOPS_ZHEMM( side[is], m, n ) / 1e9;
printf( "zhemm( %c, %c ) diff %.2g, Gflop/s %7.2f, %7.2f\n",
lapacke_side_const(side[is]), lapacke_uplo_const(uplo[iu]),
error, gflops/t1, gflops/t2 );
}}
printf( "\n" );
// ----- test ZHERK
// C = alpha*A*A^H + beta*C (no-trans) with A m*k and C m*m symmetric; or
// C = alpha*A^H*A + beta*C (trans) with A k*m and C m*m symmetric
// try upper/lower, no-trans/trans
for( int iu = 0; iu < 2; ++iu ) {
for( int it = 0; it < 3; ++it ) {
magma_zsetmatrix( n, k, A, ld, dA, ld );
magma_zsetmatrix( n, n, C, ld, dC1, ld );
magma_zsetmatrix( n, n, C, ld, dC2, ld );
t1 = magma_sync_wtime( 0 );
magma_zherk( uplo[iu], trans[it], n, k, dalpha, dA, ld, dbeta, dC1, ld );
t1 = magma_sync_wtime( 0 ) - t1;
t2 = magma_sync_wtime( 0 );
cublasZherk( handle, cublas_uplo_const(uplo[iu]), cublas_trans_const(trans[it]),
n, k, &dalpha, dA, ld, &dbeta, dC2, ld );
t2 = magma_sync_wtime( 0 ) - t2;
// check results, storing diff between magma and cuda call in C2
cublasZaxpy( handle, ld*n, &c_neg_one, dC1, 1, dC2, 1 );
magma_zgetmatrix( n, n, dC2, ld, C2, ld );
error = lapackf77_zlange( "F", &n, &n, C2, &ld, work );
total_error += error;
gflops = FLOPS_ZHERK( k, n ) / 1e9;
printf( "zherk( %c, %c ) diff %.2g, Gflop/s %7.2f, %7.2f\n",
lapacke_uplo_const(uplo[iu]), lapacke_trans_const(trans[it]),
error, gflops/t1, gflops/t2 );
}}
printf( "\n" );
// ----- test ZHER2K
// C = alpha*A*B^H + ^alpha*B*A^H + beta*C (no-trans) with A,B n*k; C n*n symmetric; or
// C = alpha*A^H*B + ^alpha*B^H*A + beta*C (trans) with A,B k*n; C n*n symmetric
// try upper/lower, no-trans/trans
for( int iu = 0; iu < 2; ++iu ) {
for( int it = 0; it < 3; ++it ) {
bool nt = (trans[it] == MagmaNoTrans);
magma_zsetmatrix( (nt ? n : k), (nt ? n : k), A, ld, dA, ld );
magma_zsetmatrix( n, n, C, ld, dC1, ld );
magma_zsetmatrix( n, n, C, ld, dC2, ld );
t1 = magma_sync_wtime( 0 );
magma_zher2k( uplo[iu], trans[it], n, k, alpha, dA, ld, dB, ld, dbeta, dC1, ld );
t1 = magma_sync_wtime( 0 ) - t1;
t2 = magma_sync_wtime( 0 );
cublasZher2k( handle, cublas_uplo_const(uplo[iu]), cublas_trans_const(trans[it]),
n, k, &alpha, dA, ld, dB, ld, &dbeta, dC2, ld );
t2 = magma_sync_wtime( 0 ) - t2;
// check results, storing diff between magma and cuda call in C2
cublasZaxpy( handle, ld*n, &c_neg_one, dC1, 1, dC2, 1 );
magma_zgetmatrix( n, n, dC2, ld, C2, ld );
error = lapackf77_zlange( "F", &n, &n, C2, &ld, work );
total_error += error;
gflops = FLOPS_ZHER2K( k, n ) / 1e9;
printf( "zher2k( %c, %c ) diff %.2g, Gflop/s %7.2f, %7.2f\n",
lapacke_uplo_const(uplo[iu]), lapacke_trans_const(trans[it]),
error, gflops/t1, gflops/t2 );
}}
printf( "\n" );
// ----- test ZTRMM
// C = alpha*A*C (left) with A m*m triangular; C m*n; or
// C = alpha*C*A (right) with A n*n triangular; C m*n
// try left/right, upper/lower, no-trans/trans, unit/non-unit
for( int is = 0; is < 2; ++is ) {
for( int iu = 0; iu < 2; ++iu ) {
for( int it = 0; it < 3; ++it ) {
for( int id = 0; id < 2; ++id ) {
bool left = (side[is] == MagmaLeft);
magma_zsetmatrix( (left ? m : n), (left ? m : n), A, ld, dA, ld );
magma_zsetmatrix( m, n, C, ld, dC1, ld );
magma_zsetmatrix( m, n, C, ld, dC2, ld );
t1 = magma_sync_wtime( 0 );
magma_ztrmm( side[is], uplo[iu], trans[it], diag[id], m, n, alpha, dA, ld, dC1, ld );
t1 = magma_sync_wtime( 0 ) - t1;
// note cublas does trmm out-of-place (i.e., adds output matrix C),
// but allows C=B to do in-place.
t2 = magma_sync_wtime( 0 );
cublasZtrmm( handle, cublas_side_const(side[is]), cublas_uplo_const(uplo[iu]),
cublas_trans_const(trans[it]), cublas_diag_const(diag[id]),
m, n, &alpha, dA, ld, dC2, ld, dC2, ld );
t2 = magma_sync_wtime( 0 ) - t2;
// check results, storing diff between magma and cuda call in C2
cublasZaxpy( handle, ld*n, &c_neg_one, dC1, 1, dC2, 1 );
magma_zgetmatrix( m, n, dC2, ld, C2, ld );
error = lapackf77_zlange( "F", &n, &n, C2, &ld, work );
total_error += error;
gflops = FLOPS_ZTRMM( side[is], m, n ) / 1e9;
printf( "ztrmm( %c, %c ) diff %.2g, Gflop/s %7.2f, %7.2f\n",
lapacke_uplo_const(uplo[iu]), lapacke_trans_const(trans[it]),
error, gflops/t1, gflops/t2 );
}}}}
printf( "\n" );
// ----- test ZTRSM
// solve A*X = alpha*B (left) with A m*m triangular; B m*n; or
// solve X*A = alpha*B (right) with A n*n triangular; B m*n
// try left/right, upper/lower, no-trans/trans, unit/non-unit
for( int is = 0; is < 2; ++is ) {
for( int iu = 0; iu < 2; ++iu ) {
for( int it = 0; it < 3; ++it ) {
for( int id = 0; id < 2; ++id ) {
bool left = (side[is] == MagmaLeft);
magma_zsetmatrix( (left ? m : n), (left ? m : n), LU, ld, dA, ld );
magma_zsetmatrix( m, n, C, ld, dC1, ld );
magma_zsetmatrix( m, n, C, ld, dC2, ld );
t1 = magma_sync_wtime( 0 );
magma_ztrsm( side[is], uplo[iu], trans[it], diag[id], m, n, alpha, dA, ld, dC1, ld );
t1 = magma_sync_wtime( 0 ) - t1;
t2 = magma_sync_wtime( 0 );
cublasZtrsm( handle, cublas_side_const(side[is]), cublas_uplo_const(uplo[iu]),
cublas_trans_const(trans[it]), cublas_diag_const(diag[id]),
m, n, &alpha, dA, ld, dC2, ld );
t2 = magma_sync_wtime( 0 ) - t2;
// check results, storing diff between magma and cuda call in C2
cublasZaxpy( handle, ld*n, &c_neg_one, dC1, 1, dC2, 1 );
magma_zgetmatrix( m, n, dC2, ld, C2, ld );
error = lapackf77_zlange( "F", &n, &n, C2, &ld, work );
total_error += error;
gflops = FLOPS_ZTRSM( side[is], m, n ) / 1e9;
printf( "ztrsm( %c, %c ) diff %.2g, Gflop/s %7.2f, %7.2f\n",
lapacke_uplo_const(uplo[iu]), lapacke_trans_const(trans[it]),
error, gflops/t1, gflops/t2 );
}}}}
printf( "\n" );
// cleanup
magma_free_cpu( piv );
magma_free_pinned( A );
magma_free_pinned( B );
magma_free_pinned( C );
magma_free_pinned( C2 );
magma_free_pinned( LU );
magma_free( dA );
magma_free( dB );
magma_free( dC1 );
magma_free( dC2 );
fflush( stdout );
}
if ( total_error != 0. ) {
printf( "total error %.2g -- ought to be 0 -- some test failed (see above).\n",
total_error );
}
else {
printf( "all tests passed\n" );
}
TESTING_FINALIZE();
int status = (total_error != 0.);
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing clarfb_gpu
*/
int main( int argc, char** argv )
{
TESTING_INIT();
magmaFloatComplex c_zero = MAGMA_C_ZERO;
magmaFloatComplex c_one = MAGMA_C_ONE;
magmaFloatComplex c_neg_one = MAGMA_C_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};
float 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 );
float tol = opts.tolerance * lapackf77_slamch("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 clarfb 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
magmaFloatComplex *C, *R, *V, *T, *W;
TESTING_MALLOC_CPU( C, magmaFloatComplex, ldc*N );
TESTING_MALLOC_CPU( R, magmaFloatComplex, ldc*N );
TESTING_MALLOC_CPU( V, magmaFloatComplex, ldv*K );
TESTING_MALLOC_CPU( T, magmaFloatComplex, ldt*K );
TESTING_MALLOC_CPU( W, magmaFloatComplex, ldw*K );
magmaFloatComplex_ptr dC, dV, dT, dW;
TESTING_MALLOC_DEV( dC, magmaFloatComplex, ldc*N );
TESTING_MALLOC_DEV( dV, magmaFloatComplex, ldv*K );
TESTING_MALLOC_DEV( dT, magmaFloatComplex, ldt*K );
TESTING_MALLOC_DEV( dW, magmaFloatComplex, ldw*K );
// C is M x N.
size = ldc*N;
lapackf77_clarnv( &ione, ISEED, &size, C );
//printf( "C=" ); magma_cprint( 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_clarnv( &ione, ISEED, &size, V );
if ( storev[istor] == MagmaColumnwise ) {
if ( direct[idir] == MagmaForward ) {
lapackf77_claset( MagmaUpperStr, &K, &K, &c_zero, &c_one, V, &ldv );
}
else {
lapackf77_claset( 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_claset( MagmaLowerStr, &K, &K, &c_zero, &c_one, V, &ldv );
}
else {
lapackf77_claset( MagmaUpperStr, &K, &K, &c_zero, &c_one, &V[(nv-K)*ldv], &ldv );
}
}
//printf( "# ldv %d, nv %d\n", ldv, nv );
//printf( "V=" ); magma_cprint( 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_clarnv( &ione, ISEED, &size, T );
if ( direct[idir] == MagmaForward ) {
lapackf77_claset( MagmaLowerStr, &k1, &k1, &c_zero, &c_zero, &T[1], &ldt );
}
else {
lapackf77_claset( MagmaUpperStr, &k1, &k1, &c_zero, &c_zero, &T[1*ldt], &ldt );
}
//printf( "T=" ); magma_cprint( K, K, T, ldt );
magma_csetmatrix( M, N, C, ldc, dC, ldc );
magma_csetmatrix( ldv, nv, V, ldv, dV, ldv );
magma_csetmatrix( K, K, T, ldt, dT, ldt );
lapackf77_clarfb( 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_cprint( M, N, C, ldc );
magma_clarfb_gpu( side[iside], trans[itran], direct[idir], storev[istor],
M, N, K,
dV, ldv, dT, ldt, dC, ldc, dW, ldw );
magma_cgetmatrix( M, N, dC, ldc, R, ldc );
//printf( "dHC=" ); magma_cprint( M, N, R, ldc );
// compute relative error |HC_magma - HC_lapack| / |HC_lapack|
error = lapackf77_clange( "Fro", &M, &N, C, &ldc, work );
size = ldc*N;
blasf77_caxpy( &size, &c_neg_one, C, &ione, R, &ione );
error = lapackf77_clange( "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;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing cunmqr_gpu
*/
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;
magma_int_t ione = 1;
magma_int_t m, n, k, size, info;
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t nb, ldc, lda, lwork, lwork_max, dt_size;
magmaFloatComplex *C, *R, *A, *W, *tau;
magmaFloatComplex_ptr dC, dA, dT;
magma_int_t status = 0;
magma_opts opts;
parse_opts( argc, argv, &opts );
float tol = 2. * opts.tolerance * lapackf77_slamch("E");
// test all combinations of input parameters
magma_side_t side [] = { MagmaLeft, MagmaRight };
magma_trans_t trans[] = { Magma_ConjTrans, MagmaNoTrans };
printf(" M N K side trans CPU GFlop/s (sec) GPU GFlop/s (sec) ||R||_F / ||QC||_F\n");
printf("===============================================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iside = 0; iside < 2; ++iside ) {
for( int itran = 0; itran < 2; ++itran ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
m = opts.msize[itest];
n = opts.nsize[itest];
k = opts.ksize[itest];
nb = magma_get_cgeqrf_nb( m );
ldc = ((m + 31)/32)*32;
lda = ((max(m,n) + 31)/32)*32;
gflops = FLOPS_CUNMQR( m, n, k, side[iside] ) / 1e9;
if ( side[iside] == MagmaLeft && m < k ) {
printf( "%5d %5d %5d %4c %5c skipping because side=left and m < k\n",
(int) m, (int) n, (int) k,
lapacke_side_const( side[iside] ),
lapacke_trans_const( trans[itran] ) );
continue;
}
if ( side[iside] == MagmaRight && n < k ) {
printf( "%5d %5d %5d %4c %5c skipping because side=right and n < k\n",
(int) m, (int) n, (int) k,
lapacke_side_const( side[iside] ),
lapacke_trans_const( trans[itran] ) );
continue;
}
if ( side[iside] == MagmaLeft ) {
// side = left
lwork_max = (m - k + nb)*(n + nb) + n*nb;
dt_size = ( 2*min(m,k) + ((max(m,n) + 31)/32)*32 )*nb;
}
else {
// side = right
lwork_max = (n - k + nb)*(m + nb) + m*nb;
dt_size = ( 2*min(n,k) + ((max(m,n) + 31)/32)*32 )*nb;
}
TESTING_MALLOC_CPU( C, magmaFloatComplex, ldc*n );
TESTING_MALLOC_CPU( R, magmaFloatComplex, ldc*n );
TESTING_MALLOC_CPU( A, magmaFloatComplex, lda*k );
TESTING_MALLOC_CPU( W, magmaFloatComplex, lwork_max );
TESTING_MALLOC_CPU( tau, magmaFloatComplex, k );
TESTING_MALLOC_DEV( dC, magmaFloatComplex, ldc*n );
TESTING_MALLOC_DEV( dA, magmaFloatComplex, lda*k );
TESTING_MALLOC_DEV( dT, magmaFloatComplex, dt_size );
// C is full, m x n
size = ldc*n;
lapackf77_clarnv( &ione, ISEED, &size, C );
magma_csetmatrix( m, n, C, ldc, dC, ldc );
// A is m x k (left) or n x k (right)
lda = (side[iside] == MagmaLeft ? m : n);
size = lda*k;
lapackf77_clarnv( &ione, ISEED, &size, A );
// compute QR factorization to get Householder vectors in dA, tau, dT
magma_csetmatrix( lda, k, A, lda, dA, lda );
magma_cgeqrf_gpu( lda, k, dA, lda, tau, dT, &info );
magma_cgetmatrix( lda, k, dA, lda, A, lda );
if (info != 0)
printf("magma_cgeqrf_gpu returned error %d: %s.\n",
(int) info, magma_strerror( info ));
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
cpu_time = magma_wtime();
lapackf77_cunmqr( lapack_side_const( side[iside] ), lapack_trans_const( trans[itran] ),
&m, &n, &k,
A, &lda, tau, C, &ldc, W, &lwork_max, &info );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
if (info != 0)
printf("lapackf77_cunmqr returned error %d: %s.\n",
(int) info, magma_strerror( info ));
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
// query for workspace size
lwork = -1;
magma_cunmqr_gpu( side[iside], trans[itran],
m, n, k,
dA, lda, tau, dC, ldc, W, lwork, dT, nb, &info );
if (info != 0)
printf("magma_cunmqr_gpu (lwork query) returned error %d: %s.\n",
(int) info, magma_strerror( info ));
lwork = (magma_int_t) MAGMA_C_REAL( W[0] );
if ( lwork < 0 || lwork > lwork_max )
printf("invalid lwork %d, lwork_max %d\n", (int) lwork, (int) lwork_max );
gpu_time = magma_sync_wtime( 0 ); // sync needed for L,N and R,T cases
magma_cunmqr_gpu( side[iside], trans[itran],
m, n, k,
dA, lda, tau, dC, ldc, W, lwork, dT, nb, &info );
gpu_time = magma_sync_wtime( 0 ) - gpu_time;
gpu_perf = gflops / gpu_time;
if (info != 0)
printf("magma_cunmqr_gpu returned error %d: %s.\n",
(int) info, magma_strerror( info ));
magma_cgetmatrix( m, n, dC, ldc, R, ldc );
/* =====================================================================
compute relative error |QC_magma - QC_lapack| / |QC_lapack|
=================================================================== */
error = lapackf77_clange( "Fro", &m, &n, C, &ldc, work );
size = ldc*n;
blasf77_caxpy( &size, &c_neg_one, C, &ione, R, &ione );
error = lapackf77_clange( "Fro", &m, &n, R, &ldc, work ) / error;
printf( "%5d %5d %5d %4c %5c %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %s\n",
(int) m, (int) n, (int) k,
lapacke_side_const( side[iside] ),
lapacke_trans_const( trans[itran] ),
cpu_perf, cpu_time, gpu_perf, gpu_time,
error, (error < tol ? "ok" : "failed") );
status += ! (error < tol);
TESTING_FREE_CPU( C );
TESTING_FREE_CPU( R );
TESTING_FREE_CPU( A );
TESTING_FREE_CPU( W );
TESTING_FREE_CPU( tau );
TESTING_FREE_DEV( dC );
TESTING_FREE_DEV( dA );
TESTING_FREE_DEV( dT );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}} // end iside, itran
printf( "\n" );
}
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing dormqr_gpu
*/
int main( int argc, char** argv )
{
TESTING_INIT();
real_Double_t gflops, gpu_perf, gpu_time, cpu_perf, cpu_time;
double Cnorm, error, work[1];
double c_neg_one = MAGMA_D_NEG_ONE;
magma_int_t ione = 1;
magma_int_t mm, m, n, k, size, info;
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t nb, ldc, lda, lwork, lwork_max, dt_size;
double *C, *R, *A, *hwork, *tau;
magmaDouble_ptr dC, dA, dT;
magma_int_t status = 0;
magma_opts opts;
opts.parse_opts( argc, argv );
// need slightly looser bound (60*eps instead of 30*eps) for some tests
opts.tolerance = max( 60., opts.tolerance );
double tol = opts.tolerance * lapackf77_dlamch("E");
// test all combinations of input parameters
magma_side_t side [] = { MagmaLeft, MagmaRight };
magma_trans_t trans[] = { MagmaTrans, MagmaNoTrans };
printf("%% M N K side trans CPU Gflop/s (sec) GPU Gflop/s (sec) ||R||_F / ||QC||_F\n");
printf("%%==============================================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iside = 0; iside < 2; ++iside ) {
for( int itran = 0; itran < 2; ++itran ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
m = opts.msize[itest];
n = opts.nsize[itest];
k = opts.ksize[itest];
ldc = magma_roundup( m, opts.align ); // multiple of 32 by default
// A is m x k (left) or n x k (right)
mm = (side[iside] == MagmaLeft ? m : n);
nb = magma_get_dgeqrf_nb( mm, k );
lda = magma_roundup( mm, opts.align ); // multiple of 32 by default
gflops = FLOPS_DORMQR( m, n, k, side[iside] ) / 1e9;
if ( side[iside] == MagmaLeft && m < k ) {
printf( "%5d %5d %5d %4c %5c skipping because side=left and m < k\n",
(int) m, (int) n, (int) k,
lapacke_side_const( side[iside] ),
lapacke_trans_const( trans[itran] ) );
continue;
}
if ( side[iside] == MagmaRight && n < k ) {
printf( "%5d %5d %5d %4c %5c skipping because side=right and n < k\n",
(int) m, (int) n, (int) k,
lapacke_side_const( side[iside] ),
lapacke_trans_const( trans[itran] ) );
continue;
}
if ( side[iside] == MagmaLeft ) {
// side = left
lwork_max = (m - k + nb)*(n + nb) + n*nb;
dt_size = ( 2*min(m,k) + magma_roundup( max(m,n), 32) )*nb;
}
else {
// side = right
lwork_max = (n - k + nb)*(m + nb) + m*nb;
dt_size = ( 2*min(n,k) + magma_roundup( max(m,n), 32 ) )*nb;
}
// this rounds it up slightly if needed to agree with lwork query below
lwork_max = int( real( magma_dmake_lwork( lwork_max )));
TESTING_MALLOC_CPU( C, double, ldc*n );
TESTING_MALLOC_CPU( R, double, ldc*n );
TESTING_MALLOC_CPU( A, double, lda*k );
TESTING_MALLOC_CPU( hwork, double, lwork_max );
TESTING_MALLOC_CPU( tau, double, k );
TESTING_MALLOC_DEV( dC, double, ldc*n );
TESTING_MALLOC_DEV( dA, double, lda*k );
TESTING_MALLOC_DEV( dT, double, dt_size );
// C is full, m x n
size = ldc*n;
lapackf77_dlarnv( &ione, ISEED, &size, C );
magma_dsetmatrix( m, n, C, ldc, dC, ldc );
// A is m x k (left) or n x k (right)
size = lda*k;
lapackf77_dlarnv( &ione, ISEED, &size, A );
// compute QR factorization to get Householder vectors in dA, tau, dT
magma_dsetmatrix( mm, k, A, lda, dA, lda );
magma_dgeqrf_gpu( mm, k, dA, lda, tau, dT, &info );
magma_dgetmatrix( mm, k, dA, lda, A, lda );
if (info != 0) {
printf("magma_dgeqrf_gpu returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
cpu_time = magma_wtime();
lapackf77_dormqr( lapack_side_const( side[iside] ), lapack_trans_const( trans[itran] ),
&m, &n, &k,
A, &lda, tau, C, &ldc, hwork, &lwork_max, &info );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
if (info != 0) {
printf("lapackf77_dormqr returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
// query for workspace size
lwork = -1;
magma_dormqr_gpu( side[iside], trans[itran],
m, n, k,
dA, lda, tau, dC, ldc, hwork, lwork, dT, nb, &info );
if (info != 0) {
printf("magma_dormqr_gpu (lwork query) returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
lwork = (magma_int_t) MAGMA_D_REAL( hwork[0] );
if ( lwork < 0 || lwork > lwork_max ) {
printf("Warning: optimal lwork %d > allocated lwork_max %d\n", (int) lwork, (int) lwork_max );
lwork = lwork_max;
}
// dormqr2 takes a copy of dA in CPU memory
if ( opts.version == 2 ) {
magma_dgetmatrix( mm, k, dA, lda, A, lda );
}
magmablasSetKernelStream( opts.queue );
gpu_time = magma_sync_wtime( opts.queue ); // sync needed for L,N and R,T cases
if ( opts.version == 1 ) {
magma_dormqr_gpu( side[iside], trans[itran],
m, n, k,
dA, lda, tau, dC, ldc, hwork, lwork, dT, nb, &info );
}
else if ( opts.version == 2 ) {
magma_dormqr2_gpu( side[iside], trans[itran],
m, n, k,
dA, lda, tau, dC, ldc, A, lda, &info );
}
gpu_time = magma_sync_wtime( opts.queue ) - gpu_time;
gpu_perf = gflops / gpu_time;
if (info != 0) {
printf("magma_dormqr_gpu returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
magma_dgetmatrix( m, n, dC, ldc, R, ldc );
/* =====================================================================
compute relative error |QC_magma - QC_lapack| / |QC_lapack|
=================================================================== */
size = ldc*n;
blasf77_daxpy( &size, &c_neg_one, C, &ione, R, &ione );
Cnorm = lapackf77_dlange( "Fro", &m, &n, C, &ldc, work );
error = lapackf77_dlange( "Fro", &m, &n, R, &ldc, work ) / (magma_dsqrt(m*n) * Cnorm);
printf( "%5d %5d %5d %4c %5c %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %s\n",
(int) m, (int) n, (int) k,
lapacke_side_const( side[iside] ),
lapacke_trans_const( trans[itran] ),
cpu_perf, cpu_time, gpu_perf, gpu_time,
error, (error < tol ? "ok" : "failed") );
status += ! (error < tol);
TESTING_FREE_CPU( C );
TESTING_FREE_CPU( R );
TESTING_FREE_CPU( A );
TESTING_FREE_CPU( hwork );
TESTING_FREE_CPU( tau );
TESTING_FREE_DEV( dC );
TESTING_FREE_DEV( dA );
TESTING_FREE_DEV( dT );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}} // end iside, itran
printf( "\n" );
}
opts.cleanup();
TESTING_FINALIZE();
return status;
}
extern "C" magma_int_t
magma_sgetrf_msub(
magma_trans_t trans, magma_int_t num_subs, magma_int_t num_gpus,
magma_int_t m, magma_int_t n,
magmaFloat_ptr *d_lA, size_t dlA_offset, magma_int_t ldda,
magma_int_t *ipiv,
magma_queue_t *queues,
magma_int_t *info)
{
/* -- clMAGMA (version 1.3.0) --
Univ. of Tennessee, Knoxville
Univ. of California, Berkeley
Univ. of Colorado, Denver
@date November 2014
Purpose
=======
SGETRF computes an LU factorization of a general M-by-N matrix A
using partial pivoting with row interchanges.
The factorization has the form
A = P * L * U
where P is a permutation matrix, L is lower triangular with unit
diagonal elements (lower trapezoidal if m > n), and U is upper
triangular (upper trapezoidal if m < n).
This is the right-looking Level 3 BLAS version of the algorithm.
Arguments
=========
NUM_GPUS
(input) INTEGER
The number of GPUS to be used for the factorization.
M (input) INTEGER
The number of rows of the matrix A. M >= 0.
N (input) INTEGER
The number of columns of the matrix A. N >= 0.
A (input/output) REAL array on the GPU, dimension (LDDA,N).
On entry, the M-by-N matrix to be factored.
On exit, the factors L and U from the factorization
A = P*L*U; the unit diagonal elements of L are not stored.
LDDA (input) INTEGER
The leading dimension of the array A. LDDA >= max(1,M).
IPIV (output) INTEGER array, dimension (min(M,N))
The pivot indices; for 1 <= i <= min(M,N), row i of the
matrix was interchanged with row IPIV(i).
INFO (output) INTEGER
= 0: successful exit
< 0: if INFO = -i, the i-th argument had an illegal value
or another error occured, such as memory allocation failed.
> 0: if INFO = i, U(i,i) is exactly zero. The factorization
has been completed, but the factor U is exactly
singular, and division by zero will occur if it is used
to solve a system of equations.
===================================================================== */
#define d_lAT(id,i,j) d_lAT[(id)], (((i)*nb)*lddat + (j)*nb)
#define d_lA( id,i,j) d_lA[(id)], (((i)*nb)+ldda * (j)*nb)
magma_int_t maxm, tot_subs = num_subs*num_gpus;
magma_int_t i, j, d, lddat;
/* submatrix info */
magma_int_t nb, n_local[ MagmaMaxSubs * MagmaMaxGPUs ];
magmaFloat_ptr d_lAT[ MagmaMaxSubs * MagmaMaxGPUs ];
/* local workspace per GPU */
magmaFloat_ptr d_panel[ MagmaMaxGPUs ];
magmaFloat_ptr d_lAP[ MagmaMaxGPUs ];
float *work;
/* Check arguments */
*info = 0;
if (m < 0)
*info = -2;
else if (n < 0)
*info = -3;
else if (trans == MagmaTrans && ldda < max(1,m))
*info = -5;
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
/* Quick return if possible */
if (m == 0 || n == 0)
return *info;
/* Function Body */
nb = magma_get_sgetrf_nb(m);
if (nb <= 1 || nb >= n) {
/* Use CPU code. */
magma_smalloc_cpu( &work, m * n );
if (work == NULL) {
*info = MAGMA_ERR_HOST_ALLOC;
return *info;
}
printf( "trans %c, m %d, n %d\n", lapacke_trans_const(trans), m, n );
magma_sgetmatrix( m, n, d_lA[0], 0, ldda, work, m, queues[0] );
lapackf77_sgetrf( &m, &n, work, &m, ipiv, info );
magma_ssetmatrix( m, n, work, m, d_lA[0], 0, ldda, queues[0] );
magma_free_cpu( work );
} else {
/* Use hybrid blocked code. */
maxm = ((m + 31)/32)*32;
if (tot_subs > ceil((float)n/nb)) {
printf( " * too many GPUs for the matrix size, using %d GPUs\n", (int) tot_subs );
*info = -1;
return *info;
}
/* allocate workspace for each GPU */
lddat = n/nb; /* number of block columns */
lddat = lddat/tot_subs; /* number of block columns per GPU */
lddat = nb*lddat; /* number of columns per GPU */
if (lddat * tot_subs < n) {
/* left over */
if (n-lddat*tot_subs >= nb) {
lddat += nb;
} else {
lddat += (n-lddat*tot_subs)%nb;
}
}
lddat = ((lddat+31)/32)*32; /* make it a multiple of 32 */
/* allocating workspace */
for (d=0; d < num_gpus; d++) {
//#define SINGLE_GPU_PER_CONTEXT
#ifdef SINGLE_GPU_PER_CONTEXT
if ((MAGMA_SUCCESS != magma_smalloc_mgpu( d, &d_panel[d], (2+num_gpus)*nb*maxm )) ||
(MAGMA_SUCCESS != magma_smalloc_mgpu( d, &d_lAP[d], (2+num_gpus)*nb*maxm )) ) {
#else
if ((MAGMA_SUCCESS != magma_smalloc( &d_panel[d], (2+num_gpus)*nb*maxm )) ||
(MAGMA_SUCCESS != magma_smalloc( &d_lAP[d], (2+num_gpus)*nb*maxm )) ) {
#endif
for( i=0; i < d; i++ ) {
magma_free( d_panel[i] );
magma_free( d_lAP[i] );
}
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
}
/* transposing the local matrix */
for (i=0; i < tot_subs; i++) {
/* local-n and local-ld */
n_local[i] = ((n/nb)/tot_subs)*nb;
if (i < (n/nb)%tot_subs)
n_local[i] += nb;
else if (i == (n/nb)%tot_subs)
n_local[i] += n%nb;
/* local-matrix storage */
if (trans == MagmaNoTrans) {
d_lAT[i] = d_lA[i];
} else {
if ( m == n_local[i] ) {
d_lAT[i] = d_lA[i];
magmablas_stranspose_inplace( m, d_lA[i], 0, ldda, queues[2*(i%num_gpus)+1] );
} else {
#ifdef SINGLE_GPU_PER_CONTEXT
if (MAGMA_SUCCESS != magma_smalloc_mgpu( i%num_gpus, &d_lAT[i], lddat*maxm )) {
#else
if (MAGMA_SUCCESS != magma_smalloc( &d_lAT[i], lddat*maxm )) {
#endif
for (j=0; j <= i; j++) {
magma_free( d_panel[j] );
magma_free( d_lAP[j] );
}
for (j=0; j < i; j++) {
if (d_lAT[j] != d_lA[j]) magma_free( d_lAT[j] );
}
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
magmablas_stranspose( m, n_local[i], d_lA[i], 0, ldda, d_lAT[i], 0, lddat, queues[2*(i%num_gpus)+1]);
}
}
}
if (trans == MagmaNoTrans) {
for (d=0; d < num_gpus; d++){
magma_queue_sync(queues[2*d+1]);
}
}
/* cpu workspace */
#ifdef USE_PINNED_CLMEMORY
cl_mem buffer = clCreateBuffer(gContext, CL_MEM_READ_WRITE | CL_MEM_ALLOC_HOST_PTR, sizeof(float)*maxm*nb*(1+num_gpus), NULL, NULL);
for (d=0; d < num_gpus; d++) {
work = (float*)clEnqueueMapBuffer(queues[2*d], buffer, CL_TRUE, CL_MAP_READ | CL_MAP_WRITE, 0,
sizeof(float)*maxm*nb*(1+num_gpus), 0, NULL, NULL, NULL);
}
#else
if (MAGMA_SUCCESS != magma_smalloc_cpu( &work, maxm*nb*(1+num_gpus) )) {
for(d=0; d < num_gpus; d++ ) magma_free( d_panel[d] );
for(d=0; d < tot_subs; d++ ) {
if( d_lAT[d] != d_lA[d] ) magma_free( d_lAT[d] );
}
*info = MAGMA_ERR_HOST_ALLOC;
return *info;
}
#endif
/* calling multi-gpu interface with allocated workspaces and streams */
magma_sgetrf2_msub(num_subs, num_gpus, m, n, nb, 0, d_lAT, 0, lddat, ipiv, d_lAP, d_panel, 0, work, maxm,
queues, info);
/* save on output */
for (d=0; d < tot_subs; d++) {
if (trans == MagmaNoTrans) {
//magma_scopymatrix( n_local[d], m, d_lAT[d], 0, lddat, d_lA[d], 0, ldda, queues[2*d+1] );
} else {
if (d_lAT[d] == d_lA[d]) {
magmablas_stranspose_inplace( m, d_lA[d], 0, ldda, queues[2*(d%num_gpus)+1] );
} else {
magmablas_stranspose( n_local[d], m, d_lAT[d], 0, lddat, d_lA[d], 0, ldda, queues[2*(d%num_gpus)+1] );
}
}
}
/* clean up */
for (d=0; d < num_gpus; d++) {
magma_queue_sync(queues[2*d+1]);
magma_free( d_panel[d] );
magma_free( d_lAP[d] );
d_panel[d] = d_lAP[d] = NULL;
}
for (d=0; d < tot_subs; d++) {
if (d_lAT[d] != d_lA[d]) {
magma_free( d_lAT[d] );
d_lAT[d] = NULL;
}
}
#ifdef USE_PINNED_CLMEMORY
for (d=0; d < num_gpus; d++) {
clEnqueueUnmapMemObject(queues[2*d], buffer, work, 0, NULL, NULL);
}
clReleaseMemObject( buffer );
#else
magma_free_cpu( work );
#endif
work = NULL;
}
return *info;
/* End of MAGMA_SGETRF_MSUB */
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing dormql
*/
int main( int argc, char** argv )
{
TESTING_INIT();
real_Double_t gflops, gpu_perf, gpu_time, cpu_perf, cpu_time;
double error, work[1];
double c_neg_one = MAGMA_D_NEG_ONE;
magma_int_t ione = 1;
magma_int_t mm, m, n, k, size, info;
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t nb, ldc, lda, lwork, lwork_max;
double *C, *R, *A, *W, *tau;
magma_int_t status = 0;
magma_opts opts;
parse_opts( argc, argv, &opts );
// need slightly looser bound (60*eps instead of 30*eps) for some tests
opts.tolerance = max( 60., opts.tolerance );
double tol = opts.tolerance * lapackf77_dlamch("E");
// test all combinations of input parameters
magma_side_t side [] = { MagmaLeft, MagmaRight };
magma_trans_t trans[] = { MagmaTrans, MagmaNoTrans };
printf(" M N K side trans CPU GFlop/s (sec) GPU GFlop/s (sec) ||R||_F / ||QC||_F\n");
printf("===============================================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iside = 0; iside < 2; ++iside ) {
for( int itran = 0; itran < 2; ++itran ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
m = opts.msize[itest];
n = opts.nsize[itest];
k = opts.ksize[itest];
nb = magma_get_dgeqlf_nb( m );
ldc = m;
// A is m x k (left) or n x k (right)
mm = (side[iside] == MagmaLeft ? m : n);
lda = mm;
gflops = FLOPS_DORMQL( m, n, k, side[iside] ) / 1e9;
if ( side[iside] == MagmaLeft && m < k ) {
printf( "%5d %5d %5d %4c %5c skipping because side=left and m < k\n",
(int) m, (int) n, (int) k,
lapacke_side_const( side[iside] ),
lapacke_trans_const( trans[itran] ) );
continue;
}
if ( side[iside] == MagmaRight && n < k ) {
printf( "%5d %5d %5d %4c %5c skipping because side=right and n < k\n",
(int) m, (int) n, (int) k,
lapacke_side_const( side[iside] ),
lapacke_trans_const( trans[itran] ) );
continue;
}
// need at least 2*nb*nb for geqlf
lwork_max = max( max( m*nb, n*nb ), 2*nb*nb );
TESTING_MALLOC_CPU( C, double, ldc*n );
TESTING_MALLOC_CPU( R, double, ldc*n );
TESTING_MALLOC_CPU( A, double, lda*k );
TESTING_MALLOC_CPU( W, double, lwork_max );
TESTING_MALLOC_CPU( tau, double, k );
// C is full, m x n
size = ldc*n;
lapackf77_dlarnv( &ione, ISEED, &size, C );
lapackf77_dlacpy( "Full", &m, &n, C, &ldc, R, &ldc );
size = lda*k;
lapackf77_dlarnv( &ione, ISEED, &size, A );
// compute QL factorization to get Householder vectors in A, tau
magma_dgeqlf( mm, k, A, lda, tau, W, lwork_max, &info );
if (info != 0)
printf("magma_dgeqlf returned error %d: %s.\n",
(int) info, magma_strerror( info ));
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
cpu_time = magma_wtime();
lapackf77_dormql( lapack_side_const( side[iside] ), lapack_trans_const( trans[itran] ),
&m, &n, &k,
A, &lda, tau, C, &ldc, W, &lwork_max, &info );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
if (info != 0)
printf("lapackf77_dormql returned error %d: %s.\n",
(int) info, magma_strerror( info ));
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
// query for workspace size
lwork = -1;
magma_dormql( side[iside], trans[itran],
m, n, k,
A, lda, tau, R, ldc, W, lwork, &info );
if (info != 0)
printf("magma_dormql (lwork query) returned error %d: %s.\n",
(int) info, magma_strerror( info ));
lwork = (magma_int_t) MAGMA_D_REAL( W[0] );
if ( lwork < 0 || lwork > lwork_max ) {
printf("optimal lwork %d > lwork_max %d\n", (int) lwork, (int) lwork_max );
lwork = lwork_max;
}
gpu_time = magma_wtime();
magma_dormql( side[iside], trans[itran],
m, n, k,
A, lda, tau, R, ldc, W, lwork, &info );
gpu_time = magma_wtime() - gpu_time;
gpu_perf = gflops / gpu_time;
if (info != 0)
printf("magma_dormql returned error %d: %s.\n",
(int) info, magma_strerror( info ));
/* =====================================================================
compute relative error |QC_magma - QC_lapack| / |QC_lapack|
=================================================================== */
error = lapackf77_dlange( "Fro", &m, &n, C, &ldc, work );
size = ldc*n;
blasf77_daxpy( &size, &c_neg_one, C, &ione, R, &ione );
error = lapackf77_dlange( "Fro", &m, &n, R, &ldc, work ) / error;
printf( "%5d %5d %5d %4c %5c %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %s\n",
(int) m, (int) n, (int) k,
lapacke_side_const( side[iside] ),
lapacke_trans_const( trans[itran] ),
cpu_perf, cpu_time, gpu_perf, gpu_time,
error, (error < tol ? "ok" : "failed") );
status += ! (error < tol);
TESTING_FREE_CPU( C );
TESTING_FREE_CPU( R );
TESTING_FREE_CPU( A );
TESTING_FREE_CPU( W );
TESTING_FREE_CPU( tau );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}} // end iside, itran
printf( "\n" );
}
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing ctranspose
Code is very similar to testing_csymmetrize.cpp
*/
int main( int argc, char** argv)
{
TESTING_INIT();
// OpenCL use: cl_mem , offset (two arguments);
// else use: pointer + offset (one argument).
#ifdef HAVE_clBLAS
#define d_A(i_, j_) d_A, ((i_) + (j_)*ldda)
#define d_B(i_, j_) d_B, ((i_) + (j_)*lddb)
#else
#define d_A(i_, j_) (d_A + (i_) + (j_)*ldda)
#define d_B(i_, j_) (d_B + (i_) + (j_)*lddb)
#endif
real_Double_t gbytes, gpu_perf, gpu_time, gpu_perf2=0, gpu_time2=0, cpu_perf, cpu_time;
float error, error2, work[1];
magmaFloatComplex c_neg_one = MAGMA_C_NEG_ONE;
magmaFloatComplex *h_A, *h_B, *h_R;
magmaFloatComplex_ptr d_A, d_B;
magma_int_t M, N, size, lda, ldda, ldb, lddb;
magma_int_t ione = 1;
magma_int_t status = 0;
magma_opts opts;
opts.parse_opts( argc, argv );
#ifdef COMPLEX
magma_int_t ntrans = 2;
magma_trans_t trans[] = { Magma_ConjTrans, MagmaTrans };
#else
magma_int_t ntrans = 1;
magma_trans_t trans[] = { MagmaTrans };
#endif
printf("%% Inplace transpose requires M == N.\n");
printf("%% Trans M N CPU GByte/s (ms) GPU GByte/s (ms) check Inplace GB/s (ms) check\n");
printf("%%=========================================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int itran = 0; itran < ntrans; ++itran ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
M = opts.msize[itest];
N = opts.nsize[itest];
lda = M;
ldda = magma_roundup( M, opts.align ); // multiple of 32 by default
ldb = N;
lddb = magma_roundup( N, opts.align ); // multiple of 32 by default
// load entire matrix, save entire matrix
gbytes = sizeof(magmaFloatComplex) * 2.*M*N / 1e9;
TESTING_MALLOC_CPU( h_A, magmaFloatComplex, lda*N ); // input: M x N
TESTING_MALLOC_CPU( h_B, magmaFloatComplex, ldb*M ); // output: N x M
TESTING_MALLOC_CPU( h_R, magmaFloatComplex, ldb*M ); // output: N x M
TESTING_MALLOC_DEV( d_A, magmaFloatComplex, ldda*N ); // input: M x N
TESTING_MALLOC_DEV( d_B, magmaFloatComplex, lddb*M ); // output: N x M
/* Initialize the matrix */
for( int j = 0; j < N; ++j ) {
for( int i = 0; i < M; ++i ) {
h_A[i + j*lda] = MAGMA_C_MAKE( i + j/10000., j );
}
}
for( int j = 0; j < M; ++j ) {
for( int i = 0; i < N; ++i ) {
h_B[i + j*ldb] = MAGMA_C_MAKE( i + j/10000., j );
}
}
magma_csetmatrix( N, M, h_B, ldb, d_B(0,0), lddb, opts.queue );
/* =====================================================================
Performs operation using naive out-of-place algorithm
(LAPACK doesn't implement transpose)
=================================================================== */
cpu_time = magma_wtime();
//for( int j = 1; j < N-1; ++j ) { // inset by 1 row & col
// for( int i = 1; i < M-1; ++i ) { // inset by 1 row & col
if ( trans[itran] == MagmaTrans ) {
for( int j = 0; j < N; ++j ) {
for( int i = 0; i < M; ++i ) {
h_B[j + i*ldb] = h_A[i + j*lda];
}
}
}
else {
for( int j = 0; j < N; ++j ) {
for( int i = 0; i < M; ++i ) {
h_B[j + i*ldb] = conj( h_A[i + j*lda] );
}
}
}
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gbytes / cpu_time;
/* ====================================================================
Performs operation using MAGMA, out-of-place
=================================================================== */
magma_csetmatrix( M, N, h_A, lda, d_A(0,0), ldda, opts.queue );
magma_csetmatrix( N, M, h_B, ldb, d_B(0,0), lddb, opts.queue );
gpu_time = magma_sync_wtime( opts.queue );
if ( trans[itran] == MagmaTrans ) {
//magmablas_ctranspose( M-2, N-2, d_A(1,1), ldda, d_B(1,1), lddb, opts.queue ); // inset by 1 row & col
magmablas_ctranspose( M, N, d_A(0,0), ldda, d_B(0,0), lddb, opts.queue );
}
#ifdef HAVE_CUBLAS
else {
//magmablas_ctranspose_conj( M-2, N-2, d_A(1,1), ldda, d_B(1,1), lddb, opts.queue ); // inset by 1 row & col
magmablas_ctranspose_conj( M, N, d_A(0,0), ldda, d_B(0,0), lddb, opts.queue );
}
#endif
gpu_time = magma_sync_wtime( opts.queue ) - gpu_time;
gpu_perf = gbytes / gpu_time;
/* ====================================================================
Performs operation using MAGMA, in-place
=================================================================== */
if ( M == N ) {
magma_csetmatrix( M, N, h_A, lda, d_A(0,0), ldda, opts.queue );
gpu_time2 = magma_sync_wtime( opts.queue );
if ( trans[itran] == MagmaTrans ) {
//magmablas_ctranspose_inplace( N-2, d_A(1,1), ldda, opts.queue ); // inset by 1 row & col
magmablas_ctranspose_inplace( N, d_A(0,0), ldda, opts.queue );
}
#ifdef HAVE_CUBLAS
else {
//magmablas_ctranspose_conj_inplace( N-2, d_A(1,1), ldda, opts.queue ); // inset by 1 row & col
magmablas_ctranspose_conj_inplace( N, d_A(0,0), ldda, opts.queue );
}
#endif
gpu_time2 = magma_sync_wtime( opts.queue ) - gpu_time2;
gpu_perf2 = gbytes / gpu_time2;
}
/* =====================================================================
Check the result
=================================================================== */
// check out-of-place transpose (d_B)
size = ldb*M;
magma_cgetmatrix( N, M, d_B(0,0), lddb, h_R, ldb, opts.queue );
blasf77_caxpy( &size, &c_neg_one, h_B, &ione, h_R, &ione );
error = lapackf77_clange("f", &N, &M, h_R, &ldb, work );
if ( M == N ) {
// also check in-place tranpose (d_A)
magma_cgetmatrix( N, M, d_A(0,0), ldda, h_R, ldb, opts.queue );
blasf77_caxpy( &size, &c_neg_one, h_B, &ione, h_R, &ione );
error2 = lapackf77_clange("f", &N, &M, h_R, &ldb, work );
printf("%5c %5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %6s %7.2f (%7.2f) %s\n",
lapacke_trans_const( trans[itran] ),
(int) M, (int) N,
cpu_perf, cpu_time*1000., gpu_perf, gpu_time*1000.,
(error == 0. ? "ok" : "failed"),
gpu_perf2, gpu_time2,
(error2 == 0. ? "ok" : "failed") );
status += ! (error == 0. && error2 == 0.);
}
else {
printf("%5c %5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %6s --- ( --- )\n",
lapacke_trans_const( trans[itran] ),
(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_B );
TESTING_FREE_CPU( h_R );
TESTING_FREE_DEV( d_A );
TESTING_FREE_DEV( d_B );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
}
opts.cleanup();
TESTING_FINALIZE();
return status;
}
