void SVDMatrix_magma(Tensor_core<complex<double>,2>& U, Tensor_core<double,1>& D, Tensor_core<complex<double>,2>& V)
{
if( U.rank(0)!=U.rank(1) || U.rank(1)!=D.rank(0) || D.rank(0)!=V.rank(0) || V.rank(0)!=V.rank(1) )
{
cout<<"Size is not consistent in SVDMatrix_magma! Only support square matrix."<<endl;
exit(1);
}
magma_int_t m=U.rank(0); magma_int_t n=V.rank(0);
magma_vec_t jobz(MagmaOverwriteVec); magma_int_t lda=m;
magmaDoubleComplex* u=nullptr; magma_int_t ldu=1; magma_int_t ldv=n;
magmaDoubleComplex work_test[1]; magma_int_t lwork=-1;
double* rwork; magma_int_t* iwork;
magma_dmalloc_cpu( &rwork, 5*m*m+7*m ); magma_imalloc_cpu(&iwork, 8*m);
magma_int_t info;
magma_zgesdd(jobz, m, n, (magmaDoubleComplex *) U.data(), lda, D.data(), u, ldu, (magmaDoubleComplex *) V.data(), ldv,
work_test, lwork, rwork, iwork, &info);
lwork=lround( MAGMA_Z_REAL(work_test[0]) );
magmaDoubleComplex* work; magma_zmalloc_cpu(&work, lwork);
magma_zgesdd(jobz, m, n, (magmaDoubleComplex *) U.data(), lda, D.data(), u, ldu, (magmaDoubleComplex *) V.data(), ldv,
work, lwork, rwork, iwork, &info);
magma_free_cpu(work); magma_free_cpu(rwork); magma_free_cpu(iwork);
if(info!=0)
{
cout<<"SVDMatrix_magma is not suceesful, info= "<<info<<endl;
exit(1);
}
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing zgesdd (SVD with Divide & Conquer)
Please keep code in testing_zgesdd.cpp and testing_zgesvd.cpp similar.
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gpu_time, cpu_time;
magmaDoubleComplex *h_A, *h_R, *U, *VT, *h_work;
magmaDoubleComplex dummy[1];
double *S1, *S2;
#ifdef COMPLEX
magma_int_t lrwork=0;
double *rwork;
#endif
magma_int_t *iwork;
magma_int_t M, N, N_U, M_VT, lda, ldu, ldv, n2, min_mn, max_mn, info, nb, lwork;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
magma_vec_t jobz;
magma_int_t status = 0;
MAGMA_UNUSED( max_mn ); // used only in complex
magma_opts opts;
opts.parse_opts( argc, argv );
double tol = opts.tolerance * lapackf77_dlamch("E");
jobz = opts.jobu;
magma_vec_t jobs[] = { MagmaNoVec, MagmaSomeVec, MagmaOverwriteVec, MagmaAllVec };
if ( opts.check && ! opts.all && (jobz == MagmaNoVec)) {
printf( "%% NOTE: some checks require that singular vectors are computed;\n"
"%% set jobz (option -U[NASO]) to be S, O, or A.\n\n" );
}
printf("%% jobz M N CPU time (sec) GPU time (sec) |S1-S2| |A-USV^H| |I-UU^H|/M |I-VV^H|/N S sorted\n");
printf("%%==========================================================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int ijobz = 0; ijobz < 4; ++ijobz ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
if ( opts.all ) {
jobz = jobs[ ijobz ];
}
else if ( ijobz > 0 ) {
// if not testing all, run only once, when ijobz = 0,
// but jobz come from opts (above loops).
continue;
}
M = opts.msize[itest];
N = opts.nsize[itest];
min_mn = min(M, N);
max_mn = max(M, N);
N_U = (jobz == MagmaAllVec ? M : min_mn);
M_VT = (jobz == MagmaAllVec ? N : min_mn);
lda = M;
ldu = M;
ldv = M_VT;
n2 = lda*N;
nb = magma_get_zgesvd_nb( M, N );
// x and y abbreviations used in zgesdd and dgesdd documentation
magma_int_t x = max(M,N);
magma_int_t y = min(M,N);
#ifdef COMPLEX
bool tall = (x >= int(y*17/9.)); // true if tall (m >> n) or wide (n >> m)
#else
bool tall = (x >= int(y*11/6.)); // true if tall (m >> n) or wide (n >> m)
#endif
// query or use formula for workspace size
switch( opts.svd_work ) {
case 0: { // query for workspace size
lwork = -1;
magma_zgesdd( jobz, M, N,
NULL, lda, NULL, NULL, ldu, NULL, ldv, dummy, lwork,
#ifdef COMPLEX
NULL,
#endif
NULL, &info );
lwork = (int) MAGMA_Z_REAL( dummy[0] );
break;
}
case 1: // minimum
case 2: // optimal
case 3: { // optimal (for gesdd, 2 & 3 are same; for gesvd, they differ)
// formulas from zgesdd and dgesdd documentation
bool sml = (opts.svd_work == 1); // 1 is small workspace, 2,3 are large workspace
#ifdef COMPLEX // ----------------------------------------
if (jobz == MagmaNoVec) {
if (tall) { lwork = 2*y + (2*y)*nb; }
else { lwork = 2*y + (x+y)*nb; }
}
if (jobz == MagmaOverwriteVec) {
if (tall) {
if (sml) { lwork = 2*y*y + 2*y + (2*y)*nb; }
else { lwork = y*y + x*y + 2*y + (2*y)*nb; } // not big deal
}
else {
//if (sml) { lwork = 2*y + max( (x+y)*nb, y*y + y ); }
//else { lwork = 2*y + max( (x+y)*nb, x*y + y*nb ); }
// LAPACK 3.4.2 over-estimates workspaces. For compatability, use these:
if (sml) { lwork = 2*y + max( (x+y)*nb, y*y + x ); }
else { lwork = 2*y + (x+y)*nb + x*y; }
}
}
if (jobz == MagmaSomeVec) {
if (tall) { lwork = y*y + 2*y + (2*y)*nb; }
else { lwork = 2*y + (x+y)*nb; }
}
if (jobz == MagmaAllVec) {
if (tall) {
if (sml) { lwork = y*y + 2*y + max( (2*y)*nb, x ); }
else { lwork = y*y + 2*y + max( (2*y)*nb, x*nb ); }
}
else { lwork = 2*y + (x+y)*nb; }
}
#else // REAL ----------------------------------------
if (jobz == MagmaNoVec) {
if (tall) { lwork = 3*y + max( (2*y)*nb, 7*y ); }
else { lwork = 3*y + max( (x+y)*nb, 7*y ); }
}
if (jobz == MagmaOverwriteVec) {
if (tall) {
if (sml) { lwork = y*y + 3*y + max( (2*y)*nb, 4*y*y + 4*y ); }
else { lwork = y*y + 3*y + max( max( (2*y)*nb, 4*y*y + 4*y ), y*y + y*nb ); }
}
else {
if (sml) { lwork = 3*y + max( (x+y)*nb, 4*y*y + 4*y ); }
else { lwork = 3*y + max( (x+y)*nb, 3*y*y + 4*y + x*y ); } // extra space not too important?
}
}
if (jobz == MagmaSomeVec) {
if (tall) { lwork = y*y + 3*y + max( (2*y)*nb, 3*y*y + 4*y ); }
else { lwork = 3*y + max( (x+y)*nb, 3*y*y + 4*y ); }
}
if (jobz == MagmaAllVec) {
if (tall) {
if (sml) { lwork = y*y + max( 3*y + max( (2*y)*nb, 3*y*y + 4*y ), y + x ); }
else { lwork = y*y + max( 3*y + max( (2*y)*nb, 3*y*y + 4*y ), y + x*nb ); }
// LAPACK 3.4.2 over-estimates workspaces. For compatability, use these:
//if (sml) { lwork = y*y + 3*y + max( (2*y)*nb, 3*y*y + 3*y + x ); }
//else { lwork = y*y + max( 3*y + max( (2*y)*nb, max( 3*y*y + 3*y + x, 3*y*y + 4*y )), y + x*nb ); }
}
else { lwork = 3*y + max( (x+y)*nb, 3*y*y + 4*y ); }
}
#endif
break;
}
default: {
fprintf( stderr, "Error: unknown option svd_work %d\n", (int) opts.svd_work );
return -1;
break;
}
}
TESTING_MALLOC_CPU( h_A, magmaDoubleComplex, lda*N );
TESTING_MALLOC_CPU( VT, magmaDoubleComplex, ldv*N ); // N x N (jobz=A) or min(M,N) x N
TESTING_MALLOC_CPU( U, magmaDoubleComplex, ldu*N_U ); // M x M (jobz=A) or M x min(M,N)
TESTING_MALLOC_CPU( S1, double, min_mn );
TESTING_MALLOC_CPU( S2, double, min_mn );
TESTING_MALLOC_CPU( iwork, magma_int_t, 8*min_mn );
TESTING_MALLOC_PIN( h_R, magmaDoubleComplex, lda*N );
TESTING_MALLOC_PIN( h_work, magmaDoubleComplex, lwork );
#ifdef COMPLEX
if (jobz == MagmaNoVec) {
// requires 5*min_mn, but MKL (11.1) seems to have a bug
// requiring 7*min_mn in some cases (e.g., jobz=N, m=100, n=170)
lrwork = 7*min_mn;
}
else if (tall) {
lrwork = 5*min_mn*min_mn + 5*min_mn;
}
else {
lrwork = max( 5*min_mn*min_mn + 5*min_mn,
2*max_mn*min_mn + 2*min_mn*min_mn + min_mn );
}
TESTING_MALLOC_CPU( rwork, double, lrwork );
#endif
/* Initialize the matrix */
lapackf77_zlarnv( &ione, ISEED, &n2, h_A );
lapackf77_zlacpy( MagmaFullStr, &M, &N, h_A, &lda, h_R, &lda );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
gpu_time = magma_wtime();
magma_zgesdd( jobz, M, N,
h_R, lda, S1, U, ldu, VT, ldv, h_work, lwork,
#ifdef COMPLEX
rwork,
#endif
iwork, &info );
gpu_time = magma_wtime() - gpu_time;
if (info != 0) {
printf("magma_zgesdd returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
double eps = lapackf77_dlamch( "E" );
double result[5] = { -1/eps, -1/eps, -1/eps, -1/eps, -1/eps };
if ( opts.check ) {
/* =====================================================================
Check the results following the LAPACK's [zcds]drvbd routine.
A is factored as A = U diag(S) VT and the following 4 tests computed:
(1) | A - U diag(S) VT | / ( |A| max(M,N) )
(2) | I - U^H U | / ( M )
(3) | I - VT VT^H | / ( N )
(4) S contains MNMIN nonnegative values in decreasing order.
(Return 0 if true, 1/ULP if false.)
=================================================================== */
magma_int_t izero = 0;
// get size and location of U and V^T depending on jobz
// U2=NULL and VT2=NULL if they were not computed (e.g., jobz=N)
magmaDoubleComplex *U2 = NULL;
magmaDoubleComplex *VT2 = NULL;
if ( jobz == MagmaSomeVec || jobz == MagmaAllVec ) {
U2 = U;
VT2 = VT;
}
else if ( jobz == MagmaOverwriteVec ) {
if ( M >= N ) {
U2 = h_R;
ldu = lda;
VT2 = VT;
}
else {
U2 = U;
VT2 = h_R;
ldv = lda;
}
}
// zbdt01 needs M+N
// zunt01 prefers N*(N+1) to check U; M*(M+1) to check V
magma_int_t lwork_err = M+N;
if ( U2 != NULL ) {
lwork_err = max( lwork_err, N_U*(N_U+1) );
}
if ( VT2 != NULL ) {
lwork_err = max( lwork_err, M_VT*(M_VT+1) );
}
magmaDoubleComplex *h_work_err;
TESTING_MALLOC_CPU( h_work_err, magmaDoubleComplex, lwork_err );
// zbdt01 and zunt01 need max(M,N), depending
double *rwork_err;
TESTING_MALLOC_CPU( rwork_err, double, max(M,N) );
if ( U2 != NULL && VT2 != NULL ) {
// since KD=0 (3rd arg), E is not referenced so pass NULL (9th arg)
lapackf77_zbdt01(&M, &N, &izero, h_A, &lda,
U2, &ldu, S1, NULL, VT2, &ldv,
h_work_err,
#ifdef COMPLEX
rwork_err,
#endif
&result[0]);
}
if ( U2 != NULL ) {
lapackf77_zunt01("Columns", &M, &N_U, U2, &ldu, h_work_err, &lwork_err,
#ifdef COMPLEX
rwork_err,
#endif
&result[1]);
}
if ( VT2 != NULL ) {
lapackf77_zunt01("Rows", &M_VT, &N, VT2, &ldv, h_work_err, &lwork_err,
#ifdef COMPLEX
rwork_err,
#endif
&result[2]);
}
result[3] = 0.;
for (int j=0; j < min_mn-1; j++) {
if ( S1[j] < S1[j+1] )
result[3] = 1.;
if ( S1[j] < 0. )
result[3] = 1.;
}
if (min_mn > 1 && S1[min_mn-1] < 0.)
result[3] = 1.;
result[0] *= eps;
result[1] *= eps;
result[2] *= eps;
TESTING_FREE_CPU( h_work_err );
TESTING_FREE_CPU( rwork_err );
}
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
if ( opts.lapack ) {
cpu_time = magma_wtime();
lapackf77_zgesdd( lapack_vec_const(jobz), &M, &N,
h_A, &lda, S2, U, &ldu, VT, &ldv, h_work, &lwork,
#ifdef COMPLEX
rwork,
#endif
iwork, &info);
cpu_time = magma_wtime() - cpu_time;
if (info != 0) {
printf("lapackf77_zgesdd returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
/* =====================================================================
Check the result compared to LAPACK
=================================================================== */
double work[1], c_neg_one = -1;
blasf77_daxpy(&min_mn, &c_neg_one, S1, &ione, S2, &ione);
result[4] = lapackf77_dlange("f", &min_mn, &ione, S2, &min_mn, work);
result[4] /= lapackf77_dlange("f", &min_mn, &ione, S1, &min_mn, work);
printf(" %c %5d %5d %7.2f %7.2f %8.2e",
lapack_vec_const(jobz)[0],
(int) M, (int) N, cpu_time, gpu_time, result[4] );
}
else {
printf(" %c %5d %5d --- %7.2f --- ",
lapack_vec_const(jobz)[0],
(int) M, (int) N, gpu_time );
}
if ( opts.check ) {
if ( result[0] < 0. ) { printf(" --- "); } else { printf(" %#9.3g", result[0]); }
if ( result[1] < 0. ) { printf(" --- "); } else { printf(" %#9.3g", result[1]); }
if ( result[2] < 0. ) { printf(" --- "); } else { printf(" %#9.3g", result[2]); }
bool okay = (result[0] < tol) && (result[1] < tol) && (result[2] < tol) && (result[3] == 0.) && (result[4] < tol);
printf(" %3s %s\n", (result[3] == 0. ? "yes" : "no"), (okay ? "ok" : "failed"));
status += ! okay;
}
else {
printf("\n");
}
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( VT );
TESTING_FREE_CPU( U );
TESTING_FREE_CPU( S1 );
TESTING_FREE_CPU( S2 );
TESTING_FREE_CPU( iwork );
#ifdef COMPLEX
TESTING_FREE_CPU( rwork );
#endif
TESTING_FREE_PIN( h_R );
TESTING_FREE_PIN( h_work );
fflush( stdout );
}}
if ( opts.all || opts.niter > 1 ) {
printf("\n");
}
}
opts.cleanup();
TESTING_FINALIZE();
return status;
}