// On input, LU and ipiv is LU factorization of A. On output, LU is overwritten.
// Works for any m, n.
// Uses init_matrix() to re-generate original A as needed.
// Returns error in factorization, |PA - LU| / (n |A|)
// This allocates 3 more matrices to store A, L, and U.
float get_LU_error(
magma_opts &opts,
magma_int_t M, magma_int_t N,
magmaFloatComplex *LU, magma_int_t lda,
magma_int_t *ipiv)
{
magma_int_t min_mn = min(M,N);
magma_int_t ione = 1;
magma_int_t i, j;
magmaFloatComplex alpha = MAGMA_C_ONE;
magmaFloatComplex beta = MAGMA_C_ZERO;
magmaFloatComplex *A, *L, *U;
float work[1], matnorm, residual;
TESTING_MALLOC_CPU( A, magmaFloatComplex, lda*N );
TESTING_MALLOC_CPU( L, magmaFloatComplex, M*min_mn );
TESTING_MALLOC_CPU( U, magmaFloatComplex, min_mn*N );
memset( L, 0, M*min_mn*sizeof(magmaFloatComplex) );
memset( U, 0, min_mn*N*sizeof(magmaFloatComplex) );
// set to original A
init_matrix( opts, M, N, A, lda );
lapackf77_claswp( &N, A, &lda, &ione, &min_mn, ipiv, &ione);
// copy LU to L and U, and set diagonal to 1
lapackf77_clacpy( MagmaLowerStr, &M, &min_mn, LU, &lda, L, &M );
lapackf77_clacpy( MagmaUpperStr, &min_mn, &N, LU, &lda, U, &min_mn );
for (j=0; j < min_mn; j++)
L[j+j*M] = MAGMA_C_MAKE( 1., 0. );
matnorm = lapackf77_clange("f", &M, &N, A, &lda, work);
blasf77_cgemm("N", "N", &M, &N, &min_mn,
&alpha, L, &M, U, &min_mn, &beta, LU, &lda);
for( j = 0; j < N; j++ ) {
for( i = 0; i < M; i++ ) {
LU[i+j*lda] = MAGMA_C_SUB( LU[i+j*lda], A[i+j*lda] );
}
}
residual = lapackf77_clange("f", &M, &N, LU, &lda, work);
TESTING_FREE_CPU( A );
TESTING_FREE_CPU( L );
TESTING_FREE_CPU( U );
return residual / (matnorm * N);
}
/* ////////////////////////////////////////////////////////////////////////////
-- 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, 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 *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;
parse_opts( argc, argv, &opts );
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", opts.version );
return -1;
}
float tol, eps = lapackf77_slamch("E");
tol = 10* opts.tolerance * eps;
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 = ((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);
ldwork = N*N;
if (opts.version == 2) {
ldwork = 3*N*N + min_mn;
}
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( MagmaUpperLowerStr, &M, &N, h_A, &lda, h_R, &lda );
magma_csetmatrix( M, N, h_R, lda, d_A, ldda );
// 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
=================================================================== */
gpu_time = magma_sync_wtime( 0 );
magma_cgegqr_gpu( opts.version, M, N, d_A, ldda, dwork, h_rwork, &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 ));
magma_cgetmatrix( M, N, d_A, ldda, h_R, M );
// Regenerate R
// blasf77_cgemm("t", "n", &N, &N, &M, &c_one, h_R, &M, h_A, &M, &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, &M);
blasf77_caxpy( &n2, &c_neg_one, h_A, &ione, h_R, &ione );
e5 = lapackf77_clange("i", &M, &N, h_R, &M, work) /
lapackf77_clange("i", &M, &N, h_A, &lda, work);
magma_cgetmatrix( M, N, d_A, ldda, h_R, M );
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("t", "n", &N, &N, &M, &c_one, h_R, &M, h_R, &M, &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("t", "n", &N, &N, &M, &c_one, h_A, &M, h_A, &M, &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;
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,
(e1 < tol ? "ok" : "failed"));
status += ! (e1 < 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" );
}
}
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing sgeev
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gpu_time, cpu_time;
float *h_A, *h_R, *VL, *VR, *h_work, *w1, *w2;
float *w1i, *w2i;
magmaFloatComplex *w1copy, *w2copy;
magmaFloatComplex c_neg_one = MAGMA_C_NEG_ONE;
float tnrm, result[9];
magma_int_t N, n2, lda, nb, lwork, info;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
float ulp, ulpinv, error;
magma_int_t status = 0;
ulp = lapackf77_slamch( "P" );
ulpinv = 1./ulp;
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 );
float tol = opts.tolerance * lapackf77_slamch("E");
float tolulp = opts.tolerance * lapackf77_slamch("P");
// enable at least some minimal checks, if requested
if ( opts.check && !opts.lapack && opts.jobvl == MagmaNoVec && opts.jobvr == MagmaNoVec ) {
fprintf( stderr, "NOTE: Some checks require vectors to be computed;\n"
" set jobvl=V (option -LV), or jobvr=V (option -RV), or both.\n"
" Some checks require running lapack (-l); setting lapack.\n\n");
opts.lapack = true;
}
printf(" N CPU Time (sec) GPU Time (sec) |W_magma - W_lapack| / |W_lapack|\n");
printf("===========================================================================\n");
for( int i = 0; i < opts.ntest; ++i ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
N = opts.nsize[i];
lda = N;
n2 = lda*N;
nb = magma_get_sgehrd_nb(N);
lwork = N*(2 + nb);
// generous workspace - required by sget22
lwork = max( lwork, N*(5 + 2*N) );
TESTING_MALLOC_CPU( w1copy, magmaFloatComplex, N );
TESTING_MALLOC_CPU( w2copy, magmaFloatComplex, N );
TESTING_MALLOC_CPU( w1, float, N );
TESTING_MALLOC_CPU( w2, float, N );
TESTING_MALLOC_CPU( w1i, float, N );
TESTING_MALLOC_CPU( w2i, float, N );
TESTING_MALLOC_CPU( h_A, float, n2 );
TESTING_MALLOC_PIN( h_R, float, n2 );
TESTING_MALLOC_PIN( VL, float, n2 );
TESTING_MALLOC_PIN( VR, float, n2 );
TESTING_MALLOC_PIN( h_work, float, lwork );
/* Initialize the matrix */
lapackf77_slarnv( &ione, ISEED, &n2, h_A );
lapackf77_slacpy( MagmaUpperLowerStr, &N, &N, h_A, &lda, h_R, &lda );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
gpu_time = magma_wtime();
magma_sgeev( opts.jobvl, opts.jobvr,
N, h_R, lda, w1, w1i,
VL, lda, VR, lda,
h_work, lwork, &info );
gpu_time = magma_wtime() - gpu_time;
if (info != 0)
printf("magma_sgeev returned error %d: %s.\n",
(int) info, magma_strerror( info ));
/* =====================================================================
Check the result
=================================================================== */
if ( opts.check ) {
/* ===================================================================
* Check the result following LAPACK's [zcds]drvev routine.
* The following tests are performed:
* (1) | A * VR - VR * W | / ( n |A| )
*
* Here VR is the matrix of unit right eigenvectors.
* W is a diagonal matrix with diagonal entries W(j).
*
* (2) | |VR(i)| - 1 | and whether largest component real
*
* VR(i) denotes the i-th column of VR.
*
* (3) | A**T * VL - VL * W**T | / ( n |A| )
*
* Here VL is the matrix of unit left eigenvectors, A**T is the
* transpose of A, and W is as above.
*
* (4) | |VL(i)| - 1 | and whether largest component real
*
* VL(i) denotes the i-th column of VL.
*
* (5) W(full) = W(partial, W only) -- currently skipped
* (6) W(full) = W(partial, W and VR)
* (7) W(full) = W(partial, W and VL)
*
* W(full) denotes the eigenvalues computed when both VR and VL
* are also computed, and W(partial) denotes the eigenvalues
* computed when only W, only W and VR, or only W and VL are
* computed.
*
* (8) VR(full) = VR(partial, W and VR)
*
* VR(full) denotes the right eigenvectors computed when both VR
* and VL are computed, and VR(partial) denotes the result
* when only VR is computed.
*
* (9) VL(full) = VL(partial, W and VL)
*
* VL(full) denotes the left eigenvectors computed when both VR
* and VL are also computed, and VL(partial) denotes the result
* when only VL is computed.
*
* (1, 2) only if jobvr = V
* (3, 4) only if jobvl = V
* (5-9) only if check = 2 (option -c2)
================================================================= */
float vmx, vrmx, vtst;
// Initialize result. -1 indicates test was not run.
for( int j = 0; j < 9; ++j )
result[j] = -1.;
if ( opts.jobvr == MagmaVec ) {
// Do test 1: | A * VR - VR * W | / ( n |A| )
// Note this writes result[1] also
lapackf77_sget22( MagmaNoTransStr, MagmaNoTransStr, MagmaNoTransStr,
&N, h_A, &lda, VR, &lda, w1, w1i,
h_work, &result[0] );
result[0] *= ulp;
// Do test 2: | |VR(i)| - 1 | and whether largest component real
result[1] = -1.;
for( int j = 0; j < N; ++j ) {
tnrm = 1.;
if (w1i[j] == 0.)
tnrm = cblas_snrm2(N, &VR[j*lda], ione);
else if (w1i[j] > 0.)
tnrm = magma_slapy2( cblas_snrm2(N, &VR[j *lda], ione),
cblas_snrm2(N, &VR[(j+1)*lda], ione) );
result[1] = max( result[1], min( ulpinv, MAGMA_S_ABS(tnrm-1.)/ulp ));
if (w1i[j] > 0.) {
vmx = vrmx = 0.;
for( int jj = 0; jj < N; ++jj ) {
vtst = magma_slapy2( VR[jj+j*lda], VR[jj+(j+1)*lda]);
if (vtst > vmx)
vmx = vtst;
if ( (VR[jj + (j+1)*lda])==0. &&
MAGMA_S_ABS( VR[jj+j*lda] ) > vrmx)
{
vrmx = MAGMA_S_ABS( VR[jj+j*lda] );
}
}
if (vrmx / vmx < 1. - ulp*2.)
result[1] = ulpinv;
}
}
result[1] *= ulp;
}
if ( opts.jobvl == MagmaVec ) {
// Do test 3: | A**T * VL - VL * W**T | / ( n |A| )
// Note this writes result[3] also
lapackf77_sget22( MagmaTransStr, MagmaNoTransStr, MagmaTransStr,
&N, h_A, &lda, VL, &lda, w1, w1i,
h_work, &result[2] );
result[2] *= ulp;
// Do test 4: | |VL(i)| - 1 | and whether largest component real
result[3] = -1.;
for( int j = 0; j < N; ++j ) {
tnrm = 1.;
if (w1i[j] == 0.)
tnrm = cblas_snrm2(N, &VL[j*lda], ione);
else if (w1i[j] > 0.)
tnrm = magma_slapy2( cblas_snrm2(N, &VL[j *lda], ione),
cblas_snrm2(N, &VL[(j+1)*lda], ione) );
result[3] = max( result[3], min( ulpinv, MAGMA_S_ABS(tnrm-1.)/ulp ));
if (w1i[j] > 0.) {
vmx = vrmx = 0.;
for( int jj = 0; jj < N; ++jj ) {
vtst = magma_slapy2( VL[jj+j*lda], VL[jj+(j+1)*lda]);
if (vtst > vmx)
vmx = vtst;
if ( (VL[jj + (j+1)*lda])==0. &&
MAGMA_S_ABS( VL[jj+j*lda]) > vrmx)
{
vrmx = MAGMA_S_ABS( VL[jj+j*lda] );
}
}
if (vrmx / vmx < 1. - ulp*2.)
result[3] = ulpinv;
}
}
result[3] *= ulp;
}
}
if ( opts.check == 2 ) {
// more extensive tests
// this is really slow because it calls magma_zgeev multiple times
float *LRE, DUM;
TESTING_MALLOC_PIN( LRE, float, n2 );
lapackf77_slarnv( &ione, ISEED, &n2, h_A );
lapackf77_slacpy( MagmaUpperLowerStr, &N, &N, h_A, &lda, h_R, &lda );
// ----------
// Compute eigenvalues, left and right eigenvectors
magma_sgeev( MagmaVec, MagmaVec,
N, h_R, lda, w1, w1i,
VL, lda, VR, lda,
h_work, lwork, &info );
if (info != 0)
printf("magma_zgeev (case V, V) returned error %d: %s.\n",
(int) info, magma_strerror( info ));
// ----------
// Compute eigenvalues only
// These are not exactly equal, and not in the same order, so skip for now.
//lapackf77_slacpy( MagmaUpperLowerStr, &N, &N, h_A, &lda, h_R, &lda );
//magma_sgeev( MagmaNoVec, MagmaNoVec,
// N, h_R, lda, w2, w2i,
// &DUM, 1, &DUM, 1,
// h_work, lwork, &info );
//if (info != 0)
// printf("magma_sgeev (case N, N) returned error %d: %s.\n",
// (int) info, magma_strerror( info ));
//
//// Do test 5: W(full) = W(partial, W only)
//result[4] = 1;
//for( int j = 0; j < N; ++j )
// if ( w1[j] != w2[j] || w1i[j] != w2i[j] )
// result[4] = 0;
// ----------
// Compute eigenvalues and right eigenvectors
lapackf77_slacpy( MagmaUpperLowerStr, &N, &N, h_A, &lda, h_R, &lda );
magma_sgeev( MagmaNoVec, MagmaVec,
N, h_R, lda, w2, w2i,
&DUM, 1, LRE, lda,
h_work, lwork, &info );
if (info != 0)
printf("magma_sgeev (case N, V) returned error %d: %s.\n",
(int) info, magma_strerror( info ));
// Do test 6: W(full) = W(partial, W and VR)
result[5] = 1;
for( int j = 0; j < N; ++j )
if ( w1[j] != w2[j] || w1i[j] != w2i[j] )
result[5] = 0;
// Do test 8: VR(full) = VR(partial, W and VR)
result[7] = 1;
for( int j = 0; j < N; ++j )
for( int jj = 0; jj < N; ++jj )
if ( ! MAGMA_S_EQUAL( VR[j+jj*lda], LRE[j+jj*lda] ))
result[7] = 0;
// ----------
// Compute eigenvalues and left eigenvectors
lapackf77_slacpy( MagmaUpperLowerStr, &N, &N, h_A, &lda, h_R, &lda );
magma_sgeev( MagmaVec, MagmaNoVec,
N, h_R, lda, w2, w2i,
LRE, lda, &DUM, 1,
h_work, lwork, &info );
if (info != 0)
printf("magma_sgeev (case V, N) returned error %d: %s.\n",
(int) info, magma_strerror( info ));
// Do test 7: W(full) = W(partial, W and VL)
result[6] = 1;
for( int j = 0; j < N; ++j )
if ( w1[j] != w2[j] || w1i[j] != w2i[j] )
result[6] = 0;
// Do test 9: VL(full) = VL(partial, W and VL)
result[8] = 1;
for( int j = 0; j < N; ++j )
for( int jj = 0; jj < N; ++jj )
if ( ! MAGMA_S_EQUAL( VL[j+jj*lda], LRE[j+jj*lda] ))
result[8] = 0;
TESTING_FREE_PIN( LRE );
}
/* =====================================================================
Performs operation using LAPACK
Do this after checks, because it overwrites VL and VR.
=================================================================== */
if ( opts.lapack ) {
cpu_time = magma_wtime();
lapackf77_sgeev( &opts.jobvl, &opts.jobvr,
&N, h_A, &lda, w2, w2i,
VL, &lda, VR, &lda,
h_work, &lwork, &info );
cpu_time = magma_wtime() - cpu_time;
if (info != 0)
printf("lapackf77_sgeev returned error %d: %s.\n",
(int) info, magma_strerror( info ));
// check | W_magma - W_lapack | / | W |
// need to sort eigenvalues first
// copy them into complex vectors for ease
for( int j=0; j < N; ++j ) {
w1copy[j] = MAGMA_C_MAKE( w1[j], w1i[j] );
w2copy[j] = MAGMA_C_MAKE( w2[j], w2i[j] );
}
std::sort( w1copy, &w1copy[N], compare );
std::sort( w2copy, &w2copy[N], compare );
// adjust sorting to deal with numerical inaccuracy
// search down w2 for eigenvalue that matches w1's eigenvalue
for( int j=0; j < N; ++j ) {
for( int j2=j; j2 < N; ++j2 ) {
magmaFloatComplex diff = MAGMA_C_SUB( w1copy[j], w2copy[j2] );
float diff2 = magma_szlapy2( diff ) / max( magma_szlapy2( w1copy[j] ), tol );
if ( diff2 < 100*tol ) {
if ( j != j2 ) {
std::swap( w2copy[j], w2copy[j2] );
}
break;
}
}
}
blasf77_caxpy( &N, &c_neg_one, w2copy, &ione, w1copy, &ione );
error = cblas_scnrm2( N, w1copy, 1 );
error /= cblas_scnrm2( N, w2copy, 1 );
printf("%5d %7.2f %7.2f %.2e %s\n",
(int) N, cpu_time, gpu_time,
error, (error < tolulp ? " ok" : " failed"));
status |= ! (error < tolulp);
}
else {
printf("%5d --- %7.2f\n",
(int) N, gpu_time);
}
if ( opts.check ) {
// -1 indicates test was not run
if ( result[0] != -1 ) { printf(" | A * VR - VR * W | / ( n |A| ) = %8.2e %s\n", result[0], (result[0] < tol ? " ok" : " failed")); }
if ( result[1] != -1 ) { printf(" | |VR(i)| - 1 | = %8.2e %s\n", result[1], (result[1] < tol ? " ok" : " failed")); }
if ( result[2] != -1 ) { printf(" | A'* VL - VL * W'| / ( n |A| ) = %8.2e %s\n", result[2], (result[2] < tol ? " ok" : " failed")); }
if ( result[3] != -1 ) { printf(" | |VL(i)| - 1 | = %8.2e %s\n", result[3], (result[3] < tol ? " ok" : " failed")); }
if ( result[4] != -1 ) { printf(" W (full) == W (partial, W only) %s\n", (result[4] == 1. ? " ok" : " failed")); }
if ( result[5] != -1 ) { printf(" W (full) == W (partial, W and VR) %s\n", (result[5] == 1. ? " ok" : " failed")); }
if ( result[6] != -1 ) { printf(" W (full) == W (partial, W and VL) %s\n", (result[6] == 1. ? " ok" : " failed")); }
if ( result[7] != -1 ) { printf(" VR (full) == VR (partial, W and VR) %s\n", (result[7] == 1. ? " ok" : " failed")); }
if ( result[8] != -1 ) { printf(" VL (full) == VL (partial, W and VL) %s\n", (result[8] == 1. ? " ok" : " failed")); }
int newline = 0;
if ( result[0] != -1 ) { status |= ! (result[0] < tol); newline = 1; }
if ( result[1] != -1 ) { status |= ! (result[1] < tol); newline = 1; }
if ( result[2] != -1 ) { status |= ! (result[2] < tol); newline = 1; }
if ( result[3] != -1 ) { status |= ! (result[3] < tol); newline = 1; }
if ( result[4] != -1 ) { status |= ! (result[4] == 1.); newline = 1; }
if ( result[5] != -1 ) { status |= ! (result[5] == 1.); newline = 1; }
if ( result[6] != -1 ) { status |= ! (result[6] == 1.); newline = 1; }
if ( result[7] != -1 ) { status |= ! (result[7] == 1.); newline = 1; }
if ( result[8] != -1 ) { status |= ! (result[8] == 1.); newline = 1; }
if ( newline ) {
printf( "\n" );
}
}
TESTING_FREE_CPU( w1copy );
TESTING_FREE_CPU( w2copy );
TESTING_FREE_CPU( w1 );
TESTING_FREE_CPU( w2 );
TESTING_FREE_CPU( w1i );
TESTING_FREE_CPU( w2i );
TESTING_FREE_CPU( h_A );
TESTING_FREE_PIN( h_R );
TESTING_FREE_PIN( VL );
TESTING_FREE_PIN( VR );
TESTING_FREE_PIN( h_work );
}
if ( opts.niter > 1 ) {
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 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;
}
