// task execution code
void SCHED_zsyrk(Quark* quark)
{
magma_int_t UPPER;
magma_int_t N;
cuDoubleComplex *A1;
magma_int_t LDA;
magma_int_t K;
cuDoubleComplex *A2;
double mone = -1.0;
double one = 1.0;
quark_unpack_args_6(quark, UPPER, N, K, A1, LDA, A2);
if (UPPER) {
blasf77_zherk(MagmaUpperStr, MagmaConjTransStr, &N, &K, &mone, A1, &LDA, &one,
A2, &LDA);
} else {
blasf77_zherk(MagmaLowerStr, MagmaNoTransStr, &N, &K, &mone, A1, &LDA, &one,
A2, &LDA);
}
}
void get_QR_error(magma_int_t M, magma_int_t N, magma_int_t min_mn,
magmaDoubleComplex *h_R, magmaDoubleComplex *h_A, magma_int_t lda,
magmaDoubleComplex *tau,
magmaDoubleComplex *Q, magma_int_t ldq,
magmaDoubleComplex *R, magma_int_t ldr,
magmaDoubleComplex *h_work, magma_int_t lwork,
double *work, double *error, double *error2)
{
/* h_R:input the factorized matrix by lapack QR,
h_A:input the original matrix copy
tau: input
*/
const double d_neg_one = MAGMA_D_NEG_ONE;
const double d_one = MAGMA_D_ONE;
const magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
const magmaDoubleComplex c_one = MAGMA_Z_ONE;
const magmaDoubleComplex c_zero = MAGMA_Z_ZERO;
double Anorm;
magma_int_t info;
// generate M by K matrix Q, where K = min(M,N)
lapackf77_zlacpy( "Lower", &M, &min_mn, h_R, &lda, Q, &ldq );
lapackf77_zungqr( &M, &min_mn, &min_mn, Q, &ldq, tau, h_work, &lwork, &info );
assert( info == 0 );
// copy K by N matrix R
lapackf77_zlaset( "Lower", &min_mn, &N, &c_zero, &c_zero, R, &ldr );
lapackf77_zlacpy( "Upper", &min_mn, &N, h_R, &lda, R, &ldr );
// error = || R - Q^H*A || / (N * ||A||)
blasf77_zgemm( "Conj", "NoTrans", &min_mn, &N, &M,
&c_neg_one, Q, &ldq, h_A, &lda, &c_one, R, &ldr );
Anorm = lapackf77_zlange( "1", &M, &N, h_A, &lda, work );
*error = lapackf77_zlange( "1", &min_mn, &N, R, &ldr, work );
if ( N > 0 && Anorm > 0 )
*error /= (N*Anorm);
// set R = I (K by K identity), then R = I - Q^H*Q
// error = || I - Q^H*Q || / N
lapackf77_zlaset( "Upper", &min_mn, &min_mn, &c_zero, &c_one, R, &ldr );
blasf77_zherk( "Upper", "Conj", &min_mn, &M, &d_neg_one, Q, &ldq, &d_one, R, &ldr );
*error2 = safe_lapackf77_zlanhe( "1", "Upper", &min_mn, R, &ldr, work );
if ( N > 0 )
*error2 /= N;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing zgeqlf
*/
int main( int argc, char** argv)
{
TESTING_INIT();
const double d_neg_one = MAGMA_D_NEG_ONE;
const double d_one = MAGMA_D_ONE;
const magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
const magmaDoubleComplex c_one = MAGMA_Z_ONE;
const magmaDoubleComplex c_zero = MAGMA_Z_ZERO;
const magma_int_t ione = 1;
real_Double_t gflops, gpu_perf, gpu_time, cpu_perf=0, cpu_time=0;
double Anorm, error=0, error2=0;
magmaDoubleComplex *h_A, *h_R, *tau, *h_work, tmp[1];
magma_int_t M, N, n2, lda, lwork, info, min_mn, nb;
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t status = 0;
magma_opts opts;
parse_opts( argc, argv, &opts );
double tol = opts.tolerance * lapackf77_dlamch("E");
printf(" M N CPU GFlop/s (sec) GPU GFlop/s (sec) |L - Q^H*A| |I - Q^H*Q|\n");
printf("===============================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
M = opts.msize[itest];
N = opts.nsize[itest];
min_mn = min(M, N);
lda = M;
n2 = lda*N;
nb = magma_get_zgeqlf_nb(M);
gflops = FLOPS_ZGEQLF( M, N ) / 1e9;
// query for workspace size
lwork = -1;
lapackf77_zgeqlf(&M, &N, NULL, &M, NULL, tmp, &lwork, &info);
lwork = (magma_int_t)MAGMA_Z_REAL( tmp[0] );
lwork = max( lwork, N*nb );
lwork = max( lwork, 2*nb*nb);
TESTING_MALLOC_CPU( tau, magmaDoubleComplex, min_mn );
TESTING_MALLOC_CPU( h_A, magmaDoubleComplex, n2 );
TESTING_MALLOC_CPU( h_work, magmaDoubleComplex, lwork );
TESTING_MALLOC_PIN( h_R, magmaDoubleComplex, n2 );
/* Initialize the matrix */
lapackf77_zlarnv( &ione, ISEED, &n2, h_A );
lapackf77_zlacpy( MagmaUpperLowerStr, &M, &N, h_A, &lda, h_R, &lda );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
gpu_time = magma_wtime();
magma_zgeqlf( M, N, h_R, lda, tau, h_work, lwork, &info);
gpu_time = magma_wtime() - gpu_time;
gpu_perf = gflops / gpu_time;
if (info != 0)
printf("magma_zgeqlf returned error %d: %s.\n",
(int) info, magma_strerror( info ));
/* =====================================================================
Check the result, following zqlt01 except using the reduced Q.
This works for any M,N (square, tall, wide).
=================================================================== */
if ( opts.check ) {
magma_int_t ldq = M;
magma_int_t ldl = min_mn;
magmaDoubleComplex *Q, *L;
double *work;
TESTING_MALLOC_CPU( Q, magmaDoubleComplex, ldq*min_mn ); // M by K
TESTING_MALLOC_CPU( L, magmaDoubleComplex, ldl*N ); // K by N
TESTING_MALLOC_CPU( work, double, min_mn );
// copy M by K matrix V to Q (copying diagonal, which isn't needed) and
// copy K by N matrix L
lapackf77_zlaset( "Full", &min_mn, &N, &c_zero, &c_zero, L, &ldl );
if ( M >= N ) {
// for M=5, N=3: A = [ V V V ] <= V full block (M-N by K)
// K=N [ V V V ]
// [ ----- ]
// [ L V V ] <= V triangle (N by K, copying diagonal too)
// [ L L V ] <= L triangle (K by N)
// [ L L L ]
magma_int_t M_N = M - N;
lapackf77_zlacpy( "Full", &M_N, &min_mn, h_R, &lda, Q, &ldq );
lapackf77_zlacpy( "Upper", &N, &min_mn, &h_R[M_N], &lda, &Q[M_N], &ldq );
lapackf77_zlacpy( "Lower", &min_mn, &N, &h_R[M_N], &lda, L, &ldl );
}
else {
// for M=3, N=5: A = [ L L | L V V ] <= V triangle (K by K)
// K=M [ L L | L L V ] <= L triangle (K by M)
// [ L L | L L L ]
// ^^^============= L full block (K by N-M)
magma_int_t N_M = N - M;
lapackf77_zlacpy( "Upper", &M, &min_mn, &h_R[N_M*lda], &lda, Q, &ldq );
lapackf77_zlacpy( "Full", &min_mn, &N_M, h_R, &lda, L, &ldl );
lapackf77_zlacpy( "Lower", &min_mn, &M, &h_R[N_M*lda], &lda, &L[N_M*ldl], &ldl );
}
// generate M by K matrix Q, where K = min(M,N)
lapackf77_zungql( &M, &min_mn, &min_mn, Q, &ldq, tau, h_work, &lwork, &info );
assert( info == 0 );
// error = || L - Q^H*A || / (N * ||A||)
blasf77_zgemm( "Conj", "NoTrans", &min_mn, &N, &M,
&c_neg_one, Q, &ldq, h_A, &lda, &c_one, L, &ldl );
Anorm = lapackf77_zlange( "1", &M, &N, h_A, &lda, work );
error = lapackf77_zlange( "1", &min_mn, &N, L, &ldl, work );
if ( N > 0 && Anorm > 0 )
error /= (N*Anorm);
// set L = I (K by K identity), then L = I - Q^H*Q
// error = || I - Q^H*Q || / N
lapackf77_zlaset( "Upper", &min_mn, &min_mn, &c_zero, &c_one, L, &ldl );
blasf77_zherk( "Upper", "Conj", &min_mn, &M, &d_neg_one, Q, &ldq, &d_one, L, &ldl );
error2 = lapackf77_zlanhe( "1", "Upper", &min_mn, L, &ldl, work );
if ( N > 0 )
error2 /= N;
TESTING_FREE_CPU( Q ); Q = NULL;
TESTING_FREE_CPU( L ); L = NULL;
TESTING_FREE_CPU( work ); work = NULL;
}
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
if ( opts.lapack ) {
cpu_time = magma_wtime();
lapackf77_zgeqlf( &M, &N, h_A, &lda, tau, h_work, &lwork, &info );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
if (info != 0)
printf("lapack_zgeqlf returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
/* =====================================================================
Print performance and error.
=================================================================== */
printf("%5d %5d ", (int) M, (int) N );
if ( opts.lapack ) {
printf( "%7.2f (%7.2f)", cpu_perf, cpu_time );
}
else {
printf(" --- ( --- )" );
}
printf( " %7.2f (%7.2f) ", gpu_perf, gpu_time );
if ( opts.check ) {
bool okay = (error < tol && error2 < tol);
status += ! okay;
printf( "%11.2e %11.2e %s\n", error, error2, (okay ? "ok" : "failed") );
}
else {
printf( " ---\n" );
}
TESTING_FREE_CPU( tau );
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_work );
TESTING_FREE_PIN( h_R );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing zhegvd
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gpu_time, cpu_time;
magmaDoubleComplex *h_A, *h_R, *h_B, *h_S, *h_work;
double *rwork, *w1, *w2;
double result[4] = {0};
magma_int_t *iwork;
magma_int_t N, n2, info, nb, lwork, liwork, lda, lrwork;
magmaDoubleComplex c_zero = MAGMA_Z_ZERO;
magmaDoubleComplex c_one = MAGMA_Z_ONE;
magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
double d_one = 1.;
double d_neg_one = -1.;
//double d_ten = 10.;
//magma_int_t izero = 0;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t status = 0;
magma_opts opts;
parse_opts( argc, argv, &opts );
double tol = opts.tolerance * lapackf77_dlamch("E");
double tolulp = opts.tolerance * lapackf77_dlamch("P");
if ( opts.check && opts.jobz == MagmaNoVec ) {
fprintf( stderr, "checking results requires vectors; setting jobz=V (option -JV)\n" );
opts.jobz = MagmaVec;
}
printf("using: itype = %d, jobz = %s, uplo = %s\n",
(int) opts.itype, lapack_vec_const(opts.jobz), lapack_uplo_const(opts.uplo));
printf(" N CPU Time (sec) GPU Time(sec)\n");
printf("======================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
N = opts.nsize[itest];
lda = N;
n2 = N*lda;
nb = magma_get_zhetrd_nb(N);
lwork = 2*N*nb + N*N;
lrwork = 1 + 5*N +2*N*N;
liwork = 3 + 5*N;
TESTING_MALLOC_CPU( h_A, magmaDoubleComplex, n2 );
TESTING_MALLOC_CPU( h_B, magmaDoubleComplex, n2 );
TESTING_MALLOC_CPU( w1, double, N );
TESTING_MALLOC_CPU( w2, double, N );
TESTING_MALLOC_CPU( rwork, double, lrwork );
TESTING_MALLOC_CPU( iwork, magma_int_t, liwork );
TESTING_MALLOC_PIN( h_R, magmaDoubleComplex, n2 );
TESTING_MALLOC_PIN( h_S, magmaDoubleComplex, n2 );
TESTING_MALLOC_PIN( h_work, magmaDoubleComplex, lwork );
/* Initialize the matrix */
lapackf77_zlarnv( &ione, ISEED, &n2, h_A );
//lapackf77_zlatms( &N, &N, "U", ISEED, "P", w1, &five, &d_ten,
// &d_one, &N, &N, lapack_uplo_const(opts.uplo), h_B, &lda, h_work, &info);
//lapackf77_zlaset( "A", &N, &N, &c_zero, &c_one, h_B, &lda);
lapackf77_zlarnv( &ione, ISEED, &n2, h_B );
magma_zmake_hpd( N, h_B, lda );
lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_A, &lda, h_R, &lda );
lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_B, &lda, h_S, &lda );
/* warmup */
if ( opts.warmup ) {
magma_zhegvd( opts.itype, opts.jobz, opts.uplo,
N, h_R, lda, h_S, lda, w1,
h_work, lwork,
rwork, lrwork,
iwork, liwork,
&info );
if (info != 0)
printf("magma_zhegvd returned error %d: %s.\n",
(int) info, magma_strerror( info ));
lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_A, &lda, h_R, &lda );
lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_B, &lda, h_S, &lda );
}
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
gpu_time = magma_wtime();
magma_zhegvd( opts.itype, opts.jobz, opts.uplo,
N, h_R, lda, h_S, lda, w1,
h_work, lwork,
rwork, lrwork,
iwork, liwork,
&info );
gpu_time = magma_wtime() - gpu_time;
if (info != 0)
printf("magma_zhegvd returned error %d: %s.\n",
(int) info, magma_strerror( info ));
if ( opts.check ) {
/* =====================================================================
Check the results following the LAPACK's [zc]hegvd routine.
A x = lambda B x is solved
and the following 3 tests computed:
(1) | A Z - B Z D | / ( |A||Z| N ) (itype = 1)
| A B Z - Z D | / ( |A||Z| N ) (itype = 2)
| B A Z - Z D | / ( |A||Z| N ) (itype = 3)
(2) | I - V V' B | / ( N ) (itype = 1,2)
| B - V V' | / ( |B| N ) (itype = 3)
(3) | S(with V) - S(w/o V) | / | S |
=================================================================== */
double temp1, temp2;
//magmaDoubleComplex *tau;
if ( opts.itype == 1 || opts.itype == 2 ) {
lapackf77_zlaset( "A", &N, &N, &c_zero, &c_one, h_S, &lda);
blasf77_zgemm("N", "C", &N, &N, &N, &c_one, h_R, &lda, h_R, &lda, &c_zero, h_work, &N);
blasf77_zhemm("R", lapack_uplo_const(opts.uplo), &N, &N, &c_neg_one, h_B, &lda, h_work, &N, &c_one, h_S, &lda);
result[1] = lapackf77_zlange("1", &N, &N, h_S, &lda, rwork) / N;
}
else if ( opts.itype == 3 ) {
lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_B, &lda, h_S, &lda);
blasf77_zherk(lapack_uplo_const(opts.uplo), "N", &N, &N, &d_neg_one, h_R, &lda, &d_one, h_S, &lda);
result[1] = lapackf77_zlanhe("1", lapack_uplo_const(opts.uplo), &N, h_S, &lda, rwork) / N
/ lapackf77_zlanhe("1", lapack_uplo_const(opts.uplo), &N, h_B, &lda, rwork);
}
result[0] = 1.;
result[0] /= lapackf77_zlanhe("1", lapack_uplo_const(opts.uplo), &N, h_A, &lda, rwork);
result[0] /= lapackf77_zlange("1", &N, &N, h_R, &lda, rwork);
if ( opts.itype == 1 ) {
blasf77_zhemm("L", lapack_uplo_const(opts.uplo), &N, &N, &c_one, h_A, &lda, h_R, &lda, &c_zero, h_work, &N);
for(int i=0; i<N; ++i)
blasf77_zdscal(&N, &w1[i], &h_R[i*N], &ione);
blasf77_zhemm("L", lapack_uplo_const(opts.uplo), &N, &N, &c_neg_one, h_B, &lda, h_R, &lda, &c_one, h_work, &N);
result[0] *= lapackf77_zlange("1", &N, &N, h_work, &lda, rwork)/N;
}
else if ( opts.itype == 2 ) {
blasf77_zhemm("L", lapack_uplo_const(opts.uplo), &N, &N, &c_one, h_B, &lda, h_R, &lda, &c_zero, h_work, &N);
for(int i=0; i<N; ++i)
blasf77_zdscal(&N, &w1[i], &h_R[i*N], &ione);
blasf77_zhemm("L", lapack_uplo_const(opts.uplo), &N, &N, &c_one, h_A, &lda, h_work, &N, &c_neg_one, h_R, &lda);
result[0] *= lapackf77_zlange("1", &N, &N, h_R, &lda, rwork)/N;
}
else if ( opts.itype == 3 ) {
blasf77_zhemm("L", lapack_uplo_const(opts.uplo), &N, &N, &c_one, h_A, &lda, h_R, &lda, &c_zero, h_work, &N);
for(int i=0; i<N; ++i)
blasf77_zdscal(&N, &w1[i], &h_R[i*N], &ione);
blasf77_zhemm("L", lapack_uplo_const(opts.uplo), &N, &N, &c_one, h_B, &lda, h_work, &N, &c_neg_one, h_R, &lda);
result[0] *= lapackf77_zlange("1", &N, &N, h_R, &lda, rwork)/N;
}
/*
lapackf77_zhet21( &ione, lapack_uplo_const(opts.uplo), &N, &izero,
h_A, &lda,
w1, w1,
h_R, &lda,
h_R, &lda,
tau, h_work, rwork, &result[0] );
*/
lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_A, &lda, h_R, &lda );
lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_B, &lda, h_S, &lda );
magma_zhegvd( opts.itype, MagmaNoVec, opts.uplo,
N, h_R, lda, h_S, lda, w2,
h_work, lwork,
rwork, lrwork,
iwork, liwork,
&info );
if (info != 0)
printf("magma_zhegvd returned error %d: %s.\n",
(int) info, magma_strerror( info ));
temp1 = temp2 = 0;
for(int j=0; j<N; j++) {
temp1 = max(temp1, absv(w1[j]));
temp1 = max(temp1, absv(w2[j]));
temp2 = max(temp2, absv(w1[j]-w2[j]));
}
result[2] = temp2 / (((double)N)*temp1);
}
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
if ( opts.lapack ) {
cpu_time = magma_wtime();
lapackf77_zhegvd( &opts.itype, lapack_vec_const(opts.jobz), lapack_uplo_const(opts.uplo),
&N, h_A, &lda, h_B, &lda, w2,
h_work, &lwork,
rwork, &lrwork,
iwork, &liwork,
&info );
cpu_time = magma_wtime() - cpu_time;
if (info != 0)
printf("lapackf77_zhegvd returned error %d: %s.\n",
(int) info, magma_strerror( info ));
printf("%5d %7.2f %7.2f\n",
(int) N, cpu_time, gpu_time);
}
else {
printf("%5d --- %7.2f\n",
(int) N, gpu_time);
}
/* =====================================================================
Print execution time
=================================================================== */
if ( opts.check ) {
printf("Testing the eigenvalues and eigenvectors for correctness:\n");
if ( opts.itype==1 ) {
printf("(1) | A Z - B Z D | / (|A| |Z| N) = %8.2e %s\n", result[0], (result[0] < tol ? "ok" : "failed") );
}
else if ( opts.itype==2 ) {
printf("(1) | A B Z - Z D | / (|A| |Z| N) = %8.2e %s\n", result[0], (result[0] < tol ? "ok" : "failed") );
}
else if ( opts.itype==3 ) {
printf("(1) | B A Z - Z D | / (|A| |Z| N) = %8.2e %s\n", result[0], (result[0] < tol ? "ok" : "failed") );
}
if ( opts.itype==1 || opts.itype==2 ) {
printf("(2) | I - Z Z' B | / N = %8.2e %s\n", result[1], (result[1] < tol ? "ok" : "failed") );
}
else {
printf("(2) | B - Z Z' | / (|B| N) = %8.2e %s\n", result[1], (result[1] < tol ? "ok" : "failed") );
}
printf( "(3) | D(w/ Z) - D(w/o Z) | / |D| = %8.2e %s\n\n", result[2], (result[2] < tolulp ? "ok" : "failed") );
status += ! (result[0] < tol && result[1] < tol && result[2] < tolulp);
}
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_B );
TESTING_FREE_CPU( w1 );
TESTING_FREE_CPU( w2 );
TESTING_FREE_CPU( rwork );
TESTING_FREE_CPU( iwork );
TESTING_FREE_PIN( h_R );
TESTING_FREE_PIN( h_S );
TESTING_FREE_PIN( h_work );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing zherk
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gflops, cublas_perf, cublas_time, cpu_perf, cpu_time;
double cublas_error, Cnorm, work[1];
magma_int_t N, K;
magma_int_t Ak, An;
magma_int_t sizeA, sizeC;
magma_int_t lda, ldc, ldda, lddc;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
magmaDoubleComplex *h_A, *h_C, *h_Ccublas;
magmaDoubleComplex *d_A, *d_C;
magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
double alpha = MAGMA_D_MAKE( 0.29, -0.86 );
double beta = MAGMA_D_MAKE( -0.48, 0.38 );
magma_int_t status = 0;
magma_opts opts;
parse_opts( argc, argv, &opts );
opts.lapack |= opts.check; // check (-c) implies lapack (-l)
double tol = opts.tolerance * lapackf77_dlamch("E");
printf("If running lapack (option --lapack), CUBLAS error is computed\n"
"relative to CPU BLAS result.\n\n");
printf("uplo = %s, transA = %s\n",
lapack_uplo_const(opts.uplo), lapack_trans_const(opts.transA) );
printf(" N K CUBLAS Gflop/s (ms) CPU Gflop/s (ms) CUBLAS error\n");
printf("==================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
N = opts.nsize[itest];
K = opts.ksize[itest];
gflops = FLOPS_ZHERK(K, N) / 1e9;
if ( opts.transA == MagmaNoTrans ) {
lda = An = N;
Ak = K;
} else {
lda = An = K;
Ak = N;
}
ldc = N;
ldda = ((lda+31)/32)*32;
lddc = ((ldc+31)/32)*32;
sizeA = lda*Ak;
sizeC = ldc*N;
TESTING_MALLOC_CPU( h_A, magmaDoubleComplex, lda*Ak );
TESTING_MALLOC_CPU( h_C, magmaDoubleComplex, ldc*N );
TESTING_MALLOC_CPU( h_Ccublas, magmaDoubleComplex, ldc*N );
TESTING_MALLOC_DEV( d_A, magmaDoubleComplex, ldda*Ak );
TESTING_MALLOC_DEV( d_C, magmaDoubleComplex, lddc*N );
/* Initialize the matrices */
lapackf77_zlarnv( &ione, ISEED, &sizeA, h_A );
lapackf77_zlarnv( &ione, ISEED, &sizeC, h_C );
/* =====================================================================
Performs operation using CUBLAS
=================================================================== */
magma_zsetmatrix( An, Ak, h_A, lda, d_A, ldda );
magma_zsetmatrix( N, N, h_C, ldc, d_C, lddc );
cublas_time = magma_sync_wtime( NULL );
cublasZherk( handle, cublas_uplo_const(opts.uplo), cublas_trans_const(opts.transA), N, K,
&alpha, d_A, ldda,
&beta, d_C, lddc );
cublas_time = magma_sync_wtime( NULL ) - cublas_time;
cublas_perf = gflops / cublas_time;
magma_zgetmatrix( N, N, d_C, lddc, h_Ccublas, ldc );
/* =====================================================================
Performs operation using CPU BLAS
=================================================================== */
if ( opts.lapack ) {
cpu_time = magma_wtime();
blasf77_zherk( lapack_uplo_const(opts.uplo), lapack_trans_const(opts.transA), &N, &K,
&alpha, h_A, &lda,
&beta, h_C, &ldc );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
}
/* =====================================================================
Check the result
=================================================================== */
if ( opts.lapack ) {
// compute relative error for both magma & cublas, relative to lapack,
// |C_magma - C_lapack| / |C_lapack|
Cnorm = lapackf77_zlanhe("fro", lapack_uplo_const(opts.uplo), &N, h_C, &ldc, work);
blasf77_zaxpy( &sizeC, &c_neg_one, h_C, &ione, h_Ccublas, &ione );
cublas_error = lapackf77_zlanhe( "fro", lapack_uplo_const(opts.uplo), &N, h_Ccublas, &ldc, work ) / Cnorm;
printf("%5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %s\n",
(int) N, (int) K,
cublas_perf, 1000.*cublas_time,
cpu_perf, 1000.*cpu_time,
cublas_error, (cublas_error < tol ? "ok" : "failed"));
status += ! (cublas_error < tol);
}
else {
printf("%5d %5d %7.2f (%7.2f) --- ( --- ) --- ---\n",
(int) N, (int) K,
cublas_perf, 1000.*cublas_time);
}
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_C );
TESTING_FREE_CPU( h_Ccublas );
TESTING_FREE_DEV( d_A );
TESTING_FREE_DEV( d_C );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return status;
}