/*-------------------------------------------------------------------
* Check the orthogonality of Q
*/
static magma_int_t check_orthogonality(magma_int_t M, magma_int_t N, magmaFloatComplex *Q, magma_int_t LDQ, float eps)
{
float done = 1.0;
float mdone = -1.0;
magmaFloatComplex c_zero = MAGMA_C_ZERO;
magmaFloatComplex c_one = MAGMA_C_ONE;
float normQ, result;
magma_int_t info_ortho;
magma_int_t minMN = min(M, N);
float *work = (float *)malloc(minMN*sizeof(float));
/* Build the idendity matrix */
magmaFloatComplex *Id = (magmaFloatComplex *) malloc(minMN*minMN*sizeof(magmaFloatComplex));
lapackf77_claset("A", &minMN, &minMN, &c_zero, &c_one, Id, &minMN);
/* Perform Id - Q'Q */
if (M >= N)
blasf77_cherk("U", "C", &N, &M, &done, Q, &LDQ, &mdone, Id, &N);
else
blasf77_cherk("U", "N", &M, &N, &done, Q, &LDQ, &mdone, Id, &M);
normQ = lapackf77_clanhe("I", "U", &minMN, Id, &minMN, work);
result = normQ / (minMN * eps);
printf(" ======================================================\n");
printf(" ||Id-Q'*Q||_oo / (minMN*eps) : %15.3E \n", result );
printf(" ======================================================\n");
if ( isnan(result) || isinf(result) || (result > 60.0) ) {
printf("-- Orthogonality is suspicious ! \n");
info_ortho=1;
}
else {
printf("-- Orthogonality is CORRECT ! \n");
info_ortho=0;
}
free(work); free(Id);
return info_ortho;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing cherk
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gflops, cublas_perf, cublas_time, cpu_perf, cpu_time;
float 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};
magmaFloatComplex *h_A, *h_C, *h_Ccublas;
magmaFloatComplex *d_A, *d_C;
magmaFloatComplex c_neg_one = MAGMA_C_NEG_ONE;
float alpha = MAGMA_D_MAKE( 0.29, -0.86 );
float beta = MAGMA_D_MAKE( -0.48, 0.38 );
magma_opts opts;
parse_opts( argc, argv, &opts );
printf("If running lapack (option --lapack), MAGMA and CUBLAS error are both computed\n"
"relative to CPU BLAS result. Else, MAGMA error is computed relative to CUBLAS result.\n\n"
"uplo = %c, transA = %c\n", opts.uplo, opts.transA );
printf(" N K CUBLAS Gflop/s (ms) CPU Gflop/s (ms) CUBLAS error\n");
printf("==================================================================\n");
for( int i = 0; i < opts.ntest; ++i ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
N = opts.nsize[i];
K = opts.ksize[i];
gflops = FLOPS_CHERK(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( h_A, magmaFloatComplex, lda*Ak );
TESTING_MALLOC( h_C, magmaFloatComplex, ldc*N );
TESTING_MALLOC( h_Ccublas, magmaFloatComplex, ldc*N );
TESTING_DEVALLOC( d_A, magmaFloatComplex, ldda*Ak );
TESTING_DEVALLOC( d_C, magmaFloatComplex, lddc*N );
/* Initialize the matrices */
lapackf77_clarnv( &ione, ISEED, &sizeA, h_A );
lapackf77_clarnv( &ione, ISEED, &sizeC, h_C );
/* =====================================================================
Performs operation using CUDA-BLAS
=================================================================== */
magma_csetmatrix( An, Ak, h_A, lda, d_A, ldda );
magma_csetmatrix( N, N, h_C, ldc, d_C, lddc );
cublas_time = magma_sync_wtime( NULL );
cublasCherk( opts.uplo, 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_cgetmatrix( N, N, d_C, lddc, h_Ccublas, ldc );
/* =====================================================================
Performs operation using CPU BLAS
=================================================================== */
if ( opts.lapack ) {
cpu_time = magma_wtime();
blasf77_cherk( &opts.uplo, &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_clanhe("fro", &opts.uplo, &N, h_C, &ldc, work);
blasf77_caxpy( &sizeC, &c_neg_one, h_C, &ione, h_Ccublas, &ione );
cublas_error = lapackf77_clanhe( "fro", &opts.uplo, &N, h_Ccublas, &ldc, work ) / Cnorm;
printf("%5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e\n",
(int) N, (int) K,
cublas_perf, 1000.*cublas_time,
cpu_perf, 1000.*cpu_time,
cublas_error );
}
else {
printf("%5d %5d %7.2f (%7.2f) --- ( --- ) --- ---\n",
(int) N, (int) K,
cublas_perf, 1000.*cublas_time);
}
TESTING_FREE( h_A );
TESTING_FREE( h_C );
TESTING_FREE( h_Ccublas );
TESTING_DEVFREE( d_A );
TESTING_DEVFREE( d_C );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return 0;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing cgelqf_gpu
*/
int main( int argc, char** argv)
{
TESTING_INIT();
const float d_neg_one = MAGMA_D_NEG_ONE;
const float d_one = MAGMA_D_ONE;
const magmaFloatComplex c_neg_one = MAGMA_C_NEG_ONE;
const magmaFloatComplex c_one = MAGMA_C_ONE;
const magmaFloatComplex c_zero = MAGMA_C_ZERO;
const magma_int_t ione = 1;
real_Double_t gflops, gpu_perf, gpu_time, cpu_perf=0, cpu_time=0;
float Anorm, error=0, error2=0;
magmaFloatComplex *h_A, *h_R, *tau, *h_work, tmp[1];
magmaFloatComplex_ptr d_A;
magma_int_t M, N, n2, lda, ldda, lwork, info, min_mn, nb;
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t status = 0;
magma_opts opts;
opts.parse_opts( argc, argv );
float tol = opts.tolerance * lapackf77_slamch("E");
printf("%% M N CPU Gflop/s (sec) GPU Gflop/s (sec) |L - A*Q^H| |I - Q*Q^H|\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;
ldda = magma_roundup( M, opts.align ); // multiple of 32 by default
n2 = lda*N;
nb = magma_get_cgeqrf_nb( M, N );
gflops = FLOPS_CGELQF( M, N ) / 1e9;
// query for workspace size
lwork = -1;
lapackf77_cgelqf(&M, &N, NULL, &M, NULL, tmp, &lwork, &info);
lwork = (magma_int_t)MAGMA_C_REAL( tmp[0] );
lwork = max( lwork, M*nb );
TESTING_MALLOC_CPU( tau, magmaFloatComplex, min_mn );
TESTING_MALLOC_CPU( h_A, magmaFloatComplex, n2 );
TESTING_MALLOC_PIN( h_R, magmaFloatComplex, n2 );
TESTING_MALLOC_PIN( h_work, magmaFloatComplex, lwork );
TESTING_MALLOC_DEV( d_A, magmaFloatComplex, ldda*N );
/* Initialize the matrix */
lapackf77_clarnv( &ione, ISEED, &n2, h_A );
lapackf77_clacpy( MagmaFullStr, &M, &N, h_A, &lda, h_R, &lda );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
magma_csetmatrix( M, N, h_R, lda, d_A, ldda, opts.queue );
gpu_time = magma_wtime();
magma_cgelqf_gpu( M, N, d_A, ldda, tau, h_work, lwork, &info);
gpu_time = magma_wtime() - gpu_time;
gpu_perf = gflops / gpu_time;
if (info != 0) {
printf("magma_cgelqf_gpu returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
/* =====================================================================
Check the result, following zlqt01 except using the reduced Q.
This works for any M,N (square, tall, wide).
=================================================================== */
if ( opts.check ) {
magma_cgetmatrix( M, N, d_A, ldda, h_R, lda, opts.queue );
magma_int_t ldq = min_mn;
magma_int_t ldl = M;
magmaFloatComplex *Q, *L;
float *work;
TESTING_MALLOC_CPU( Q, magmaFloatComplex, ldq*N ); // K by N
TESTING_MALLOC_CPU( L, magmaFloatComplex, ldl*min_mn ); // M by K
TESTING_MALLOC_CPU( work, float, min_mn );
// generate K by N matrix Q, where K = min(M,N)
lapackf77_clacpy( "Upper", &min_mn, &N, h_R, &lda, Q, &ldq );
lapackf77_cunglq( &min_mn, &N, &min_mn, Q, &ldq, tau, h_work, &lwork, &info );
assert( info == 0 );
// copy N by K matrix L
lapackf77_claset( "Upper", &M, &min_mn, &c_zero, &c_zero, L, &ldl );
lapackf77_clacpy( "Lower", &M, &min_mn, h_R, &lda, L, &ldl );
// error = || L - A*Q^H || / (N * ||A||)
blasf77_cgemm( "NoTrans", "Conj", &M, &min_mn, &N,
&c_neg_one, h_A, &lda, Q, &ldq, &c_one, L, &ldl );
Anorm = lapackf77_clange( "1", &M, &N, h_A, &lda, work );
error = lapackf77_clange( "1", &M, &min_mn, L, &ldl, work );
if ( N > 0 && Anorm > 0 )
error /= (N*Anorm);
// set L = I (K by K), then L = I - Q*Q^H
// error = || I - Q*Q^H || / N
lapackf77_claset( "Upper", &min_mn, &min_mn, &c_zero, &c_one, L, &ldl );
blasf77_cherk( "Upper", "NoTrans", &min_mn, &N, &d_neg_one, Q, &ldq, &d_one, L, &ldl );
error2 = safe_lapackf77_clanhe( "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_cgelqf( &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_cgelqf 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);
printf( "error %.4g, error2 %.4g, tol %.4g, okay %d\n", error, error2, tol, okay );
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_PIN( h_R );
TESTING_FREE_PIN( h_work );
TESTING_FREE_DEV( d_A );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
opts.cleanup();
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing cherk
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gflops, cublas_perf, cublas_time, cpu_perf, cpu_time;
float 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};
magmaFloatComplex *h_A, *h_C, *h_Ccublas;
magmaFloatComplex_ptr d_A, d_C;
magmaFloatComplex c_neg_one = MAGMA_C_NEG_ONE;
float alpha = MAGMA_D_MAKE( 0.29, -0.86 );
float 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)
float tol = opts.tolerance * lapackf77_slamch("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_CHERK(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, magmaFloatComplex, lda*Ak );
TESTING_MALLOC_CPU( h_C, magmaFloatComplex, ldc*N );
TESTING_MALLOC_CPU( h_Ccublas, magmaFloatComplex, ldc*N );
TESTING_MALLOC_DEV( d_A, magmaFloatComplex, ldda*Ak );
TESTING_MALLOC_DEV( d_C, magmaFloatComplex, lddc*N );
/* Initialize the matrices */
lapackf77_clarnv( &ione, ISEED, &sizeA, h_A );
lapackf77_clarnv( &ione, ISEED, &sizeC, h_C );
/* =====================================================================
Performs operation using CUBLAS
=================================================================== */
magma_csetmatrix( An, Ak, h_A, lda, d_A, ldda );
magma_csetmatrix( N, N, h_C, ldc, d_C, lddc );
cublas_time = magma_sync_wtime( NULL );
cublasCherk( opts.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_cgetmatrix( N, N, d_C, lddc, h_Ccublas, ldc );
/* =====================================================================
Performs operation using CPU BLAS
=================================================================== */
if ( opts.lapack ) {
cpu_time = magma_wtime();
blasf77_cherk( 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 ) {
#ifdef MAGMA_WITH_MKL
// MKL (11.1.2) has bug in multi-threaded clanhe; use single thread to work around
int threads = magma_get_lapack_numthreads();
magma_set_lapack_numthreads( 1 );
#endif
// compute relative error for both magma & cublas, relative to lapack,
// |C_magma - C_lapack| / |C_lapack|
Cnorm = lapackf77_clanhe("fro", lapack_uplo_const(opts.uplo), &N, h_C, &ldc, work);
blasf77_caxpy( &sizeC, &c_neg_one, h_C, &ione, h_Ccublas, &ione );
cublas_error = lapackf77_clanhe( "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);
#ifdef MAGMA_WITH_MKL
// end single thread to work around MKL bug
magma_set_lapack_numthreads( threads );
#endif
}
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;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing chegvd
*/
int main( int argc, char** argv)
{
TESTING_CUDA_INIT();
cuFloatComplex *h_A, *h_R, *h_B, *h_S, *h_work;
float *rwork, *w1, *w2;
magma_int_t *iwork;
float gpu_time, cpu_time;
magma_timestr_t start, end;
/* Matrix size */
magma_int_t N=0, n2;
magma_int_t size[4] = {1024,2048,4100,6001};
magma_int_t i, itype, info;
magma_int_t ione = 1, izero = 0;
magma_int_t five = 5;
cuFloatComplex c_zero = MAGMA_C_ZERO;
cuFloatComplex c_one = MAGMA_C_ONE;
cuFloatComplex c_neg_one = MAGMA_C_NEG_ONE;
float d_one = 1.;
float d_neg_one = -1.;
float d_ten = 10.;
magma_int_t ISEED[4] = {0,0,0,1};
const char *uplo = MagmaLowerStr;
const char *jobz = MagmaVectorsStr;
itype = 1;
magma_int_t checkres;
float result[4];
int flagN = 0;
if (argc != 1){
for(i = 1; i<argc; i++){
if (strcmp("-N", argv[i])==0){
N = atoi(argv[++i]);
if (N>0){
printf(" testing_chegvd -N %d\n\n", (int) N);
flagN=1;
}
else {
printf("\nUsage: \n");
printf(" testing_chegvd -N %d\n\n", (int) N);
exit(1);
}
}
if (strcmp("-itype", argv[i])==0){
itype = atoi(argv[++i]);
if (itype>0 && itype <= 3){
printf(" testing_chegvd -itype %d\n\n", (int) itype);
}
else {
printf("\nUsage: \n");
printf(" testing_chegvd -itype %d\n\n", (int) itype);
exit(1);
}
}
if (strcmp("-L", argv[i])==0){
uplo = MagmaLowerStr;
printf(" testing_chegvd -L");
}
if (strcmp("-U", argv[i])==0){
uplo = MagmaUpperStr;
printf(" testing_chegvd -U");
}
}
} else {
printf("\nUsage: \n");
printf(" testing_chegvd -L/U -N %d -itype %d\n\n", 1024, 1);
}
if(!flagN)
N = size[3];
checkres = getenv("MAGMA_TESTINGS_CHECK") != NULL;
n2 = N * N;
/* Allocate host memory for the matrix */
TESTING_MALLOC( h_A, cuFloatComplex, n2);
TESTING_MALLOC( h_B, cuFloatComplex, n2);
TESTING_MALLOC( w1, float , N);
TESTING_MALLOC( w2, float , N);
TESTING_HOSTALLOC(h_R, cuFloatComplex, n2);
TESTING_HOSTALLOC(h_S, cuFloatComplex, n2);
magma_int_t nb = magma_get_chetrd_nb(N);
magma_int_t lwork = 2*N*nb + N*N;
magma_int_t lrwork = 1 + 5*N +2*N*N;
magma_int_t liwork = 3 + 5*N;
TESTING_HOSTALLOC(h_work, cuFloatComplex, lwork);
TESTING_MALLOC( rwork, float, lrwork);
TESTING_MALLOC( iwork, magma_int_t, liwork);
printf(" N CPU Time(s) GPU Time(s) \n");
printf("===================================\n");
for(i=0; i<4; i++){
if (!flagN){
N = size[i];
n2 = N*N;
}
/* Initialize the matrix */
lapackf77_clarnv( &ione, ISEED, &n2, h_A );
//lapackf77_clatms( &N, &N, "U", ISEED, "P", w1, &five, &d_ten,
// &d_one, &N, &N, uplo, h_B, &N, h_work, &info);
//lapackf77_claset( "A", &N, &N, &c_zero, &c_one, h_B, &N);
lapackf77_clarnv( &ione, ISEED, &n2, h_B );
/* increase the diagonal */
{
magma_int_t i, j;
for(i=0; i<N; i++) {
MAGMA_C_SET2REAL( h_B[i*N+i], MAGMA_C_REAL(h_B[i*N+i]) + 1.*N );
MAGMA_C_SET2REAL( h_A[i*N+i], MAGMA_C_REAL(h_A[i*N+i]) );
}
}
lapackf77_clacpy( MagmaUpperLowerStr, &N, &N, h_A, &N, h_R, &N );
lapackf77_clacpy( MagmaUpperLowerStr, &N, &N, h_B, &N, h_S, &N );
magma_chegvd(itype, jobz[0], uplo[0],
N, h_R, N, h_S, N, w1,
h_work, lwork,
rwork, lrwork,
iwork, liwork,
&info);
lapackf77_clacpy( MagmaUpperLowerStr, &N, &N, h_A, &N, h_R, &N );
lapackf77_clacpy( MagmaUpperLowerStr, &N, &N, h_B, &N, h_S, &N );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
start = get_current_time();
magma_chegvd(itype, jobz[0], uplo[0],
N, h_R, N, h_S, N, w1,
h_work, lwork,
rwork, lrwork,
iwork, liwork,
&info);
end = get_current_time();
gpu_time = GetTimerValue(start,end)/1000.;
if ( checkres ) {
/* =====================================================================
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 |
=================================================================== */
float temp1, temp2;
cuFloatComplex *tau;
if (itype == 1 || itype == 2){
lapackf77_claset( "A", &N, &N, &c_zero, &c_one, h_S, &N);
blasf77_cgemm("N", "C", &N, &N, &N, &c_one, h_R, &N, h_R, &N, &c_zero, h_work, &N);
blasf77_chemm("R", uplo, &N, &N, &c_neg_one, h_B, &N, h_work, &N, &c_one, h_S, &N);
result[1]= lapackf77_clange("1", &N, &N, h_S, &N, rwork) / N;
}
else if (itype == 3){
lapackf77_clacpy( MagmaUpperLowerStr, &N, &N, h_B, &N, h_S, &N);
blasf77_cherk(uplo, "N", &N, &N, &d_neg_one, h_R, &N, &d_one, h_S, &N);
result[1]= lapackf77_clanhe("1",uplo, &N, h_S, &N, rwork) / N / lapackf77_clanhe("1",uplo, &N, h_B, &N, rwork);
}
result[0] = 1.;
result[0] /= lapackf77_clanhe("1",uplo, &N, h_A, &N, rwork);
result[0] /= lapackf77_clange("1",&N , &N, h_R, &N, rwork);
if (itype == 1){
blasf77_chemm("L", uplo, &N, &N, &c_one, h_A, &N, h_R, &N, &c_zero, h_work, &N);
for(int i=0; i<N; ++i)
blasf77_csscal(&N, &w1[i], &h_R[i*N], &ione);
blasf77_chemm("L", uplo, &N, &N, &c_neg_one, h_B, &N, h_R, &N, &c_one, h_work, &N);
result[0] *= lapackf77_clange("1", &N, &N, h_work, &N, rwork)/N;
}
else if (itype == 2){
blasf77_chemm("L", uplo, &N, &N, &c_one, h_B, &N, h_R, &N, &c_zero, h_work, &N);
for(int i=0; i<N; ++i)
blasf77_csscal(&N, &w1[i], &h_R[i*N], &ione);
blasf77_chemm("L", uplo, &N, &N, &c_one, h_A, &N, h_work, &N, &c_neg_one, h_R, &N);
result[0] *= lapackf77_clange("1", &N, &N, h_R, &N, rwork)/N;
}
else if (itype == 3){
blasf77_chemm("L", uplo, &N, &N, &c_one, h_A, &N, h_R, &N, &c_zero, h_work, &N);
for(int i=0; i<N; ++i)
blasf77_csscal(&N, &w1[i], &h_R[i*N], &ione);
blasf77_chemm("L", uplo, &N, &N, &c_one, h_B, &N, h_work, &N, &c_neg_one, h_R, &N);
result[0] *= lapackf77_clange("1", &N, &N, h_R, &N, rwork)/N;
}
/* lapackf77_chet21(&ione, uplo, &N, &izero,
h_A, &N,
w1, w1,
h_R, &N,
h_R, &N,
tau, h_work, rwork, &result[0]);
*/
lapackf77_clacpy( MagmaUpperLowerStr, &N, &N, h_A, &N, h_R, &N );
lapackf77_clacpy( MagmaUpperLowerStr, &N, &N, h_B, &N, h_S, &N );
magma_chegvd(itype, 'N', uplo[0],
N, h_R, N, h_S, N, w2,
h_work, lwork,
rwork, lrwork,
iwork, liwork,
&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 / temp1;
}
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
start = get_current_time();
lapackf77_chegvd(&itype, jobz, uplo,
&N, h_A, &N, h_B, &N, w2,
h_work, &lwork,
rwork, &lrwork,
iwork, &liwork,
&info);
end = get_current_time();
if (info < 0)
printf("Argument %d of chegvd had an illegal value.\n", (int) -info);
cpu_time = GetTimerValue(start,end)/1000.;
/* =====================================================================
Print execution time
=================================================================== */
printf("%5d %6.2f %6.2f\n",
(int) N, cpu_time, gpu_time);
if ( checkres ){
printf("Testing the eigenvalues and eigenvectors for correctness:\n");
if(itype==1)
printf("(1) | A Z - B Z D | / (|A| |Z| N) = %e\n", result[0]);
else if(itype==2)
printf("(1) | A B Z - Z D | / (|A| |Z| N) = %e\n", result[0]);
else if(itype==3)
printf("(1) | B A Z - Z D | / (|A| |Z| N) = %e\n", result[0]);
if(itype==1 || itype ==2)
printf("(2) | I - Z Z' B | / N = %e\n", result[1]);
else
printf("(2) | B - Z Z' | / (|B| N) = %e\n", result[1]);
printf("(3) | D(w/ Z)-D(w/o Z)|/ |D| = %e\n\n", result[2]);
}
if (flagN)
break;
}
/* Memory clean up */
TESTING_FREE( h_A);
TESTING_FREE( h_B);
TESTING_FREE( w1);
TESTING_FREE( w2);
TESTING_FREE( rwork);
TESTING_FREE( iwork);
TESTING_HOSTFREE(h_work);
TESTING_HOSTFREE( h_R);
TESTING_HOSTFREE( h_S);
/* Shutdown */
TESTING_CUDA_FINALIZE();
}