// Initialize matrix to random & symmetrize.
// Having this in separate function ensures the same ISEED is always used,
// so we can re-generate the identical matrix.
void init_matrix( int m, int n, magmaDoubleComplex *h_A, magma_int_t lda )
{
assert( m == n );
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t n2 = lda*n;
lapackf77_zlarnv( &ione, ISEED, &n2, h_A );
magma_zmake_hermitian( n, h_A, lda );
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing zhegvdx
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gpu_time;
magmaDoubleComplex *h_A, *h_R, *h_B, *h_S, *h_work;
#if defined(PRECISION_z) || defined(PRECISION_c)
double *rwork;
magma_int_t lrwork;
#endif
/* Matrix size */
double *w1, *w2, result[2]={0,0};
magma_int_t *iwork;
magma_int_t N, n2, info, lwork, liwork;
magmaDoubleComplex c_zero = MAGMA_Z_ZERO;
magmaDoubleComplex c_one = MAGMA_Z_ONE;
magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
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");
magma_range_t range = MagmaRangeAll;
if (opts.fraction != 1)
range = MagmaRangeI;
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, range = %s, uplo = %s, opts.check = %d, fraction = %6.4f\n",
(int) opts.itype, lapack_vec_const(opts.jobz), lapack_range_const(range), lapack_uplo_const(opts.uplo),
(int) opts.check, opts.fraction);
printf(" N M GPU Time (sec)\n");
printf("============================\n");
magma_int_t threads = magma_get_parallel_numthreads();
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
N = opts.nsize[itest];
n2 = N*N;
#if defined(PRECISION_z) || defined(PRECISION_c)
lwork = magma_zbulge_get_lq2(N, threads) + 2*N + N*N;
lrwork = 1 + 5*N +2*N*N;
#else
lwork = magma_zbulge_get_lq2(N, threads) + 1 + 6*N + 2*N*N;
#endif
liwork = 3 + 5*N;
/* Allocate host memory for the matrix */
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( 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 );
#if defined(PRECISION_z) || defined(PRECISION_c)
TESTING_MALLOC_PIN( rwork, double, lrwork);
#endif
/* Initialize the matrix */
lapackf77_zlarnv( &ione, ISEED, &n2, h_A );
lapackf77_zlarnv( &ione, ISEED, &n2, h_B );
magma_zmake_hpd( N, h_B, N );
magma_zmake_hermitian( N, h_A, N );
magma_int_t m1 = 0;
double vl = 0;
double vu = 0;
magma_int_t il = 0;
magma_int_t iu = 0;
if (range == MagmaRangeI) {
il = 1;
iu = (int) (opts.fraction*N);
}
// ==================================================================
// Warmup using MAGMA
// ==================================================================
if (opts.warmup) {
lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_A, &N, h_R, &N );
lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_B, &N, h_S, &N );
magma_zhegvdx_2stage(opts.itype, opts.jobz, range, opts.uplo,
N, h_R, N, h_S, N, vl, vu, il, iu, &m1, w1,
h_work, lwork,
#if defined(PRECISION_z) || defined(PRECISION_c)
rwork, lrwork,
#endif
iwork, liwork,
&info);
}
// ===================================================================
// Performs operation using MAGMA
// ===================================================================
lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_A, &N, h_R, &N );
lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_B, &N, h_S, &N );
gpu_time = magma_wtime();
magma_zhegvdx_2stage(opts.itype, opts.jobz, range, opts.uplo,
N, h_R, N, h_S, N, vl, vu, il, iu, &m1, w1,
h_work, lwork,
#if defined(PRECISION_z) || defined(PRECISION_c)
rwork, lrwork,
#endif
iwork, liwork,
&info);
gpu_time = magma_wtime() - gpu_time;
if ( opts.check && opts.jobz != MagmaNoVec ) {
/* =====================================================================
Check the results following the LAPACK's [zc]hegvdx 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) | S(with V) - S(w/o V) | / | S |
=================================================================== */
#if defined(PRECISION_d) || defined(PRECISION_s)
double *rwork = h_work + N*N;
#endif
result[0] = 1.;
result[0] /= lapackf77_zlanhe("1", lapack_uplo_const(opts.uplo), &N, h_A, &N, rwork);
result[0] /= lapackf77_zlange("1", &N, &m1, h_R, &N, rwork);
if (opts.itype == 1) {
blasf77_zhemm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_one, h_A, &N, h_R, &N, &c_zero, h_work, &N);
for(int i=0; i<m1; ++i)
blasf77_zdscal(&N, &w1[i], &h_R[i*N], &ione);
blasf77_zhemm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_neg_one, h_B, &N, h_R, &N, &c_one, h_work, &N);
result[0] *= lapackf77_zlange("1", &N, &m1, h_work, &N, rwork)/N;
}
else if (opts.itype == 2) {
blasf77_zhemm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_one, h_B, &N, h_R, &N, &c_zero, h_work, &N);
for(int i=0; i<m1; ++i)
blasf77_zdscal(&N, &w1[i], &h_R[i*N], &ione);
blasf77_zhemm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_one, h_A, &N, h_work, &N, &c_neg_one, h_R, &N);
result[0] *= lapackf77_zlange("1", &N, &m1, h_R, &N, rwork)/N;
}
else if (opts.itype == 3) {
blasf77_zhemm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_one, h_A, &N, h_R, &N, &c_zero, h_work, &N);
for(int i=0; i<m1; ++i)
blasf77_zdscal(&N, &w1[i], &h_R[i*N], &ione);
blasf77_zhemm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_one, h_B, &N, h_work, &N, &c_neg_one, h_R, &N);
result[0] *= lapackf77_zlange("1", &N, &m1, h_R, &N, rwork)/N;
}
lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_A, &N, h_R, &N );
lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_B, &N, h_S, &N );
magma_int_t m2 = m1;
lapackf77_zhegvd(&opts.itype, "N", lapack_uplo_const(opts.uplo), &N,
h_R, &N, h_S, &N, w2,
h_work, &lwork,
#if defined(PRECISION_z) || defined(PRECISION_c)
rwork, &lrwork,
#endif
iwork, &liwork,
&info);
double maxw=0, diff=0;
for(int j=0; j<m2; j++) {
maxw = max(maxw, fabs(w1[j]));
maxw = max(maxw, fabs(w2[j]));
diff = max(diff, fabs(w1[j] - w2[j]));
}
result[1] = diff / (m2*maxw);
}
/* =====================================================================
Print execution time
=================================================================== */
printf("%5d %5d %7.2f\n",
(int) N, (int) m1, gpu_time);
if ( opts.check && opts.jobz != MagmaNoVec ) {
printf("Testing the eigenvalues and eigenvectors for correctness:\n");
if (opts.itype==1) {
printf(" | 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(" | 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(" | B A Z - Z D | / (|A| |Z| N) = %8.2e %s\n", result[0], (result[0] < tol ? "ok" : "failed"));
}
printf( " | D(w/ Z) - D(w/o Z) | / |D| = %8.2e %s\n\n", result[1], (result[1] < tolulp ? "ok" : "failed"));
status += ! (result[0] < tol && result[1] < tolulp);
}
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_B );
TESTING_FREE_CPU( w1 );
TESTING_FREE_CPU( w2 );
TESTING_FREE_CPU( iwork );
TESTING_FREE_PIN( h_R );
TESTING_FREE_PIN( h_S );
TESTING_FREE_PIN( h_work );
#if defined(PRECISION_z) || defined(PRECISION_c)
TESTING_FREE_PIN( rwork );
#endif
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
/* Shutdown */
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing magma_zhemm_mgpu
*/
int main( int argc, char** argv)
{
TESTING_INIT();
magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
magmaDoubleComplex calpha = MAGMA_Z_MAKE( 3.456, 5.678 );
magmaDoubleComplex cbeta = MAGMA_Z_MAKE( 1.234, 2.456 );
real_Double_t gflops, gpu_perf=0., cpu_perf=0., gpu_time=0., cpu_time=0.;
real_Double_t gpu_perf2=0., gpu_time2=0.;
double error=0., errorbis=0., work[1];
magmaDoubleComplex *hA, *hX, *hB, *hR;
magmaDoubleComplex_ptr dA[MagmaMaxGPUs], dX[MagmaMaxGPUs], dB[MagmaMaxGPUs], dwork[MagmaMaxGPUs], hwork[MagmaMaxGPUs+1];
magmaDoubleComplex_ptr dA2;
magma_int_t M, N, size, lda, ldda, msize, nb, nstream;
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");
// default values
nb = (opts.nb > 0 ? opts.nb : 64);
nstream = (opts.nstream > 0 ? opts.nstream : 2);
magma_int_t gnode[MagmaMaxGPUs][MagmaMaxGPUs+2];
magma_int_t nbcmplx = 0;
magma_buildconnection_mgpu(gnode, &nbcmplx, opts.ngpu);
printf("Initializing communication pattern... GPU-ncmplx %d\n\n", (int) nbcmplx);
for (int i=0; i < nbcmplx; ++i) {
int myngpu = gnode[i][MagmaMaxGPUs];
printf("cmplx %d has %d gpu ", i, myngpu);
for(int j=0; j < myngpu; ++j)
printf(" %d", (int) gnode[i][j]);
printf("\n");
}
magma_int_t nbevents = 2;
magma_queue_t streams[MagmaMaxGPUs][20];
magma_event_t redevents[MagmaMaxGPUs][20];
magma_event_t redevents2[MagmaMaxGPUs][MagmaMaxGPUs*MagmaMaxGPUs+10];
for( int d = 0; d < opts.ngpu; ++d ) {
for( magma_int_t i = 0; i < nstream; ++i ) {
magma_queue_create( &streams[d][i] );
}
for( magma_int_t i = 0; i < nbevents; ++i ) {
cudaEventCreateWithFlags(&redevents[d][i], cudaEventDisableTiming);
cudaEventCreateWithFlags(&redevents2[d][i], cudaEventDisableTiming);
}
}
printf( "nb %d, ngpu %d, nstream %d version %d\n", (int) nb, (int) opts.ngpu, (int) nstream, (int) opts.version );
printf(" M N nb offset CPU GFlop/s (sec) GPU GFlop/s (sec) CUBLAS hemm (sec) ||R|| / ||A||*||X||\n");
printf("=========================================================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
M = opts.msize[itest];
N = opts.nsize[itest];
for( int offset = 0; offset < N; offset += min(N,nb) ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
msize = M - offset;
lda = M;
ldda = ((M + 31)/32)*32;
size = lda*M;
gflops = FLOPS_ZHEMM( MagmaLeft, (double)msize, (double)N ) / 1e9;
magma_int_t dworksiz = ldda*N*3;
magma_int_t hworksiz = lda*N;
TESTING_MALLOC_CPU( hA, magmaDoubleComplex, lda*M );
TESTING_MALLOC_CPU( hX, magmaDoubleComplex, lda*N );
TESTING_MALLOC_CPU( hB, magmaDoubleComplex, lda*N );
TESTING_MALLOC_PIN( hR, magmaDoubleComplex, lda*N );
for( int d = 0; d < opts.ngpu; ++d ) {
magma_int_t mlocal = ((M / nb) / opts.ngpu + 1) * nb;
magma_setdevice( d );
TESTING_MALLOC_DEV( dA[d], magmaDoubleComplex, ldda*mlocal );
TESTING_MALLOC_DEV( dX[d], magmaDoubleComplex, ldda*N );
TESTING_MALLOC_DEV( dB[d], magmaDoubleComplex, ldda*N );
TESTING_MALLOC_DEV( dwork[d], magmaDoubleComplex, dworksiz );
TESTING_MALLOC_PIN( hwork[d], magmaDoubleComplex, hworksiz );
}
TESTING_MALLOC_PIN( hwork[opts.ngpu], magmaDoubleComplex, lda*N );
if ( opts.check ) {
magma_setdevice( 0 );
TESTING_MALLOC_DEV( dA2, magmaDoubleComplex, ldda*M );
}
lapackf77_zlarnv( &ione, iseed, &size, hA );
magma_zmake_hermitian( M, hA, lda );
size = lda*N;
lapackf77_zlarnv( &ione, iseed, &size, hX );
lapackf77_zlarnv( &ione, iseed, &size, hB );
lapackf77_zlacpy( "Full", &M, &N, hB, &lda, hR, &lda );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
magma_zsetmatrix_1D_col_bcyclic( M, M, hA, lda, dA, ldda, opts.ngpu, nb );
for( int d = 0; d < opts.ngpu; ++d ) {
magma_setdevice( d );
//magmablasSetKernelStream( streams[ d ][ 0 ] );
magma_zsetmatrix( M, N, hX, lda, dX[d], ldda );
//if (d == 0) magma_zsetmatrix( M, N, hB, lda, dB[d], ldda ); // this is wrong coz when offset != 0 the gpu who do the beta*C may be not 0 so this should be related to stdev(starting device who own i=0 first col)
magma_zsetmatrix( M, N, hB, lda, dB[d], ldda );
}
//memset(hR, 0, lda*N*sizeof(magmaDoubleComplex));
trace_init( 1, opts.ngpu, nstream, (magma_queue_t*) streams );
//magma_int_t offset = 0; //nb;
gpu_time = magma_sync_wtime(0);
magmablas_zhemm_mgpu_com(
MagmaLeft, MagmaLower, msize, N,
calpha, dA, ldda, offset,
dX, ldda,
cbeta, dB, ldda, dwork, dworksiz, hR, lda, hwork, hworksiz,
opts.ngpu, nb, streams, nstream, redevents2, nbevents, gnode, nbcmplx);
gpu_time = magma_sync_wtime(0) - gpu_time;
gpu_perf = gflops / gpu_time;
#ifdef TRACING
char buf[80];
snprintf( buf, sizeof(buf), "zhemm-m%d-n%d-nb%d-stream%d-ngpu%d-run%d.svg",
(int) M, (int) N, (int) nb, (int) nstream, (int) opts.ngpu, (int) iter );
trace_finalize( buf, "trace.css" );
#endif
/* ====================================================================
Performs operation using CUBLAS
=================================================================== */
if ( opts.check && iter == 0 ) {
magma_setdevice( 0 );
magmablasSetKernelStream( 0 );
magma_zsetmatrix( M, M, hA, lda, dA2, ldda );
magma_zsetmatrix( M, N, hX, lda, dX[0], ldda );
magma_zsetmatrix( M, N, hB, lda, dwork[0], ldda );
gpu_time2 = magma_sync_wtime(0);
magma_zhemm(
MagmaLeft, MagmaLower, msize, N,
calpha, dA2+offset*ldda+offset, ldda,
dX[0], ldda,
cbeta, dwork[0], ldda );
gpu_time2 = magma_sync_wtime(0) - gpu_time2;
gpu_perf2 = gflops / gpu_time2;
}
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
if ( opts.check ) {
// store ||A||*||X||
errorbis = lapackf77_zlange("fro", &msize, &msize, hA+offset*lda+offset, &lda, work );
errorbis *= lapackf77_zlange("fro", &msize, &N, hX, &lda, work );
//printf( "A =" ); magma_zprint( M, M, hA, lda );
//printf( "X =" ); magma_zprint( M, N, hX, lda );
//printf( "B =" ); magma_zprint( M, N, hB, lda );
cpu_time = magma_wtime();
blasf77_zhemm( "Left", "Lower", &msize, &N,
&calpha, hA+offset*lda+offset, &lda,
hX, &lda,
&cbeta, hB, &lda );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
/*
trace_file = fopen("AJETE/C", "w");
for (int j = 0; j < N; j++)
for (int i = 0; i < siz; i++)
fprintf(trace_file, "%10d%10d%40.30e\n", i+1, j+1, hB[j*lda+i]);
fclose(trace_file);
*/
magma_int_t firstprint=0;
for(magma_int_t dev=0; dev < opts.ngpu; ++dev) {
magma_setdevice( dev );
magma_zgetmatrix( M, N, dB[dev], ldda, hR, lda );
// compute relative error ||R||/||A||*||X||, where R := B_magma - B_lapack = R - B
size = lda*N;
blasf77_zaxpy( &size, &c_neg_one, hB, &ione, hR, &ione );
error = lapackf77_zlange("fro", &msize, &N, hR, &lda, work) / errorbis;
//printf( "R =" ); magma_zprint( M, N, hR, lda );
if (firstprint == 0) {
printf( "%5d %5d %5d %5d %7.1f (%7.4f) %7.1f (%7.4f) %7.1f (%7.4f) %8.2e %s\n",
(int) M, (int) N, (int) nb, (int) offset,
cpu_perf, cpu_time,
gpu_perf, gpu_time,
gpu_perf2, gpu_time2,
error, (error < tol ? "ok" : "failed") );
}
else {
printf( "%89s %8.2e %s\n", " ",
error, (error < tol ? "ok" : "failed") );
}
status += ! (error < tol);
firstprint =1;
}
} else {
printf( "%5d %5d %5d %5d --- ( --- ) %7.1f (%7.4f) --- ( --- ) ---\n",
(int) M, (int) N, (int) nb, (int) offset,
gpu_perf, gpu_time );
}
TESTING_FREE_CPU( hA );
TESTING_FREE_CPU( hX );
TESTING_FREE_CPU( hB );
TESTING_FREE_PIN( hR );
for( int d = 0; d < opts.ngpu; ++d ) {
magma_setdevice( d );
TESTING_FREE_DEV( dA[d] );
TESTING_FREE_DEV( dX[d] );
TESTING_FREE_DEV( dB[d] );
TESTING_FREE_DEV( dwork[d] );
TESTING_FREE_PIN( hwork[d] );
}
TESTING_FREE_PIN( hwork[opts.ngpu] );
if ( opts.check ) {
magma_setdevice( 0 );
TESTING_FREE_DEV( dA2 );
}
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
} // offset
printf( "\n" );
}
for( int d = 0; d < opts.ngpu; ++d ) {
magma_setdevice( d );
for( magma_int_t i = 0; i < nstream; ++i ) {
magma_queue_destroy( streams[d][i] );
}
for( magma_int_t i = 0; i < nbevents; ++i ) {
magma_event_destroy( redevents[d][i] );
magma_event_destroy( redevents2[d][i] );
}
}
TESTING_FINALIZE();
return status;
}
int main(int argc, char **argv)
{
TESTING_INIT();
real_Double_t gflops, magma_perf, magma_time, cublas_perf, cublas_time, cpu_perf, cpu_time;
double magma_error, cublas_error, work[1];
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t N, lda, sizeA, sizeX, sizeY, blocks, ldwork;
magma_int_t incx = 1;
magma_int_t incy = 1;
magma_int_t nb = 64;
magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
magmaDoubleComplex alpha = MAGMA_Z_MAKE( 1.5, -2.3 );
magmaDoubleComplex beta = MAGMA_Z_MAKE( -0.6, 0.8 );
magmaDoubleComplex *A, *X, *Y, *Ycublas, *Ymagma;
magmaDoubleComplex *dA, *dX, *dY, *dC_work;
magma_opts opts;
parse_opts( argc, argv, &opts );
printf(" N MAGMA Gflop/s (ms) CUBLAS Gflop/s (ms) CPU Gflop/s (ms) MAGMA error 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];
lda = ((N + 31)/32)*32;
sizeA = N*lda;
sizeX = N*incx;
sizeY = N*incy;
gflops = FLOPS_ZHEMV( N ) / 1e9;
TESTING_MALLOC( A, magmaDoubleComplex, sizeA );
TESTING_MALLOC( X, magmaDoubleComplex, sizeX );
TESTING_MALLOC( Y, magmaDoubleComplex, sizeY );
TESTING_MALLOC( Ycublas, magmaDoubleComplex, sizeY );
TESTING_MALLOC( Ymagma, magmaDoubleComplex, sizeY );
TESTING_DEVALLOC( dA, magmaDoubleComplex, sizeA );
TESTING_DEVALLOC( dX, magmaDoubleComplex, sizeX );
TESTING_DEVALLOC( dY, magmaDoubleComplex, sizeY );
blocks = (N + nb - 1) / nb;
ldwork = lda * (blocks + 1);
TESTING_DEVALLOC( dC_work, magmaDoubleComplex, ldwork );
/* Initialize the matrix */
lapackf77_zlarnv( &ione, ISEED, &sizeA, A );
magma_zmake_hermitian( N, A, lda );
lapackf77_zlarnv( &ione, ISEED, &sizeX, X );
lapackf77_zlarnv( &ione, ISEED, &sizeY, Y );
/* =====================================================================
Performs operation using CUBLAS
=================================================================== */
magma_zsetmatrix( N, N, A, lda, dA, lda );
magma_zsetvector( N, X, incx, dX, incx );
magma_zsetvector( N, Y, incy, dY, incy );
cublas_time = magma_sync_wtime( 0 );
cublasZhemv( opts.uplo, N, alpha, dA, lda, dX, incx, beta, dY, incy );
cublas_time = magma_sync_wtime( 0 ) - cublas_time;
cublas_perf = gflops / cublas_time;
magma_zgetvector( N, dY, incy, Ycublas, incy );
/* =====================================================================
Performs operation using MAGMA BLAS
=================================================================== */
magma_zsetvector( N, Y, incy, dY, incy );
magma_time = magma_sync_wtime( 0 );
#if (GPUSHMEM >= 200)
magmablas_zhemv2( opts.uplo, N, alpha, dA, lda, dX, incx, beta, dY, incy, dC_work, ldwork );
#else
magmablas_zhemv( opts.uplo, N, alpha, dA, lda, dX, incx, beta, dY, incy );
#endif
magma_time = magma_sync_wtime( 0 ) - magma_time;
magma_perf = gflops / magma_time;
magma_zgetvector( N, dY, incy, Ymagma, incy );
/* =====================================================================
Performs operation using CPU BLAS
=================================================================== */
cpu_time = magma_wtime();
blasf77_zhemv( &opts.uplo, &N, &alpha, A, &lda, X, &incx, &beta, Y, &incy );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
/* =====================================================================
Check the result
=================================================================== */
blasf77_zaxpy( &N, &c_neg_one, Y, &incy, Ymagma, &incy );
magma_error = lapackf77_zlange( "M", &N, &ione, Ymagma, &N, work ) / N;
blasf77_zaxpy( &N, &c_neg_one, Y, &incy, Ycublas, &incy );
cublas_error = lapackf77_zlange( "M", &N, &ione, Ycublas, &N, work ) / N;
printf("%5d %7.2f (%7.2f) %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %8.2e\n",
(int) N,
magma_perf, 1000.*magma_time,
cublas_perf, 1000.*cublas_time,
cpu_perf, 1000.*cpu_time,
magma_error, cublas_error );
TESTING_FREE( A );
TESTING_FREE( X );
TESTING_FREE( Y );
TESTING_FREE( Ycublas );
TESTING_FREE( Ymagma );
TESTING_DEVFREE( dA );
TESTING_DEVFREE( dX );
TESTING_DEVFREE( dY );
TESTING_DEVFREE( dC_work );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return 0;
}
// --------------------
int main(int argc, char **argv)
{
TESTING_INIT();
real_Double_t gflops, cpu_time=0, cpu_perf=0, gpu_time, gpu_perf, mgpu_time, mgpu_perf, cuda_time, cuda_perf;
double Ynorm, error=0, error2=0, work[1];
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
magma_int_t n_local[MagmaMaxGPUs];
magma_int_t N, Noffset, lda, ldda, blocks, lhwork, ldwork, matsize, vecsize;
magma_int_t incx = 1;
magmaDoubleComplex alpha = MAGMA_Z_MAKE( 1.5, -2.3 );
magmaDoubleComplex beta = MAGMA_Z_MAKE( -0.6, 0.8 );
magmaDoubleComplex *A, *X, *Y, *Ylapack, *Ycublas, *Ymagma, *Ymagma1, *hwork;
magmaDoubleComplex_ptr dA, dX, dY;
magmaDoubleComplex_ptr d_lA[MagmaMaxGPUs], dwork[MagmaMaxGPUs];
magma_device_t dev;
magma_queue_t queues[MagmaMaxGPUs];
magma_int_t status = 0;
magma_opts opts;
opts.parse_opts( argc, argv );
opts.ngpu = abs( opts.ngpu ); // always uses multi-GPU code
double tol = opts.tolerance * lapackf77_dlamch("E");
magma_int_t nb = 64; // required by magmablas_zhemv_mgpu implementation
for( dev=0; dev < opts.ngpu; ++dev ) {
magma_queue_create( dev, &queues[dev] );
}
// currently, tests all offsets in the offsets array;
// comment out loop below to test a specific offset.
magma_int_t offset = opts.offset;
magma_int_t offsets[] = { 0, 1, 31, 32, 33, 63, 64, 65, 100, 200 };
magma_int_t noffsets = sizeof(offsets) / sizeof(*offsets);
printf("%% uplo = %s, ngpu %d, block size = %d, offset %d\n",
lapack_uplo_const(opts.uplo), (int) opts.ngpu, (int) nb, (int) offset );
printf( "%% BLAS CUBLAS MAGMA 1 GPU MAGMA MGPU Error rel Error rel\n"
"%% N offset Gflop/s (msec) Gflop/s (msec) Gflop/s (msec) Gflop/s (msec) to CUBLAS to LAPACK\n"
"%%==================================================================================================================\n" );
for( int itest = 0; itest < opts.ntest; ++itest ) {
// comment out these two lines & end of loop to test a specific offset
for( int ioffset=0; ioffset < noffsets; ioffset += 1 ) {
offset = offsets[ioffset];
for( int iter = 0; iter < opts.niter; ++iter ) {
N = opts.nsize[itest];
Noffset = N + offset;
lda = Noffset;
ldda = magma_roundup( Noffset, opts.align ); // multiple of 32 by default
matsize = Noffset*ldda;
vecsize = (Noffset-1)*incx + 1;
gflops = FLOPS_ZHEMV( N ) / 1e9;
blocks = magma_ceildiv( N + (offset % nb), nb );
lhwork = N*opts.ngpu;
ldwork = ldda*(blocks + 1);
TESTING_MALLOC_CPU( A, magmaDoubleComplex, matsize );
TESTING_MALLOC_CPU( Y, magmaDoubleComplex, vecsize );
TESTING_MALLOC_CPU( Ycublas, magmaDoubleComplex, vecsize );
TESTING_MALLOC_CPU( Ymagma, magmaDoubleComplex, vecsize );
TESTING_MALLOC_CPU( Ymagma1, magmaDoubleComplex, vecsize );
TESTING_MALLOC_CPU( Ylapack, magmaDoubleComplex, vecsize );
TESTING_MALLOC_PIN( X, magmaDoubleComplex, vecsize );
TESTING_MALLOC_PIN( hwork, magmaDoubleComplex, lhwork );
magma_setdevice( opts.device );
TESTING_MALLOC_DEV( dA, magmaDoubleComplex, matsize );
TESTING_MALLOC_DEV( dX, magmaDoubleComplex, vecsize );
TESTING_MALLOC_DEV( dY, magmaDoubleComplex, vecsize );
// TODO make magma_zmalloc_bcyclic helper function?
for( dev=0; dev < opts.ngpu; dev++ ) {
n_local[dev] = ((Noffset/nb)/opts.ngpu)*nb;
if (dev < (Noffset/nb) % opts.ngpu)
n_local[dev] += nb;
else if (dev == (Noffset/nb) % opts.ngpu)
n_local[dev] += Noffset % nb;
magma_setdevice( dev );
TESTING_MALLOC_DEV( d_lA[dev], magmaDoubleComplex, ldda*n_local[dev] );
TESTING_MALLOC_DEV( dwork[dev], magmaDoubleComplex, ldwork );
}
//////////////////////////////////////////////////////////////////////////
/* Initialize the matrix */
lapackf77_zlarnv( &ione, ISEED, &matsize, A );
magma_zmake_hermitian( Noffset, A, lda );
lapackf77_zlarnv( &ione, ISEED, &vecsize, X );
lapackf77_zlarnv( &ione, ISEED, &vecsize, Y );
/* =====================================================================
Performs operation using CUBLAS
=================================================================== */
magma_setdevice( opts.device );
magma_zsetmatrix( Noffset, Noffset, A, lda, dA, ldda, opts.queue );
magma_zsetvector( Noffset, X, incx, dX, incx, opts.queue );
magma_zsetvector( Noffset, Y, incx, dY, incx, opts.queue );
cuda_time = magma_sync_wtime(0);
cublasZhemv( opts.handle, cublas_uplo_const(opts.uplo), N,
&alpha, dA + offset + offset*ldda, ldda,
dX + offset, incx,
&beta, dY + offset, incx );
cuda_time = magma_sync_wtime(0) - cuda_time;
cuda_perf = gflops / cuda_time;
magma_zgetvector( Noffset, dY, incx, Ycublas, incx, opts.queue );
/* =====================================================================
Performs operation using MAGMABLAS (1 GPU)
=================================================================== */
magma_setdevice( opts.device );
magma_zsetvector( Noffset, Y, incx, dY, incx, opts.queue );
gpu_time = magma_sync_wtime( opts.queue );
magmablas_zhemv_work( opts.uplo, N,
alpha, dA + offset + offset*ldda, ldda,
dX + offset, incx,
beta, dY + offset, incx, dwork[ opts.device ], ldwork,
opts.queue );
gpu_time = magma_sync_wtime( opts.queue ) - gpu_time;
gpu_perf = gflops / gpu_time;
magma_zgetvector( Noffset, dY, incx, Ymagma1, incx, opts.queue );
/* =====================================================================
Performs operation using MAGMABLAS (multi-GPU)
=================================================================== */
magma_zsetmatrix_1D_col_bcyclic( Noffset, Noffset, A, lda, d_lA, ldda, opts.ngpu, nb, queues );
blasf77_zcopy( &Noffset, Y, &incx, Ymagma, &incx );
// workspaces do NOT need to be zero -- set to NAN to prove
for( dev=0; dev < opts.ngpu; ++dev ) {
magma_setdevice( dev );
magmablas_zlaset( MagmaFull, ldwork, 1, MAGMA_Z_NAN, MAGMA_Z_NAN, dwork[dev], ldwork, opts.queue );
}
lapackf77_zlaset( "Full", &lhwork, &ione, &MAGMA_Z_NAN, &MAGMA_Z_NAN, hwork, &lhwork );
mgpu_time = magma_sync_wtime(0);
magma_int_t info;
info = magmablas_zhemv_mgpu(
opts.uplo, N,
alpha,
d_lA, ldda, offset,
X + offset, incx,
beta,
Ymagma + offset, incx,
hwork, lhwork,
dwork, ldwork,
opts.ngpu, nb, queues );
if (info != 0) {
printf("magmablas_zhemv_mgpu returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
info = magmablas_zhemv_mgpu_sync(
opts.uplo, N,
alpha,
d_lA, ldda, offset,
X + offset, incx,
beta,
Ymagma + offset, incx,
hwork, lhwork,
dwork, ldwork,
opts.ngpu, nb, queues );
if (info != 0) {
printf("magmablas_zhemv_sync returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
mgpu_time = magma_sync_wtime(0) - mgpu_time;
mgpu_perf = gflops / mgpu_time;
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
if ( opts.lapack ) {
blasf77_zcopy( &Noffset, Y, &incx, Ylapack, &incx );
cpu_time = magma_wtime();
blasf77_zhemv( lapack_uplo_const(opts.uplo), &N,
&alpha, A + offset + offset*lda, &lda,
X + offset, &incx,
&beta, Ylapack + offset, &incx );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
/* =====================================================================
Compute the Difference LAPACK vs. Magma
=================================================================== */
Ynorm = lapackf77_zlange( "F", &Noffset, &ione, Ylapack, &Noffset, work );
blasf77_zaxpy( &Noffset, &c_neg_one, Ymagma, &incx, Ylapack, &incx );
error2 = lapackf77_zlange( "F", &Noffset, &ione, Ylapack, &Noffset, work ) / Ynorm;
}
/* =====================================================================
Compute the Difference Cublas vs. Magma
=================================================================== */
Ynorm = lapackf77_zlange( "F", &Noffset, &ione, Ycublas, &Noffset, work );
blasf77_zaxpy( &Noffset, &c_neg_one, Ymagma, &incx, Ycublas, &incx );
error = lapackf77_zlange( "F", &Noffset, &ione, Ycublas, &Noffset, work ) / Ynorm;
bool okay = (error < tol && error2 < tol);
status += ! okay;
if ( opts.lapack ) {
printf( "%5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %8.2e %s\n",
(int) N, (int) offset,
cpu_perf, cpu_time*1000.,
cuda_perf, cuda_time*1000.,
gpu_perf, gpu_time*1000.,
mgpu_perf, mgpu_time*1000.,
error, error2, (okay ? "ok" : "failed") );
}
else {
printf( "%5d %5d --- ( --- ) %7.2f (%7.2f) %7.2f (%7.2f) %7.2f (%7.2f) %8.2e --- %s\n",
(int) N, (int) offset,
cuda_perf, cuda_time*1000.,
gpu_perf, gpu_time*1000.,
mgpu_perf, mgpu_time*1000.,
error, (okay ? "ok" : "failed") );
}
/* Free Memory */
TESTING_FREE_CPU( A );
TESTING_FREE_CPU( Y );
TESTING_FREE_CPU( Ycublas );
TESTING_FREE_CPU( Ymagma );
TESTING_FREE_CPU( Ymagma1 );
TESTING_FREE_CPU( Ylapack );
TESTING_FREE_PIN( X );
TESTING_FREE_PIN( hwork );
magma_setdevice( opts.device );
TESTING_FREE_DEV( dA );
TESTING_FREE_DEV( dX );
TESTING_FREE_DEV( dY );
for( dev=0; dev < opts.ngpu; dev++ ) {
magma_setdevice( dev );
TESTING_FREE_DEV( d_lA[dev] );
TESTING_FREE_DEV( dwork[dev] );
}
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
// comment out these two lines line & top of loop test a specific offset
} // end for ioffset
printf( "\n" );
}
for( dev=0; dev < opts.ngpu; ++dev ) {
magma_queue_destroy( queues[dev] );
}
opts.cleanup();
TESTING_FINALIZE();
return status;
}
int main(int argc, char **argv)
{
TESTING_INIT();
real_Double_t gflops, magma_perf, magma_time, cpu_perf, cpu_time;
double magma_error, work[1];
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t N, lda, ldda, sizeA, sizeX, sizeY, blocks, ldwork;
magma_int_t incx = 1;
magma_int_t incy = 1;
magma_int_t nb = 64;
magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
magmaDoubleComplex alpha = MAGMA_Z_MAKE( 1.5, -2.3 );
magmaDoubleComplex beta = MAGMA_Z_MAKE( -0.6, 0.8 );
magmaDoubleComplex *A, *X, *Y, *Ymagma;
magmaDoubleComplex *dA, *dX, *dY, *dwork;
magma_int_t status = 0;
magma_opts opts;
parse_opts( argc, argv, &opts );
double tol = opts.tolerance * lapackf77_dlamch("E");
printf("uplo = %s\n", lapack_uplo_const(opts.uplo) );
printf(" N MAGMA Gflop/s (ms) CPU Gflop/s (ms) MAGMA error\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;
ldda = ((N + 31)/32)*32;
sizeA = N*lda;
sizeX = N*incx;
sizeY = N*incy;
gflops = FLOPS_ZSYMV( N ) / 1e9;
TESTING_MALLOC_CPU( A, magmaDoubleComplex, sizeA );
TESTING_MALLOC_CPU( X, magmaDoubleComplex, sizeX );
TESTING_MALLOC_CPU( Y, magmaDoubleComplex, sizeY );
TESTING_MALLOC_CPU( Ymagma, magmaDoubleComplex, sizeY );
TESTING_MALLOC_DEV( dA, magmaDoubleComplex, ldda*N );
TESTING_MALLOC_DEV( dX, magmaDoubleComplex, sizeX );
TESTING_MALLOC_DEV( dY, magmaDoubleComplex, sizeY );
blocks = (N + nb - 1) / nb;
ldwork = ldda*blocks;
TESTING_MALLOC_DEV( dwork, magmaDoubleComplex, ldwork );
magmablas_zlaset( MagmaFull, ldwork, 1, MAGMA_Z_NAN, MAGMA_Z_NAN, dwork, ldwork );
magmablas_zlaset( MagmaFull, ldda, N, MAGMA_Z_NAN, MAGMA_Z_NAN, dA, ldda );
/* Initialize the matrix */
lapackf77_zlarnv( &ione, ISEED, &sizeA, A );
magma_zmake_hermitian( N, A, lda );
lapackf77_zlarnv( &ione, ISEED, &sizeX, X );
lapackf77_zlarnv( &ione, ISEED, &sizeY, Y );
/* Note: CUBLAS does not implement zsymv */
/* =====================================================================
Performs operation using MAGMABLAS
=================================================================== */
magma_zsetmatrix( N, N, A, lda, dA, ldda );
magma_zsetvector( N, X, incx, dX, incx );
magma_zsetvector( N, Y, incy, dY, incy );
//magma_zprint_gpu( ldda, blocks, dwork, ldda );
magma_time = magma_sync_wtime( 0 );
magmablas_zsymv_work( opts.uplo, N, alpha, dA, ldda, dX, incx, beta, dY, incy, dwork, ldwork );
// TODO provide option to test non-work interface
//magmablas_zsymv( opts.uplo, N, alpha, dA, ldda, dX, incx, beta, dY, incy );
magma_time = magma_sync_wtime( 0 ) - magma_time;
magma_perf = gflops / magma_time;
magma_zgetvector( N, dY, incy, Ymagma, incy );
//magma_zprint_gpu( ldda, blocks, dwork, ldda );
/* =====================================================================
Performs operation using CPU BLAS
=================================================================== */
cpu_time = magma_wtime();
lapackf77_zsymv( lapack_uplo_const(opts.uplo), &N, &alpha, A, &lda, X, &incx, &beta, Y, &incy );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
/* =====================================================================
Check the result
=================================================================== */
blasf77_zaxpy( &N, &c_neg_one, Y, &incy, Ymagma, &incy );
magma_error = lapackf77_zlange( "M", &N, &ione, Ymagma, &N, work ) / N;
printf("%5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %s\n",
(int) N,
magma_perf, 1000.*magma_time,
cpu_perf, 1000.*cpu_time,
magma_error, (magma_error < tol ? "ok" : "failed"));
status += ! (magma_error < tol);
TESTING_FREE_CPU( A );
TESTING_FREE_CPU( X );
TESTING_FREE_CPU( Y );
TESTING_FREE_CPU( Ymagma );
TESTING_FREE_DEV( dA );
TESTING_FREE_DEV( dX );
TESTING_FREE_DEV( dY );
TESTING_FREE_DEV( dwork );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing zhetrd_he2hb
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gflops, gpu_time, gpu_perf;
magmaDoubleComplex *h_A, *h_R, *h_work;
magmaDoubleComplex *tau;
double *D, *E;
magma_int_t N, n2, lda, ldda, lwork, ldt, info, nstream;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t status = 0;
// TODO add these options to parse_opts
magma_int_t NE = 0;
magma_int_t distblk = 0;
magma_opts opts;
opts.parse_opts( argc, argv );
magma_int_t WANTZ = (opts.jobz == MagmaVec);
double tol = opts.tolerance * lapackf77_dlamch("E");
if (opts.nb == 0)
opts.nb = 64; //magma_get_zhetrd_he2hb_nb(N);
if (NE < 1)
NE = N; //64; //magma_get_zhetrd_he2hb_nb(N);
nstream = max(3, opts.ngpu+2);
magma_queue_t streams[MagmaMaxGPUs][20];
magmaDoubleComplex_ptr da[MagmaMaxGPUs], dT1[MagmaMaxGPUs];
if ((distblk == 0) || (distblk < opts.nb))
distblk = max(256, opts.nb);
printf("%% ngpu %d, distblk %d, NB %d, nstream %d\n",
(int) opts.ngpu, (int) distblk, (int) opts.nb, (int) nstream);
for( magma_int_t dev = 0; dev < opts.ngpu; ++dev ) {
magma_setdevice( dev );
for( int i = 0; i < nstream; ++i ) {
magma_queue_create( &streams[dev][i] );
}
}
magma_setdevice( 0 );
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
N = opts.nsize[itest];
lda = N;
ldt = N;
ldda = magma_roundup( N, opts.align ); // multiple of 32 by default
n2 = lda*N;
/* We suppose the magma NB is bigger than lapack NB */
lwork = N*opts.nb;
//gflops = ....?
/* Allocate host memory for the matrix */
TESTING_MALLOC_CPU( tau, magmaDoubleComplex, N-1 );
TESTING_MALLOC_PIN( h_A, magmaDoubleComplex, lda*N );
TESTING_MALLOC_PIN( h_R, magmaDoubleComplex, lda*N );
TESTING_MALLOC_PIN( h_work, magmaDoubleComplex, lwork );
TESTING_MALLOC_PIN( D, double, N );
TESTING_MALLOC_PIN( E, double, N );
for( magma_int_t dev = 0; dev < opts.ngpu; ++dev ) {
magma_int_t mlocal = ((N / distblk) / opts.ngpu + 1) * distblk;
magma_setdevice( dev );
TESTING_MALLOC_DEV( da[dev], magmaDoubleComplex, ldda*mlocal );
TESTING_MALLOC_DEV( dT1[dev], magmaDoubleComplex, N*opts.nb );
}
/* ====================================================================
Initialize the matrix
=================================================================== */
lapackf77_zlarnv( &ione, ISEED, &n2, h_A );
magma_zmake_hermitian( N, h_A, lda );
lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_A, &lda, h_R, &lda );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
/* Copy the matrix to the GPU */
magma_zsetmatrix_1D_col_bcyclic( N, N, h_R, lda, da, ldda, opts.ngpu, distblk);
//magmaDoubleComplex_ptr dabis;
//TESTING_MALLOC_DEV( dabis, magmaDoubleComplex, ldda*N );
//magma_zsetmatrix(N, N, h_R, lda, dabis, ldda);
for (int count=0; count < 1; ++count) {
magma_setdevice(0);
gpu_time = magma_wtime();
if (opts.version == 30) {
// see src/obsolete and magmablas/obsolete
printf( "magma_zhetrd_he2hb_mgpu_spec not compiled\n" );
//magma_zhetrd_he2hb_mgpu_spec(
// opts.uplo, N, opts.nb, h_R, lda, tau, h_work, lwork,
// da, ldda, dT1, opts.nb, opts.ngpu, distblk,
// streams, nstream, opts.nthread, &info);
} else {
nstream = 3;
magma_zhetrd_he2hb_mgpu(
opts.uplo, N, opts.nb, h_R, lda, tau, h_work, lwork,
da, ldda, dT1, opts.nb, opts.ngpu, distblk,
streams, nstream, opts.nthread, &info);
}
// magma_zhetrd_he2hb(opts.uplo, N, opts.nb, h_R, lda, tau, h_work, lwork, dT1[0], &info);
gpu_time = magma_wtime() - gpu_time;
printf(" Finish BAND N %d NB %d dist %d ngpu %d version %d timing= %f\n",
N, opts.nb, distblk, opts.ngpu, opts.version, gpu_time);
}
magma_setdevice(0);
for( magma_int_t dev = 0; dev < opts.ngpu; ++dev ) {
magma_setdevice(dev);
magma_device_sync();
}
magma_setdevice(0);
magmablasSetKernelStream( NULL );
// todo neither of these is declared in headers
// magma_zhetrd_bhe2trc_v5(opts.nthread, WANTZ, opts.uplo, NE, N, opts.nb, h_R, lda, D, E, dT1[0], ldt);
// magma_zhetrd_bhe2trc(opts.nthread, WANTZ, opts.uplo, NE, N, opts.nb, h_R, lda, D, E, dT1[0], ldt);
// todo where is this timer started?
// gpu_time = magma_wtime() - gpu_time;
// todo what are the gflops?
gpu_perf = gflops / gpu_time;
if (info != 0)
printf("magma_zhetrd_he2hb returned error %d: %s.\n",
(int) info, magma_strerror( info ));
/* =====================================================================
Print performance and error.
=================================================================== */
#if defined(CHECKEIG)
#if defined(PRECISION_z) || defined(PRECISION_d)
if ( opts.check ) {
printf(" Total N %5d flops %6.2f timing %6.2f seconds\n", (int) N, gpu_perf, gpu_time );
double nrmI=0.0, nrm1=0.0, nrm2=0.0;
int lwork2 = 256*N;
magmaDoubleComplex *work2, *AINIT;
double *rwork2, *D2;
// TODO free this memory !
magma_zmalloc_cpu( &work2, lwork2 );
magma_dmalloc_cpu( &rwork2, N );
magma_dmalloc_cpu( &D2, N );
magma_zmalloc_cpu( &AINIT, N*lda );
memcpy(AINIT, h_A, N*lda*sizeof(magmaDoubleComplex));
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
cpu_time = magma_wtime();
int nt = min(12, opts.nthread);
magma_set_lapack_numthreads(nt);
lapackf77_zheev( "N", "L", &N, h_A, &lda, D2, work2, &lwork2,
#ifdef COMPLEX
rwork2,
#endif
&info );
///* call eigensolver for our resulting tridiag [D E] and for Q */
//dstedc_withZ('V', N, D, E, h_R, lda);
////dsterf_( &N, D, E, &info);
cpu_time = magma_wtime() - cpu_time;
printf(" Finish CHECK - EIGEN timing= %f threads %d\n", cpu_time, nt);
/* compare result */
cmp_vals(N, D2, D, &nrmI, &nrm1, &nrm2);
magmaDoubleComplex *WORKAJETER;
double *RWORKAJETER, *RESU;
// TODO free this memory !
magma_zmalloc_cpu( &WORKAJETER, (2* N * N + N) );
magma_dmalloc_cpu( &RWORKAJETER, N );
magma_dmalloc_cpu( &RESU, 10 );
int MATYPE;
memset(RESU, 0, 10*sizeof(double));
MATYPE=3;
double NOTHING=0.0;
cpu_time = magma_wtime();
// check results
zcheck_eig_( lapack_vec_const(opts.jobz), &MATYPE, &N, &opts.nb,
AINIT, &lda, &NOTHING, &NOTHING, D2, D,
h_R, &lda, WORKAJETER, RWORKAJETER, RESU );
cpu_time = magma_wtime() - cpu_time;
printf(" Finish CHECK - results timing= %f\n", cpu_time);
magma_set_lapack_numthreads(1);
printf("\n");
printf(" ================================================================================================================\n");
printf(" ==> INFO voici threads=%d N=%d NB=%d WANTZ=%d\n", (int) opts.nthread, (int) N, (int) opts.nb, (int) WANTZ);
printf(" ================================================================================================================\n");
printf(" DSBTRD : %15s \n", "STATblgv9withQ ");
printf(" ================================================================================================================\n");
if (WANTZ > 0)
printf(" | A - U S U' | / ( |A| n ulp ) : %15.3E \n", RESU[0]);
if (WANTZ > 0)
printf(" | I - U U' | / ( n ulp ) : %15.3E \n", RESU[1]);
printf(" | D1 - EVEIGS | / (|D| ulp) : %15.3E \n", RESU[2]);
printf(" max | D1 - EVEIGS | : %15.3E \n", RESU[6]);
printf(" ================================================================================================================\n\n\n");
printf(" ****************************************************************************************************************\n");
printf(" * Hello here are the norm Infinite (max)=%8.2e norm one (sum)=%8.2e norm2(sqrt)=%8.2e *\n", nrmI, nrm1, nrm2);
printf(" ****************************************************************************************************************\n\n");
}
#endif // PRECISION_z || PRECISION_d
#endif // CHECKEIG
printf(" Total N %5d flops %6.2f timing %6.2f seconds\n", (int) N, 0.0, gpu_time );
printf("%%===========================================================================\n\n\n");
TESTING_FREE_CPU( tau );
TESTING_FREE_PIN( h_A );
TESTING_FREE_PIN( h_R );
TESTING_FREE_PIN( h_work );
TESTING_FREE_PIN( D );
TESTING_FREE_PIN( E );
for( magma_int_t dev = 0; dev < opts.ngpu; ++dev ) {
magma_setdevice( dev );
TESTING_FREE_DEV( da[dev] );
TESTING_FREE_DEV( dT1[dev] );
}
magma_setdevice( 0 );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
for( magma_int_t dev = 0; dev < opts.ngpu; ++dev ) {
for( int i = 0; i < nstream; ++i ) {
magma_queue_destroy( streams[dev][i] );
}
}
TESTING_FINALIZE();
return status;
}
int main(int argc, char **argv)
{
TESTING_INIT();
const magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
const magma_int_t ione = 1;
real_Double_t atomics_perf=0, atomics_time=0;
real_Double_t gflops, magma_perf=0, magma_time=0, cublas_perf, cublas_time, cpu_perf, cpu_time;
double magma_error=0, atomics_error=0, cublas_error, work[1];
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t N, lda, ldda, sizeA, sizeX, sizeY, blocks, ldwork;
magma_int_t incx = 1;
magma_int_t incy = 1;
magma_int_t nb = 64;
magmaDoubleComplex alpha = MAGMA_Z_MAKE( 1.5, -2.3 );
magmaDoubleComplex beta = MAGMA_Z_MAKE( -0.6, 0.8 );
magmaDoubleComplex *A, *X, *Y, *Yatomics, *Ycublas, *Ymagma;
magmaDoubleComplex_ptr dA, dX, dY, dwork;
magma_int_t status = 0;
magma_opts opts;
opts.parse_opts( argc, argv );
double tol = opts.tolerance * lapackf77_dlamch("E");
printf("%% uplo = %s\n", lapack_uplo_const(opts.uplo) );
printf("%% N MAGMA Gflop/s (ms) Atomics Gflop/s CUBLAS Gflop/s CPU Gflop/s MAGMA error Atomics CUBLAS\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;
ldda = magma_roundup( N, opts.align ); // multiple of 32 by default
sizeA = N*lda;
sizeX = N*incx;
sizeY = N*incy;
gflops = FLOPS_ZHEMV( N ) / 1e9;
TESTING_MALLOC_CPU( A, magmaDoubleComplex, sizeA );
TESTING_MALLOC_CPU( X, magmaDoubleComplex, sizeX );
TESTING_MALLOC_CPU( Y, magmaDoubleComplex, sizeY );
TESTING_MALLOC_CPU( Yatomics, magmaDoubleComplex, sizeY );
TESTING_MALLOC_CPU( Ycublas, magmaDoubleComplex, sizeY );
TESTING_MALLOC_CPU( Ymagma, magmaDoubleComplex, sizeY );
TESTING_MALLOC_DEV( dA, magmaDoubleComplex, ldda*N );
TESTING_MALLOC_DEV( dX, magmaDoubleComplex, sizeX );
TESTING_MALLOC_DEV( dY, magmaDoubleComplex, sizeY );
blocks = magma_ceildiv( N, nb );
ldwork = ldda*blocks;
TESTING_MALLOC_DEV( dwork, magmaDoubleComplex, ldwork );
magmablas_zlaset( MagmaFull, ldwork, 1, MAGMA_Z_NAN, MAGMA_Z_NAN, dwork, ldwork );
magmablas_zlaset( MagmaFull, ldda, N, MAGMA_Z_NAN, MAGMA_Z_NAN, dA, ldda );
/* Initialize the matrix */
lapackf77_zlarnv( &ione, ISEED, &sizeA, A );
magma_zmake_hermitian( N, A, lda );
// should not use data from the opposite triangle -- fill with NAN to check
magma_int_t N1 = N-1;
if ( opts.uplo == MagmaUpper ) {
lapackf77_zlaset( "Lower", &N1, &N1, &MAGMA_Z_NAN, &MAGMA_Z_NAN, &A[1], &lda );
}
else {
lapackf77_zlaset( "Upper", &N1, &N1, &MAGMA_Z_NAN, &MAGMA_Z_NAN, &A[lda], &lda );
}
lapackf77_zlarnv( &ione, ISEED, &sizeX, X );
lapackf77_zlarnv( &ione, ISEED, &sizeY, Y );
/* =====================================================================
Performs operation using CUBLAS
=================================================================== */
magma_zsetmatrix( N, N, A, lda, dA, ldda );
magma_zsetvector( N, X, incx, dX, incx );
magma_zsetvector( N, Y, incy, dY, incy );
magmablasSetKernelStream( opts.queue ); // opts.handle also uses opts.queue
cublas_time = magma_sync_wtime( opts.queue );
#ifdef HAVE_CUBLAS
cublasZhemv( opts.handle, cublas_uplo_const(opts.uplo),
N, &alpha, dA, ldda, dX, incx, &beta, dY, incy );
#else
magma_zhemv( opts.uplo, N, alpha, dA, 0, ldda, dX, 0, incx, beta, dY, 0, incy, opts.queue );
#endif
cublas_time = magma_sync_wtime( opts.queue ) - cublas_time;
cublas_perf = gflops / cublas_time;
magma_zgetvector( N, dY, incy, Ycublas, incy );
/* =====================================================================
Performs operation using CUBLAS - using atomics
=================================================================== */
#ifdef HAVE_CUBLAS
cublasSetAtomicsMode( opts.handle, CUBLAS_ATOMICS_ALLOWED );
magma_zsetvector( N, Y, incy, dY, incy );
// sync on queue doesn't work -- need device sync or use NULL stream -- bug in CUBLAS?
atomics_time = magma_sync_wtime( NULL /*opts.queue*/ );
cublasZhemv( opts.handle, cublas_uplo_const(opts.uplo),
N, &alpha, dA, ldda, dX, incx, &beta, dY, incy );
atomics_time = magma_sync_wtime( NULL /*opts.queue*/ ) - atomics_time;
atomics_perf = gflops / atomics_time;
magma_zgetvector( N, dY, incy, Yatomics, incy );
cublasSetAtomicsMode( opts.handle, CUBLAS_ATOMICS_NOT_ALLOWED );
#endif
/* =====================================================================
Performs operation using MAGMABLAS
=================================================================== */
#ifdef HAVE_CUBLAS
magma_zsetvector( N, Y, incy, dY, incy );
magma_time = magma_sync_wtime( opts.queue );
if ( opts.version == 1 ) {
magmablas_zhemv_work( opts.uplo, N, alpha, dA, ldda, dX, incx, beta, dY, incy, dwork, ldwork, opts.queue );
}
else {
// non-work interface (has added overhead)
magmablas_zhemv( opts.uplo, N, alpha, dA, ldda, dX, incx, beta, dY, incy );
}
magma_time = magma_sync_wtime( opts.queue ) - magma_time;
magma_perf = gflops / magma_time;
magma_zgetvector( N, dY, incy, Ymagma, incy );
#endif
/* =====================================================================
Performs operation using CPU BLAS
=================================================================== */
cpu_time = magma_wtime();
blasf77_zhemv( lapack_uplo_const(opts.uplo), &N, &alpha, A, &lda, X, &incx, &beta, Y, &incy );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
/* =====================================================================
Check the result
=================================================================== */
blasf77_zaxpy( &N, &c_neg_one, Y, &incy, Ycublas, &incy );
cublas_error = lapackf77_zlange( "M", &N, &ione, Ycublas, &N, work ) / N;
#ifdef HAVE_CUBLAS
blasf77_zaxpy( &N, &c_neg_one, Y, &incy, Yatomics, &incy );
atomics_error = lapackf77_zlange( "M", &N, &ione, Yatomics, &N, work ) / N;
blasf77_zaxpy( &N, &c_neg_one, Y, &incy, Ymagma, &incy );
magma_error = lapackf77_zlange( "M", &N, &ione, Ymagma, &N, work ) / N;
#endif
bool okay = (magma_error < tol && cublas_error < tol && atomics_error < tol);
status += ! okay;
printf("%5d %7.2f (%7.2f) %7.2f (%7.2f) %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %8.2e %8.2e %s\n",
(int) N,
magma_perf, 1000.*magma_time,
atomics_perf, 1000.*atomics_time,
cublas_perf, 1000.*cublas_time,
cpu_perf, 1000.*cpu_time,
magma_error, cublas_error, atomics_error,
(okay ? "ok" : "failed"));
TESTING_FREE_CPU( A );
TESTING_FREE_CPU( X );
TESTING_FREE_CPU( Y );
TESTING_FREE_CPU( Ycublas );
TESTING_FREE_CPU( Yatomics );
TESTING_FREE_CPU( Ymagma );
TESTING_FREE_DEV( dA );
TESTING_FREE_DEV( dX );
TESTING_FREE_DEV( dY );
TESTING_FREE_DEV( dwork );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
opts.cleanup();
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing zhetrd_he2hb
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gpu_time, gpu_perf, gflops;
magmaDoubleComplex *h_A, *h_R, *h_work, *dT1;
magmaDoubleComplex *tau;
double *D, *E;
/* Matrix size */
magma_int_t N, n2, lda, lwork, lwork0; //ldt
magma_int_t info;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
#if defined(CHECKEIG)
#if defined(PRECISION_z) || defined(PRECISION_d)
magma_int_t WANTZ=0;
magma_int_t THREADS=1;
#endif
#endif
magma_int_t NE = 0;
magma_int_t NB = 0;
magma_int_t ngpu = 1;
magma_opts opts;
opts.parse_opts( argc, argv );
NB = opts.nb;
if (NB < 1)
NB = 64; //64; //magma_get_zhetrd_he2hb_nb(N);
// what is NE ?
if (NE < 1)
NE = 64; //N; //magma_get_zhetrd_he2hb_nb(N); // N not yet initialized
printf("%% N GPU GFlop/s \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;
//ldt = N;
n2 = lda*N;
gflops = FLOPS_ZHETRD( N ) / 1e9;
/* We suppose the magma NB is bigger than lapack NB */
lwork0 = N*NB;
/* Allocate host memory for the matrix */
TESTING_MALLOC_CPU( h_A, magmaDoubleComplex, lda*N );
TESTING_MALLOC_CPU( tau, magmaDoubleComplex, N-1 );
TESTING_MALLOC_PIN( h_R, magmaDoubleComplex, lda*N );
TESTING_MALLOC_PIN( h_work, magmaDoubleComplex, lwork0 );
TESTING_MALLOC_PIN( D, double, N );
TESTING_MALLOC_PIN( E, double, N );
//TESTING_MALLOC_DEV( dT1, magmaDoubleComplex, (2*min(N,N) + roundup( N, 32 ))*NB );
TESTING_MALLOC_DEV( dT1, magmaDoubleComplex, (N*NB) );
// if (WANTZ) gflops = 2.0*gflops;
/* ====================================================================
Initialize the matrix
=================================================================== */
lapackf77_zlarnv( &ione, ISEED, &n2, h_A );
magma_zmake_hermitian( N, h_A, lda );
lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_A, &lda, h_R, &lda );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
magma_device_t cdev;
magma_getdevice( &cdev );
gpu_time = magma_wtime();
/*
magma_zhetrd_he2hb( opts.uplo, N, NB, h_R, lda, tau, h_work, lwork0, dT1, THREADS, &info);
tband = magma_wtime - gpu_time();
printf(" Finish BAND N %d NB %d ngpu %d timing= %f\n", N, NB, ngpu, tband);
magma_zhetrd_bhe2trc_v5(THREADS, WANTZ, opts.uplo, NE, N, NB, h_R, lda, D, E, dT1, ldt);
*/
/*
magma_zhetrd_he2hb( opts.uplo, N, NB, h_R, lda, tau, h_work, lwork, dT1, THREADS, &info);
tband = magma_wtime - gpu_time();
printf(" Finish BAND N %d NB %d ngpu %d timing= %f\n", N, NB, ngpu, tband);
magma_zhetrd_bhe2trc(THREADS, WANTZ, opts.uplo, NE, N, NB, h_R, lda, D, E, dT1, ldt);
*/
magma_range_t range = MagmaRangeAll;
magma_int_t m1 = 0;
double vl = 0;
double vu = 0;
magma_int_t il = 0;
magma_int_t iu = 0;
if (opts.fraction == 0) {
il = max( 1, magma_int_t(0.1*N) );
iu = max( 1, magma_int_t(0.3*N) );
}
else {
il = 1;
iu = max( 1, magma_int_t(opts.fraction*N) );
}
magmaDoubleComplex *hh_work;
magma_int_t *iwork;
magma_int_t /*nb,*/ /*lwork,*/ liwork;
magma_int_t threads = magma_get_parallel_numthreads();
#ifdef COMPLEX
double *rwork;
magma_int_t lrwork;
#endif
magma_zheevdx_getworksize(N, threads, (opts.jobz == MagmaVec),
&lwork,
#ifdef COMPLEX
&lrwork,
#endif
&liwork);
TESTING_MALLOC_PIN( hh_work, magmaDoubleComplex, lwork );
TESTING_MALLOC_CPU( iwork, magma_int_t, liwork );
#ifdef COMPLEX
TESTING_MALLOC_PIN( rwork, double, lrwork );
#endif
if (ngpu == 1) {
printf("calling zheevdx_2stage 1 GPU\n");
magma_zheevdx_2stage( opts.jobz, range, opts.uplo, N,
h_R, lda,
vl, vu, il, iu,
&m1, D,
hh_work, lwork,
#ifdef COMPLEX
rwork, lrwork,
#endif
iwork, liwork,
&info);
} else {
printf("calling zheevdx_2stage_m %d GPU\n", (int) ngpu);
magma_zheevdx_2stage_m(ngpu, opts.jobz, range, opts.uplo, N,
h_R, lda,
vl, vu, il, iu,
&m1, D,
hh_work, lwork,
#ifdef COMPLEX
rwork, lrwork,
#endif
iwork, liwork,
&info);
}
magma_setdevice( cdev );
gpu_time = magma_wtime() - gpu_time;
gpu_perf = gflops / gpu_time;
/* =====================================================================
Check the factorization
=================================================================== */
/*
if ( opts.check ) {
FILE *fp;
printf("Writing input matrix in matlab_i_mat.txt ...\n");
fp = fopen ("matlab_i_mat.txt", "w");
if ( fp == NULL ) {
printf("Couldn't open output file\n");
return -1;
}
for (j=0; j < N; j++) {
for (k=0; k < N; k++) {
#ifdef COMPLEX
fprintf(fp, "%5d %5d %11.8f %11.8f\n", k+1, j+1,
h_A[k+j*lda].x, h_A[k+j*lda].y);
#else
fprintf(fp, "%5d %5d %11.8f\n", k+1, j+1, h_A[k+j*lda]);
#endif
}
}
fclose( fp );
printf("Writing output matrix in matlab_o_mat.txt ...\n");
fp = fopen ("matlab_o_mat.txt", "w");
if ( fp == NULL ) {
printf("Couldn't open output file\n");
return -1;
}
for (j=0; j < N; j++) {
for (k=0; k < N; k++) {
#ifdef COMPLEX
fprintf(fp, "%5d %5d %11.8f %11.8f\n", k+1, j+1,
h_R[k+j*lda].x, h_R[k+j*lda].y);
#else
fprintf(fp, "%5d %5d %11.8f\n", k+1, j+1, h_R[k+j*lda]);
#endif
}
}
fclose( fp );
}
*/
/* =====================================================================
Print performance and error.
=================================================================== */
#if defined(CHECKEIG)
#if defined(PRECISION_z) || defined(PRECISION_d)
if ( opts.check ) {
printf(" Total N %5d gflops %6.2f timing %6.2f seconds\n", (int) N, gpu_perf, gpu_time );
const char* JOBZ;
if (WANTZ == 0)
JOBZ = MagmaNoVecStr;
else
JOBZ = MagmaVecStr;
double nrmI=0.0, nrm1=0.0, nrm2=0.0;
int lwork2 = 256*N;
magmaDoubleComplex *work2, *AINIT;
double *rwork2, *D2;
// TODO free this memory !
magma_zmalloc_cpu( &work2, lwork2 );
magma_dmalloc_cpu( &rwork2, N );
magma_dmalloc_cpu( &D2, N );
magma_zmalloc_cpu( &AINIT, N*lda );
memcpy(AINIT, h_A, N*lda*sizeof(magmaDoubleComplex));
/* compute the eigenvalues using lapack routine to be able to compare to it and used as ref */
cpu_time = magma_wtime();
i= min(12, THREADS);
magma_set_lapack_numthreads( i );
lapackf77_zheev( "N", "L", &N, h_A, &lda, D2, work2, &lwork2,
#ifdef COMPLEX
rwork2,
#endif
&info );
///* call eigensolver for our resulting tridiag [D E] and for Q */
//dstedc_withZ('V', N, D, E, h_R, lda);
////dsterf_( &N, D, E, &info);
////
cpu_time = magma_wtime() - cpu_time;
printf(" Finish CHECK - EIGEN timing= %f threads %d\n", cpu_time, i);
/*
for (i=0; i < 10; i++)
printf(" voici lpk D[%d] %8.2e\n", i, D2[i]);
*/
//magmaDoubleComplex mydz=0.0, mydo=1.0;
//magmaDoubleComplex *Z;
// magma_zmalloc_cpu( &Z, N*lda );
// dgemm_("N", "N", &N, &N, &N, &mydo, h_R, &lda, h_A, &lda, &mydz, Z, &lda);
/* compare result */
cmp_vals(N, D2, D, &nrmI, &nrm1, &nrm2);
magmaDoubleComplex *WORKAJETER;
double *RWORKAJETER, *RESU;
// TODO free this memory !
magma_zmalloc_cpu( &WORKAJETER, (2* N * N + N) );
magma_dmalloc_cpu( &RWORKAJETER, N );
magma_dmalloc_cpu( &RESU, 10 );
int MATYPE;
memset(RESU, 0, 10*sizeof(double));
MATYPE=3;
double NOTHING=0.0;
cpu_time = magma_wtime();
// check results
zcheck_eig_( JOBZ, &MATYPE, &N, &NB, AINIT, &lda, &NOTHING, &NOTHING, D2, D, h_R, &lda, WORKAJETER, RWORKAJETER, RESU );
cpu_time = magma_wtime() - cpu_time;
printf(" Finish CHECK - results timing= %f\n", cpu_time);
magma_set_lapack_numthreads( 1 );
printf("\n");
printf(" ================================================================================================================\n");
printf(" ==> INFO voici threads=%d N=%d NB=%d WANTZ=%d\n", (int) THREADS, (int) N, (int) NB, (int) WANTZ);
printf(" ================================================================================================================\n");
printf(" DSBTRD : %15s \n", "STATblgv9withQ ");
printf(" ================================================================================================================\n");
if (WANTZ > 0)
printf(" | A - U S U' | / ( |A| n ulp ) : %15.3E \n", RESU[0]);
if (WANTZ > 0)
printf(" | I - U U' | / ( n ulp ) : %15.3E \n", RESU[1]);
printf(" | D1 - EVEIGS | / (|D| ulp) : %15.3E \n", RESU[2]);
printf(" max | D1 - EVEIGS | : %15.3E \n", RESU[6]);
printf(" ================================================================================================================\n\n\n");
printf(" ****************************************************************************************************************\n");
printf(" * Hello here are the norm Infinite (max)=%8.2e norm one (sum)=%8.2e norm2(sqrt)=%8.2e *\n", nrmI, nrm1, nrm2);
printf(" ****************************************************************************************************************\n\n");
}
#endif
#endif
printf(" Total N %5d gflops %6.2f timing %6.2f seconds\n", (int) N, gpu_perf, gpu_time );
printf("%%===========================================================================\n\n\n");
/* Memory clean up */
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( tau );
TESTING_FREE_PIN( h_R );
TESTING_FREE_PIN( h_work );
TESTING_FREE_PIN( D );
TESTING_FREE_PIN( E );
TESTING_FREE_DEV( dT1 );
/* TODO - not all memory has been freed inside loop */
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return EXIT_SUCCESS;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing zheevd_gpu
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gpu_time, cpu_time;
magmaDoubleComplex *h_A, *h_R, *d_R, *h_work, aux_work[1];
#ifdef COMPLEX
double *rwork, aux_rwork[1];
magma_int_t lrwork;
#endif
double *w1, *w2, result[4]={0, 0, 0, 0}, eps;
magma_int_t *iwork, aux_iwork[1];
magma_int_t N, n2, info, lwork, liwork, lda, ldda;
magma_int_t izero = 0;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
eps = lapackf77_dlamch( "E" );
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");
// checking NoVec requires LAPACK
opts.lapack |= (opts.check && opts.jobz == MagmaNoVec);
printf("using: jobz = %s, uplo = %s\n",
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];
n2 = N*N;
lda = N;
ldda = roundup( N, opts.roundup ); // by default, round to multiple of 32
// query for workspace sizes
magma_zheevd_gpu( opts.jobz, opts.uplo,
N, NULL, ldda, NULL,
NULL, lda,
aux_work, -1,
#ifdef COMPLEX
aux_rwork, -1,
#endif
aux_iwork, -1,
&info );
lwork = (magma_int_t) MAGMA_Z_REAL( aux_work[0] );
#ifdef COMPLEX
lrwork = (magma_int_t) aux_rwork[0];
#endif
liwork = aux_iwork[0];
/* Allocate host memory for the matrix */
TESTING_MALLOC_CPU( h_A, magmaDoubleComplex, N*lda );
TESTING_MALLOC_CPU( w1, double, N );
TESTING_MALLOC_CPU( w2, double, N );
#ifdef COMPLEX
TESTING_MALLOC_CPU( rwork, double, lrwork );
#endif
TESTING_MALLOC_CPU( iwork, magma_int_t, liwork );
TESTING_MALLOC_PIN( h_R, magmaDoubleComplex, N*lda );
TESTING_MALLOC_PIN( h_work, magmaDoubleComplex, lwork );
TESTING_MALLOC_DEV( d_R, magmaDoubleComplex, N*ldda );
/* Initialize the matrix */
lapackf77_zlarnv( &ione, ISEED, &n2, h_A );
magma_zmake_hermitian( N, h_A, N );
magma_zsetmatrix( N, N, h_A, lda, d_R, ldda );
/* warm up run */
if ( opts.warmup ) {
magma_zheevd_gpu( opts.jobz, opts.uplo,
N, d_R, ldda, w1,
h_R, lda,
h_work, lwork,
#ifdef COMPLEX
rwork, lrwork,
#endif
iwork, liwork,
&info );
if (info != 0)
printf("magma_zheevd_gpu returned error %d: %s.\n",
(int) info, magma_strerror( info ));
magma_zsetmatrix( N, N, h_A, lda, d_R, ldda );
}
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
gpu_time = magma_wtime();
magma_zheevd_gpu( opts.jobz, opts.uplo,
N, d_R, ldda, w1,
h_R, lda,
h_work, lwork,
#ifdef COMPLEX
rwork, lrwork,
#endif
iwork, liwork,
&info );
gpu_time = magma_wtime() - gpu_time;
if (info != 0)
printf("magma_zheevd_gpu returned error %d: %s.\n",
(int) info, magma_strerror( info ));
if ( opts.check && opts.jobz != MagmaNoVec ) {
/* =====================================================================
Check the results following the LAPACK's [zcds]drvst routine.
A is factored as A = U S U' and the following 3 tests computed:
(1) | A - U S U' | / ( |A| N )
(2) | I - U'U | / ( N )
(3) | S(with U) - S(w/o U) | / | S |
=================================================================== */
magma_zgetmatrix( N, N, d_R, ldda, h_R, lda );
magmaDoubleComplex *work;
TESTING_MALLOC_CPU( work, magmaDoubleComplex, 2*N*N );
// e=NULL is unused since kband=0; tau=NULL is unused since itype=1
lapackf77_zhet21( &ione, lapack_uplo_const(opts.uplo), &N, &izero,
h_A, &lda,
w1, NULL,
h_R, &lda,
h_R, &lda,
NULL, work,
#ifdef COMPLEX
rwork,
#endif
&result[0] );
result[0] *= eps;
result[1] *= eps;
TESTING_FREE_CPU( work ); work=NULL;
// Disable eigenvalue check which calls routine again --
// it obscures whether error occurs in first call above or in this call.
// But see comparison to LAPACK below.
//
//magma_zsetmatrix( N, N, h_A, lda, d_R, ldda );
//magma_zheevd_gpu( MagmaNoVec, opts.uplo,
// N, d_R, ldda, w2,
// h_R, lda,
// h_work, lwork,
// #ifdef COMPLEX
// rwork, lrwork,
// #endif
// iwork, liwork,
// &info);
//if (info != 0)
// printf("magma_zheevd_gpu returned error %d: %s.\n",
// (int) info, magma_strerror( info ));
//
//double maxw=0, diff=0;
//for( int j=0; j < N; j++ ) {
// maxw = max(maxw, fabs(w1[j]));
// maxw = max(maxw, fabs(w2[j]));
// diff = max(diff, fabs(w1[j] - w2[j]));
//}
//result[2] = diff / (N*maxw);
}
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
if ( opts.lapack ) {
cpu_time = magma_wtime();
lapackf77_zheevd( lapack_vec_const(opts.jobz), lapack_uplo_const(opts.uplo),
&N, h_A, &lda, w2,
h_work, &lwork,
#ifdef COMPLEX
rwork, &lrwork,
#endif
iwork, &liwork,
&info);
cpu_time = magma_wtime() - cpu_time;
if (info != 0)
printf("lapackf77_zheevd returned error %d: %s.\n",
(int) info, magma_strerror( info ));
// compare eigenvalues
double maxw=0, diff=0;
for( int j=0; j < N; j++ ) {
maxw = max(maxw, fabs(w1[j]));
maxw = max(maxw, fabs(w2[j]));
diff = max(diff, fabs(w1[j] - w2[j]));
}
result[3] = diff / (N*maxw);
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 && opts.jobz != MagmaNoVec ) {
printf("Testing the factorization A = U S U' for correctness:\n");
printf(" | A - U S U' | / (|A| N) = %8.2e %s\n", result[0], (result[0] < tol ? "ok" : "failed") );
printf(" | I - U'U | / N = %8.2e %s\n", result[1], (result[1] < tol ? "ok" : "failed") );
//printf(" | S(w/ U) - S(w/o U) | / |S| = %8.2e %s\n\n", result[2], (result[2] < tolulp ? "ok" : "failed") );
status += ! (result[0] < tol && result[1] < tol); // && result[2] < tolulp)
}
if ( opts.lapack ) {
printf(" | S_magma - S_lapack | / |S| = %8.2e %s\n\n", result[3], (result[3] < tolulp ? "ok" : "failed") );
status += ! (result[3] < tolulp);
}
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( w1 );
TESTING_FREE_CPU( w2 );
#ifdef COMPLEX
TESTING_FREE_CPU( rwork );
#endif
TESTING_FREE_CPU( iwork );
TESTING_FREE_PIN( h_R );
TESTING_FREE_PIN( h_work );
TESTING_FREE_DEV( d_R );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing zhegvdx
*/
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 *w1, *w2, vl=0, vu=0;
double result[2] = {0};
magma_int_t *iwork;
magma_int_t N, n2, info, il, iu, m1, m2, nb, lwork, liwork;
magmaDoubleComplex c_zero = MAGMA_Z_ZERO;
magmaDoubleComplex c_one = MAGMA_Z_ONE;
magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
#if defined(PRECISION_z) || defined(PRECISION_c)
double *rwork;
magma_int_t lrwork;
#endif
//double d_one = 1.;
//double d_ten = 10.;
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, check = %d, fraction = %6.4f\n",
(int) opts.itype, lapack_vec_const(opts.jobz), lapack_uplo_const(opts.uplo),
(int) opts.check, opts.fraction);
printf(" N M 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];
n2 = N*N;
nb = magma_get_zhetrd_nb(N);
#if defined(PRECISION_z) || defined(PRECISION_c)
lwork = 2*N*nb + N*N;
lrwork = 1 + 5*N +2*N*N;
#else
lwork = 1 + 6*N*nb + 2* N*N;
#endif
liwork = 3 + 5*N;
if ( opts.fraction == 0 ) {
il = N / 10;
iu = N / 5+il;
}
else {
il = 1;
iu = (int) (opts.fraction*N);
if (iu < 1) iu = 1;
}
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( 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 );
#if defined(PRECISION_z) || defined(PRECISION_c)
TESTING_MALLOC_PIN( rwork, double, lrwork);
#endif
/* Initialize the matrix */
lapackf77_zlarnv( &ione, ISEED, &n2, h_A );
lapackf77_zlarnv( &ione, ISEED, &n2, h_B );
magma_zmake_hpd( N, h_B, N );
magma_zmake_hermitian( N, h_A, N );
// ==================================================================
// Warmup using MAGMA
// ==================================================================
if(opts.warmup){
lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_A, &N, h_R, &N );
lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_B, &N, h_S, &N );
magma_zhegvdx( opts.itype, opts.jobz, MagmaRangeI, opts.uplo,
N, h_R, N, h_S, N, vl, vu, il, iu, &m1, w1,
h_work, lwork,
#if defined(PRECISION_z) || defined(PRECISION_c)
rwork, lrwork,
#endif
iwork, liwork,
&info );
if (info != 0)
printf("magma_zhegvdx returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_A, &N, h_R, &N );
lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_B, &N, h_S, &N );
gpu_time = magma_wtime();
magma_zhegvdx( opts.itype, opts.jobz, MagmaRangeI, opts.uplo,
N, h_R, N, h_S, N, vl, vu, il, iu, &m1, w1,
h_work, lwork,
#if defined(PRECISION_z) || defined(PRECISION_c)
rwork, lrwork,
#endif
iwork, liwork,
&info );
gpu_time = magma_wtime() - gpu_time;
if (info != 0)
printf("magma_zhegvdx returned error %d: %s.\n",
(int) info, magma_strerror( info ));
if ( opts.check ) {
/* =====================================================================
Check the results following the LAPACK's [zc]hegvdx 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) | S(with V) - S(w/o V) | / | S |
=================================================================== */
#if defined(PRECISION_d) || defined(PRECISION_s)
double *rwork = h_work + N*N;
#endif
double temp1, temp2;
result[0] = 1.;
result[0] /= lapackf77_zlanhe("1", lapack_uplo_const(opts.uplo), &N, h_A, &N, rwork);
result[0] /= lapackf77_zlange("1", &N, &m1, h_R, &N, rwork);
if (opts.itype == 1) {
blasf77_zhemm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_one, h_A, &N, h_R, &N, &c_zero, h_work, &N);
for(int i=0; i < m1; ++i)
blasf77_zdscal(&N, &w1[i], &h_R[i*N], &ione);
blasf77_zhemm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_neg_one, h_B, &N, h_R, &N, &c_one, h_work, &N);
result[0] *= lapackf77_zlange("1", &N, &m1, h_work, &N, rwork)/N;
}
else if (opts.itype == 2) {
blasf77_zhemm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_one, h_B, &N, h_R, &N, &c_zero, h_work, &N);
for(int i=0; i < m1; ++i)
blasf77_zdscal(&N, &w1[i], &h_R[i*N], &ione);
blasf77_zhemm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_one, h_A, &N, h_work, &N, &c_neg_one, h_R, &N);
result[0] *= lapackf77_zlange("1", &N, &m1, h_R, &N, rwork)/N;
}
else if (opts.itype == 3) {
blasf77_zhemm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_one, h_A, &N, h_R, &N, &c_zero, h_work, &N);
for(int i=0; i < m1; ++i)
blasf77_zdscal(&N, &w1[i], &h_R[i*N], &ione);
blasf77_zhemm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_one, h_B, &N, h_work, &N, &c_neg_one, h_R, &N);
result[0] *= lapackf77_zlange("1", &N, &m1, h_R, &N, rwork)/N;
}
lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_A, &N, h_R, &N );
lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_B, &N, h_S, &N );
magma_zhegvdx( opts.itype, MagmaNoVec, MagmaRangeI, opts.uplo,
N, h_R, N, h_S, N, vl, vu, il, iu, &m2, w2,
h_work, lwork,
#if defined(PRECISION_z) || defined(PRECISION_c)
rwork, lrwork,
#endif
iwork, liwork,
&info );
if (info != 0)
printf("magma_zhegvdx returned error %d: %s.\n",
(int) info, magma_strerror( info ));
temp1 = temp2 = 0;
for(int j=0; j < m2; j++) {
temp1 = max(temp1, absv(w1[j]));
temp1 = max(temp1, absv(w2[j]));
temp2 = max(temp2, absv(w1[j]-w2[j]));
}
result[1] = temp2 / (((double)m2)*temp1);
}
/* =====================================================================
Print execution time
=================================================================== */
printf("%5d %5d %7.2f\n",
(int) N, (int) m1, gpu_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"));
}
printf( "(2) | D(w/ Z) - D(w/o Z) | / |D| = %8.2e %s\n\n", result[1], (result[1] < tolulp ? "ok" : "failed"));
status += ! (result[0] < tol && result[1] < tolulp);
}
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_B );
TESTING_FREE_CPU( w1 );
TESTING_FREE_CPU( w2 );
TESTING_FREE_CPU( iwork );
TESTING_FREE_PIN( h_R );
TESTING_FREE_PIN( h_S );
TESTING_FREE_PIN( h_work );
#if defined(PRECISION_z) || defined(PRECISION_c)
TESTING_FREE_PIN( rwork );
#endif
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing zhetrd_he2hb
*/
int main( int argc, char** argv)
{
TESTING_INIT_MGPU();
magma_timestr_t start, end;
double eps, flops, gpu_perf, gpu_time;
magmaDoubleComplex *h_A, *h_R, *h_work;
magmaDoubleComplex *tau;
double *D, *E;
/* Matrix size */
magma_int_t N = 0, n2, lda, lwork,ldt;
magma_int_t size[10] = {1024,2048,3072,4032,5184,6016,7040,8064,9088,10112};
magma_int_t i, info, checkres, once = 0;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
char *uplo = (char *)MagmaLowerStr;
magma_int_t ngpu = magma_num_gpus();
magma_int_t nstream = max(3,ngpu+1);
magma_int_t WANTZ=0;
magma_int_t THREADS=1;
magma_int_t NE = 0;
magma_int_t NB = 0;
magma_int_t distblk =0;
magma_int_t ver =0;
checkres = 0; //getenv("MAGMA_TESTINGS_CHECK") != NULL;
if (argc != 1){
for(i = 1; i<argc; i++){
if (strcmp("-N", argv[i])==0) {
N = atoi(argv[++i]);
once = 1;
}
else if (strcmp("-NB", argv[i])==0) {
NB = atoi(argv[++i]);
}
else if (strcmp("-D", argv[i])==0) {
distblk = atoi(argv[++i]);
}
else if (strcmp("-threads", argv[i])==0) {
THREADS = atoi(argv[++i]);
}
else if (strcmp("-wantz", argv[i])==0) {
WANTZ = atoi(argv[++i]);
}
else if (strcmp("-NE", argv[i])==0) {
NE = atoi(argv[++i]);
}
else if ( strcmp("-c", argv[i]) == 0 ) {
checkres = 1;
}
else if ( strcmp("-v", argv[i]) == 0 && i+1 < argc ) {
ver = atoi( argv[++i] );
}
else if ( strcmp("-nstream", argv[i]) == 0 && i+1 < argc ) {
nstream = atoi( argv[++i] );
magma_assert( nstream > 0 && nstream <= 20,
"error: -nstream %s is invalid; must be > 0 and <= 20.\n", argv[i] );
}
else if ( strcmp("-ngpu", argv[i]) == 0 && i+1 < argc ) {
ngpu = atoi( argv[++i] );
magma_assert( ngpu > 0 || ngpu > MagmaMaxGPUs, "error: -ngpu %s is invalid; must be > 0.\n", argv[i] );
}
else if (strcmp("-U", argv[i])==0)
uplo = (char *)MagmaUpperStr;
else if (strcmp("-L", argv[i])==0)
uplo = (char *)MagmaLowerStr;
}
if ( N > 0 )
printf(" testing_zhetrd_he2hb -L|U -N %d -NB %d -wantz %d -threads %d check %d \n\n", N, NB, WANTZ, THREADS, checkres);
else
{
printf("\nUsage: \n");
printf(" testing_zhetrd_he2hb -L|U -N %d -NB -wantz -threads \n\n", 1024);
exit(1);
}
}
else {
printf("\nUsage: \n");
printf(" testing_zhetrd_he2hb -L|U -N %d\n\n", 1024);
N = size[9];
}
eps = lapackf77_dlamch( "E" );
lda = N;
ldt = N;
n2 = lda * N;
if(NB<1)
NB = 64; //64; //magma_get_zhetrd_he2hb_nb(N);
if(NE<1)
NE = N; //64; //magma_get_zhetrd_he2hb_nb(N);
/* We suppose the magma NB is bigger than lapack NB */
lwork = N*NB;
/* Allocate host memory for the matrix */
TESTING_MALLOC_CPU( tau, magmaDoubleComplex, N-1 );
TESTING_MALLOC_PIN( h_A, magmaDoubleComplex, lda*N );
TESTING_MALLOC_PIN( h_R, magmaDoubleComplex, lda*N );
TESTING_MALLOC_PIN( h_work, magmaDoubleComplex, lwork );
TESTING_MALLOC_PIN( D, double, N );
TESTING_MALLOC_PIN( E, double, N );
nstream = max(3,ngpu+2);
magma_queue_t streams[MagmaMaxGPUs][20];
magmaDoubleComplex *da[MagmaMaxGPUs],*dT1[MagmaMaxGPUs];
magma_int_t ldda = ((N+31)/32)*32;
if((distblk==0)||(distblk<NB))
distblk = max(256,NB);
printf("voici ngpu %d distblk %d NB %d nstream %d\n ",ngpu,distblk,NB,nstream);
for( magma_int_t dev = 0; dev < ngpu; ++dev ) {
magma_int_t mlocal = ((N / distblk) / ngpu + 1) * distblk;
magma_setdevice( dev );
TESTING_MALLOC_DEV( da[dev], magmaDoubleComplex, ldda*mlocal );
TESTING_MALLOC_DEV( dT1[dev], magmaDoubleComplex, N*NB );
for( int i = 0; i < nstream; ++i ) {
magma_queue_create( &streams[dev][i] );
}
}
magma_setdevice( 0 );
for(i=0; i<10; i++){
if ( !once ) {
N = size[i];
}
lda = N;
n2 = N*lda;
/* ====================================================================
Initialize the matrix
=================================================================== */
lapackf77_zlarnv( &ione, ISEED, &n2, h_A );
magma_zmake_hermitian( N, h_A, lda );
lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_A, &lda, h_R, &lda );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
/* Copy the matrix to the GPU */
magma_zsetmatrix_1D_col_bcyclic( N, N, h_R, lda, da, ldda, ngpu, distblk);
//magmaDoubleComplex *dabis;
// TESTING_MALLOC_DEV( dabis, magmaDoubleComplex, ldda*N );
// magma_zsetmatrix(N,N,h_R,lda,dabis,ldda);
for (int count=0; count<1;++count){
magma_setdevice(0);
start = get_current_time();
if(ver==30){
magma_zhetrd_he2hb_mgpu_spec(uplo[0], N, NB, h_R, lda, tau, h_work, lwork, da, ldda, dT1, NB, ngpu, distblk, streams, nstream, THREADS, &info);
}else{
nstream =3;
magma_zhetrd_he2hb_mgpu(uplo[0], N, NB, h_R, lda, tau, h_work, lwork, da, ldda, dT1, NB, ngpu, distblk, streams, nstream, THREADS, &info);
}
// magma_zhetrd_he2hb(uplo[0], N, NB, h_R, lda, tau, h_work, lwork, dT1[0], &info);
end = get_current_time();
printf(" Finish BAND N %d NB %d dist %d ngpu %d version %d timing= %f\n", N, NB, distblk, ngpu, ver, GetTimerValue(start,end) / 1000.);
}
magma_setdevice(0);
//goto fin;
//return 0;
for( magma_int_t dev = 0; dev < ngpu; ++dev ) {
magma_setdevice(dev);
cudaDeviceSynchronize();
}
magma_setdevice(0);
magmablasSetKernelStream( NULL );
magma_zhetrd_bhe2trc_v5(THREADS, WANTZ, uplo[0], NE, N, NB, h_R, lda, D, E, dT1[0], ldt);
// magma_zhetrd_bhe2trc(THREADS, WANTZ, uplo[0], NE, N, NB, h_R, lda, D, E, dT1[0], ldt);
end = get_current_time();
if (info != 0)
printf("magma_zhetrd_he2hb returned error %d: %s.\n",
(int) info, magma_strerror( info ));
gpu_perf = flops / GetTimerValue(start,end);
gpu_time = GetTimerValue(start,end) / 1000.;
/* =====================================================================
Print performance and error.
=================================================================== */
#if defined(CHECKEIG)
#if defined(PRECISION_z) || defined(PRECISION_d)
if ( checkres ) {
printf(" Total N %5d flops %6.2f timing %6.2f seconds\n", (int) N, gpu_perf, gpu_time );
char JOBZ;
if(WANTZ==0)
JOBZ='N';
else
JOBZ = 'V';
double nrmI=0.0, nrm1=0.0, nrm2=0.0;
int lwork2 = 256*N;
magmaDoubleComplex *work2 = (magmaDoubleComplex *) malloc (lwork2*sizeof(magmaDoubleComplex));
double *rwork2 = (double *) malloc (N*sizeof(double));
double *D2 = (double *) malloc (N*sizeof(double));
magmaDoubleComplex *AINIT = (magmaDoubleComplex *) malloc (N*lda*sizeof(magmaDoubleComplex));
memcpy(AINIT, h_A, N*lda*sizeof(magmaDoubleComplex));
/* compute the eigenvalues using lapack routine to be able to compare to it and used as ref */
start = get_current_time();
i= min(12,THREADS);
#if defined(USEMKL)
mkl_set_num_threads( i );
#endif
#if defined(USEACML)
omp_set_num_threads(i);
#endif
#if defined(PRECISION_z) || defined (PRECISION_c)
lapackf77_zheev( "N", "L", &N, h_A, &lda, D2, work2, &lwork2, rwork2, &info );
#else
lapackf77_dsyev( "N", "L", &N, h_A, &lda, D2, work2, &lwork2, &info );
#endif
///* call eigensolver for our resulting tridiag [D E] and for Q */
//dstedc_withZ('V', N, D, E, h_R, lda);
////dsterf_( &N, D, E, &info);
////
end = get_current_time();
printf(" Finish CHECK - EIGEN timing= %f threads %d\n", GetTimerValue(start,end) / 1000., i);
/* compare result */
cmp_vals(N, D2, D, &nrmI, &nrm1, &nrm2);
magmaDoubleComplex *WORKAJETER;
double *RWORKAJETER, *RESU;
WORKAJETER = (magmaDoubleComplex *) malloc( (2* N * N + N) * sizeof(magmaDoubleComplex) );
RWORKAJETER = (double *) malloc( N * sizeof(double) );
RESU = (double *) malloc(10*sizeof(double));
int MATYPE;
memset(RESU,0,10*sizeof(double));
MATYPE=3;
double NOTHING=0.0;
start = get_current_time();
// check results
zcheck_eig_(&JOBZ, &MATYPE, &N, &NB, AINIT, &lda, &NOTHING, &NOTHING, D2 , D, h_R, &lda, WORKAJETER, RWORKAJETER, RESU );
end = get_current_time();
printf(" Finish CHECK - results timing= %f\n", GetTimerValue(start,end) / 1000.);
#if defined(USEMKL)
mkl_set_num_threads( 1 );
#endif
#if defined(USEACML)
omp_set_num_threads(1);
#endif
printf("\n");
printf(" ================================================================================================================\n");
printf(" ==> INFO voici threads=%d N=%d NB=%d WANTZ=%d\n", (int) THREADS, (int) N, (int) NB, (int) WANTZ);
printf(" ================================================================================================================\n");
printf(" DSBTRD : %15s \n", "STATblgv9withQ ");
printf(" ================================================================================================================\n");
if(WANTZ>0)
printf(" | A - U S U' | / ( |A| n ulp ) : %15.3E \n",RESU[0]);
if(WANTZ>0)
printf(" | I - U U' | / ( n ulp ) : %15.3E \n", RESU[1]);
printf(" | D1 - EVEIGS | / (|D| ulp) : %15.3E \n", RESU[2]);
printf(" max | D1 - EVEIGS | : %15.3E \n", RESU[6]);
printf(" ================================================================================================================\n\n\n");
printf(" ****************************************************************************************************************\n");
printf(" * Hello here are the norm Infinite (max)=%8.2e norm one (sum)=%8.2e norm2(sqrt)=%8.2e *\n", nrmI, nrm1, nrm2);
printf(" ****************************************************************************************************************\n\n");
}
#endif
#endif
printf(" Total N %5d flops %6.2f timing %6.2f seconds\n", (int) N, 0.0, gpu_time );
printf("============================================================================\n\n\n");
if ( once )
break;
}
//fin:
/* Memory clean up */
TESTING_FREE_CPU( tau );
TESTING_FREE_PIN( h_A );
TESTING_FREE_PIN( h_R );
TESTING_FREE_PIN( h_work );
TESTING_FREE_PIN( D );
TESTING_FREE_PIN( E );
for( magma_int_t dev = 0; dev < ngpu; ++dev ) {
magma_setdevice( dev );
TESTING_FREE_DEV( da[dev] );
TESTING_FREE_DEV( dT1[dev] );
for( int i = 0; i < nstream; ++i ) {
magma_queue_destroy( streams[dev][i] );
}
}
magma_setdevice( 0 );
/* Shutdown */
TESTING_FINALIZE_MGPU();
return EXIT_SUCCESS;
}
