// 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, magmaFloatComplex *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_clarnv( &ione, ISEED, &n2, h_A );
magma_cmake_hermitian( n, h_A, lda );
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing chegvdx
*/
int main( int argc, char** argv)
{
TESTING_INIT();
/* Constants */
const magmaFloatComplex c_zero = MAGMA_C_ZERO;
const magmaFloatComplex c_one = MAGMA_C_ONE;
const magmaFloatComplex c_neg_one = MAGMA_C_NEG_ONE;
const magma_int_t ione = 1;
/* Local variables */
real_Double_t gpu_time;
magmaFloatComplex *h_A, *h_R, *h_B, *h_S, *h_work;
#ifdef COMPLEX
float *rwork;
magma_int_t lrwork;
#endif
float *w1, *w2, result[2]={0,0};
magma_int_t *iwork;
magma_int_t N, n2, info, lda, lwork, liwork;
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");
float tolulp = opts.tolerance * lapackf77_slamch("P");
magma_range_t range = MagmaRangeAll;
if (opts.fraction != 1)
range = MagmaRangeI;
// pass ngpu = -1 to test multi-GPU code using 1 gpu
magma_int_t abs_ngpu = abs( opts.ngpu );
printf("%% itype = %d, jobz = %s, range = %s, uplo = %s, fraction = %6.4f, ngpu = %d\n",
int(opts.itype), lapack_vec_const(opts.jobz), lapack_range_const(range), lapack_uplo_const(opts.uplo),
opts.fraction, int(abs_ngpu) );
if (opts.itype == 1) {
printf("%% N M GPU Time (sec) |AZ-BZD| |D - D_magma|\n");
}
else if (opts.itype == 2) {
printf("%% N M GPU Time (sec) |ABZ-ZD| |D - D_magma|\n");
}
else if (opts.itype == 3) {
printf("%% N M GPU Time (sec) |BAZ-ZD| |D - D_magma|\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];
lda = N;
n2 = lda*N;
// TODO: test vl-vu range
magma_int_t m1 = 0;
float vl = 0;
float 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) );
}
magma_cheevdx_getworksize(N, threads, (opts.jobz == MagmaVec),
&lwork,
#ifdef COMPLEX
&lrwork,
#endif
&liwork);
/* Allocate host memory for the matrix */
TESTING_MALLOC_CPU( h_A, magmaFloatComplex, n2 );
TESTING_MALLOC_CPU( h_B, magmaFloatComplex, n2 );
TESTING_MALLOC_CPU( w1, float, N );
TESTING_MALLOC_CPU( w2, float, N );
TESTING_MALLOC_CPU( iwork, magma_int_t, liwork );
TESTING_MALLOC_PIN( h_R, magmaFloatComplex, n2 );
TESTING_MALLOC_PIN( h_S, magmaFloatComplex, n2 );
TESTING_MALLOC_PIN( h_work, magmaFloatComplex, max( lwork, N*N )); // check needs N*N
#ifdef COMPLEX
TESTING_MALLOC_PIN( rwork, float, lrwork);
#endif
/* Initialize the matrix */
lapackf77_clarnv( &ione, ISEED, &n2, h_A );
lapackf77_clarnv( &ione, ISEED, &n2, h_B );
magma_cmake_hpd( N, h_B, lda );
magma_cmake_hermitian( N, h_A, lda );
lapackf77_clacpy( MagmaFullStr, &N, &N, h_A, &lda, h_R, &lda );
lapackf77_clacpy( MagmaFullStr, &N, &N, h_B, &lda, h_S, &lda );
// ===================================================================
// Performs operation using MAGMA
// ===================================================================
gpu_time = magma_wtime();
if (opts.ngpu == 1) {
magma_chegvdx_2stage( opts.itype, opts.jobz, range, opts.uplo,
N, h_R, lda, h_S, lda, vl, vu, il, iu, &m1, w1,
h_work, lwork,
#ifdef COMPLEX
rwork, lrwork,
#endif
iwork, liwork,
&info );
}
else {
magma_chegvdx_2stage_m( abs_ngpu, opts.itype, opts.jobz, range, opts.uplo,
N, h_R, lda, h_S, lda, vl, vu, il, iu, &m1, w1,
h_work, lwork,
#ifdef COMPLEX
rwork, lrwork,
#endif
iwork, liwork,
&info );
}
gpu_time = magma_wtime() - gpu_time;
if (info != 0) {
printf("magma_chegvdx_2stage 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) | D(with V, magma) - D(w/o V, lapack) | / | D |
=================================================================== */
#ifdef REAL
float *rwork = h_work + N*N;
#endif
if ( opts.jobz != MagmaNoVec ) {
result[0] = 1.;
result[0] /= safe_lapackf77_clanhe("1", lapack_uplo_const(opts.uplo), &N, h_A, &lda, rwork);
result[0] /= lapackf77_clange("1", &N, &m1, h_R, &lda, rwork);
if (opts.itype == 1) {
blasf77_chemm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_one, h_A, &lda, h_R, &lda, &c_zero, h_work, &N);
for (int i=0; i < m1; ++i)
blasf77_csscal(&N, &w1[i], &h_R[i*N], &ione);
blasf77_chemm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_neg_one, h_B, &lda, h_R, &lda, &c_one, h_work, &N);
result[0] *= lapackf77_clange("1", &N, &m1, h_work, &N, rwork)/N;
}
else if (opts.itype == 2) {
blasf77_chemm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_one, h_B, &lda, h_R, &lda, &c_zero, h_work, &N);
for (int i=0; i < m1; ++i)
blasf77_csscal(&N, &w1[i], &h_R[i*N], &ione);
blasf77_chemm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_one, h_A, &lda, h_work, &N, &c_neg_one, h_R, &lda);
result[0] *= lapackf77_clange("1", &N, &m1, h_R, &lda, rwork)/N;
}
else if (opts.itype == 3) {
blasf77_chemm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_one, h_A, &lda, h_R, &lda, &c_zero, h_work, &N);
for (int i=0; i < m1; ++i)
blasf77_csscal(&N, &w1[i], &h_R[i*N], &ione);
blasf77_chemm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_one, h_B, &lda, h_work, &N, &c_neg_one, h_R, &lda);
result[0] *= lapackf77_clange("1", &N, &m1, h_R, &lda, rwork)/N;
}
}
lapackf77_clacpy( MagmaFullStr, &N, &N, h_A, &lda, h_R, &lda );
lapackf77_clacpy( MagmaFullStr, &N, &N, h_B, &lda, h_S, &lda );
lapackf77_chegvd( &opts.itype, "N", lapack_uplo_const(opts.uplo), &N,
h_R, &lda, h_S, &lda, w2,
h_work, &lwork,
#ifdef COMPLEX
rwork, &lrwork,
#endif
iwork, &liwork,
&info );
if (info != 0) {
printf("lapackf77_chegvd returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
float maxw=0, diff=0;
for (int j=0; j < m1; j++) {
maxw = max(maxw, fabs(w1[j]));
maxw = max(maxw, fabs(w2[j]));
diff = max(diff, fabs(w1[j] - w2[j]));
}
result[1] = diff / (m1*maxw);
}
/* =====================================================================
Print execution time
=================================================================== */
printf("%5d %5d %9.4f ",
(int) N, (int) m1, gpu_time);
if ( opts.check ) {
bool okay = (result[1] < tolulp);
if ( opts.jobz != MagmaNoVec ) {
okay = okay && (result[0] < tol);
printf(" %8.2e", result[0] );
}
else {
printf(" --- ");
}
printf(" %8.2e %s\n", result[1], (okay ? "ok" : "failed"));
status += ! okay;
}
else {
printf(" ---\n");
}
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 );
#ifdef COMPLEX
TESTING_FREE_PIN( rwork );
#endif
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
opts.cleanup();
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;
float 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;
magmaFloatComplex c_neg_one = MAGMA_C_NEG_ONE;
magmaFloatComplex alpha = MAGMA_C_MAKE( 1.5, -2.3 );
magmaFloatComplex beta = MAGMA_C_MAKE( -0.6, 0.8 );
magmaFloatComplex *A, *X, *Y, *Ycublas, *Ymagma;
magmaFloatComplex *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_CHEMV( N ) / 1e9;
TESTING_MALLOC( A, magmaFloatComplex, sizeA );
TESTING_MALLOC( X, magmaFloatComplex, sizeX );
TESTING_MALLOC( Y, magmaFloatComplex, sizeY );
TESTING_MALLOC( Ycublas, magmaFloatComplex, sizeY );
TESTING_MALLOC( Ymagma, magmaFloatComplex, sizeY );
TESTING_DEVALLOC( dA, magmaFloatComplex, sizeA );
TESTING_DEVALLOC( dX, magmaFloatComplex, sizeX );
TESTING_DEVALLOC( dY, magmaFloatComplex, sizeY );
blocks = (N + nb - 1) / nb;
ldwork = lda * (blocks + 1);
TESTING_DEVALLOC( dC_work, magmaFloatComplex, ldwork );
/* Initialize the matrix */
lapackf77_clarnv( &ione, ISEED, &sizeA, A );
magma_cmake_hermitian( N, A, lda );
lapackf77_clarnv( &ione, ISEED, &sizeX, X );
lapackf77_clarnv( &ione, ISEED, &sizeY, Y );
/* =====================================================================
Performs operation using CUBLAS
=================================================================== */
magma_csetmatrix( N, N, A, lda, dA, lda );
magma_csetvector( N, X, incx, dX, incx );
magma_csetvector( N, Y, incy, dY, incy );
cublas_time = magma_sync_wtime( 0 );
cublasChemv( 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_cgetvector( N, dY, incy, Ycublas, incy );
/* =====================================================================
Performs operation using MAGMA BLAS
=================================================================== */
magma_csetvector( N, Y, incy, dY, incy );
magma_time = magma_sync_wtime( 0 );
#if (GPUSHMEM >= 200)
magmablas_chemv2( opts.uplo, N, alpha, dA, lda, dX, incx, beta, dY, incy, dC_work, ldwork );
#else
magmablas_chemv( 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_cgetvector( N, dY, incy, Ymagma, incy );
/* =====================================================================
Performs operation using CPU BLAS
=================================================================== */
cpu_time = magma_wtime();
blasf77_chemv( &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_caxpy( &N, &c_neg_one, Y, &incy, Ymagma, &incy );
magma_error = lapackf77_clange( "M", &N, &ione, Ymagma, &N, work ) / N;
blasf77_caxpy( &N, &c_neg_one, Y, &incy, Ycublas, &incy );
cublas_error = lapackf77_clange( "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;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing chegst
*/
int main( int argc, char** argv)
{
TESTING_INIT();
// Constants
const magmaFloatComplex c_neg_one = MAGMA_C_NEG_ONE;
const magma_int_t ione = 1;
// Local variables
real_Double_t gpu_time, cpu_time;
magmaFloatComplex *h_A, *h_B, *h_R;
magmaFloatComplex_ptr d_A, d_B;
float Anorm, error, work[1];
magma_int_t N, n2, lda, ldda, info;
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t status = 0;
magma_opts opts;
opts.parse_opts( argc, argv );
opts.lapack |= opts.check; // check (-c) implies lapack (-l)
float tol = opts.tolerance * lapackf77_slamch("E");
printf("%% uplo = %s\n", lapack_uplo_const(opts.uplo) );
printf("%% itype N CPU time (sec) GPU time (sec) |R| \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( lda, opts.align );
n2 = N*lda;
TESTING_MALLOC_CPU( h_A, magmaFloatComplex, lda*N );
TESTING_MALLOC_CPU( h_B, magmaFloatComplex, lda*N );
TESTING_MALLOC_PIN( h_R, magmaFloatComplex, lda*N );
TESTING_MALLOC_DEV( d_A, magmaFloatComplex, ldda*N );
TESTING_MALLOC_DEV( d_B, magmaFloatComplex, ldda*N );
/* ====================================================================
Initialize the matrix
=================================================================== */
lapackf77_clarnv( &ione, ISEED, &n2, h_A );
lapackf77_clarnv( &ione, ISEED, &n2, h_B );
magma_cmake_hermitian( N, h_A, lda );
magma_cmake_hpd( N, h_B, lda );
magma_cpotrf( opts.uplo, N, h_B, lda, &info );
if (info != 0) {
printf("magma_cpotrf returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
magma_csetmatrix( N, N, h_A, lda, d_A, ldda );
magma_csetmatrix( N, N, h_B, lda, d_B, ldda );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
gpu_time = magma_wtime();
magma_chegst_gpu( opts.itype, opts.uplo, N, d_A, ldda, d_B, ldda, &info );
gpu_time = magma_wtime() - gpu_time;
if (info != 0) {
printf("magma_chegst_gpu returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
if ( opts.lapack ) {
cpu_time = magma_wtime();
lapackf77_chegst( &opts.itype, lapack_uplo_const(opts.uplo),
&N, h_A, &lda, h_B, &lda, &info );
cpu_time = magma_wtime() - cpu_time;
if (info != 0) {
printf("lapackf77_chegst returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
magma_cgetmatrix( N, N, d_A, ldda, h_R, lda );
blasf77_caxpy( &n2, &c_neg_one, h_A, &ione, h_R, &ione );
Anorm = safe_lapackf77_clanhe("f", lapack_uplo_const(opts.uplo), &N, h_A, &lda, work );
error = safe_lapackf77_clanhe("f", lapack_uplo_const(opts.uplo), &N, h_R, &lda, work )
/ Anorm;
bool okay = (error < tol);
status += ! okay;
printf("%3d %5d %7.2f %7.2f %8.2e %s\n",
(int) opts.itype, (int) N, cpu_time, gpu_time,
error, (okay ? "ok" : "failed"));
}
else {
printf("%3d %5d --- %7.2f\n",
(int) opts.itype, (int) N, gpu_time );
}
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_B );
TESTING_FREE_PIN( h_R );
TESTING_FREE_DEV( d_A );
TESTING_FREE_DEV( d_B );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
opts.cleanup();
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing chetrd_he2hb
*/
int main( int argc, char** argv)
{
TESTING_INIT_MGPU();
real_Double_t gpu_time, gpu_perf, gflops;
magmaFloatComplex *h_A, *h_R, *h_work, *dT1;
magmaFloatComplex *tau;
float *D, *E;
/* Matrix size */
magma_int_t N, n2, lda, lwork, ldt, lwork0;
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;
parse_opts( argc, argv, &opts );
NB = opts.nb;
if (NB < 1)
NB = 64; //64; //magma_get_chetrd_he2hb_nb(N);
// what is NE ?
if (NE < 1)
NE = 64; //N; //magma_get_chetrd_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 = N*lda;
gflops = FLOPS_CHETRD( 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, magmaFloatComplex, lda*N );
TESTING_MALLOC_CPU( tau, magmaFloatComplex, N-1 );
TESTING_MALLOC_PIN( h_R, magmaFloatComplex, lda*N );
TESTING_MALLOC_PIN( h_work, magmaFloatComplex, lwork0 );
TESTING_MALLOC_PIN( D, float, N );
TESTING_MALLOC_PIN( E, float, N );
//TESTING_MALLOC_DEV( dT1, magmaFloatComplex, (2*min(N,N)+(N+31)/32*32)*NB );
TESTING_MALLOC_DEV( dT1, magmaFloatComplex, (N*NB) );
// if (WANTZ) gflops = 2.0*gflops;
/* ====================================================================
Initialize the matrix
=================================================================== */
lapackf77_clarnv( &ione, ISEED, &n2, h_A );
magma_cmake_hermitian( N, h_A, lda );
lapackf77_clacpy( 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_chetrd_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_chetrd_bhe2trc_v5(THREADS, WANTZ, opts.uplo, NE, N, NB, h_R, lda, D, E, dT1, ldt);
*/
/*
magma_chetrd_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_chetrd_bhe2trc(THREADS, WANTZ, opts.uplo, NE, N, NB, h_R, lda, D, E, dT1, ldt);
*/
magma_range_t range = MagmaRangeAll;
magma_int_t fraction_ev = 100;
magma_int_t il, iu, m1;
float vl=0., vu=0.;
if (fraction_ev == 0) {
il = N / 10;
iu = N / 5+il;
}
else {
il = 1;
iu = (int)(fraction_ev*N);
if (iu < 1) iu = 1;
}
magmaFloatComplex *hh_work;
magma_int_t *iwork;
magma_int_t nb, /*lwork,*/ liwork;
magma_int_t threads = magma_get_parallel_numthreads();
#if defined(PRECISION_z) || defined(PRECISION_c)
float *rwork;
magma_int_t lrwork;
lwork = magma_cbulge_get_lq2(N, threads) + 2*N + N*N;
lrwork = 1 + 5*N +2*N*N;
TESTING_MALLOC_PIN( rwork, float, lrwork );
#else
lwork = magma_cbulge_get_lq2(N, threads) + 1 + 6*N + 2*N*N;
#endif
liwork = 3 + 5*N;
nb = magma_get_chetrd_nb(N);
TESTING_MALLOC_PIN( hh_work, magmaFloatComplex, lwork );
TESTING_MALLOC_CPU( iwork, magma_int_t, liwork );
if (ngpu == 1) {
printf("calling cheevdx_2stage 1 GPU\n");
magma_cheevdx_2stage( opts.jobz, range, opts.uplo, N,
h_R, lda,
vl, vu, il, iu,
&m1, D,
hh_work, lwork,
#if defined(PRECISION_z) || defined(PRECISION_c)
rwork, lrwork,
#endif
iwork, liwork,
&info);
} else {
printf("calling cheevdx_2stage_m %d GPU\n", (int) ngpu);
magma_cheevdx_2stage_m(ngpu, opts.jobz, range, opts.uplo, N,
h_R, lda,
vl, vu, il, iu,
&m1, D,
hh_work, lwork,
#if defined(PRECISION_z) || defined(PRECISION_c)
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"); exit(1); }
for (j=0; j < N; j++) {
for (k=0; k < N; k++) {
#if defined(PRECISION_z) || defined(PRECISION_c)
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"); exit(1); }
for (j=0; j < N; j++) {
for (k=0; k < N; k++) {
#if defined(PRECISION_z) || defined(PRECISION_c)
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 );
char JOBZ;
if (WANTZ == 0)
JOBZ = 'N';
else
JOBZ = 'V';
float nrmI=0.0, nrm1=0.0, nrm2=0.0;
int lwork2 = 256*N;
magmaFloatComplex *work2, *AINIT;
float *rwork2, *D2;
// TODO free this memory !
magma_cmalloc_cpu( &work2, lwork2 );
magma_smalloc_cpu( &rwork2, N );
magma_smalloc_cpu( &D2, N );
magma_cmalloc_cpu( &AINIT, N*lda );
memcpy(AINIT, h_A, N*lda*sizeof(magmaFloatComplex));
/* 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);
#if defined(USEMKL)
mkl_set_num_threads( i );
#endif
#if defined(USEACML)
omp_set_num_threads(i);
#endif
lapackf77_cheev( "N", "L", &N, h_A, &lda, D2, work2, &lwork2,
#if defined(PRECISION_z) || defined (PRECISION_c)
rwork2,
#endif
&info );
///* call eigensolver for our resulting tridiag [D E] and for Q */
//dstedc_withZ('V', N, D, E, h_R, lda);
////ssterf_( &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]);
*/
//magmaFloatComplex mydz=0.0, mydo=1.0;
//magmaFloatComplex *Z;
// magma_cmalloc_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);
magmaFloatComplex *WORKAJETER;
float *RWORKAJETER, *RESU;
// TODO free this memory !
magma_cmalloc_cpu( &WORKAJETER, (2* N * N + N) );
magma_smalloc_cpu( &RWORKAJETER, N );
magma_smalloc_cpu( &RESU, 10 );
int MATYPE;
memset(RESU, 0, 10*sizeof(float));
MATYPE=3;
float NOTHING=0.0;
cpu_time = magma_wtime();
// check results
ccheck_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);
#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 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_MGPU();
return EXIT_SUCCESS;
}
int main(int argc, char **argv)
{
TESTING_INIT();
const magmaFloatComplex c_neg_one = MAGMA_C_NEG_ONE;
const magma_int_t ione = 1;
real_Double_t gflops, magma_perf, magma_time, cpu_perf, cpu_time;
float magma_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;
magmaFloatComplex alpha = MAGMA_C_MAKE( 1.5, -2.3 );
magmaFloatComplex beta = MAGMA_C_MAKE( -0.6, 0.8 );
magmaFloatComplex *A, *X, *Y, *Ymagma;
magmaFloatComplex_ptr dA, dX, dY, dwork;
magma_int_t status = 0;
magma_opts opts;
parse_opts( argc, argv, &opts );
float tol = opts.tolerance * lapackf77_slamch("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_CSYMV( N ) / 1e9;
TESTING_MALLOC_CPU( A, magmaFloatComplex, sizeA );
TESTING_MALLOC_CPU( X, magmaFloatComplex, sizeX );
TESTING_MALLOC_CPU( Y, magmaFloatComplex, sizeY );
TESTING_MALLOC_CPU( Ymagma, magmaFloatComplex, sizeY );
TESTING_MALLOC_DEV( dA, magmaFloatComplex, ldda*N );
TESTING_MALLOC_DEV( dX, magmaFloatComplex, sizeX );
TESTING_MALLOC_DEV( dY, magmaFloatComplex, sizeY );
blocks = (N + nb - 1) / nb;
ldwork = ldda*blocks;
TESTING_MALLOC_DEV( dwork, magmaFloatComplex, ldwork );
magmablas_claset( MagmaFull, ldwork, 1, MAGMA_C_NAN, MAGMA_C_NAN, dwork, ldwork );
magmablas_claset( MagmaFull, ldda, N, MAGMA_C_NAN, MAGMA_C_NAN, dA, ldda );
/* Initialize the matrix */
lapackf77_clarnv( &ione, ISEED, &sizeA, A );
magma_cmake_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_claset( "Lower", &N1, &N1, &MAGMA_C_NAN, &MAGMA_C_NAN, &A[1], &lda );
}
else {
lapackf77_claset( "Upper", &N1, &N1, &MAGMA_C_NAN, &MAGMA_C_NAN, &A[lda], &lda );
}
lapackf77_clarnv( &ione, ISEED, &sizeX, X );
lapackf77_clarnv( &ione, ISEED, &sizeY, Y );
/* Note: CUBLAS does not implement csymv */
/* =====================================================================
Performs operation using MAGMABLAS
=================================================================== */
magma_csetmatrix( N, N, A, lda, dA, ldda );
magma_csetvector( N, X, incx, dX, incx );
magma_csetvector( N, Y, incy, dY, incy );
magma_time = magma_sync_wtime( 0 );
if ( opts.version == 1 ) {
magmablas_csymv_work( opts.uplo, N, alpha, dA, ldda, dX, incx, beta, dY, incy, dwork, ldwork, opts.queue );
}
else {
// non-work interface (has added overhead)
magmablas_csymv( 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_cgetvector( N, dY, incy, Ymagma, incy );
/* =====================================================================
Performs operation using CPU BLAS
=================================================================== */
cpu_time = magma_wtime();
lapackf77_csymv( 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_caxpy( &N, &c_neg_one, Y, &incy, Ymagma, &incy );
magma_error = lapackf77_clange( "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 cheevd
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gpu_time, cpu_time;
magmaFloatComplex *h_A, *h_R, *h_work, aux_work[1];
float *rwork, *w1, *w2, result[3], eps, aux_rwork[1];
magma_int_t *iwork, aux_iwork[1];
magma_int_t N, n2, info, lwork, lrwork, liwork, lda;
magma_int_t izero = 0;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
eps = lapackf77_slamch( "E" );
magma_int_t status = 0;
magma_opts opts;
parse_opts( argc, argv, &opts );
float tol = opts.tolerance * lapackf77_slamch("E");
float tolulp = opts.tolerance * lapackf77_slamch("P");
if ( opts.check && opts.jobz == MagmaNoVec ) {
fprintf( stderr, "checking results requires vectors; setting jobz=V (option -JV)\n" );
opts.jobz = MagmaVec;
}
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;
// query for workspace sizes
magma_cheevd( opts.jobz, opts.uplo,
N, NULL, lda, NULL,
aux_work, -1,
aux_rwork, -1,
aux_iwork, -1,
&info );
lwork = (magma_int_t) MAGMA_C_REAL( aux_work[0] );
lrwork = (magma_int_t) aux_rwork[0];
liwork = aux_iwork[0];
/* Allocate host memory for the matrix */
TESTING_MALLOC_CPU( h_A, magmaFloatComplex, N*lda );
TESTING_MALLOC_CPU( w1, float, N );
TESTING_MALLOC_CPU( w2, float, N );
TESTING_MALLOC_CPU( rwork, float, lrwork );
TESTING_MALLOC_CPU( iwork, magma_int_t, liwork );
TESTING_MALLOC_PIN( h_R, magmaFloatComplex, N*lda );
TESTING_MALLOC_PIN( h_work, magmaFloatComplex, lwork );
/* Initialize the matrix */
lapackf77_clarnv( &ione, ISEED, &n2, h_A );
magma_cmake_hermitian( N, h_A, N );
lapackf77_clacpy( MagmaUpperLowerStr, &N, &N, h_A, &lda, h_R, &lda );
/* warm up run */
if ( opts.warmup ) {
magma_cheevd( opts.jobz, opts.uplo,
N, h_R, lda, w1,
h_work, lwork,
rwork, lrwork,
iwork, liwork,
&info );
if (info != 0)
printf("magma_cheevd returned error %d: %s.\n",
(int) info, magma_strerror( info ));
lapackf77_clacpy( MagmaUpperLowerStr, &N, &N, h_A, &lda, h_R, &lda );
}
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
gpu_time = magma_wtime();
magma_cheevd( opts.jobz, opts.uplo,
N, h_R, lda, w1,
h_work, lwork,
rwork, lrwork,
iwork, liwork,
&info );
gpu_time = magma_wtime() - gpu_time;
if (info != 0)
printf("magma_cheevd returned error %d: %s.\n",
(int) info, magma_strerror( info ));
if ( opts.check ) {
/* =====================================================================
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 |
=================================================================== */
float temp1, temp2;
// tau=NULL is unused since itype=1
lapackf77_chet21( &ione, lapack_uplo_const(opts.uplo), &N, &izero,
h_A, &lda,
w1, w1,
h_R, &lda,
h_R, &lda,
NULL, h_work, rwork, &result[0] );
lapackf77_clacpy( MagmaUpperLowerStr, &N, &N, h_A, &lda, h_R, &lda );
magma_cheevd( MagmaNoVec, opts.uplo,
N, h_R, lda, w2,
h_work, lwork,
rwork, lrwork,
iwork, liwork,
&info );
if (info != 0)
printf("magma_cheevd returned error %d: %s.\n",
(int) info, magma_strerror( info ));
temp1 = temp2 = 0;
for( int j=0; j<N; j++ ) {
temp1 = max(temp1, fabs(w1[j]));
temp1 = max(temp1, fabs(w2[j]));
temp2 = max(temp2, fabs(w1[j]-w2[j]));
}
result[2] = temp2 / (((float)N)*temp1);
}
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
if ( opts.lapack ) {
cpu_time = magma_wtime();
lapackf77_cheevd( lapack_vec_const(opts.jobz), lapack_uplo_const(opts.uplo),
&N, h_A, &lda, w2,
h_work, &lwork,
rwork, &lrwork,
iwork, &liwork,
&info );
cpu_time = magma_wtime() - cpu_time;
if (info != 0)
printf("lapackf77_cheevd returned error %d: %s.\n",
(int) info, magma_strerror( info ));
printf("%5d %7.2f %7.2f\n",
(int) N, cpu_time, gpu_time);
}
else {
printf("%5d --- %7.2f\n",
(int) N, gpu_time);
}
/* =====================================================================
Print execution time
=================================================================== */
if ( opts.check ) {
printf("Testing the factorization A = U S U' for correctness:\n");
printf("(1) | A - U S U' | / (|A| N) = %8.2e %s\n", result[0]*eps, (result[0]*eps < tol ? "ok" : "failed") );
printf("(2) | I - U'U | / N = %8.2e %s\n", result[1]*eps, (result[1]*eps < tol ? "ok" : "failed") );
printf("(3) | S(w/ U) - S(w/o U) | / |S| = %8.2e %s\n\n", result[2] , (result[2] < tolulp ? "ok" : "failed") );
status += ! (result[0]*eps < tol && result[1]*eps < tol && result[2] < tolulp);
}
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( w1 );
TESTING_FREE_CPU( w2 );
TESTING_FREE_CPU( rwork );
TESTING_FREE_CPU( iwork );
TESTING_FREE_PIN( h_R );
TESTING_FREE_PIN( h_work );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return status;
}
int main(int argc, char **argv)
{
TESTING_INIT();
magma_setdevice(0);
magma_timestr_t start, end;
float flops, magma_perf, cuda_perf, error, work[1];
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
magmaFloatComplex c_neg_one = MAGMA_C_NEG_ONE;
magma_int_t n_local[4];
FILE *fp ;
magma_int_t N, m, i, j, lda, LDA, M;
magma_int_t matsize;
magma_int_t vecsize;
magma_int_t istart = 64;
magma_int_t incx = 1;
char uplo = MagmaLower;
magmaFloatComplex alpha = MAGMA_C_MAKE(1., 0.); // MAGMA_C_MAKE( 1.5, -2.3 );
magmaFloatComplex beta = MAGMA_C_MAKE(0., 0.); // MAGMA_C_MAKE( -0.6, 0.8 );
magmaFloatComplex *A, *X, *Y[4], *Ycublas, *Ymagma;
magmaFloatComplex *dA, *dX[4], *dY[4], *d_lA[4], *dYcublas ;
magma_queue_t stream[4][10];
magmaFloatComplex *C_work;
magmaFloatComplex *dC_work[4];
int max_num_gpus;
magma_int_t num_gpus = 1, nb;
magma_int_t blocks, lwork;
magma_int_t offset = 0;
M = 0;
N = 0;
if (argc != 1){
for(i = 1; i<argc; i++){
if (strcmp("-N", argv[i])==0)
{
N = atoi(argv[++i]);
istart = N;
}
else if (strcmp("-M", argv[i])==0)
M = atoi(argv[++i]);
else if (strcmp("-NGPU", argv[i])==0)
num_gpus = atoi(argv[++i]);
else if (strcmp("-offset", argv[i])==0)
offset = atoi(argv[++i]);
}
if ( M == 0 ) {
M = N;
}
if ( N == 0 ) {
N = M;
}
if (M>0 && N>0)
{ printf(" testing_chemv_mgpu -M %d -N %d -NGPU %d\n\n", (int) M, (int) N, (int) num_gpus);
printf(" in %c side \n", uplo);
}
else
{
printf("\nUsage: \n");
printf(" testing_chemv_mgpu -M %d -N %d -NGPU %d\n\n",
1024, 1024, 1);
exit(1);
}
}
else {
#if defined(PRECISION_z)
M = N = 8000;
#else
M = N = 12480;
#endif
num_gpus = 2;
offset = 0;
printf("\nUsage: \n");
printf(" testing_chemv_mgpu -M %d -N %d -NGPU %d\n\n", (int) M, (int) N, (int) num_gpus);
}
//////////////////////////////////////////////////////////////////////////
cudaGetDeviceCount(&max_num_gpus);
if (num_gpus > max_num_gpus){
printf("More GPUs requested than available. Have to change it.\n");
num_gpus = max_num_gpus;
}
printf("Number of GPUs to be used = %d\n", (int) num_gpus);
for(int i=0; i< num_gpus; i++)
{
magma_queue_create(&stream[i][0]);
}
LDA = ((N+31)/32)*32;
matsize = N*LDA;
vecsize = N*incx;
nb = 32;
//nb = 64;
printf("block size = %d\n", (int) nb);
TESTING_MALLOC_CPU( A, magmaFloatComplex, matsize );
TESTING_MALLOC_CPU( X, magmaFloatComplex, vecsize );
TESTING_MALLOC_CPU( Ycublas, magmaFloatComplex, vecsize );
TESTING_MALLOC_CPU( Ymagma, magmaFloatComplex, vecsize );
for(i=0; i<num_gpus; i++)
{
TESTING_MALLOC_CPU( Y[i], magmaFloatComplex, vecsize );
}
magma_setdevice(0);
TESTING_MALLOC_DEV( dA, magmaFloatComplex, matsize );
TESTING_MALLOC_DEV( dYcublas, magmaFloatComplex, vecsize );
for(i=0; i<num_gpus; i++)
{
n_local[i] = ((N/nb)/num_gpus)*nb;
if (i < (N/nb)%num_gpus)
n_local[i] += nb;
else if (i == (N/nb)%num_gpus)
n_local[i] += N%nb;
magma_setdevice(i);
TESTING_MALLOC_DEV( d_lA[i], magmaFloatComplex, LDA*n_local[i] );// potentially bugged
TESTING_MALLOC_DEV( dX[i], magmaFloatComplex, vecsize );
TESTING_MALLOC_DEV( dY[i], magmaFloatComplex, vecsize );
printf("device %2d n_local = %4d\n", (int) i, (int) n_local[i]);
}
magma_setdevice(0);
//////////////////////////////////////////////////////////////////////////
/* Initialize the matrix */
lapackf77_clarnv( &ione, ISEED, &matsize, A );
magma_cmake_hermitian( N, A, LDA );
blocks = N / nb + (N % nb != 0);
lwork = LDA * (blocks + 1);
TESTING_MALLOC_CPU( C_work, magmaFloatComplex, lwork );
for(i=0; i<num_gpus; i++){
magma_setdevice(i);
TESTING_MALLOC_DEV( dC_work[i], magmaFloatComplex, lwork );
//fillZero(dC_work[i], lwork);
}
magma_setdevice(0);
//////////////////////////////////////////////////////////////////////////
fp = fopen ("results_chemv_mgpu.csv", "w") ;
if( fp == NULL ){ printf("Couldn't open output file\n"); exit(1);}
printf("CHEMV magmaFloatComplex precision\n\n");
printf( " n CUBLAS,Gflop/s MAGMABLAS,Gflop/s \"error\"\n"
"==============================================================\n");
fprintf(fp, " n CUBLAS,Gflop/s MAGMABLAS,Gflop/s \"error\"\n"
"==============================================================\n");
// for( offset = 0; offset< N; offset ++ )
for(int size = istart ; size <= N ; size += 128)
{
// printf("offset = %d ", offset);
m = size ;
// m = N;
// lda = ((m+31)/32)*32;//
lda = LDA;
flops = FLOPS( (float)m ) / 1e6;
printf( "N %5d ", (int) m );
fprintf( fp, "%5d, ", (int) m );
vecsize = m * incx;
lapackf77_clarnv( &ione, ISEED, &vecsize, X );
lapackf77_clarnv( &ione, ISEED, &vecsize, Y[0] );
/* =====================================================================
Performs operation using CUDA-BLAS
=================================================================== */
magma_setdevice(0);
magma_csetmatrix_1D_col_bcyclic(m, m, A, LDA, d_lA, lda, num_gpus, nb);
magma_setdevice(0);
magma_csetmatrix( m, m, A, LDA, dA, lda );
magma_csetvector( m, Y[0], incx, dYcublas, incx );
for(i=0; i<num_gpus; i++){
magma_setdevice(i);
magma_csetvector( m, X, incx, dX[i], incx );
magma_csetvector( m, Y[0], incx, dY[i], incx );
blocks = m / nb + (m % nb != 0);
magma_csetmatrix( lda, blocks, C_work, LDA, dC_work[i], lda );
}
magma_setdevice(0);
start = get_current_time();
cublasChemv( uplo, m-offset, alpha, dA + offset + offset * lda, lda, dX[0] + offset, incx, beta, dYcublas + offset, incx );
end = get_current_time();
magma_cgetvector( m, dYcublas, incx, Ycublas, incx );
cuda_perf = flops / GetTimerValue(start,end);
printf( "%11.2f", cuda_perf );
fprintf(fp, "%11.2f,", cuda_perf );
magma_setdevice(0);
start = get_current_time();
if(nb == 32)
{
magmablas_chemv2_mgpu_32_offset( uplo, m, alpha, d_lA, lda, dX, incx, beta, dY, incx,
dC_work, lwork, num_gpus, nb, offset);
}
else // nb = 64
{
magmablas_chemv2_mgpu_offset( uplo, m, alpha, d_lA, lda, dX, incx, beta, dY, incx,
dC_work, lwork, num_gpus, nb, offset);
}
for(i=1; i<num_gpus; i++)
{
magma_setdevice(i);
cudaDeviceSynchronize();
}
end = get_current_time();
magma_perf = flops / GetTimerValue(start,end);
printf( "%11.2f", magma_perf );
fprintf(fp, "%11.2f,", magma_perf );
for(i=0; i<num_gpus; i++)
{
magma_setdevice(i);
magma_cgetvector( m, dY[i], incx, Y[i], incx );
}
magma_setdevice(0);
#ifdef validate
for( j= offset;j<m;j++)
{
for(i=1; i<num_gpus; i++)
{
// printf("Y[%d][%d] = %15.14f\n", i, j, Y[i][j].x);
#if defined(PRECISION_z) || defined(PRECISION_c)
Y[0][j].x = Y[0][j].x + Y[i][j].x;
Y[0][j].y = Y[0][j].y + Y[i][j].y;
#else
Y[0][j] = Y[0][j] + Y[i][j];
#endif
}
}
/*
#if defined(PRECISION_z) || defined(PRECISION_c)
for( j=offset;j<m;j++)
{
if(Y[0][j].x != Ycublas[j].x)
{
printf("Y-multi[%d] = %f, %f\n", j, Y[0][j].x, Y[0][j].y );
printf("Ycublas[%d] = %f, %f\n", j, Ycublas[j].x, Ycublas[j].y);
}
}
#else
for( j=offset;j<m;j++)
{
if(Y[0][j] != Ycublas[j])
{
printf("Y-multi[%d] = %f\n", j, Y[0][j] );
printf("Ycublas[%d] = %f\n", j, Ycublas[j]);
}
}
#endif
*/
/* =====================================================================
Computing the Difference Cublas VS Magma
=================================================================== */
magma_int_t nw = m - offset ;
blasf77_caxpy( &nw, &c_neg_one, Y[0] + offset, &incx, Ycublas + offset, &incx);
error = lapackf77_clange( "M", &nw, &ione, Ycublas + offset, &nw, work );
#if 0
printf( "\t\t %8.6e", error / m );
fprintf( fp, "\t\t %8.6e", error / m );
/*
* Extra check with cblas vs magma
*/
cblas_ccopy( m, Y, incx, Ycublas, incx );
cblas_chemv( CblasColMajor, CblasLower, m,
CBLAS_SADDR(alpha), A, LDA, X, incx,
CBLAS_SADDR(beta), Ycublas, incx );
blasf77_caxpy( &m, &c_neg_one, Ymagma, &incx, Ycublas, &incx);
error = lapackf77_clange( "M", &m, &ione, Ycublas, &m, work );
#endif
printf( "\t\t %8.6e", error / m );
fprintf( fp, "\t\t %8.6e", error / m );
#endif
printf("\n");
fprintf(fp, "\n");
}
fclose( fp ) ;
/* Free Memory */
TESTING_FREE_CPU( A );
TESTING_FREE_CPU( X );
TESTING_FREE_CPU( Ycublas );
TESTING_FREE_CPU( Ymagma );
TESTING_FREE_CPU( C_work );
magma_setdevice(0);
TESTING_FREE_DEV( dA );
TESTING_FREE_DEV( dYcublas );
for(i=0; i<num_gpus; i++)
{
TESTING_FREE_CPU( Y[i] );
magma_setdevice(i);
TESTING_FREE_DEV( d_lA[i] );
TESTING_FREE_DEV( dX[i] );
TESTING_FREE_DEV( dY[i] );
TESTING_FREE_DEV( dC_work[i] );
}
magma_setdevice(0);
///////////////////////////////////////////////////////////
/* Free device */
TESTING_FINALIZE();
return 0;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing chegvdx
*/
int main( int argc, char** argv)
{
TESTING_INIT_MGPU();
real_Double_t mgpu_time;
magmaFloatComplex *h_A, *h_Ainit, *h_B, *h_Binit, *h_work;
#if defined(PRECISION_z) || defined(PRECISION_c)
float *rwork;
magma_int_t lrwork;
#endif
float *w1, result=0;
magma_int_t *iwork;
magma_int_t N, n2, info, lwork, liwork;
magmaFloatComplex c_zero = MAGMA_C_ZERO;
magmaFloatComplex c_one = MAGMA_C_ONE;
magmaFloatComplex c_neg_one = MAGMA_C_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 );
float tol = opts.tolerance * lapackf77_slamch("E");
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: ngpu = %d, itype = %d, jobz = %s, range = %s, uplo = %s, opts.check = %d, fraction = %6.4f\n",
(int) opts.ngpu, (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 ngpu MGPU 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_cbulge_get_lq2(N, threads) + 2*N + N*N;
lrwork = 1 + 5*N +2*N*N;
#else
lwork = magma_cbulge_get_lq2(N, threads) + 1 + 6*N + 2*N*N;
#endif
liwork = 3 + 5*N;
//magma_int_t NB = 96;//magma_bulge_get_nb(N);
//magma_int_t sizvblg = magma_cbulge_get_lq2(N, threads);
//magma_int_t siz = max(sizvblg,n2)+2*(N*NB+N)+24*N;
/* Allocate host memory for the matrix */
TESTING_MALLOC_PIN( h_A, magmaFloatComplex, n2 );
TESTING_MALLOC_PIN( h_B, magmaFloatComplex, n2 );
TESTING_MALLOC_PIN( h_work, magmaFloatComplex, lwork );
#if defined(PRECISION_z) || defined(PRECISION_c)
TESTING_MALLOC_PIN( rwork, float, lrwork);
#endif
TESTING_MALLOC_CPU( w1, float, N );
TESTING_MALLOC_CPU( iwork, magma_int_t, liwork);
/* Initialize the matrix */
lapackf77_clarnv( &ione, ISEED, &n2, h_A );
lapackf77_clarnv( &ione, ISEED, &n2, h_B );
magma_cmake_hpd( N, h_B, N );
magma_cmake_hermitian( N, h_A, N );
if ( opts.warmup || opts.check ) {
TESTING_MALLOC_CPU( h_Ainit, magmaFloatComplex, n2 );
TESTING_MALLOC_CPU( h_Binit, magmaFloatComplex, n2 );
lapackf77_clacpy( MagmaUpperLowerStr, &N, &N, h_A, &N, h_Ainit, &N );
lapackf77_clacpy( MagmaUpperLowerStr, &N, &N, h_B, &N, h_Binit, &N );
}
magma_int_t m1 = 0;
float vl = 0;
float vu = 0;
magma_int_t il = 0;
magma_int_t iu = 0;
if (range == MagmaRangeI) {
il = 1;
iu = (int) (opts.fraction*N);
}
if ( opts.warmup ) {
// ==================================================================
// Warmup using MAGMA. I prefer to use smalltest to warmup A-
// ==================================================================
magma_chegvdx_2stage_m(opts.ngpu, opts.itype, opts.jobz, range, opts.uplo,
N, h_A, N, h_B, N, vl, vu, il, iu, &m1, w1,
h_work, lwork,
#if defined(PRECISION_z) || defined(PRECISION_c)
rwork, lrwork,
#endif
iwork, liwork,
&info);
lapackf77_clacpy( MagmaUpperLowerStr, &N, &N, h_Ainit, &N, h_A, &N );
lapackf77_clacpy( MagmaUpperLowerStr, &N, &N, h_Binit, &N, h_B, &N );
}
// ===================================================================
// Performs operation using MAGMA
// ===================================================================
mgpu_time = magma_wtime();
magma_chegvdx_2stage_m(opts.ngpu, opts.itype, opts.jobz, range, opts.uplo,
N, h_A, N, h_B, N, vl, vu, il, iu, &m1, w1,
h_work, lwork,
#if defined(PRECISION_z) || defined(PRECISION_c)
rwork, lrwork,
#endif
iwork, liwork,
&info);
mgpu_time = magma_wtime() - mgpu_time;
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)
// ===================================================================
#if defined(PRECISION_d) || defined(PRECISION_s)
float *rwork = h_work + N*N;
#endif
result = 1.;
result /= lapackf77_clanhe("1", lapack_uplo_const(opts.uplo), &N, h_Ainit, &N, rwork);
result /= lapackf77_clange("1", &N , &m1, h_A, &N, rwork);
if (opts.itype == 1) {
blasf77_chemm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_one, h_Ainit, &N, h_A, &N, &c_zero, h_work, &N);
for(int i=0; i<m1; ++i)
blasf77_csscal(&N, &w1[i], &h_A[i*N], &ione);
blasf77_chemm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_neg_one, h_Binit, &N, h_A, &N, &c_one, h_work, &N);
result *= lapackf77_clange("1", &N, &m1, h_work, &N, rwork)/N;
}
else if (opts.itype == 2) {
blasf77_chemm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_one, h_Binit, &N, h_A, &N, &c_zero, h_work, &N);
for(int i=0; i<m1; ++i)
blasf77_csscal(&N, &w1[i], &h_A[i*N], &ione);
blasf77_chemm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_one, h_Ainit, &N, h_work, &N, &c_neg_one, h_A, &N);
result *= lapackf77_clange("1", &N, &m1, h_A, &N, rwork)/N;
}
else if (opts.itype == 3) {
blasf77_chemm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_one, h_Ainit, &N, h_A, &N, &c_zero, h_work, &N);
for(int i=0; i<m1; ++i)
blasf77_csscal(&N, &w1[i], &h_A[i*N], &ione);
blasf77_chemm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_one, h_Binit, &N, h_work, &N, &c_neg_one, h_A, &N);
result *= lapackf77_clange("1", &N, &m1, h_A, &N, rwork)/N;
}
}
// ===================================================================
// Print execution time
// ===================================================================
printf("%5d %5d %4d %7.2f\n",
(int) N, (int) m1, (int) opts.ngpu, mgpu_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, (result < tol ? "ok" : "failed") );
}
else if (opts.itype==2) {
printf("(1) | A B Z - Z D | / (|A| |Z| N) = %8.2e %s\n", result, (result < tol ? "ok" : "failed") );
}
else if (opts.itype==3) {
printf("(1) | B A Z - Z D | / (|A| |Z| N) = %8.2e %s\n", result, (result < tol ? "ok" : "failed") );
}
printf("\n");
status += ! (result < tol);
}
TESTING_FREE_PIN( h_A );
TESTING_FREE_PIN( h_B );
TESTING_FREE_PIN( h_work );
#if defined(PRECISION_z) || defined(PRECISION_c)
TESTING_FREE_PIN( rwork );
#endif
TESTING_FREE_CPU( w1 );
TESTING_FREE_CPU( iwork );
if ( opts.warmup || opts.check ) {
TESTING_FREE_CPU( h_Ainit );
TESTING_FREE_CPU( h_Binit );
}
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
/* Shutdown */
TESTING_FINALIZE_MGPU();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing chegvd
*/
int main( int argc, char** argv)
{
TESTING_INIT_MGPU();
magmaFloatComplex *h_A, *h_Ainit, *h_B, *h_Binit, *h_work;
#if defined(PRECISION_z) || defined(PRECISION_c)
float *rwork;
#endif
float *w1, *w2, result;
magma_int_t *iwork;
real_Double_t mgpu_time, gpu_time, cpu_time;
/* Matrix size */
magma_int_t N, n2, nb;
magma_int_t info;
magma_int_t ione = 1;
magmaFloatComplex c_zero = MAGMA_C_ZERO;
magmaFloatComplex c_one = MAGMA_C_ONE;
magmaFloatComplex c_neg_one = MAGMA_C_NEG_ONE;
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t status = 0;
magma_opts opts;
parse_opts( argc, argv, &opts );
float tol = opts.tolerance * lapackf77_slamch("E");
float tolulp = opts.tolerance * lapackf77_slamch("P");
if ( opts.check && opts.jobz == MagmaNoVec ) {
fprintf( stderr, "checking results requires vectors; setting jobz=V (option -JV)\n" );
opts.jobz = MagmaVec;
}
printf("using: ngpu = %d, itype = %d, jobz = %s, uplo = %s, check = %d\n",
(int) opts.ngpu, (int) opts.itype,
lapack_vec_const(opts.jobz), lapack_uplo_const(opts.uplo), (int) opts.check);
printf(" N CPU Time (sec) GPU Time (sec) MGPU Time (sec)\n");
printf("=========================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
// TODO define lda
N = opts.nsize[itest];
n2 = N*N;
nb = magma_get_chetrd_nb(N);
#if defined(PRECISION_z) || defined(PRECISION_c)
magma_int_t lwork = max( N + N*nb, 2*N + N*N );
magma_int_t lrwork = 1 + 5*N +2*N*N;
#else
magma_int_t lwork = max( 2*N + N*nb, 1 + 6*N + 2*N*N );
#endif
magma_int_t liwork = 3 + 5*N;
TESTING_MALLOC_PIN( h_A, magmaFloatComplex, n2 );
TESTING_MALLOC_PIN( h_B, magmaFloatComplex, n2 );
TESTING_MALLOC_PIN( h_work, magmaFloatComplex, lwork );
#if defined(PRECISION_z) || defined(PRECISION_c)
TESTING_MALLOC_PIN( rwork, float, lrwork );
#endif
TESTING_MALLOC_CPU( w1, float, N );
TESTING_MALLOC_CPU( w2, float, N );
TESTING_MALLOC_CPU( iwork, magma_int_t, liwork );
/* Initialize the matrix */
lapackf77_clarnv( &ione, ISEED, &n2, h_A );
lapackf77_clarnv( &ione, ISEED, &n2, h_B );
magma_cmake_hpd( N, h_B, N );
magma_cmake_hermitian( N, h_A, N );
if ( opts.warmup || opts.check ) {
TESTING_MALLOC_CPU( h_Ainit, magmaFloatComplex, n2 );
TESTING_MALLOC_CPU( h_Binit, magmaFloatComplex, n2 );
lapackf77_clacpy( MagmaUpperLowerStr, &N, &N, h_A, &N, h_Ainit, &N );
lapackf77_clacpy( MagmaUpperLowerStr, &N, &N, h_B, &N, h_Binit, &N );
}
if (opts.warmup) {
// ==================================================================
// Warmup using MAGMA.
// ==================================================================
magma_chegvd_m( opts.ngpu, opts.itype, opts.jobz, opts.uplo,
N, h_A, N, h_B, N, w1,
h_work, lwork,
#if defined(PRECISION_z) || defined(PRECISION_c)
rwork, lrwork,
#endif
iwork, liwork,
&info);
lapackf77_clacpy( MagmaUpperLowerStr, &N, &N, h_Ainit, &N, h_A, &N );
lapackf77_clacpy( MagmaUpperLowerStr, &N, &N, h_Binit, &N, h_B, &N );
}
// ===================================================================
// Performs operation using MAGMA
// ===================================================================
mgpu_time = magma_wtime();
magma_chegvd_m( opts.ngpu, opts.itype, opts.jobz, opts.uplo,
N, h_A, N, h_B, N, w1,
h_work, lwork,
#if defined(PRECISION_z) || defined(PRECISION_c)
rwork, lrwork,
#endif
iwork, liwork,
&info);
mgpu_time = magma_wtime() - mgpu_time;
if (info != 0)
printf("magma_chegvd_m returned error %d: %s.\n",
(int) info, magma_strerror( info ));
if ( opts.check ) {
/* =====================================================================
Check the results following the LAPACK's [zc]hegvd routine.
A x = lambda B x is solved
and the following 3 tests computed:
(1) | A Z - B Z D | / ( |A||Z| N ) (itype = 1)
| A B Z - Z D | / ( |A||Z| N ) (itype = 2)
| B A Z - Z D | / ( |A||Z| N ) (itype = 3)
=================================================================== */
#if defined(PRECISION_d) || defined(PRECISION_s)
float *rwork = h_work + N*N;
#endif
result = 1.;
result /= lapackf77_clanhe("1", lapack_uplo_const(opts.uplo), &N, h_Ainit, &N, rwork);
result /= lapackf77_clange("1", &N, &N, h_A, &N, rwork);
if (opts.itype == 1) {
blasf77_chemm("L", lapack_uplo_const(opts.uplo), &N, &N, &c_one, h_Ainit, &N, h_A, &N, &c_zero, h_work, &N);
for(int i=0; i<N; ++i)
blasf77_csscal(&N, &w1[i], &h_A[i*N], &ione);
blasf77_chemm("L", lapack_uplo_const(opts.uplo), &N, &N, &c_neg_one, h_Binit, &N, h_A, &N, &c_one, h_work, &N);
result *= lapackf77_clange("1", &N, &N, h_work, &N, rwork)/N;
}
else if (opts.itype == 2) {
blasf77_chemm("L", lapack_uplo_const(opts.uplo), &N, &N, &c_one, h_Binit, &N, h_A, &N, &c_zero, h_work, &N);
for(int i=0; i<N; ++i)
blasf77_csscal(&N, &w1[i], &h_A[i*N], &ione);
blasf77_chemm("L", lapack_uplo_const(opts.uplo), &N, &N, &c_one, h_Ainit, &N, h_work, &N, &c_neg_one, h_A, &N);
result *= lapackf77_clange("1", &N, &N, h_A, &N, rwork)/N;
}
else if (opts.itype == 3) {
blasf77_chemm("L", lapack_uplo_const(opts.uplo), &N, &N, &c_one, h_Ainit, &N, h_A, &N, &c_zero, h_work, &N);
for(int i=0; i<N; ++i)
blasf77_csscal(&N, &w1[i], &h_A[i*N], &ione);
blasf77_chemm("L", lapack_uplo_const(opts.uplo), &N, &N, &c_one, h_Binit, &N, h_work, &N, &c_neg_one, h_A, &N);
result *= lapackf77_clange("1", &N, &N, h_A, &N, rwork)/N;
}
lapackf77_clacpy( MagmaUpperLowerStr, &N, &N, h_Ainit, &N, h_A, &N );
lapackf77_clacpy( MagmaUpperLowerStr, &N, &N, h_Binit, &N, h_B, &N );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
gpu_time = magma_wtime();
magma_chegvd(opts.itype, opts.jobz, opts.uplo,
N, h_A, N, h_B, N, w2,
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_chegvd returned error %d: %s.\n",
(int) info, magma_strerror( info ));
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
cpu_time = magma_wtime();
lapackf77_chegvd(&opts.itype, lapack_vec_const(opts.jobz), lapack_uplo_const(opts.uplo),
&N, h_Ainit, &N, h_Binit, &N, w2,
h_work, &lwork,
#if defined(PRECISION_z) || defined(PRECISION_c)
rwork, &lrwork,
#endif
iwork, &liwork,
&info);
cpu_time = magma_wtime() - cpu_time;
if (info != 0)
printf("lapackf77_chegvd returned error %d: %s.\n",
(int) info, magma_strerror( info ));
float temp1 = 0;
float temp2 = 0;
for(int j=0; j<N; j++) {
temp1 = max(temp1, fabs(w1[j]));
temp1 = max(temp1, fabs(w2[j]));
temp2 = max(temp2, fabs(w1[j]-w2[j]));
}
float result2 = temp2 / (((float)N)*temp1);
/* =====================================================================
Print execution time
=================================================================== */
printf("%5d %7.2f %7.2f %7.2f\n",
(int) N, cpu_time, gpu_time, mgpu_time);
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, (result < tol ? "ok" : "failed") );
}
else if (opts.itype==2) {
printf("(1) | A B Z - Z D | / (|A| |Z| N) = %8.2e %s\n", result, (result < tol ? "ok" : "failed") );
}
else if (opts.itype==3) {
printf("(1) | B A Z - Z D | / (|A| |Z| N) = %8.2e %s\n", result, (result < tol ? "ok" : "failed") );
}
printf( "(3) | D(MGPU)-D(LAPACK) |/ |D| = %8.2e %s\n\n", result2, (result2 < tolulp ? "ok" : "failed") );
status += ! (result < tol && result2 < tolulp);
}
else {
printf("%5d --- --- %7.2f\n",
(int) N, mgpu_time);
}
/* Memory clean up */
TESTING_FREE_PIN( h_A );
TESTING_FREE_PIN( h_B );
TESTING_FREE_PIN( h_work );
#if defined(PRECISION_z) || defined(PRECISION_c)
TESTING_FREE_PIN( rwork );
#endif
TESTING_FREE_CPU( w1 );
TESTING_FREE_CPU( w2 );
TESTING_FREE_CPU( iwork );
if ( opts.warmup || opts.check ) {
TESTING_FREE_CPU( h_Ainit );
TESTING_FREE_CPU( h_Binit );
}
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
/* Shutdown */
TESTING_FINALIZE_MGPU();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing chegvdx
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gpu_time /*cpu_time*/;
magmaFloatComplex *h_A, *h_R, *h_B, *h_S, *h_work;
float *w1, *w2, vl=0, vu=0;
float result[2] = {0};
magma_int_t *iwork;
magma_int_t N, n2, info, il, iu, m1, nb, lwork, liwork;
magmaFloatComplex c_zero = MAGMA_C_ZERO;
magmaFloatComplex c_one = MAGMA_C_ONE;
magmaFloatComplex c_neg_one = MAGMA_C_NEG_ONE;
#ifdef COMPLEX
float *rwork;
magma_int_t lrwork;
#endif
//float d_one = 1.;
//float 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 );
float tol = opts.tolerance * lapackf77_slamch("E");
//float tolulp = opts.tolerance * lapackf77_slamch("P");
// TODO: how to check NoVec? This doesn't have an equivalent call to LAPACK.
if ( opts.check && opts.jobz == MagmaNoVec ) {
fprintf( stderr, "WARNING: cannot currently check results when not computing vectors (option -JN).\n" );
}
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_chetrd_nb(N);
#ifdef COMPLEX
lwork = max( N + N*nb, 2*N + N*N );
lrwork = 1 + 5*N +2*N*N;
#else
lwork = max( 2*N + N*nb, 1 + 6*N + 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, magmaFloatComplex, n2 );
TESTING_MALLOC_CPU( h_B, magmaFloatComplex, n2 );
TESTING_MALLOC_CPU( w1, float, N );
TESTING_MALLOC_CPU( w2, float, N );
TESTING_MALLOC_CPU( iwork, magma_int_t, liwork );
TESTING_MALLOC_PIN( h_R, magmaFloatComplex, n2 );
TESTING_MALLOC_PIN( h_S, magmaFloatComplex, n2 );
TESTING_MALLOC_PIN( h_work, magmaFloatComplex, lwork );
#ifdef COMPLEX
TESTING_MALLOC_PIN( rwork, float, lrwork);
#endif
/* Initialize the matrix */
lapackf77_clarnv( &ione, ISEED, &n2, h_A );
lapackf77_clarnv( &ione, ISEED, &n2, h_B );
magma_cmake_hpd( N, h_B, N );
magma_cmake_hermitian( N, h_A, N );
// ==================================================================
// Warmup using MAGMA
// ==================================================================
if ( opts.warmup ) {
lapackf77_clacpy( MagmaFullStr, &N, &N, h_A, &N, h_R, &N );
lapackf77_clacpy( MagmaFullStr, &N, &N, h_B, &N, h_S, &N );
magma_chegvdx( opts.itype, opts.jobz, MagmaRangeI, opts.uplo,
N, h_R, N, h_S, N, vl, vu, il, iu, &m1, w1,
h_work, lwork,
#ifdef COMPLEX
rwork, lrwork,
#endif
iwork, liwork,
&info );
if (info != 0)
printf("magma_chegvdx returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
lapackf77_clacpy( MagmaFullStr, &N, &N, h_A, &N, h_R, &N );
lapackf77_clacpy( MagmaFullStr, &N, &N, h_B, &N, h_S, &N );
gpu_time = magma_wtime();
magma_chegvdx( opts.itype, opts.jobz, MagmaRangeI, opts.uplo,
N, h_R, N, h_S, N, vl, vu, il, iu, &m1, w1,
h_work, lwork,
#ifdef COMPLEX
rwork, lrwork,
#endif
iwork, liwork,
&info );
gpu_time = magma_wtime() - gpu_time;
if (info != 0)
printf("magma_chegvdx returned error %d: %s.\n",
(int) info, magma_strerror( info ));
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) | D(with V) - D(w/o V) | / | D |
=================================================================== */
#ifdef REAL
float *rwork = h_work + N*N;
#endif
result[0] = 1.;
result[0] /= lapackf77_clanhe("1", lapack_uplo_const(opts.uplo), &N, h_A, &N, rwork);
result[0] /= lapackf77_clange("1", &N, &m1, h_R, &N, rwork);
if (opts.itype == 1) {
blasf77_chemm("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_csscal(&N, &w1[i], &h_R[i*N], &ione);
blasf77_chemm("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_clange("1", &N, &m1, h_work, &N, rwork)/N;
}
else if (opts.itype == 2) {
blasf77_chemm("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_csscal(&N, &w1[i], &h_R[i*N], &ione);
blasf77_chemm("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_clange("1", &N, &m1, h_R, &N, rwork)/N;
}
else if (opts.itype == 3) {
blasf77_chemm("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_csscal(&N, &w1[i], &h_R[i*N], &ione);
blasf77_chemm("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_clange("1", &N, &m1, h_R, &N, rwork)/N;
}
// Disable eigenvalue check which calls routine again --
// it obscures whether error occurs in first call above or in this call.
// TODO: add comparison to LAPACK, as in testing_chegvd.cpp?
//
//lapackf77_clacpy( MagmaFullStr, &N, &N, h_A, &N, h_R, &N );
//lapackf77_clacpy( MagmaFullStr, &N, &N, h_B, &N, h_S, &N );
//
//magma_int_t m2;
//magma_chegvdx( opts.itype, MagmaNoVec, MagmaRangeI, opts.uplo,
// N, h_R, N, h_S, N, vl, vu, il, iu, &m2, w2,
// h_work, lwork,
// #ifdef COMPLEX
// rwork, lrwork,
// #endif
// iwork, liwork,
// &info );
//if (info != 0)
// printf("magma_chegvdx returned error %d: %s.\n",
// (int) info, magma_strerror( info ));
//
//float 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 );
#ifdef COMPLEX
TESTING_FREE_PIN( rwork );
#endif
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing cheevd
*/
int main( int argc, char** argv)
{
TESTING_INIT();
/* Constants */
const float d_zero = 0;
const magma_int_t izero = 0;
const magma_int_t ione = 1;
/* Local variables */
real_Double_t gpu_time, cpu_time;
magmaFloatComplex *h_A, *h_R, *h_Z, *h_work, aux_work[1];
#ifdef COMPLEX
float *rwork, aux_rwork[1];
magma_int_t lrwork;
#endif
float *w1, *w2, result[4]={0, 0, 0, 0}, eps, abstol;
magma_int_t *iwork, *isuppz, *ifail, aux_iwork[1];
magma_int_t N, n2, info, lwork, liwork, lda;
magma_int_t ISEED[4] = {0,0,0,1};
eps = lapackf77_slamch( "E" );
magma_int_t status = 0;
magma_opts opts;
opts.parse_opts( argc, argv );
// checking NoVec requires LAPACK
opts.lapack |= (opts.check && opts.jobz == MagmaNoVec);
magma_range_t range = MagmaRangeAll;
if (opts.fraction != 1)
range = MagmaRangeI;
#ifdef REAL
if (opts.version == 3 || opts.version == 4) {
printf("%% magma_cheevr and magma_cheevx are not available for real precisions (single, float).\n");
return status;
}
#endif
float tol = opts.tolerance * lapackf77_slamch("E");
float tolulp = opts.tolerance * lapackf77_slamch("P");
// pass ngpu = -1 to test multi-GPU code using 1 gpu
magma_int_t abs_ngpu = abs( opts.ngpu );
printf("%% jobz = %s, range = %s, uplo = %s, fraction = %6.4f, ngpu = %d\n",
lapack_vec_const(opts.jobz), lapack_range_const(range), lapack_uplo_const(opts.uplo),
opts.fraction, int(abs_ngpu) );
printf("%% N CPU Time (sec) GPU Time (sec) |S-S_magma| |A-USU^H| |I-U^H U|\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;
abstol = 0; // auto, in cheevr
// TODO: test vl-vu range
magma_int_t m1 = 0;
float vl = 0;
float 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) );
}
// query for workspace sizes
if ( opts.version == 1 || opts.version == 2 ) {
magma_cheevd( opts.jobz, opts.uplo,
N, NULL, lda, NULL, // A, w
aux_work, -1,
#ifdef COMPLEX
aux_rwork, -1,
#endif
aux_iwork, -1,
&info );
}
else if ( opts.version == 3 ) {
#ifdef COMPLEX
magma_cheevr( opts.jobz, range, opts.uplo,
N, NULL, lda, // A
vl, vu, il, iu, abstol,
&m1, NULL, // w
NULL, lda, NULL, // Z, isuppz
aux_work, -1,
#ifdef COMPLEX
aux_rwork, -1,
#endif
aux_iwork, -1,
&info );
#endif
}
else if ( opts.version == 4 ) {
#ifdef COMPLEX
magma_cheevx( opts.jobz, range, opts.uplo,
N, NULL, lda, // A
vl, vu, il, iu, abstol,
&m1, NULL, // w
NULL, lda, // Z
aux_work, -1,
#ifdef COMPLEX
aux_rwork,
#endif
aux_iwork,
NULL, // ifail
&info );
// cheevx doesn't query rwork, iwork; set them for consistency
aux_rwork[0] = float(7*N);
aux_iwork[0] = float(5*N);
#endif
}
lwork = (magma_int_t) MAGMA_C_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, magmaFloatComplex, N*lda );
TESTING_MALLOC_CPU( w1, float, N );
TESTING_MALLOC_CPU( w2, float, N );
#ifdef COMPLEX
TESTING_MALLOC_CPU( rwork, float, lrwork );
#endif
TESTING_MALLOC_CPU( iwork, magma_int_t, liwork );
TESTING_MALLOC_PIN( h_R, magmaFloatComplex, N*lda );
TESTING_MALLOC_PIN( h_work, magmaFloatComplex, lwork );
if (opts.version == 3) {
TESTING_MALLOC_CPU( h_Z, magmaFloatComplex, N*lda );
TESTING_MALLOC_CPU( isuppz, magma_int_t, 2*max(1,N) );
}
if (opts.version == 4) {
TESTING_MALLOC_CPU( h_Z, magmaFloatComplex, N*lda );
TESTING_MALLOC_CPU( ifail, magma_int_t, N );
}
/* Clear eigenvalues, for |S-S_magma| check when fraction < 1. */
lapackf77_slaset( "Full", &N, &ione, &d_zero, &d_zero, w1, &N );
lapackf77_slaset( "Full", &N, &ione, &d_zero, &d_zero, w2, &N );
/* Initialize the matrix */
lapackf77_clarnv( &ione, ISEED, &n2, h_A );
magma_cmake_hermitian( N, h_A, lda );
lapackf77_clacpy( MagmaFullStr, &N, &N, h_A, &lda, h_R, &lda );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
gpu_time = magma_wtime();
if (opts.version == 1) {
if (opts.ngpu == 1) {
magma_cheevd( opts.jobz, opts.uplo,
N, h_R, lda, w1,
h_work, lwork,
#ifdef COMPLEX
rwork, lrwork,
#endif
iwork, liwork,
&info );
}
else {
//printf( "magma_cheevd_m, ngpu %d (%d)\n", opts.ngpu, abs_ngpu );
magma_cheevd_m( abs_ngpu, opts.jobz, opts.uplo,
N, h_R, lda, w1,
h_work, lwork,
#ifdef COMPLEX
rwork, lrwork,
#endif
iwork, liwork,
&info );
}
}
else if ( opts.version == 2 ) { // version 2: cheevdx computes selected eigenvalues/vectors
if (opts.ngpu == 1) {
magma_cheevdx( opts.jobz, range, opts.uplo,
N, h_R, lda,
vl, vu, il, iu,
&m1, w1,
h_work, lwork,
#ifdef COMPLEX
rwork, lrwork,
#endif
iwork, liwork,
&info );
}
else {
//printf( "magma_cheevdx_m, ngpu %d (%d)\n", opts.ngpu, abs_ngpu );
magma_cheevdx_m( abs_ngpu, opts.jobz, range, opts.uplo,
N, h_R, lda,
vl, vu, il, iu,
&m1, w1,
h_work, lwork,
#ifdef COMPLEX
rwork, lrwork,
#endif
iwork, liwork,
&info );
}
//printf( "il %d, iu %d, m1 %d\n", il, iu, m1 );
}
else if ( opts.version == 3 ) { // version 3: MRRR, computes selected eigenvalues/vectors
// only complex version available
#ifdef COMPLEX
magma_cheevr( opts.jobz, range, opts.uplo,
N, h_R, lda,
vl, vu, il, iu, abstol,
&m1, w1,
h_Z, lda, isuppz,
h_work, lwork,
#ifdef COMPLEX
rwork, lrwork,
#endif
iwork, liwork,
&info );
lapackf77_clacpy( "Full", &N, &N, h_Z, &lda, h_R, &lda );
#endif
}
else if ( opts.version == 4 ) { // version 3: cheevx (QR iteration), computes selected eigenvalues/vectors
// only complex version available
#ifdef COMPLEX
magma_cheevx( opts.jobz, range, opts.uplo,
N, h_R, lda,
vl, vu, il, iu, abstol,
&m1, w1,
h_Z, lda,
h_work, lwork,
#ifdef COMPLEX
rwork, /*lrwork,*/
#endif
iwork, /*liwork,*/
ifail,
&info );
lapackf77_clacpy( "Full", &N, &N, h_Z, &lda, h_R, &lda );
#endif
}
gpu_time = magma_wtime() - gpu_time;
if (info != 0) {
printf("magma_cheevd returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
bool okay = true;
if ( opts.check && opts.jobz != MagmaNoVec ) {
/* =====================================================================
Check the results following the LAPACK's [zcds]drvst routine.
A is factored as A = U S U^H and the following 3 tests computed:
(1) | A - U S U^H | / ( |A| N )
(2) | I - U^H U | / ( N )
(3) | S(with U) - S(w/o U) | / | S | // currently disabled, but compares to LAPACK
=================================================================== */
magmaFloatComplex *work;
TESTING_MALLOC_CPU( work, magmaFloatComplex, 2*N*N );
// e=NULL is unused since kband=0; tau=NULL is unused since itype=1
lapackf77_chet21( &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 third eigenvalue check that calls routine again --
// it obscures whether error occurs in first call above or in this call.
// But see comparison to LAPACK below.
//
//lapackf77_clacpy( MagmaFullStr, &N, &N, h_A, &lda, h_R, &lda );
//magma_cheevd( MagmaNoVec, opts.uplo,
// N, h_R, lda, w2,
// h_work, lwork,
// #ifdef COMPLEX
// rwork, lrwork,
// #endif
// iwork, liwork,
// &info );
//if (info != 0) {
// printf("magma_cheevd returned error %d: %s.\n",
// (int) info, magma_strerror( info ));
//}
//
//float 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();
if ( opts.version == 1 || opts.version == 2 ) {
lapackf77_cheevd( 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 );
}
else if ( opts.version == 3 ) {
lapackf77_cheevr( lapack_vec_const(opts.jobz),
lapack_range_const(range),
lapack_uplo_const(opts.uplo),
&N, h_A, &lda,
&vl, &vu, &il, &iu, &abstol,
&m1, w2,
h_Z, &lda, isuppz,
h_work, &lwork,
#ifdef COMPLEX
rwork, &lrwork,
#endif
iwork, &liwork,
&info );
lapackf77_clacpy( "Full", &N, &N, h_Z, &lda, h_A, &lda );
}
else if ( opts.version == 4 ) {
lapackf77_cheevx( lapack_vec_const(opts.jobz),
lapack_range_const(range),
lapack_uplo_const(opts.uplo),
&N, h_A, &lda,
&vl, &vu, &il, &iu, &abstol,
&m1, w2,
h_Z, &lda,
h_work, &lwork,
#ifdef COMPLEX
rwork,
#endif
iwork,
ifail,
&info );
lapackf77_clacpy( "Full", &N, &N, h_Z, &lda, h_A, &lda );
}
cpu_time = magma_wtime() - cpu_time;
if (info != 0) {
printf("lapackf77_cheevd returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
// compare eigenvalues
float 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);
okay = okay && (result[3] < tolulp);
printf("%5d %9.4f %9.4f %8.2e ",
(int) N, cpu_time, gpu_time, result[3] );
}
else {
printf("%5d --- %9.4f --- ",
(int) N, gpu_time);
}
// print error checks
if ( opts.check && opts.jobz != MagmaNoVec ) {
okay = okay && (result[0] < tol) && (result[1] < tol);
printf(" %8.2e %8.2e", result[0], result[1] );
}
else {
printf(" --- --- ");
}
printf(" %s\n", (okay ? "ok" : "failed"));
status += ! okay;
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 );
if ( opts.version == 3 ) {
TESTING_FREE_CPU( h_Z );
TESTING_FREE_CPU( isuppz );
}
if ( opts.version == 4 ) {
TESTING_FREE_CPU( h_Z );
TESTING_FREE_CPU( ifail );
}
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
opts.cleanup();
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing chesv_nopiv_gpu
*/
int main(int argc, char **argv)
{
TESTING_INIT();
real_Double_t gflops, cpu_perf, cpu_time, gpu_perf, gpu_time;
float error, Rnorm, Anorm, Xnorm, *work;
magmaFloatComplex c_one = MAGMA_C_ONE;
magmaFloatComplex c_neg_one = MAGMA_C_NEG_ONE;
magmaFloatComplex *h_A, *h_B, *h_X, temp, *hwork;
magmaFloatComplex_ptr d_A, d_B;
magma_int_t *ipiv;
magma_int_t N, nrhs, lda, ldb, ldda, lddb, info, sizeA, sizeB, lwork;
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 );
float tol = opts.tolerance * lapackf77_slamch("E");
nrhs = opts.nrhs;
printf(" N NRHS CPU GFlop/s (sec) GPU GFlop/s (sec) ||B - AX|| / N*||A||*||X||\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;
ldb = lda;
ldda = ((N+31)/32)*32;
lddb = ldda;
gflops = ( FLOPS_CGETRF( N, N ) + FLOPS_CGETRS( N, nrhs ) ) / 1e9;
TESTING_MALLOC_CPU( h_A, magmaFloatComplex, lda*N );
TESTING_MALLOC_CPU( h_B, magmaFloatComplex, ldb*nrhs );
TESTING_MALLOC_CPU( h_X, magmaFloatComplex, ldb*nrhs );
TESTING_MALLOC_CPU( work, float, N );
TESTING_MALLOC_CPU( ipiv, magma_int_t, N );
TESTING_MALLOC_DEV( d_A, magmaFloatComplex, ldda*N );
TESTING_MALLOC_DEV( d_B, magmaFloatComplex, lddb*nrhs );
/* Initialize the matrices */
sizeA = lda*N;
sizeB = ldb*nrhs;
lapackf77_clarnv( &ione, ISEED, &sizeA, h_A );
lapackf77_clarnv( &ione, ISEED, &sizeB, h_B );
bool nopiv = true;
if ( nopiv ) {
magma_cmake_hpd( N, h_A, lda ); // SPD / HPD does not require pivoting
}
else {
magma_cmake_hermitian( N, h_A, lda ); // symmetric/Hermitian generally requires pivoting
}
magma_csetmatrix( N, N, h_A, lda, d_A, ldda );
magma_csetmatrix( N, nrhs, h_B, ldb, d_B, lddb );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
gpu_time = magma_wtime();
magma_chesv_nopiv_gpu( opts.uplo, N, nrhs, d_A, ldda, d_B, lddb, &info );
gpu_time = magma_wtime() - gpu_time;
gpu_perf = gflops / gpu_time;
if (info != 0)
printf("magma_cgesv_gpu returned error %d: %s.\n",
(int) info, magma_strerror( info ));
//=====================================================================
// Residual
//=====================================================================
magma_cgetmatrix( N, nrhs, d_B, lddb, h_X, ldb );
Anorm = lapackf77_clange("I", &N, &N, h_A, &lda, work);
Xnorm = lapackf77_clange("I", &N, &nrhs, h_X, &ldb, work);
blasf77_cgemm( MagmaNoTransStr, MagmaNoTransStr, &N, &nrhs, &N,
&c_one, h_A, &lda,
h_X, &ldb,
&c_neg_one, h_B, &ldb);
Rnorm = lapackf77_clange("I", &N, &nrhs, h_B, &ldb, work);
error = Rnorm/(N*Anorm*Xnorm);
status += ! (error < tol);
/* ====================================================================
Performs operation using LAPACK
=================================================================== */
if ( opts.lapack ) {
lwork = -1;
lapackf77_chesv( lapack_uplo_const(opts.uplo), &N,&nrhs,
h_A, &lda, ipiv, h_B, &ldb, &temp, &lwork, &info );
lwork = (magma_int_t) MAGMA_C_REAL( temp );
TESTING_MALLOC_PIN( hwork, magmaFloatComplex, lwork );
cpu_time = magma_wtime();
lapackf77_chesv( lapack_uplo_const(opts.uplo), &N, &nrhs,
h_A, &lda, ipiv, h_B, &ldb, hwork, &lwork, &info );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
if (info != 0)
printf("lapackf77_chesv returned error %d: %s.\n",
(int) info, magma_strerror( info ));
printf( "%5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %s\n",
(int) N, (int) nrhs, cpu_perf, cpu_time, gpu_perf, gpu_time,
error, (error < tol ? "ok" : "failed"));
TESTING_FREE_CPU( hwork );
}
else {
printf( "%5d %5d --- ( --- ) %7.2f (%7.2f) %8.2e %s\n",
(int) N, (int) nrhs, gpu_perf, gpu_time,
error, (error < tol ? "ok" : "failed"));
}
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_B );
TESTING_FREE_CPU( h_X );
TESTING_FREE_CPU( work );
TESTING_FREE_CPU( ipiv );
TESTING_FREE_DEV( d_A );
TESTING_FREE_DEV( d_B );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return status;
}
