// ------------------------------------------------------------
// Solve A * X = B, where A and X are stored in CPU host memory.
// Internally, MAGMA transfers data to the GPU device
// and uses a hybrid CPU + GPU algorithm.
void cpu_interface( magma_int_t n, magma_int_t nrhs )
{
magmaDoubleComplex *A=NULL, *X=NULL;
magma_int_t *ipiv=NULL;
magma_int_t lda = n;
magma_int_t ldx = lda;
magma_int_t info = 0;
// magma_*malloc_cpu routines for CPU memory are type-safe and align to memory boundaries,
// but you can use malloc or new if you prefer.
magma_zmalloc_cpu( &A, lda*n );
magma_zmalloc_cpu( &X, ldx*nrhs );
magma_imalloc_cpu( &ipiv, n );
if ( A == NULL || X == NULL || ipiv == NULL ) {
fprintf( stderr, "malloc failed\n" );
goto cleanup;
}
// Replace these with your code to initialize A and X
zfill_matrix( n, n, A, lda );
zfill_rhs( n, nrhs, X, ldx );
magma_zgesv( n, 1, A, lda, ipiv, X, lda, &info );
if ( info != 0 ) {
fprintf( stderr, "magma_zgesv failed with info=%d\n", info );
}
// TODO: use result in X
cleanup:
magma_free_cpu( A );
magma_free_cpu( X );
magma_free_cpu( ipiv );
}
void MAGMAF_ZGESV ( magma_int_t *n, magma_int_t *nrhs,
cuDoubleComplex *A, magma_int_t *lda, magma_int_t *ipiv,
cuDoubleComplex *B, magma_int_t *ldb, magma_int_t *info)
{
magma_zgesv( *n, *nrhs,
A, *lda, ipiv,
B, *ldb,
info);
}
BOOST_FORCEINLINE result_type operator()(A0& a0, A1& a1, A2& a2) const
{
result_type that;
nt2_la_int n = std::min(nt2::height(a0),nt2::width(a0));
nt2_la_int lda = n;
nt2_la_int nhrs = nt2::width(a2);
nt2_la_int ldb = a2.leading_size();
a1.resize(nt2::of_size(n,1));
magma_zgesv(n,nhrs,(cuDoubleComplex*)a0.data(),lda,a1.data()
,(cuDoubleComplex*)a2.data(),ldb,&that);
return that;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing zgesv
*/
int main(int argc, char **argv)
{
real_Double_t gflops, cpu_perf, cpu_time, gpu_perf, gpu_time;
double error, Rnorm, Anorm, Xnorm, *work;
magmaDoubleComplex c_one = MAGMA_Z_ONE;
magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
magmaDoubleComplex *h_A, *h_LU, *h_B, *h_X;
magma_int_t *ipiv;
magma_int_t N, nrhs, lda, ldb, info, sizeA, sizeB;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t status = 0;
/* Initialize */
magma_queue_t queue[2];
magma_device_t device[ MagmaMaxGPUs ];
int num = 0;
magma_err_t err;
magma_init();
magma_opts opts;
parse_opts( argc, argv, &opts );
double tol = opts.tolerance * lapackf77_dlamch("E");
nrhs = opts.nrhs;
err = magma_get_devices( device, MagmaMaxGPUs, &num );
if ( err != 0 || num < 1 ) {
fprintf( stderr, "magma_get_devices failed: %d\n", err );
exit(-1);
}
// Create two queues on device opts.device
err = magma_queue_create( device[ opts.device ], &queue[0] );
if ( err != 0 ) {
fprintf( stderr, "magma_queue_create failed: %d\n", err );
exit(-1);
}
err = magma_queue_create( device[ opts.device ], &queue[1] );
if ( err != 0 ) {
fprintf( stderr, "magma_queue_create failed: %d\n", err );
exit(-1);
}
printf("ngpu %d\n", (int) opts.ngpu );
printf(" N NRHS CPU Gflop/s (sec) GPU GFlop/s (sec) ||B - AX|| / N*||A||*||X||\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;
ldb = lda;
gflops = ( FLOPS_ZGETRF( N, N ) + FLOPS_ZGETRS( N, nrhs ) ) / 1e9;
TESTING_MALLOC_CPU( h_A, magmaDoubleComplex, lda*N );
TESTING_MALLOC_CPU( h_LU, magmaDoubleComplex, lda*N );
TESTING_MALLOC_CPU( h_B, magmaDoubleComplex, ldb*nrhs );
TESTING_MALLOC_CPU( h_X, magmaDoubleComplex, ldb*nrhs );
TESTING_MALLOC_CPU( work, double, N );
TESTING_MALLOC_CPU( ipiv, magma_int_t, N );
/* Initialize the matrices */
sizeA = lda*N;
sizeB = ldb*nrhs;
lapackf77_zlarnv( &ione, ISEED, &sizeA, h_A );
lapackf77_zlarnv( &ione, ISEED, &sizeB, h_B );
// copy A to LU and B to X; save A and B for residual
lapackf77_zlacpy( "F", &N, &N, h_A, &lda, h_LU, &lda );
lapackf77_zlacpy( "F", &N, &nrhs, h_B, &ldb, h_X, &ldb );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
gpu_time = magma_wtime();
magma_zgesv( N, nrhs, h_LU, lda, ipiv, h_X, ldb, &info, queue );
gpu_time = magma_wtime() - gpu_time;
gpu_perf = gflops / gpu_time;
if (info != 0)
printf("magma_zgesv returned error %d: %s.\n",
(int) info, magma_strerror( info ));
//=====================================================================
// Residual
//=====================================================================
Anorm = lapackf77_zlange("I", &N, &N, h_A, &lda, work);
Xnorm = lapackf77_zlange("I", &N, &nrhs, h_X, &ldb, work);
blasf77_zgemm( MagmaNoTransStr, MagmaNoTransStr, &N, &nrhs, &N,
&c_one, h_A, &lda,
h_X, &ldb,
&c_neg_one, h_B, &ldb);
Rnorm = lapackf77_zlange("I", &N, &nrhs, h_B, &ldb, work);
error = Rnorm/(N*Anorm*Xnorm);
status |= ! (error < tol);
/* ====================================================================
Performs operation using LAPACK
=================================================================== */
if ( opts.lapack ) {
cpu_time = magma_wtime();
lapackf77_zgesv( &N, &nrhs, h_A, &lda, ipiv, h_B, &ldb, &info );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
if (info != 0)
printf("lapackf77_zgesv 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 ? "" : " failed"));
}
else {
printf( "%5d %5d --- ( --- ) %7.2f (%7.2f) %8.2e%s\n",
(int) N, (int) nrhs, gpu_perf, gpu_time,
error, (error < tol ? "" : " failed"));
}
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_LU );
TESTING_FREE_CPU( h_B );
TESTING_FREE_CPU( h_X );
TESTING_FREE_CPU( work );
TESTING_FREE_CPU( ipiv );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
magma_queue_destroy( queue[0] );
magma_queue_destroy( queue[1] );
magma_finalize();
return status;
}