/* ////////////////////////////////////////////////////////////////////////////
-- testing csr matrix add
*/
int main( int argc, char** argv )
{
TESTING_INIT();
magma_queue_t queue;
magma_queue_create( /*devices[ opts->device ],*/ &queue );
real_Double_t res;
magma_z_sparse_matrix A, B, B2, C, A_d, B_d, C_d;
magmaDoubleComplex one = MAGMA_Z_MAKE(1.0, 0.0);
magmaDoubleComplex zero = MAGMA_Z_MAKE(0.0, 0.0);
magmaDoubleComplex mone = MAGMA_Z_MAKE(-1.0, 0.0);
magma_int_t i=1;
if ( strcmp("LAPLACE2D", argv[i]) == 0 && i+1 < argc ) { // Laplace test
i++;
magma_int_t laplace_size = atoi( argv[i] );
magma_zm_5stencil( laplace_size, &A, queue );
} else { // file-matrix test
magma_z_csr_mtx( &A, argv[i], queue );
}
printf( "# matrix info: %d-by-%d with %d nonzeros\n",
(int) A.num_rows,(int) A.num_cols,(int) A.nnz );
i++;
if ( strcmp("LAPLACE2D", argv[i]) == 0 && i+1 < argc ) { // Laplace test
i++;
magma_int_t laplace_size = atoi( argv[i] );
magma_zm_5stencil( laplace_size, &B, queue );
} else { // file-matrix test
magma_z_csr_mtx( &B, argv[i], queue );
}
printf( "# matrix info: %d-by-%d with %d nonzeros\n",
(int) B.num_rows,(int) B.num_cols,(int) B.nnz );
magma_z_mtransfer( A, &A_d, Magma_CPU, Magma_DEV, queue );
magma_z_mtransfer( B, &B_d, Magma_CPU, Magma_DEV, queue );
magma_zcuspaxpy( &one, A_d, &one, B_d, &C_d, queue );
magma_z_mfree(&B_d, queue );
magma_zcuspaxpy( &mone, A_d, &one, C_d, &B_d, queue );
magma_z_mtransfer( B_d, &B2, Magma_DEV, Magma_CPU, queue );
magma_z_mfree(&A_d, queue );
magma_z_mfree(&B_d, queue );
magma_z_mfree(&C_d, queue );
// check difference
magma_zmdiff( B, B2, &res, queue );
printf("# ||A-B||_F = %8.2e\n", res);
if ( res < .000001 )
printf("# tester matrix add: ok\n");
else
printf("# tester matrix add: failed\n");
magma_z_mfree(&A, queue );
magma_z_mfree(&B, queue );
magma_z_mfree(&B2, queue );
magma_queue_destroy( queue );
TESTING_FINALIZE();
return 0;
}
/* ////////////////////////////////////////////////////////////////////////////
-- testing any solver
*/
int main( int argc, char** argv )
{
magma_int_t info = 0;
TESTING_INIT();
magma_zopts zopts;
magma_queue_t queue=NULL;
magma_queue_create( &queue );
real_Double_t res;
magma_z_matrix Z={Magma_CSR}, Z2={Magma_CSR}, A={Magma_CSR}, A2={Magma_CSR},
AT={Magma_CSR}, AT2={Magma_CSR}, B={Magma_CSR};
int i=1;
CHECK( magma_zparse_opts( argc, argv, &zopts, &i, queue ));
B.blocksize = zopts.blocksize;
B.alignment = zopts.alignment;
while( i < argc ) {
if ( strcmp("LAPLACE2D", argv[i]) == 0 && i+1 < argc ) { // Laplace test
i++;
magma_int_t laplace_size = atoi( argv[i] );
CHECK( magma_zm_5stencil( laplace_size, &Z, queue ));
} else { // file-matrix test
CHECK( magma_z_csr_mtx( &Z, argv[i], queue ));
}
printf("%% matrix info: %d-by-%d with %d nonzeros\n",
int(Z.num_rows), int(Z.num_cols), int(Z.nnz) );
// convert to be non-symmetric
CHECK( magma_zmconvert( Z, &A, Magma_CSR, Magma_CSRL, queue ));
CHECK( magma_zmconvert( Z, &B, Magma_CSR, Magma_CSRU, queue ));
// transpose
CHECK( magma_zmtranspose( A, &AT, queue ));
// quite some conversions
//ELL
CHECK( magma_zmconvert( AT, &AT2, Magma_CSR, Magma_ELL, queue ));
magma_zmfree(&AT, queue );
CHECK( magma_zmconvert( AT2, &AT, Magma_ELL, Magma_CSR, queue ));
magma_zmfree(&AT2, queue );
//ELLPACKT
CHECK( magma_zmconvert( AT, &AT2, Magma_CSR, Magma_ELLPACKT, queue ));
magma_zmfree(&AT, queue );
CHECK( magma_zmconvert( AT2, &AT, Magma_ELLPACKT, Magma_CSR, queue ));
magma_zmfree(&AT2, queue );
//ELLRT
AT2.blocksize = 8;
AT2.alignment = 8;
CHECK( magma_zmconvert( AT, &AT2, Magma_CSR, Magma_ELLRT, queue ));
magma_zmfree(&AT, queue );
CHECK( magma_zmconvert( AT2, &AT, Magma_ELLRT, Magma_CSR, queue ));
magma_zmfree(&AT2, queue );
//SELLP
AT2.blocksize = 8;
AT2.alignment = 8;
CHECK( magma_zmconvert( AT, &AT2, Magma_CSR, Magma_SELLP, queue ));
magma_zmfree(&AT, queue );
CHECK( magma_zmconvert( AT2, &AT, Magma_SELLP, Magma_CSR, queue ));
magma_zmfree(&AT2, queue );
//ELLD
CHECK( magma_zmconvert( AT, &AT2, Magma_CSR, Magma_ELLD, queue ));
magma_zmfree(&AT, queue );
CHECK( magma_zmconvert( AT2, &AT, Magma_ELLD, Magma_CSR, queue ));
magma_zmfree(&AT2, queue );
//CSRCOO
CHECK( magma_zmconvert( AT, &AT2, Magma_CSR, Magma_CSRCOO, queue ));
magma_zmfree(&AT, queue );
CHECK( magma_zmconvert( AT2, &AT, Magma_CSRCOO, Magma_CSR, queue ));
magma_zmfree(&AT2, queue );
//CSRLIST
CHECK( magma_zmconvert( AT, &AT2, Magma_CSR, Magma_CSRLIST, queue ));
magma_zmfree(&AT, queue );
CHECK( magma_zmconvert( AT2, &AT, Magma_CSRLIST, Magma_CSR, queue ));
magma_zmfree(&AT2, queue );
//CSRD
CHECK( magma_zmconvert( AT, &AT2, Magma_CSR, Magma_CSRD, queue ));
magma_zmfree(&AT, queue );
CHECK( magma_zmconvert( AT2, &AT, Magma_CSRD, Magma_CSR, queue ));
magma_zmfree(&AT2, queue );
// transpose
CHECK( magma_zmtranspose( AT, &A2, queue ));
CHECK( magma_zmdiff( A, A2, &res, queue));
printf("%% ||A-A2||_F = %8.2e\n", res);
if ( res < .000001 )
printf("%% conversion tester: ok\n");
else
printf("%% conversion tester: failed\n");
CHECK( magma_zmlumerge( A2, B, &Z2, queue ));
CHECK( magma_zmdiff( Z, Z2, &res, queue));
printf("%% ||Z-Z2||_F = %8.2e\n", res);
if ( res < .000001 )
printf("%% LUmerge tester: ok\n");
else
printf("%% LUmerge tester: failed\n");
magma_zmfree(&A, queue );
magma_zmfree(&A2, queue );
magma_zmfree(&AT, queue );
magma_zmfree(&AT2, queue );
magma_zmfree(&B, queue );
magma_zmfree(&Z2, queue );
magma_zmfree(&Z, queue );
i++;
}
cleanup:
magma_zmfree(&A, queue );
magma_zmfree(&A2, queue );
magma_zmfree(&AT, queue );
magma_zmfree(&AT2, queue );
magma_zmfree(&B, queue );
magma_zmfree(&Z2, queue );
magma_zmfree(&Z, queue );
magma_queue_destroy( queue );
TESTING_FINALIZE();
return info;
}
/* ////////////////////////////////////////////////////////////////////////////
-- testing any solver
*/
int main( int argc, char** argv )
{
magma_int_t info = 0;
TESTING_INIT();
magma_queue_t queue=NULL;
magma_queue_create( 0, &queue );
magmaDoubleComplex one = MAGMA_Z_MAKE(1.0, 0.0);
magmaDoubleComplex zero = MAGMA_Z_MAKE(0.0, 0.0);
magma_z_matrix A={Magma_CSR}, B_d={Magma_CSR};
magma_z_matrix x={Magma_CSR}, b={Magma_CSR};
int i=1;
while( i < argc ) {
if ( strcmp("LAPLACE2D", argv[i]) == 0 && i+1 < argc ) { // Laplace test
i++;
magma_int_t laplace_size = atoi( argv[i] );
CHECK( magma_zm_5stencil( laplace_size, &A, queue ));
} else { // file-matrix test
CHECK( magma_z_csr_mtx( &A, argv[i], queue ));
}
printf( "\n# matrix info: %d-by-%d with %d nonzeros\n\n",
int(A.num_rows), int(A.num_cols), int(A.nnz) );
magma_int_t n = A.num_rows;
CHECK( magma_zmtransfer( A, &B_d, Magma_CPU, Magma_DEV, queue ));
// vectors and initial guess
CHECK( magma_zvinit( &b, Magma_DEV, A.num_cols, 1, zero, queue ));
CHECK( magma_zvinit( &x, Magma_DEV, A.num_cols, 1, one, queue ));
CHECK( magma_zprint_vector( b, 90, 10, queue ));
CHECK( magma_zprint_matrix( A, queue ));
printf("\n\n\n");
CHECK( magma_zprint_matrix( B_d, queue ));
double res;
res = magma_dznrm2(n, b.dval, 1, queue );
printf("norm0: %f\n", res);
CHECK( magma_z_spmv( one, B_d, x, zero, b, queue )); // b = A x
CHECK( magma_zprint_vector( b, 0, 100, queue ));
CHECK( magma_zprint_vector( b, b.num_rows-10, 10, queue ));
res = magma_dznrm2( n, b.dval, 1, queue );
printf("norm: %f\n", res);
CHECK( magma_zresidual( B_d, x, b, &res, queue));
printf("res: %f\n", res);
magma_zmfree(&B_d, queue );
magma_zmfree(&A, queue );
magma_zmfree(&x, queue );
magma_zmfree(&b, queue );
i++;
}
cleanup:
magma_zmfree(&A, queue );
magma_zmfree(&B_d, queue );
magma_zmfree(&x, queue );
magma_zmfree(&b, queue );
magma_queue_destroy( queue );
magma_finalize();
return info;
}
/* ////////////////////////////////////////////////////////////////////////////
-- testing any solver
*/
int main( int argc, char** argv )
{
magma_int_t info = 0;
TESTING_INIT();
magma_zopts zopts;
magma_queue_t queue=NULL;
magma_queue_create( /*devices[ opts->device ],*/ &queue );
magmaDoubleComplex one = MAGMA_Z_MAKE(1.0, 0.0);
magmaDoubleComplex zero = MAGMA_Z_MAKE(0.0, 0.0);
magma_z_matrix A={Magma_CSR}, B={Magma_CSR}, B_d={Magma_CSR};
magma_z_matrix x={Magma_CSR}, b={Magma_CSR};
int i=1;
CHECK( magma_zparse_opts( argc, argv, &zopts, &i, queue ));
B.blocksize = zopts.blocksize;
B.alignment = zopts.alignment;
if ( zopts.solver_par.solver != Magma_PCG &&
zopts.solver_par.solver != Magma_PGMRES &&
zopts.solver_par.solver != Magma_PBICGSTAB &&
zopts.solver_par.solver != Magma_ITERREF &&
zopts.solver_par.solver != Magma_LOBPCG )
zopts.precond_par.solver = Magma_NONE;
CHECK( magma_zsolverinfo_init( &zopts.solver_par, &zopts.precond_par, queue ));
while( i < argc ) {
if ( strcmp("LAPLACE2D", argv[i]) == 0 && i+1 < argc ) { // Laplace test
i++;
magma_int_t laplace_size = atoi( argv[i] );
CHECK( magma_zm_5stencil( laplace_size, &A, queue ));
} else { // file-matrix test
CHECK( magma_z_csr_mtx( &A, argv[i], queue ));
}
printf( "\n# matrix info: %d-by-%d with %d nonzeros\n\n",
(int) A.num_rows,(int) A.num_cols,(int) A.nnz );
// for the eigensolver case
zopts.solver_par.ev_length = A.num_rows;
CHECK( magma_zeigensolverinfo_init( &zopts.solver_par, queue ));
// scale matrix
CHECK( magma_zmscale( &A, zopts.scaling, queue ));
CHECK( magma_zmconvert( A, &B, Magma_CSR, zopts.output_format, queue ));
CHECK( magma_zmtransfer( B, &B_d, Magma_CPU, Magma_DEV, queue ));
// vectors and initial guess
CHECK( magma_zvinit( &b, Magma_DEV, A.num_cols, 1, one, queue ));
//magma_zvinit( &x, Magma_DEV, A.num_cols, 1, one, queue );
//magma_z_spmv( one, B_d, x, zero, b, queue ); // b = A x
//magma_zmfree(&x, queue );
CHECK( magma_zvinit( &x, Magma_DEV, A.num_cols, 1, zero, queue ));
info = magma_z_solver( B_d, b, &x, &zopts, queue );
if( info != 0 ){
printf("error: solver returned: %s (%d).\n",
magma_strerror( info ), info );
}
magma_zsolverinfo( &zopts.solver_par, &zopts.precond_par, queue );
magma_zmfree(&B_d, queue );
magma_zmfree(&B, queue );
magma_zmfree(&A, queue );
magma_zmfree(&x, queue );
magma_zmfree(&b, queue );
i++;
}
cleanup:
magma_zmfree(&B_d, queue );
magma_zmfree(&B, queue );
magma_zmfree(&A, queue );
magma_zmfree(&x, queue );
magma_zmfree(&b, queue );
magma_zsolverinfo_free( &zopts.solver_par, &zopts.precond_par, queue );
magma_queue_destroy( queue );
TESTING_FINALIZE();
return info;
}
/* ////////////////////////////////////////////////////////////////////////////
-- testing csr matrix add
*/
int main( int argc, char** argv )
{
magma_int_t info = 0;
TESTING_INIT();
magma_queue_t queue=NULL;
magma_queue_create( &queue );
real_Double_t res;
magma_z_matrix A={Magma_CSR}, B={Magma_CSR}, B2={Magma_CSR},
A_d={Magma_CSR}, B_d={Magma_CSR}, C_d={Magma_CSR};
magmaDoubleComplex one = MAGMA_Z_MAKE(1.0, 0.0);
magmaDoubleComplex mone = MAGMA_Z_MAKE(-1.0, 0.0);
magma_int_t i=1;
if ( strcmp("LAPLACE2D", argv[i]) == 0 && i+1 < argc ) { // Laplace test
i++;
magma_int_t laplace_size = atoi( argv[i] );
CHECK( magma_zm_5stencil( laplace_size, &A, queue ));
} else { // file-matrix test
CHECK( magma_z_csr_mtx( &A, argv[i], queue ));
}
printf("%% matrix info: %d-by-%d with %d nonzeros\n",
int(A.num_rows), int(A.num_cols), int(A.nnz) );
i++;
if ( strcmp("LAPLACE2D", argv[i]) == 0 && i+1 < argc ) { // Laplace test
i++;
magma_int_t laplace_size = atoi( argv[i] );
CHECK( magma_zm_5stencil( laplace_size, &B, queue ));
} else { // file-matrix test
CHECK( magma_z_csr_mtx( &B, argv[i], queue ));
}
printf("%% matrix info: %d-by-%d with %d nonzeros\n",
int(B.num_rows), int(B.num_cols), int(B.nnz) );
CHECK( magma_zmtransfer( A, &A_d, Magma_CPU, Magma_DEV, queue ));
CHECK( magma_zmtransfer( B, &B_d, Magma_CPU, Magma_DEV, queue ));
CHECK( magma_zcuspaxpy( &one, A_d, &one, B_d, &C_d, queue ));
magma_zmfree(&B_d, queue );
CHECK( magma_zcuspaxpy( &mone, A_d, &one, C_d, &B_d, queue ));
CHECK( magma_zmtransfer( B_d, &B2, Magma_DEV, Magma_CPU, queue ));
magma_zmfree(&A_d, queue );
magma_zmfree(&B_d, queue );
magma_zmfree(&C_d, queue );
// check difference
CHECK( magma_zmdiff( B, B2, &res, queue ));
printf("%% ||A-B||_F = %8.2e\n", res);
if ( res < .000001 )
printf("%% tester matrix add: ok\n");
else
printf("%% tester matrix add: failed\n");
magma_zmfree(&A, queue );
magma_zmfree(&B, queue );
magma_zmfree(&B2, queue );
cleanup:
magma_zmfree(&A_d, queue );
magma_zmfree(&B_d, queue );
magma_zmfree(&C_d, queue );
magma_zmfree(&A, queue );
magma_zmfree(&B, queue );
magma_zmfree(&B2, queue );
magma_queue_destroy( queue );
TESTING_FINALIZE();
return info;
}
/* ////////////////////////////////////////////////////////////////////////////
-- testing sparse matrix vector product
*/
int main( int argc, char** argv )
{
TESTING_INIT();
magma_queue_t queue;
magma_queue_create( /*devices[ opts->device ],*/ &queue );
magma_z_sparse_matrix hA, hA_SELLP, hA_ELL, dA, dA_SELLP, dA_ELL;
hA_SELLP.blocksize = 8;
hA_SELLP.alignment = 8;
real_Double_t start, end, res;
magma_int_t *pntre;
magmaDoubleComplex c_one = MAGMA_Z_MAKE(1.0, 0.0);
magmaDoubleComplex c_zero = MAGMA_Z_MAKE(0.0, 0.0);
magma_int_t i, j;
for( i = 1; i < argc; ++i ) {
if ( strcmp("--blocksize", argv[i]) == 0 ) {
hA_SELLP.blocksize = atoi( argv[++i] );
} else if ( strcmp("--alignment", argv[i]) == 0 ) {
hA_SELLP.alignment = atoi( argv[++i] );
} else
break;
}
printf( "\n# usage: ./run_zspmv"
" [ --blocksize %d --alignment %d (for SELLP) ]"
" matrices \n\n", (int) hA_SELLP.blocksize, (int) hA_SELLP.alignment );
while( i < argc ) {
if ( strcmp("LAPLACE2D", argv[i]) == 0 && i+1 < argc ) { // Laplace test
i++;
magma_int_t laplace_size = atoi( argv[i] );
magma_zm_5stencil( laplace_size, &hA, queue );
} else { // file-matrix test
magma_z_csr_mtx( &hA, argv[i], queue );
}
printf( "\n# matrix info: %d-by-%d with %d nonzeros\n\n",
(int) hA.num_rows,(int) hA.num_cols,(int) hA.nnz );
real_Double_t FLOPS = 2.0*hA.nnz/1e9;
magma_z_vector hx, hy, dx, dy, hrefvec, hcheck;
// init CPU vectors
magma_z_vinit( &hx, Magma_CPU, hA.num_rows, c_zero, queue );
magma_z_vinit( &hy, Magma_CPU, hA.num_rows, c_zero, queue );
// init DEV vectors
magma_z_vinit( &dx, Magma_DEV, hA.num_rows, c_one, queue );
magma_z_vinit( &dy, Magma_DEV, hA.num_rows, c_zero, queue );
#ifdef MAGMA_WITH_MKL
// calling MKL with CSR
pntre = (magma_int_t*)malloc( (hA.num_rows+1)*sizeof(magma_int_t) );
pntre[0] = 0;
for (j=0; j<hA.num_rows; j++ ) {
pntre[j] = hA.row[j+1];
}
MKL_INT num_rows = hA.num_rows;
MKL_INT num_cols = hA.num_cols;
MKL_INT nnz = hA.nnz;
MKL_INT *col;
TESTING_MALLOC_CPU( col, MKL_INT, nnz );
for( magma_int_t t=0; t < hA.nnz; ++t ) {
col[ t ] = hA.col[ t ];
}
MKL_INT *row;
TESTING_MALLOC_CPU( row, MKL_INT, num_rows );
for( magma_int_t t=0; t < hA.num_rows; ++t ) {
row[ t ] = hA.col[ t ];
}
start = magma_wtime();
for (j=0; j<10; j++ ) {
mkl_zcsrmv( "N", &num_rows, &num_cols,
MKL_ADDR(&c_one), "GFNC", MKL_ADDR(hA.val),
col, row, pntre,
MKL_ADDR(hx.val),
MKL_ADDR(&c_zero), MKL_ADDR(hy.val) );
}
end = magma_wtime();
printf( "\n > MKL : %.2e seconds %.2e GFLOP/s (CSR).\n",
(end-start)/10, FLOPS*10/(end-start) );
TESTING_FREE_CPU( row );
TESTING_FREE_CPU( col );
free(pntre);
#endif // MAGMA_WITH_MKL
// copy matrix to GPU
magma_z_mtransfer( hA, &dA, Magma_CPU, Magma_DEV, queue );
// SpMV on GPU (CSR) -- this is the reference!
start = magma_sync_wtime( queue );
for (j=0; j<10; j++)
magma_z_spmv( c_one, dA, dx, c_zero, dy, queue );
end = magma_sync_wtime( queue );
printf( " > MAGMA: %.2e seconds %.2e GFLOP/s (standard CSR).\n",
(end-start)/10, FLOPS*10/(end-start) );
magma_z_mfree(&dA, queue );
magma_z_vtransfer( dy, &hrefvec , Magma_DEV, Magma_CPU, queue );
// convert to ELL and copy to GPU
magma_z_mconvert( hA, &hA_ELL, Magma_CSR, Magma_ELL, queue );
magma_z_mtransfer( hA_ELL, &dA_ELL, Magma_CPU, Magma_DEV, queue );
magma_z_mfree(&hA_ELL, queue );
magma_z_vfree( &dy, queue );
magma_z_vinit( &dy, Magma_DEV, hA.num_rows, c_zero, queue );
// SpMV on GPU (ELL)
start = magma_sync_wtime( queue );
for (j=0; j<10; j++)
magma_z_spmv( c_one, dA_ELL, dx, c_zero, dy, queue );
end = magma_sync_wtime( queue );
printf( " > MAGMA: %.2e seconds %.2e GFLOP/s (standard ELL).\n",
(end-start)/10, FLOPS*10/(end-start) );
magma_z_mfree(&dA_ELL, queue );
magma_z_vtransfer( dy, &hcheck , Magma_DEV, Magma_CPU, queue );
res = 0.0;
for(magma_int_t k=0; k<hA.num_rows; k++ )
res=res + MAGMA_Z_REAL(hcheck.val[k]) - MAGMA_Z_REAL(hrefvec.val[k]);
if ( res < .000001 )
printf("# tester spmv ELL: ok\n");
else
printf("# tester spmv ELL: failed\n");
magma_z_vfree( &hcheck, queue );
// convert to SELLP and copy to GPU
magma_z_mconvert( hA, &hA_SELLP, Magma_CSR, Magma_SELLP, queue );
magma_z_mtransfer( hA_SELLP, &dA_SELLP, Magma_CPU, Magma_DEV, queue );
magma_z_mfree(&hA_SELLP, queue );
magma_z_vfree( &dy, queue );
magma_z_vinit( &dy, Magma_DEV, hA.num_rows, c_zero, queue );
// SpMV on GPU (SELLP)
start = magma_sync_wtime( queue );
for (j=0; j<10; j++)
magma_z_spmv( c_one, dA_SELLP, dx, c_zero, dy, queue );
end = magma_sync_wtime( queue );
printf( " > MAGMA: %.2e seconds %.2e GFLOP/s (SELLP).\n",
(end-start)/10, FLOPS*10/(end-start) );
magma_z_vtransfer( dy, &hcheck , Magma_DEV, Magma_CPU, queue );
res = 0.0;
for(magma_int_t k=0; k<hA.num_rows; k++ )
res=res + MAGMA_Z_REAL(hcheck.val[k]) - MAGMA_Z_REAL(hrefvec.val[k]);
printf("# |x-y|_F = %8.2e\n", res);
if ( res < .000001 )
printf("# tester spmv SELL-P: ok\n");
else
printf("# tester spmv SELL-P: failed\n");
magma_z_vfree( &hcheck, queue );
magma_z_mfree(&dA_SELLP, queue );
// SpMV on GPU (CUSPARSE - CSR)
// CUSPARSE context //
cusparseHandle_t cusparseHandle = 0;
cusparseStatus_t cusparseStatus;
cusparseStatus = cusparseCreate(&cusparseHandle);
cusparseSetStream( cusparseHandle, queue );
cusparseMatDescr_t descr = 0;
cusparseStatus = cusparseCreateMatDescr(&descr);
cusparseSetMatType(descr,CUSPARSE_MATRIX_TYPE_GENERAL);
cusparseSetMatIndexBase(descr,CUSPARSE_INDEX_BASE_ZERO);
magmaDoubleComplex alpha = c_one;
magmaDoubleComplex beta = c_zero;
magma_z_vfree( &dy, queue );
magma_z_vinit( &dy, Magma_DEV, hA.num_rows, c_zero, queue );
// copy matrix to GPU
magma_z_mtransfer( hA, &dA, Magma_CPU, Magma_DEV, queue );
start = magma_sync_wtime( queue );
for (j=0; j<10; j++)
cusparseStatus =
cusparseZcsrmv(cusparseHandle,CUSPARSE_OPERATION_NON_TRANSPOSE,
hA.num_rows, hA.num_cols, hA.nnz, &alpha, descr,
dA.dval, dA.drow, dA.dcol, dx.dval, &beta, dy.dval);
end = magma_sync_wtime( queue );
if (cusparseStatus != 0) printf("error in cuSPARSE CSR\n");
printf( " > CUSPARSE: %.2e seconds %.2e GFLOP/s (CSR).\n",
(end-start)/10, FLOPS*10/(end-start) );
cusparseMatDescr_t descrA;
cusparseStatus = cusparseCreateMatDescr(&descrA);
if (cusparseStatus != 0) printf("error\n");
cusparseHybMat_t hybA;
cusparseStatus = cusparseCreateHybMat( &hybA );
if (cusparseStatus != 0) printf("error\n");
magma_z_vtransfer( dy, &hcheck , Magma_DEV, Magma_CPU, queue );
res = 0.0;
for(magma_int_t k=0; k<hA.num_rows; k++ )
res=res + MAGMA_Z_REAL(hcheck.val[k]) - MAGMA_Z_REAL(hrefvec.val[k]);
printf("# |x-y|_F = %8.2e\n", res);
if ( res < .000001 )
printf("# tester spmv cuSPARSE CSR: ok\n");
else
printf("# tester spmv cuSPARSE CSR: failed\n");
magma_z_vfree( &hcheck, queue );
magma_z_vfree( &dy, queue );
magma_z_vinit( &dy, Magma_DEV, hA.num_rows, c_zero, queue );
cusparseZcsr2hyb(cusparseHandle, hA.num_rows, hA.num_cols,
descrA, dA.dval, dA.drow, dA.dcol,
hybA, 0, CUSPARSE_HYB_PARTITION_AUTO);
start = magma_sync_wtime( queue );
for (j=0; j<10; j++)
cusparseStatus =
cusparseZhybmv( cusparseHandle, CUSPARSE_OPERATION_NON_TRANSPOSE,
&alpha, descrA, hybA,
dx.dval, &beta, dy.dval);
end = magma_sync_wtime( queue );
if (cusparseStatus != 0) printf("error in cuSPARSE HYB\n");
printf( " > CUSPARSE: %.2e seconds %.2e GFLOP/s (HYB).\n",
(end-start)/10, FLOPS*10/(end-start) );
magma_z_vtransfer( dy, &hcheck , Magma_DEV, Magma_CPU, queue );
res = 0.0;
for(magma_int_t k=0; k<hA.num_rows; k++ )
res=res + MAGMA_Z_REAL(hcheck.val[k]) - MAGMA_Z_REAL(hrefvec.val[k]);
printf("# |x-y|_F = %8.2e\n", res);
if ( res < .000001 )
printf("# tester spmv cuSPARSE HYB: ok\n");
else
printf("# tester spmv cuSPARSE HYB: failed\n");
magma_z_vfree( &hcheck, queue );
cusparseDestroyMatDescr( descrA );
cusparseDestroyHybMat( hybA );
cusparseDestroy( cusparseHandle );
magma_z_mfree(&dA, queue );
printf("\n\n");
// free CPU memory
magma_z_mfree(&hA, queue );
magma_z_vfree(&hx, queue );
magma_z_vfree(&hy, queue );
magma_z_vfree(&hrefvec, queue );
// free GPU memory
magma_z_vfree(&dx, queue );
magma_z_vfree(&dy, queue );
i++;
}
magma_queue_destroy( queue );
TESTING_FINALIZE();
return 0;
}
/* ////////////////////////////////////////////////////////////////////////////
-- testing any solver
*/
int main( int argc, char** argv )
{
magma_int_t info = 0;
TESTING_CHECK( magma_init() );
magma_print_environment();
magma_zopts zopts;
magma_queue_t queue=NULL;
magma_queue_create( 0, &queue );
real_Double_t res;
magma_z_matrix A={Magma_CSR}, A2={Magma_CSR},
A3={Magma_CSR}, A4={Magma_CSR}, A5={Magma_CSR};
int i=1;
TESTING_CHECK( magma_zparse_opts( argc, argv, &zopts, &i, queue ));
while( i < argc ) {
if ( strcmp("LAPLACE2D", argv[i]) == 0 && i+1 < argc ) { // Laplace test
i++;
magma_int_t laplace_size = atoi( argv[i] );
TESTING_CHECK( magma_zm_5stencil( laplace_size, &A, queue ));
} else { // file-matrix test
TESTING_CHECK( magma_z_csr_mtx( &A, argv[i], queue ));
}
printf("%% matrix info: %lld-by-%lld with %lld nonzeros\n",
(long long) A.num_rows, (long long) A.num_cols, (long long) A.nnz );
// filename for temporary matrix storage
const char *filename = "testmatrix.mtx";
// write to file
TESTING_CHECK( magma_zwrite_csrtomtx( A, filename, queue ));
// read from file
TESTING_CHECK( magma_z_csr_mtx( &A2, filename, queue ));
// delete temporary matrix
unlink( filename );
//visualize
printf("A2:\n");
TESTING_CHECK( magma_zprint_matrix( A2, queue ));
//visualize
TESTING_CHECK( magma_zmconvert(A2, &A4, Magma_CSR, Magma_CSRL, queue ));
printf("A4:\n");
TESTING_CHECK( magma_zprint_matrix( A4, queue ));
TESTING_CHECK( magma_zmconvert(A4, &A5, Magma_CSR, Magma_ELL, queue ));
printf("A5:\n");
TESTING_CHECK( magma_zprint_matrix( A5, queue ));
// pass it to another application and back
magma_int_t m, n;
magma_index_t *row, *col;
magmaDoubleComplex *val=NULL;
TESTING_CHECK( magma_zcsrget( A2, &m, &n, &row, &col, &val, queue ));
TESTING_CHECK( magma_zcsrset( m, n, row, col, val, &A3, queue ));
TESTING_CHECK( magma_zmdiff( A, A2, &res, queue ));
printf("%% ||A-B||_F = %8.2e\n", res);
if ( res < .000001 )
printf("%% tester IO: ok\n");
else
printf("%% tester IO: failed\n");
TESTING_CHECK( magma_zmdiff( A, A3, &res, queue ));
printf("%% ||A-B||_F = %8.2e\n", res);
if ( res < .000001 )
printf("%% tester matrix interface: ok\n");
else
printf("%% tester matrix interface: failed\n");
magma_zmfree(&A, queue );
magma_zmfree(&A2, queue );
magma_zmfree(&A4, queue );
magma_zmfree(&A5, queue );
i++;
}
magma_queue_destroy( queue );
TESTING_CHECK( magma_finalize() );
return info;
}
/* ////////////////////////////////////////////////////////////////////////////
-- testing any solver
*/
int main( int argc, char** argv )
{
TESTING_INIT();
magma_zopts zopts;
magma_queue_t queue;
magma_queue_create( /*devices[ opts->device ],*/ &queue );
int i=1;
magma_zparse_opts( argc, argv, &zopts, &i, queue );
real_Double_t res;
magma_z_sparse_matrix A, A2, A3, A4, A5;
while( i < argc ) {
if ( strcmp("LAPLACE2D", argv[i]) == 0 && i+1 < argc ) { // Laplace test
i++;
magma_int_t laplace_size = atoi( argv[i] );
magma_zm_5stencil( laplace_size, &A, queue );
} else { // file-matrix test
magma_z_csr_mtx( &A, argv[i], queue );
}
printf( "# matrix info: %d-by-%d with %d nonzeros\n",
(int) A.num_rows,(int) A.num_cols,(int) A.nnz );
// filename for temporary matrix storage
const char *filename = "testmatrix.mtx";
// write to file
write_z_csrtomtx( A, filename, queue );
// read from file
magma_z_csr_mtx( &A2, filename, queue );
// delete temporary matrix
unlink( filename );
//visualize
printf("A2:\n");
magma_z_mvisu( A2, queue );
//visualize
magma_z_mconvert(A2, &A4, Magma_CSR, Magma_CSRL, queue );
printf("A4:\n");
magma_z_mvisu( A4, queue );
magma_z_mconvert(A4, &A5, Magma_CSR, Magma_ELL, queue );
printf("A5:\n");
magma_z_mvisu( A5, queue );
// pass it to another application and back
magma_int_t m, n;
magma_index_t *row, *col;
magmaDoubleComplex *val;
magma_zcsrget( A2, &m, &n, &row, &col, &val, queue );
magma_zcsrset( m, n, row, col, val, &A3, queue );
magma_zmdiff( A, A2, &res, queue );
printf("# ||A-B||_F = %8.2e\n", res);
if ( res < .000001 )
printf("# tester IO: ok\n");
else
printf("# tester IO: failed\n");
magma_zmdiff( A, A3, &res, queue );
printf("# ||A-B||_F = %8.2e\n", res);
if ( res < .000001 )
printf("# tester matrix interface: ok\n");
else
printf("# tester matrix interface: failed\n");
magma_z_mfree(&A, queue );
magma_z_mfree(&A2, queue );
magma_z_mfree(&A4, queue );
magma_z_mfree(&A5, queue );
i++;
}
magma_queue_destroy( queue );
TESTING_FINALIZE();
return 0;
}
