/* ////////////////////////////////////////////////////////////////////////////
-- Debugging file
*/
int main( int argc, char** argv)
{
TESTING_INIT();
magma_d_solver_par solver_par;
magma_d_preconditioner precond_par;
solver_par.epsilon = 10e-16;
solver_par.maxiter = 1000;
solver_par.verbose = 0;
precond_par.solver = Magma_JACOBI;
magma_dsolverinfo_init( &solver_par, &precond_par );
double one = MAGMA_D_MAKE(1.0, 0.0);
double zero = MAGMA_D_MAKE(0.0, 0.0);
magma_d_sparse_matrix A, B, B_d;
magma_d_vector x, b;
// generate matrix of desired structure and size
magma_int_t n=10; // size is n*n
magma_int_t nn = n*n;
magma_int_t offdiags = 2;
magma_index_t *diag_offset;
double *diag_vals;
magma_dmalloc_cpu( &diag_vals, offdiags+1 );
magma_index_malloc_cpu( &diag_offset, offdiags+1 );
diag_offset[0] = 0;
diag_offset[1] = 1;
diag_offset[2] = n;
diag_vals[0] = MAGMA_D_MAKE( 4.0, 0.0 );
diag_vals[1] = MAGMA_D_MAKE( -1.0, 0.0 );
diag_vals[2] = MAGMA_D_MAKE( -1.0, 0.0 );
magma_dmgenerator( nn, offdiags, diag_offset, diag_vals, &A );
// convert marix into desired format
B.storage_type = Magma_SELLC;
B.blocksize = 8;
B.alignment = 8;
// scale matrix
magma_dmscale( &A, Magma_UNITDIAG );
magma_d_mconvert( A, &B, Magma_CSR, B.storage_type );
magma_d_mtransfer( B, &B_d, Magma_CPU, Magma_DEV );
// vectors and initial guess
magma_d_vinit( &b, Magma_DEV, A.num_cols, one );
magma_d_vinit( &x, Magma_DEV, A.num_cols, one );
magma_d_spmv( one, B_d, x, zero, b ); // b = A x
magma_d_vfree(&x);
magma_d_vinit( &x, Magma_DEV, A.num_cols, zero );
// solver
magma_dpcg( B_d, b, &x, &solver_par, &precond_par );
// solverinfo
magma_dsolverinfo( &solver_par, &precond_par );
magma_dsolverinfo_free( &solver_par, &precond_par );
magma_d_mfree(&B_d);
magma_d_mfree(&B);
magma_d_mfree(&A);
magma_d_vfree(&x);
magma_d_vfree(&b);
TESTING_FINALIZE();
return 0;
}
/* ////////////////////////////////////////////////////////////////////////////
-- testing any solver
*/
int main( int argc, char** argv )
{
TESTING_INIT();
magma_dopts zopts;
magma_queue_t queue;
magma_queue_create( /*devices[ opts->device ],*/ &queue );
int i=1;
magma_dparse_opts( argc, argv, &zopts, &i, queue );
real_Double_t res;
magma_d_sparse_matrix Z, A, AT, A2, B, B_d;
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] );
magma_dm_5stencil( laplace_size, &Z, queue );
} else { // file-matrix test
magma_d_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 );
// scale matrix
magma_dmscale( &Z, zopts.scaling, queue );
// remove nonzeros in matrix
magma_dmcsrcompressor( &Z, queue );
// convert to be non-symmetric
magma_d_mconvert( Z, &A, Magma_CSR, Magma_CSRL, queue );
// transpose
magma_d_mtranspose( A, &AT, queue );
// convert, copy back and forth to check everything works
magma_d_mconvert( AT, &B, Magma_CSR, zopts.output_format, queue );
magma_d_mfree(&AT, queue );
magma_d_mtransfer( B, &B_d, Magma_CPU, Magma_DEV, queue );
magma_d_mfree(&B, queue );
magma_dmcsrcompressor_gpu( &B_d, queue );
magma_d_mtransfer( B_d, &B, Magma_DEV, Magma_CPU, queue );
magma_d_mfree(&B_d, queue );
magma_d_mconvert( B, &AT, zopts.output_format,Magma_CSR, queue );
magma_d_mfree(&B, queue );
// transpose back
magma_d_mtranspose( AT, &A2, queue );
magma_d_mfree(&AT, queue );
magma_dmdiff( A, A2, &res, queue);
printf("# ||A-B||_F = %8.2e\n", res);
if ( res < .000001 )
printf("# tester: ok\n");
else
printf("# tester: failed\n");
magma_d_mfree(&A, queue );
magma_d_mfree(&A2, queue );
magma_d_mfree(&Z, queue );
i++;
}
magma_queue_destroy( queue );
TESTING_FINALIZE();
return 0;
}
extern "C" magma_int_t
magma_dmlumerge(
magma_d_sparse_matrix L,
magma_d_sparse_matrix U,
magma_d_sparse_matrix *A,
magma_queue_t queue ){
if( L.storage_type == Magma_CSR && U.storage_type == Magma_CSR ){
if( L.memory_location == Magma_CPU && U.memory_location == Magma_CPU ){
magma_d_mtransfer( L, A, Magma_CPU, Magma_CPU, queue );
magma_free_cpu( A->col );
magma_free_cpu( A->val );
// make sure it is strictly lower triangular
magma_int_t z = 0;
for(magma_int_t i=0; i<A->num_rows; i++){
for(magma_int_t j=L.row[i]; j<L.row[i+1]; j++){
if( L.col[j] < i ){// make sure it is strictly lower triangular
z++;
}
}
for(magma_int_t j=U.row[i]; j<U.row[i+1]; j++){
z++;
}
}
A->nnz = z;
// fill A with the new structure;
magma_int_t stat_cpu = 0;
stat_cpu += magma_index_malloc_cpu( &A->col, A->nnz );
stat_cpu += magma_dmalloc_cpu( &A->val, A->nnz );
if( stat_cpu != 0 ){
magma_d_mfree( A, queue );
printf("error: memory allocation.\n");
return MAGMA_ERR_HOST_ALLOC;
}
z = 0;
for(magma_int_t i=0; i<A->num_rows; i++){
A->row[i] = z;
for(magma_int_t j=L.row[i]; j<L.row[i+1]; j++){
if( L.col[j] < i ){// make sure it is strictly lower triangular
A->col[z] = L.col[j];
A->val[z] = L.val[j];
z++;
}
}
for(magma_int_t j=U.row[i]; j<U.row[i+1]; j++){
A->col[z] = U.col[j];
A->val[z] = U.val[j];
z++;
}
}
A->row[A->num_rows] = z;
A->nnz = z;
return MAGMA_SUCCESS;
}
else{
printf("error: matrix not on CPU.\n");
return MAGMA_SUCCESS;
}
}
else{
printf("error: matrix not on CPU.\n");
return MAGMA_SUCCESS;
}
}
/* ////////////////////////////////////////////////////////////////////////////
-- testing any solver
*/
int main( int argc, char** argv )
{
TESTING_INIT();
magma_dopts zopts;
magma_queue_t queue;
magma_queue_create( /*devices[ opts->device ],*/ &queue );
int i=1;
magma_dparse_opts( argc, argv, &zopts, &i, queue );
real_Double_t res;
magma_d_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_dm_5stencil( laplace_size, &A, queue );
} else { // file-matrix test
magma_d_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_d_csrtomtx( A, filename, queue );
// read from file
magma_d_csr_mtx( &A2, filename, queue );
// delete temporary matrix
unlink( filename );
//visualize
printf("A2:\n");
magma_d_mvisu( A2, queue );
//visualize
magma_d_mconvert(A2, &A4, Magma_CSR, Magma_CSRL, queue );
printf("A4:\n");
magma_d_mvisu( A4, queue );
magma_d_mconvert(A4, &A5, Magma_CSR, Magma_ELL, queue );
printf("A5:\n");
magma_d_mvisu( A5, queue );
// pass it to another application and back
magma_int_t m, n;
magma_index_t *row, *col;
double *val;
magma_dcsrget( A2, &m, &n, &row, &col, &val, queue );
magma_dcsrset( m, n, row, col, val, &A3, queue );
magma_dmdiff( 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_dmdiff( 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_d_mfree(&A, queue );
magma_d_mfree(&A2, queue );
magma_d_mfree(&A4, queue );
magma_d_mfree(&A5, queue );
i++;
}
magma_queue_destroy( queue );
TESTING_FINALIZE();
return 0;
}
extern "C" magma_int_t
magma_dpastixsetup(
magma_d_sparse_matrix A, magma_d_vector b,
magma_d_preconditioner *precond,
magma_queue_t queue )
{
#if defined(HAVE_PASTIX)
#if defined(PRECISION_d)
pastix_data_t *pastix_data = NULL; /* Pointer to a storage structure needed by pastix */
pastix_int_t ncol; /* Size of the matrix */
pastix_int_t *colptr = NULL; /* Indexes of first element of each column in row and values */
pastix_int_t *rows = NULL; /* Row of each element of the matrix */
pastix_float_t *values = NULL; /* Value of each element of the matrix */
pastix_float_t *rhs = NULL; /* right hand side */
pastix_int_t *iparm = NULL; /* integer parameters for pastix */
double *dparm = NULL; /* floating parameters for pastix */
pastix_int_t *perm = NULL; /* Permutation tabular */
pastix_int_t *invp = NULL; /* Reverse permutation tabular */
pastix_int_t mat_type;
magma_d_sparse_matrix A_h1, B;
magma_d_vector diag, c_t, b_h;
magma_d_vinit( &c_t, Magma_CPU, A.num_rows, MAGMA_D_ZERO, queue );
magma_d_vinit( &diag, Magma_CPU, A.num_rows, MAGMA_D_ZERO, queue );
magma_d_vtransfer( b, &b_h, A.memory_location, Magma_CPU, queue );
if ( A.storage_type != Magma_CSR ) {
magma_d_mtransfer( A, &A_h1, A.memory_location, Magma_CPU, queue );
magma_d_mconvert( A_h1, &B, A_h1.storage_type, Magma_CSR, queue );
}
else {
magma_d_mtransfer( A, &B, A.memory_location, Magma_CPU, queue );
}
rhs = (pastix_float_t*) b_h.dval;
ncol = B.num_rows;
colptr = B.drow;
rows = B.dcol;
values = (pastix_float_t*) B.dval;
mat_type = API_SYM_NO;
iparm = (pastix_int_t*)malloc(IPARM_SIZE*sizeof(pastix_int_t));
dparm = (pastix_float_t*)malloc(DPARM_SIZE*sizeof(pastix_float_t));
/*******************************************/
/* Initialize parameters to default values */
/*******************************************/
iparm[IPARM_MODIFY_PARAMETER] = API_NO;
pastix(&pastix_data, MPI_COMM_WORLD,
ncol, colptr, rows, values,
perm, invp, rhs, 1, iparm, dparm);
iparm[IPARM_THREAD_NBR] = 16;
iparm[IPARM_SYM] = mat_type;
iparm[IPARM_FACTORIZATION] = API_FACT_LU;
iparm[IPARM_VERBOSE] = API_VERBOSE_YES;
iparm[IPARM_ORDERING] = API_ORDER_SCOTCH;
iparm[IPARM_INCOMPLETE] = API_NO;
iparm[IPARM_RHS_MAKING] = API_RHS_B;
//iparm[IPARM_AMALGAMATION] = 5;
iparm[IPARM_LEVEL_OF_FILL] = 0;
/* if (incomplete == 1)
{
dparm[DPARM_EPSILON_REFINEMENT] = 1e-7;
}
*/
/*
* Matrix needs :
* - to be in fortran numbering
* - to have only the lower triangular part in symmetric case
* - to have a graph with a symmetric structure in unsymmetric case
* If those criteria are not matched, the csc will be reallocated and changed.
*/
iparm[IPARM_MATRIX_VERIFICATION] = API_YES;
perm = (pastix_int_t*)malloc(ncol*sizeof(pastix_int_t));
invp = (pastix_int_t*)malloc(ncol*sizeof(pastix_int_t));
/*******************************************/
/* Step 1 - Ordering / Scotch */
/* Perform it only when the pattern of */
/* matrix change. */
/* eg: mesh refinement */
/* In many cases users can simply go from */
/* API_TASK_ORDERING to API_TASK_ANALYSE */
/* in one call. */
/*******************************************/
/*******************************************/
/* Step 2 - Symbolic factorization */
/* Perform it only when the pattern of */
/* matrix change. */
/*******************************************/
/*******************************************/
/* Step 3 - Mapping and Compute scheduling */
/* Perform it only when the pattern of */
/* matrix change. */
/*******************************************/
/*******************************************/
/* Step 4 - Numerical Factorisation */
/* Perform it each time the values of the */
/* matrix changed. */
/*******************************************/
iparm[IPARM_START_TASK] = API_TASK_ORDERING;
iparm[IPARM_END_TASK] = API_TASK_NUMFACT;
pastix(&pastix_data, MPI_COMM_WORLD,
ncol, colptr, rows, values,
perm, invp, NULL, 1, iparm, dparm);
precond->int_array_1 = (magma_int_t*) perm;
precond->int_array_2 = (magma_int_t*) invp;
precond->M.dval = (double*) values;
precond->M.dcol = (magma_int_t*) colptr;
precond->M.drow = (magma_int_t*) rows;
precond->M.num_rows = A.num_rows;
precond->M.num_cols = A.num_cols;
precond->M.memory_location = Magma_CPU;
precond->pastix_data = pastix_data;
precond->iparm = iparm;
precond->dparm = dparm;
if ( A.storage_type != Magma_CSR) {
magma_d_mfree( &A_h1, queue );
}
magma_d_vfree( &b_h, queue );
magma_d_mfree( &B, queue );
#else
printf( "error: only double precision supported yet.\n");
#endif
#else
printf( "error: pastix not available.\n");
#endif
return MAGMA_SUCCESS;
}
/* ////////////////////////////////////////////////////////////////////////////
-- 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_d_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;
double c_one = MAGMA_D_MAKE(1.0, 0.0);
double c_zero = MAGMA_D_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_dspmv"
" [ --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_dm_5stencil( laplace_size, &hA, queue );
} else { // file-matrix test
magma_d_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_d_vector hx, hy, dx, dy, hrefvec, hcheck;
// init CPU vectors
magma_d_vinit( &hx, Magma_CPU, hA.num_rows, c_zero, queue );
magma_d_vinit( &hy, Magma_CPU, hA.num_rows, c_zero, queue );
// init DEV vectors
magma_d_vinit( &dx, Magma_DEV, hA.num_rows, c_one, queue );
magma_d_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_dcsrmv( "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_d_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_d_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_d_mfree(&dA, queue );
magma_d_vtransfer( dy, &hrefvec , Magma_DEV, Magma_CPU, queue );
// convert to ELL and copy to GPU
magma_d_mconvert( hA, &hA_ELL, Magma_CSR, Magma_ELL, queue );
magma_d_mtransfer( hA_ELL, &dA_ELL, Magma_CPU, Magma_DEV, queue );
magma_d_mfree(&hA_ELL, queue );
magma_d_vfree( &dy, queue );
magma_d_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_d_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_d_mfree(&dA_ELL, queue );
magma_d_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_D_REAL(hcheck.val[k]) - MAGMA_D_REAL(hrefvec.val[k]);
if ( res < .000001 )
printf("# tester spmv ELL: ok\n");
else
printf("# tester spmv ELL: failed\n");
magma_d_vfree( &hcheck, queue );
// convert to SELLP and copy to GPU
magma_d_mconvert( hA, &hA_SELLP, Magma_CSR, Magma_SELLP, queue );
magma_d_mtransfer( hA_SELLP, &dA_SELLP, Magma_CPU, Magma_DEV, queue );
magma_d_mfree(&hA_SELLP, queue );
magma_d_vfree( &dy, queue );
magma_d_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_d_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_d_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_D_REAL(hcheck.val[k]) - MAGMA_D_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_d_vfree( &hcheck, queue );
magma_d_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);
double alpha = c_one;
double beta = c_zero;
magma_d_vfree( &dy, queue );
magma_d_vinit( &dy, Magma_DEV, hA.num_rows, c_zero, queue );
// copy matrix to GPU
magma_d_mtransfer( hA, &dA, Magma_CPU, Magma_DEV, queue );
start = magma_sync_wtime( queue );
for (j=0; j<10; j++)
cusparseStatus =
cusparseDcsrmv(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_d_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_D_REAL(hcheck.val[k]) - MAGMA_D_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_d_vfree( &hcheck, queue );
magma_d_vfree( &dy, queue );
magma_d_vinit( &dy, Magma_DEV, hA.num_rows, c_zero, queue );
cusparseDcsr2hyb(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 =
cusparseDhybmv( 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_d_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_D_REAL(hcheck.val[k]) - MAGMA_D_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_d_vfree( &hcheck, queue );
cusparseDestroyMatDescr( descrA );
cusparseDestroyHybMat( hybA );
cusparseDestroy( cusparseHandle );
magma_d_mfree(&dA, queue );
printf("\n\n");
// free CPU memory
magma_d_mfree(&hA, queue );
magma_d_vfree(&hx, queue );
magma_d_vfree(&hy, queue );
magma_d_vfree(&hrefvec, queue );
// free GPU memory
magma_d_vfree(&dx, queue );
magma_d_vfree(&dy, queue );
i++;
}
magma_queue_destroy( queue );
TESTING_FINALIZE();
return 0;
}
extern "C" magma_int_t
magma_dcumiccsetup(
magma_d_sparse_matrix A, magma_d_preconditioner *precond,
magma_queue_t queue )
{
magma_d_sparse_matrix hA, hACSR, U, hD, hR, hAt;
magma_d_mtransfer( A, &hA, A.memory_location, Magma_CPU, queue );
U.diagorder_type = Magma_VALUE;
magma_d_mconvert( hA, &hACSR, hA.storage_type, Magma_CSR, queue );
magma_d_mconvert( hACSR, &U, Magma_CSR, Magma_CSRL, queue );
magma_d_mfree( &hACSR, queue );
magma_d_mtransfer(U, &(precond->M), Magma_CPU, Magma_DEV, queue );
// CUSPARSE context //
cusparseHandle_t cusparseHandle;
cusparseStatus_t cusparseStatus;
cusparseStatus = cusparseCreate(&cusparseHandle);
cusparseSetStream( cusparseHandle, queue );
if (cusparseStatus != 0) printf("error in Handle.\n");
cusparseMatDescr_t descrA;
cusparseStatus = cusparseCreateMatDescr(&descrA);
if (cusparseStatus != 0) printf("error in MatrDescr.\n");
cusparseStatus =
cusparseSetMatType(descrA,CUSPARSE_MATRIX_TYPE_SYMMETRIC);
if (cusparseStatus != 0) printf("error in MatrType.\n");
cusparseStatus =
cusparseSetMatDiagType (descrA, CUSPARSE_DIAG_TYPE_NON_UNIT);
if (cusparseStatus != 0) printf("error in DiagType.\n");
cusparseStatus =
cusparseSetMatIndexBase(descrA,CUSPARSE_INDEX_BASE_ZERO);
if (cusparseStatus != 0) printf("error in IndexBase.\n");
cusparseStatus =
cusparseSetMatFillMode(descrA,CUSPARSE_FILL_MODE_LOWER);
if (cusparseStatus != 0) printf("error in fillmode.\n");
cusparseStatus =
cusparseCreateSolveAnalysisInfo( &(precond->cuinfo) );
if (cusparseStatus != 0) printf("error in info.\n");
// end CUSPARSE context //
cusparseStatus =
cusparseDcsrsm_analysis( cusparseHandle,
CUSPARSE_OPERATION_NON_TRANSPOSE,
precond->M.num_rows, precond->M.nnz, descrA,
precond->M.dval, precond->M.drow, precond->M.dcol,
precond->cuinfo);
if (cusparseStatus != 0) printf("error in analysis IC.\n");
cusparseStatus =
cusparseDcsric0( cusparseHandle, CUSPARSE_OPERATION_NON_TRANSPOSE,
precond->M.num_rows, descrA,
precond->M.dval,
precond->M.drow,
precond->M.dcol,
precond->cuinfo);
cusparseStatus =
cusparseDestroySolveAnalysisInfo( precond->cuinfo );
if (cusparseStatus != 0) printf("error in info-free.\n");
if (cusparseStatus != 0) printf("error in ICC.\n");
cusparseMatDescr_t descrL;
cusparseStatus = cusparseCreateMatDescr(&descrL);
if (cusparseStatus != 0) printf("error in MatrDescr.\n");
cusparseStatus =
cusparseSetMatType(descrL,CUSPARSE_MATRIX_TYPE_TRIANGULAR);
if (cusparseStatus != 0) printf("error in MatrType.\n");
cusparseStatus =
cusparseSetMatDiagType (descrL, CUSPARSE_DIAG_TYPE_NON_UNIT);
if (cusparseStatus != 0) printf("error in DiagType.\n");
cusparseStatus =
cusparseSetMatIndexBase(descrL,CUSPARSE_INDEX_BASE_ZERO);
if (cusparseStatus != 0) printf("error in IndexBase.\n");
cusparseStatus =
cusparseSetMatFillMode(descrL,CUSPARSE_FILL_MODE_LOWER);
if (cusparseStatus != 0) printf("error in fillmode.\n");
cusparseStatus = cusparseCreateSolveAnalysisInfo(&precond->cuinfoL);
if (cusparseStatus != 0) printf("error in info.\n");
cusparseStatus =
cusparseDcsrsm_analysis(cusparseHandle,
CUSPARSE_OPERATION_NON_TRANSPOSE, precond->M.num_rows,
precond->M.nnz, descrL,
precond->M.dval, precond->M.drow, precond->M.dcol, precond->cuinfoL );
if (cusparseStatus != 0) printf("error in analysis L.\n");
cusparseDestroyMatDescr( descrL );
cusparseMatDescr_t descrU;
cusparseStatus = cusparseCreateMatDescr(&descrU);
if (cusparseStatus != 0) printf("error in MatrDescr.\n");
cusparseStatus =
cusparseSetMatType(descrU,CUSPARSE_MATRIX_TYPE_TRIANGULAR);
if (cusparseStatus != 0) printf("error in MatrType.\n");
cusparseStatus =
cusparseSetMatDiagType (descrU, CUSPARSE_DIAG_TYPE_NON_UNIT);
if (cusparseStatus != 0) printf("error in DiagType.\n");
cusparseStatus =
cusparseSetMatIndexBase(descrU,CUSPARSE_INDEX_BASE_ZERO);
if (cusparseStatus != 0) printf("error in IndexBase.\n");
cusparseStatus =
cusparseSetMatFillMode(descrU,CUSPARSE_FILL_MODE_LOWER);
if (cusparseStatus != 0) printf("error in fillmode.\n");
cusparseStatus = cusparseCreateSolveAnalysisInfo(&precond->cuinfoU);
if (cusparseStatus != 0) printf("error in info.\n");
cusparseStatus =
cusparseDcsrsm_analysis(cusparseHandle,
CUSPARSE_OPERATION_TRANSPOSE, precond->M.num_rows,
precond->M.nnz, descrU,
precond->M.dval, precond->M.drow, precond->M.dcol, precond->cuinfoU );
if (cusparseStatus != 0) printf("error in analysis U.\n");
cusparseDestroyMatDescr( descrU );
cusparseDestroyMatDescr( descrA );
cusparseDestroy( cusparseHandle );
magma_d_mfree(&U, queue );
magma_d_mfree(&hA, queue );
/*
// to enable also the block-asynchronous iteration for the triangular solves
magma_d_mtransfer( precond->M, &hA, Magma_DEV, Magma_CPU, queue );
hA.storage_type = Magma_CSR;
magma_dcsrsplit( 256, hA, &hD, &hR, queue );
magma_d_mtransfer( hD, &precond->LD, Magma_CPU, Magma_DEV, queue );
magma_d_mtransfer( hR, &precond->L, Magma_CPU, Magma_DEV, queue );
magma_d_mfree(&hD, queue );
magma_d_mfree(&hR, queue );
magma_d_cucsrtranspose( hA, &hAt, queue );
magma_dcsrsplit( 256, hAt, &hD, &hR, queue );
magma_d_mtransfer( hD, &precond->UD, Magma_CPU, Magma_DEV, queue );
magma_d_mtransfer( hR, &precond->U, Magma_CPU, Magma_DEV, queue );
magma_d_mfree(&hD, queue );
magma_d_mfree(&hR, queue );
magma_d_mfree(&hA, queue );
magma_d_mfree(&hAt, queue );
*/
return MAGMA_SUCCESS;
}
extern "C" magma_int_t
magma_dcumilusetup(
magma_d_sparse_matrix A, magma_d_preconditioner *precond,
magma_queue_t queue )
{
//magma_d_mvisu(A, queue );
// copy matrix into preconditioner parameter
magma_d_sparse_matrix hA, hACSR;
magma_d_mtransfer( A, &hA, A.memory_location, Magma_CPU, queue );
magma_d_mconvert( hA, &hACSR, hA.storage_type, Magma_CSR, queue );
magma_d_mtransfer(hACSR, &(precond->M), Magma_CPU, Magma_DEV, queue );
magma_d_mfree( &hA, queue );
magma_d_mfree( &hACSR, queue );
// CUSPARSE context //
cusparseHandle_t cusparseHandle;
cusparseStatus_t cusparseStatus;
cusparseStatus = cusparseCreate(&cusparseHandle);
cusparseSetStream( cusparseHandle, queue );
if (cusparseStatus != 0) printf("error in Handle.\n");
cusparseMatDescr_t descrA;
cusparseStatus = cusparseCreateMatDescr(&descrA);
if (cusparseStatus != 0) printf("error in MatrDescr.\n");
cusparseStatus =
cusparseSetMatType(descrA,CUSPARSE_MATRIX_TYPE_GENERAL);
if (cusparseStatus != 0) printf("error in MatrType.\n");
cusparseStatus =
cusparseSetMatDiagType (descrA, CUSPARSE_DIAG_TYPE_NON_UNIT);
if (cusparseStatus != 0) printf("error in DiagType.\n");
cusparseStatus =
cusparseSetMatIndexBase(descrA,CUSPARSE_INDEX_BASE_ZERO);
if (cusparseStatus != 0) printf("error in IndexBase.\n");
cusparseStatus =
cusparseCreateSolveAnalysisInfo( &(precond->cuinfo) );
if (cusparseStatus != 0) printf("error in info.\n");
// end CUSPARSE context //
cusparseStatus =
cusparseDcsrsm_analysis( cusparseHandle,
CUSPARSE_OPERATION_NON_TRANSPOSE,
precond->M.num_rows, precond->M.nnz, descrA,
precond->M.dval, precond->M.drow, precond->M.dcol,
precond->cuinfo);
if (cusparseStatus != 0)
printf("error in analysis:%d\n", cusparseStatus);
cusparseStatus =
cusparseDcsrilu0( cusparseHandle, CUSPARSE_OPERATION_NON_TRANSPOSE,
precond->M.num_rows, descrA,
precond->M.dval,
precond->M.drow,
precond->M.dcol,
precond->cuinfo);
if (cusparseStatus != 0)
printf("error in ILU:%d\n", cusparseStatus);
cusparseStatus =
cusparseDestroySolveAnalysisInfo( precond->cuinfo );
if (cusparseStatus != 0) printf("error in info-free.\n");
cusparseDestroyMatDescr( descrA );
magma_d_sparse_matrix hL, hU;
magma_d_mtransfer( precond->M, &hA, Magma_DEV, Magma_CPU, queue );
hL.diagorder_type = Magma_UNITY;
magma_d_mconvert( hA, &hL , Magma_CSR, Magma_CSRL, queue );
hU.diagorder_type = Magma_VALUE;
magma_d_mconvert( hA, &hU , Magma_CSR, Magma_CSRU, queue );
magma_d_mtransfer( hL, &(precond->L), Magma_CPU, Magma_DEV, queue );
magma_d_mtransfer( hU, &(precond->U), Magma_CPU, Magma_DEV, queue );
cusparseMatDescr_t descrL;
cusparseStatus = cusparseCreateMatDescr(&descrL);
if (cusparseStatus != 0) printf("error in MatrDescr.\n");
cusparseStatus =
cusparseSetMatType(descrL,CUSPARSE_MATRIX_TYPE_TRIANGULAR);
if (cusparseStatus != 0) printf("error in MatrType.\n");
cusparseStatus =
cusparseSetMatDiagType (descrL, CUSPARSE_DIAG_TYPE_UNIT);
if (cusparseStatus != 0) printf("error in DiagType.\n");
cusparseStatus =
cusparseSetMatIndexBase(descrL,CUSPARSE_INDEX_BASE_ZERO);
if (cusparseStatus != 0) printf("error in IndexBase.\n");
cusparseStatus =
cusparseSetMatFillMode(descrL,CUSPARSE_FILL_MODE_LOWER);
if (cusparseStatus != 0) printf("error in fillmode.\n");
cusparseStatus = cusparseCreateSolveAnalysisInfo(&precond->cuinfoL);
if (cusparseStatus != 0) printf("error in info.\n");
cusparseStatus =
cusparseDcsrsm_analysis(cusparseHandle,
CUSPARSE_OPERATION_NON_TRANSPOSE, precond->L.num_rows,
precond->L.nnz, descrL,
precond->L.dval, precond->L.drow, precond->L.dcol, precond->cuinfoL );
if (cusparseStatus != 0) printf("error in analysis.\n");
cusparseDestroyMatDescr( descrL );
cusparseMatDescr_t descrU;
cusparseStatus = cusparseCreateMatDescr(&descrU);
if (cusparseStatus != 0) printf("error in MatrDescr.\n");
cusparseStatus =
cusparseSetMatType(descrU,CUSPARSE_MATRIX_TYPE_TRIANGULAR);
if (cusparseStatus != 0) printf("error in MatrType.\n");
cusparseStatus =
cusparseSetMatDiagType (descrU, CUSPARSE_DIAG_TYPE_NON_UNIT);
if (cusparseStatus != 0) printf("error in DiagType.\n");
cusparseStatus =
cusparseSetMatIndexBase(descrU,CUSPARSE_INDEX_BASE_ZERO);
if (cusparseStatus != 0) printf("error in IndexBase.\n");
cusparseStatus =
cusparseSetMatFillMode(descrU,CUSPARSE_FILL_MODE_UPPER);
if (cusparseStatus != 0) printf("error in fillmode.\n");
cusparseStatus = cusparseCreateSolveAnalysisInfo(&precond->cuinfoU);
if (cusparseStatus != 0) printf("error in info.\n");
cusparseStatus =
cusparseDcsrsm_analysis(cusparseHandle,
CUSPARSE_OPERATION_NON_TRANSPOSE, precond->U.num_rows,
precond->U.nnz, descrU,
precond->U.dval, precond->U.drow, precond->U.dcol, precond->cuinfoU );
if (cusparseStatus != 0) printf("error in analysis.\n");
cusparseDestroyMatDescr( descrU );
magma_d_mfree(&hA, queue );
magma_d_mfree(&hL, queue );
magma_d_mfree(&hU, queue );
cusparseDestroy( cusparseHandle );
return MAGMA_SUCCESS;
}
extern "C" magma_int_t
magma_dcuspmm( magma_d_sparse_matrix A, magma_d_sparse_matrix B,
magma_d_sparse_matrix *AB ){
if( A.memory_location == Magma_DEV
&& B.memory_location == Magma_DEV
&& ( A.storage_type == Magma_CSR ||
A.storage_type == Magma_CSRCOO )
&& ( B.storage_type == Magma_CSR ||
B.storage_type == Magma_CSRCOO ) ){
magma_d_sparse_matrix C;
C.num_rows = A.num_rows;
C.num_cols = A.num_cols;
C.storage_type = A.storage_type;
C.memory_location = A.memory_location;
// CUSPARSE context //
cusparseHandle_t handle;
cusparseStatus_t cusparseStatus;
cusparseStatus = cusparseCreate(&handle);
if(cusparseStatus != 0) printf("error in Handle.\n");
cusparseMatDescr_t descrA;
cusparseMatDescr_t descrB;
cusparseMatDescr_t descrC;
cusparseStatus = cusparseCreateMatDescr(&descrA);
cusparseStatus = cusparseCreateMatDescr(&descrB);
cusparseStatus = cusparseCreateMatDescr(&descrC);
if(cusparseStatus != 0) printf("error in MatrDescr.\n");
cusparseStatus =
cusparseSetMatType(descrA,CUSPARSE_MATRIX_TYPE_GENERAL);
cusparseSetMatType(descrB,CUSPARSE_MATRIX_TYPE_GENERAL);
cusparseSetMatType(descrC,CUSPARSE_MATRIX_TYPE_GENERAL);
if(cusparseStatus != 0) printf("error in MatrType.\n");
cusparseStatus =
cusparseSetMatIndexBase(descrA,CUSPARSE_INDEX_BASE_ZERO);
cusparseSetMatIndexBase(descrB,CUSPARSE_INDEX_BASE_ZERO);
cusparseSetMatIndexBase(descrC,CUSPARSE_INDEX_BASE_ZERO);
if(cusparseStatus != 0) printf("error in IndexBase.\n");
// multiply A and B on the device
magma_int_t baseC;
// nnzTotalDevHostPtr points to host memory
magma_index_t *nnzTotalDevHostPtr = (magma_index_t*) &C.nnz;
cusparseSetPointerMode(handle, CUSPARSE_POINTER_MODE_HOST);
magma_index_malloc( &C.row, (A.num_rows + 1) );
cusparseXcsrgemmNnz(handle, CUSPARSE_OPERATION_NON_TRANSPOSE,
CUSPARSE_OPERATION_NON_TRANSPOSE,
A.num_rows, A.num_rows, A.num_rows,
descrA, A.nnz, A.row, A.col,
descrB, B.nnz, B.row, B.col,
descrC, C.row, nnzTotalDevHostPtr );
if (NULL != nnzTotalDevHostPtr){
C.nnz = *nnzTotalDevHostPtr;
}else{
// workaround as nnz and base C are magma_int_t
magma_index_t base_t, nnz_t;
magma_index_getvector( 1, C.row+C.num_rows, 1, &nnz_t, 1 );
magma_index_getvector( 1, C.row, 1, &base_t, 1 );
C.nnz = (magma_int_t) nnz_t;
baseC = (magma_int_t) base_t;
C.nnz -= baseC;
}
magma_index_malloc( &C.col, C.nnz );
magma_dmalloc( &C.val, C.nnz );
cusparseDcsrgemm(handle, CUSPARSE_OPERATION_NON_TRANSPOSE,
CUSPARSE_OPERATION_NON_TRANSPOSE,
A.num_rows, A.num_rows, A.num_rows,
descrA, A.nnz,
A.val, A.row, A.col,
descrB, B.nnz,
B.val, B.row, B.col,
descrC,
C.val, C.row, C.col);
cusparseDestroyMatDescr( descrA );
cusparseDestroyMatDescr( descrB );
cusparseDestroyMatDescr( descrC );
cusparseDestroy( handle );
// end CUSPARSE context //
magma_d_mtransfer( C, AB, Magma_DEV, Magma_DEV );
magma_d_mfree( &C );
return MAGMA_SUCCESS;
}
else{
printf("error: CSRMM only supported on device and CSR format.\n");
return MAGMA_SUCCESS;
}
}
/* ////////////////////////////////////////////////////////////////////////////
-- testing any solver
*/
int main( int argc, char** argv )
{
TESTING_INIT();
magma_dopts zopts;
magma_queue_t queue;
magma_queue_create( /*devices[ opts->device ],*/ &queue );
int i=1;
magma_dparse_opts( argc, argv, &zopts, &i, queue );
real_Double_t res;
magma_d_sparse_matrix Z, Z2, A, A2, AT, AT2, B;
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] );
magma_dm_5stencil( laplace_size, &Z, queue );
} else { // file-matrix test
magma_d_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
magma_d_mconvert( Z, &A, Magma_CSR, Magma_CSRL, queue );
magma_d_mconvert( Z, &B, Magma_CSR, Magma_CSRU, queue );
// transpose
magma_d_mtranspose( A, &AT, queue );
// quite some conversions
//ELL
magma_d_mconvert( AT, &AT2, Magma_CSR, Magma_ELL, queue );
magma_d_mfree(&AT, queue );
magma_d_mconvert( AT2, &AT, Magma_ELL, Magma_CSR, queue );
magma_d_mfree(&AT2, queue );
//ELLPACKT
magma_d_mconvert( AT, &AT2, Magma_CSR, Magma_ELLPACKT, queue );
magma_d_mfree(&AT, queue );
magma_d_mconvert( AT2, &AT, Magma_ELLPACKT, Magma_CSR, queue );
magma_d_mfree(&AT2, queue );
//ELLRT
AT2.blocksize = 8;
AT2.alignment = 8;
magma_d_mconvert( AT, &AT2, Magma_CSR, Magma_ELLRT, queue );
magma_d_mfree(&AT, queue );
magma_d_mconvert( AT2, &AT, Magma_ELLRT, Magma_CSR, queue );
magma_d_mfree(&AT2, queue );
//SELLP
AT2.blocksize = 8;
AT2.alignment = 8;
magma_d_mconvert( AT, &AT2, Magma_CSR, Magma_SELLP, queue );
magma_d_mfree(&AT, queue );
magma_d_mconvert( AT2, &AT, Magma_SELLP, Magma_CSR, queue );
magma_d_mfree(&AT2, queue );
//ELLD
magma_d_mconvert( AT, &AT2, Magma_CSR, Magma_ELLD, queue );
magma_d_mfree(&AT, queue );
magma_d_mconvert( AT2, &AT, Magma_ELLD, Magma_CSR, queue );
magma_d_mfree(&AT2, queue );
//CSRCOO
magma_d_mconvert( AT, &AT2, Magma_CSR, Magma_CSRCOO, queue );
magma_d_mfree(&AT, queue );
magma_d_mconvert( AT2, &AT, Magma_CSRCOO, Magma_CSR, queue );
magma_d_mfree(&AT2, queue );
//CSRD
magma_d_mconvert( AT, &AT2, Magma_CSR, Magma_CSRD, queue );
magma_d_mfree(&AT, queue );
magma_d_mconvert( AT2, &AT, Magma_CSRD, Magma_CSR, queue );
magma_d_mfree(&AT2, queue );
//BCSR
magma_d_mconvert( AT, &AT2, Magma_CSR, Magma_BCSR, queue );
magma_d_mfree(&AT, queue );
magma_d_mconvert( AT2, &AT, Magma_BCSR, Magma_CSR, queue );
magma_d_mfree(&AT2, queue );
// transpose
magma_d_mtranspose( AT, &A2, queue );
magma_dmdiff( 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");
magma_dmlumerge( A2, B, &Z2, queue );
magma_dmdiff( 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_d_mfree(&A, queue );
magma_d_mfree(&A2, queue );
magma_d_mfree(&AT, queue );
magma_d_mfree(&AT2, queue );
magma_d_mfree(&B, queue );
magma_d_mfree(&Z2, queue );
magma_d_mfree(&Z, queue );
i++;
}
magma_queue_destroy( queue );
TESTING_FINALIZE();
return 0;
}
/* ////////////////////////////////////////////////////////////////////////////
-- running magma_dlobpcg
*/
int main( int argc, char** argv)
{
TESTING_INIT();
magma_d_solver_par solver_par;
solver_par.epsilon = 1e-5;
solver_par.maxiter = 1000;
solver_par.verbose = 0;
solver_par.num_eigenvalues = 32;
solver_par.solver = Magma_LOBPCG;
magma_d_preconditioner precond_par;
precond_par.solver = Magma_JACOBI;
int precond = 0;
int format = 0;
int scale = 0;
magma_scale_t scaling = Magma_NOSCALE;
magma_d_sparse_matrix A, B, dA;
B.blocksize = 8;
B.alignment = 8;
B.storage_type = Magma_CSR;
int i;
for( i = 1; i < argc; ++i ) {
if ( strcmp("--format", argv[i]) == 0 ) {
format = atoi( argv[++i] );
switch( format ) {
case 0: B.storage_type = Magma_CSR; break;
case 1: B.storage_type = Magma_ELL; break;
case 2: B.storage_type = Magma_ELLRT; break;
case 3: B.storage_type = Magma_SELLP; break;
}
}else if ( strcmp("--mscale", argv[i]) == 0 ) {
scale = atoi( argv[++i] );
switch( scale ) {
case 0: scaling = Magma_NOSCALE; break;
case 1: scaling = Magma_UNITDIAG; break;
case 2: scaling = Magma_UNITROW; break;
}
}else if ( strcmp("--precond", argv[i]) == 0 ) {
format = atoi( argv[++i] );
switch( precond ) {
case 0: precond_par.solver = Magma_JACOBI; break;
}
}else if ( strcmp("--blocksize", argv[i]) == 0 ) {
B.blocksize = atoi( argv[++i] );
}else if ( strcmp("--alignment", argv[i]) == 0 ) {
B.alignment = atoi( argv[++i] );
}else if ( strcmp("--verbose", argv[i]) == 0 ) {
solver_par.verbose = atoi( argv[++i] );
} else if ( strcmp("--maxiter", argv[i]) == 0 ) {
solver_par.maxiter = atoi( argv[++i] );
} else if ( strcmp("--tol", argv[i]) == 0 ) {
sscanf( argv[++i], "%lf", &solver_par.epsilon );
} else if ( strcmp("--eigenvalues", argv[i]) == 0 ) {
solver_par.num_eigenvalues = atoi( argv[++i] );
} else
break;
}
printf( "\n# usage: ./run_dlobpcg"
" [ --format %d (0=CSR, 1=ELL, 2=ELLRT, 4=SELLP)"
" [ --blocksize %d --alignment %d ]"
" --mscale %d (0=no, 1=unitdiag, 2=unitrownrm)"
" --verbose %d (0=summary, k=details every k iterations)"
" --maxiter %d --tol %.2e"
" --preconditioner %d (0=Jacobi) "
" --eigenvalues %d ]"
" matrices \n\n", format, (int) B.blocksize, (int) B.alignment,
(int) scale,
(int) solver_par.verbose,
(int) solver_par.maxiter, solver_par.epsilon, precond,
(int) solver_par.num_eigenvalues);
while( i < argc ){
magma_d_csr_mtx( &A, argv[i] );
printf( "\n# matrix info: %d-by-%d with %d nonzeros\n\n",
(int) A.num_rows,(int) A.num_cols,(int) A.nnz );
// scale initial guess
magma_dmscale( &A, scaling );
solver_par.ev_length = A.num_cols;
magma_d_sparse_matrix A2;
A2.storage_type = Magma_SELLC;
A2.blocksize = 8;
A2.alignment = 4;
magma_d_mconvert( A, &A2, Magma_CSR, A2.storage_type );
// copy matrix to GPU
magma_d_mtransfer( A2, &dA, Magma_CPU, Magma_DEV);
magma_dsolverinfo_init( &solver_par, &precond_par ); // inside the loop!
// as the matrix size has influence on the EV-length
real_Double_t gpu_time;
// Find the blockSize smallest eigenvalues and corresponding eigen-vectors
gpu_time = magma_wtime();
magma_dlobpcg( dA, &solver_par );
gpu_time = magma_wtime() - gpu_time;
printf("Time (sec) = %7.2f\n", gpu_time);
printf("solver runtime (sec) = %7.2f\n", solver_par.runtime );
magma_dsolverinfo_free( &solver_par, &precond_par );
magma_d_mfree( &dA );
magma_d_mfree( &A2 );
magma_d_mfree( &A );
i++;
}
TESTING_FINALIZE();
return 0;
}
extern "C" magma_int_t
magma_d_cucsrtranspose(
magma_d_sparse_matrix A,
magma_d_sparse_matrix *B,
magma_queue_t queue )
{
// for symmetric matrices: convert to csc using cusparse
if( A.storage_type == Magma_CSR && A.memory_location == Magma_DEV ) {
magma_d_sparse_matrix C;
magma_d_mtransfer( A, &C, Magma_DEV, Magma_DEV, queue );
// CUSPARSE context //
cusparseHandle_t handle;
cusparseStatus_t cusparseStatus;
cusparseStatus = cusparseCreate(&handle);
cusparseSetStream( handle, queue );
if (cusparseStatus != 0) printf("error in Handle.\n");
cusparseMatDescr_t descrA;
cusparseMatDescr_t descrB;
cusparseStatus = cusparseCreateMatDescr(&descrA);
cusparseStatus = cusparseCreateMatDescr(&descrB);
if (cusparseStatus != 0) printf("error in MatrDescr.\n");
cusparseStatus =
cusparseSetMatType(descrA,CUSPARSE_MATRIX_TYPE_GENERAL);
cusparseSetMatType(descrB,CUSPARSE_MATRIX_TYPE_GENERAL);
if (cusparseStatus != 0) printf("error in MatrType.\n");
cusparseStatus =
cusparseSetMatIndexBase(descrA,CUSPARSE_INDEX_BASE_ZERO);
cusparseSetMatIndexBase(descrB,CUSPARSE_INDEX_BASE_ZERO);
if (cusparseStatus != 0) printf("error in IndexBase.\n");
cusparseStatus =
cusparseDcsr2csc( handle, A.num_rows, A.num_rows, A.nnz,
A.dval, A.drow, A.dcol, C.dval, C.dcol, C.drow,
CUSPARSE_ACTION_NUMERIC,
CUSPARSE_INDEX_BASE_ZERO);
if (cusparseStatus != 0)
printf("error in transpose: %d.\n", cusparseStatus);
cusparseDestroyMatDescr( descrA );
cusparseDestroyMatDescr( descrB );
cusparseDestroy( handle );
magma_d_mtransfer( C, B, Magma_DEV, Magma_DEV, queue );
if( A.fill_mode == Magma_FULL ){
B->fill_mode = Magma_FULL;
}
else if( A.fill_mode == Magma_LOWER ){
B->fill_mode = Magma_UPPER;
}
else if ( A.fill_mode == Magma_UPPER ){
B->fill_mode = Magma_LOWER;
}
// end CUSPARSE context //
return MAGMA_SUCCESS;
}else if( A.storage_type == Magma_CSR && A.memory_location == Magma_CPU ){
magma_d_sparse_matrix A_d, B_d;
magma_d_mtransfer( A, &A_d, A.memory_location, Magma_DEV, queue );
magma_d_cucsrtranspose( A_d, &B_d, queue );
magma_d_mtransfer( B_d, B, Magma_DEV, A.memory_location, queue );
magma_d_mfree( &A_d, queue );
magma_d_mfree( &B_d, queue );
return MAGMA_SUCCESS;
}else {
magma_d_sparse_matrix ACSR, BCSR;
magma_d_mconvert( A, &ACSR, A.storage_type, Magma_CSR, queue );
magma_d_cucsrtranspose( ACSR, &BCSR, queue );
magma_d_mconvert( BCSR, B, Magma_CSR, A.storage_type, queue );
magma_d_mfree( &ACSR, queue );
magma_d_mfree( &BCSR, queue );
return MAGMA_SUCCESS;
}
}
/* ////////////////////////////////////////////////////////////////////////////
-- Debugging file
*/
int main( int argc, char** argv)
{
TESTING_INIT();
magma_d_solver_par solver_par;
magma_d_preconditioner precond_par;
solver_par.epsilon = 10e-16;
solver_par.maxiter = 1000;
solver_par.verbose = 0;
solver_par.restart = 30;
solver_par.num_eigenvalues = 0;
solver_par.ortho = Magma_CGS;
double one = MAGMA_D_MAKE(1.0, 0.0);
double zero = MAGMA_D_MAKE(0.0, 0.0);
magma_d_sparse_matrix A, B, B_d;
magma_d_vector x, b;
// generate matrix of desired structure and size
magma_int_t n=100; // size is n*n
magma_int_t nn = n*n;
magma_int_t offdiags = 2;
magma_index_t *diag_offset;
double *diag_vals;
magma_dmalloc_cpu( &diag_vals, offdiags+1 );
magma_index_malloc_cpu( &diag_offset, offdiags+1 );
diag_offset[0] = 0;
diag_offset[1] = 1;
diag_offset[2] = n;
diag_vals[0] = MAGMA_D_MAKE( 4.1, 0.0 );
diag_vals[1] = MAGMA_D_MAKE( -1.0, 0.0 );
diag_vals[2] = MAGMA_D_MAKE( -1.0, 0.0 );
magma_dmgenerator( nn, offdiags, diag_offset, diag_vals, &A );
// convert marix into desired format
B.storage_type = Magma_SELLC;
B.blocksize = 8;
B.alignment = 8;
// scale matrix
magma_dmscale( &A, Magma_UNITDIAG );
magma_d_mconvert( A, &B, Magma_CSR, B.storage_type );
magma_d_mtransfer( B, &B_d, Magma_CPU, Magma_DEV );
// test CG ####################################
// vectors and initial guess
magma_d_vinit( &b, Magma_DEV, A.num_cols, one );
magma_d_vinit( &x, Magma_DEV, A.num_cols, one );
magma_d_spmv( one, B_d, x, zero, b ); // b = A x
magma_d_vfree(&x);
magma_d_vinit( &x, Magma_DEV, A.num_cols, zero );
magma_dsolverinfo_init( &solver_par, &precond_par );
// solver
magma_dcg_res( B_d, b, &x, &solver_par );
// solverinfo
magma_dsolverinfo( &solver_par, &precond_par );
if( solver_par.numiter > 150 ){
printf("error: test not passed!\n"); exit(-1);
}
magma_dsolverinfo_free( &solver_par, &precond_par );
magma_d_vfree(&x);
magma_d_vfree(&b);
// test PCG Jacobi ############################
// vectors and initial guess
magma_d_vinit( &b, Magma_DEV, A.num_cols, one );
magma_d_vinit( &x, Magma_DEV, A.num_cols, one );
magma_d_spmv( one, B_d, x, zero, b ); // b = A x
magma_d_vfree(&x);
magma_d_vinit( &x, Magma_DEV, A.num_cols, zero );
magma_dsolverinfo_init( &solver_par, &precond_par );
// Preconditioner
precond_par.solver = Magma_JACOBI;
magma_d_precondsetup( B_d, b, &precond_par );
// solver
magma_dpcg( B_d, b, &x, &solver_par, &precond_par );
// solverinfo
magma_dsolverinfo( &solver_par, &precond_par );
if( solver_par.numiter > 150 ){
printf("error: test not passed!\n"); exit(-1);
}
magma_dsolverinfo_free( &solver_par, &precond_par );
magma_d_vfree(&x);
magma_d_vfree(&b);
// test PCG IC ################################
// vectors and initial guess
magma_d_vinit( &b, Magma_DEV, A.num_cols, one );
magma_d_vinit( &x, Magma_DEV, A.num_cols, one );
magma_d_spmv( one, B_d, x, zero, b ); // b = A x
magma_d_vfree(&x);
magma_d_vinit( &x, Magma_DEV, A.num_cols, zero );
magma_dsolverinfo_init( &solver_par, &precond_par );
// Preconditioner
precond_par.solver = Magma_ICC;
magma_d_precondsetup( B_d, b, &precond_par );
// solver
magma_dpcg( B_d, b, &x, &solver_par, &precond_par );
// solverinfo
magma_dsolverinfo( &solver_par, &precond_par );
if( solver_par.numiter > 150 ){
printf("error: test not passed!\n"); exit(-1);
}
magma_dsolverinfo_free( &solver_par, &precond_par );
magma_d_vfree(&x);
magma_d_vfree(&b);
// test PCG IC ################################
// vectors and initial guess
magma_d_vinit( &b, Magma_DEV, A.num_cols, one );
magma_d_vinit( &x, Magma_DEV, A.num_cols, one );
magma_d_spmv( one, B_d, x, zero, b ); // b = A x
magma_d_vfree(&x);
magma_d_vinit( &x, Magma_DEV, A.num_cols, zero );
magma_dsolverinfo_init( &solver_par, &precond_par );
// Preconditioner
precond_par.solver = Magma_ICC;
magma_d_precondsetup( B_d, b, &precond_par );
// solver
magma_dpcg( B_d, b, &x, &solver_par, &precond_par );
// solverinfo
magma_dsolverinfo( &solver_par, &precond_par );
if( solver_par.numiter > 150 ){
printf("error: test not passed!\n"); exit(-1);
}
magma_dsolverinfo_free( &solver_par, &precond_par );
magma_d_vfree(&x);
magma_d_vfree(&b);
// test BICGSTAB ####################################
// vectors and initial guess
magma_d_vinit( &b, Magma_DEV, A.num_cols, one );
magma_d_vinit( &x, Magma_DEV, A.num_cols, one );
magma_d_spmv( one, B_d, x, zero, b ); // b = A x
magma_d_vfree(&x);
magma_d_vinit( &x, Magma_DEV, A.num_cols, zero );
magma_dsolverinfo_init( &solver_par, &precond_par );
// solver
magma_dbicgstab( B_d, b, &x, &solver_par );
// solverinfo
magma_dsolverinfo( &solver_par, &precond_par );
if( solver_par.numiter > 150 ){
printf("error: test not passed!\n"); exit(-1);
}
magma_dsolverinfo_free( &solver_par, &precond_par );
magma_d_vfree(&x);
magma_d_vfree(&b);
// test PBICGSTAB Jacobi ############################
// vectors and initial guess
magma_d_vinit( &b, Magma_DEV, A.num_cols, one );
magma_d_vinit( &x, Magma_DEV, A.num_cols, one );
magma_d_spmv( one, B_d, x, zero, b ); // b = A x
magma_d_vfree(&x);
magma_d_vinit( &x, Magma_DEV, A.num_cols, zero );
magma_dsolverinfo_init( &solver_par, &precond_par );
// Preconditioner
precond_par.solver = Magma_JACOBI;
magma_d_precondsetup( B_d, b, &precond_par );
// solver
magma_dpbicgstab( B_d, b, &x, &solver_par, &precond_par );
// solverinfo
magma_dsolverinfo( &solver_par, &precond_par );
if( solver_par.numiter > 150 ){
printf("error: test not passed!\n"); exit(-1);
}
magma_dsolverinfo_free( &solver_par, &precond_par );
magma_d_vfree(&x);
magma_d_vfree(&b);
/*
// test PBICGSTAB ILU ###############################
// vectors and initial guess
magma_d_vinit( &b, Magma_DEV, A.num_cols, one );
magma_d_vinit( &x, Magma_DEV, A.num_cols, one );
magma_d_spmv( one, B_d, x, zero, b ); // b = A x
magma_d_vfree(&x);
magma_d_vinit( &x, Magma_DEV, A.num_cols, zero );
magma_dsolverinfo_init( &solver_par, &precond_par );
// Preconditioner
precond_par.solver = Magma_ILU;
magma_d_precondsetup( B_d, b, &precond_par );
// solver
magma_dpbicgstab( B_d, b, &x, &solver_par, &precond_par );
// solverinfo
magma_dsolverinfo( &solver_par, &precond_par );
if( solver_par.numiter > 150 ){
printf("error: test not passed!\n"); exit(-1);
}
magma_dsolverinfo_free( &solver_par, &precond_par );
magma_d_vfree(&x);
magma_d_vfree(&b);
// test PBICGSTAB ILU ###############################
// vectors and initial guess
magma_d_vinit( &b, Magma_DEV, A.num_cols, one );
magma_d_vinit( &x, Magma_DEV, A.num_cols, one );
magma_d_spmv( one, B_d, x, zero, b ); // b = A x
magma_d_vfree(&x);printf("here\n");
magma_d_vinit( &x, Magma_DEV, A.num_cols, zero );
magma_dsolverinfo_init( &solver_par, &precond_par );
// Preconditioner
precond_par.solver = Magma_ILU;
magma_d_precondsetup( B_d, b, &precond_par );
// solver
magma_dpbicgstab( B_d, b, &x, &solver_par, &precond_par );
// solverinfo
magma_dsolverinfo( &solver_par, &precond_par );
if( solver_par.numiter > 150 ){
printf("error: test not passed!\n"); exit(-1);
}
magma_dsolverinfo_free( &solver_par, &precond_par );
magma_d_vfree(&x);
magma_d_vfree(&b);
// test GMRES ####################################
// vectors and initial guess
magma_d_vinit( &b, Magma_DEV, A.num_cols, one );
magma_d_vinit( &x, Magma_DEV, A.num_cols, one );
magma_d_spmv( one, B_d, x, zero, b ); // b = A x
magma_d_vfree(&x);
magma_d_vinit( &x, Magma_DEV, A.num_cols, zero );
magma_dsolverinfo_init( &solver_par, &precond_par );
// solver
magma_dgmres( B_d, b, &x, &solver_par );
// solverinfo
magma_dsolverinfo( &solver_par, &precond_par );
magma_dsolverinfo_free( &solver_par, &precond_par );
magma_d_vfree(&x);
magma_d_vfree(&b);
// test PGMRES Jacobi ############################
// vectors and initial guess
magma_d_vinit( &b, Magma_DEV, A.num_cols, one );
magma_d_vinit( &x, Magma_DEV, A.num_cols, one );
magma_d_spmv( one, B_d, x, zero, b ); // b = A x
magma_d_vfree(&x);
magma_d_vinit( &x, Magma_DEV, A.num_cols, zero );
magma_dsolverinfo_init( &solver_par, &precond_par );
// Preconditioner
precond_par.solver = Magma_JACOBI;
magma_d_precondsetup( B_d, b, &precond_par );
// solver
magma_dpgmres( B_d, b, &x, &solver_par, &precond_par );
// solverinfo
magma_dsolverinfo( &solver_par, &precond_par );
magma_dsolverinfo_free( &solver_par, &precond_par );
magma_d_vfree(&x);
magma_d_vfree(&b);*/
// test PGMRES ILU ###############################
// vectors and initial guess
magma_d_vinit( &b, Magma_DEV, A.num_cols, one );
magma_d_vinit( &x, Magma_DEV, A.num_cols, one );
magma_d_spmv( one, B_d, x, zero, b ); // b = A x
magma_d_vfree(&x);
magma_d_vinit( &x, Magma_DEV, A.num_cols, zero );
magma_dsolverinfo_init( &solver_par, &precond_par );
// Preconditioner
precond_par.solver = Magma_ILU;
magma_d_precondsetup( B_d, b, &precond_par );
// solver
magma_dpgmres( B_d, b, &x, &solver_par, &precond_par );
// solverinfo
magma_dsolverinfo( &solver_par, &precond_par );
if( solver_par.numiter > 150 ){
printf("error: test not passed!\n"); exit(-1);
}
magma_dsolverinfo_free( &solver_par, &precond_par );
magma_d_vfree(&x);
magma_d_vfree(&b);
printf("all tests passed.\n");
magma_d_mfree(&B_d);
magma_d_mfree(&B);
magma_d_mfree(&A);
TESTING_FINALIZE();
return 0;
}