extern "C" magma_int_t
magma_z_spmv(
magmaDoubleComplex alpha,
magma_z_sparse_matrix A,
magma_z_vector x,
magmaDoubleComplex beta,
magma_z_vector y,
magma_queue_t queue )
{
// set queue for old dense routines
magma_queue_t orig_queue;
magmablasGetKernelStream( &orig_queue );
if ( A.memory_location != x.memory_location ||
x.memory_location != y.memory_location ) {
printf("error: linear algebra objects are not located in same memory!\n");
printf("memory locations are: %d %d %d\n",
A.memory_location, x.memory_location, y.memory_location );
magmablasSetKernelStream( orig_queue );
return MAGMA_ERR_INVALID_PTR;
}
// DEV case
if ( A.memory_location == Magma_DEV ) {
if ( A.num_cols == x.num_rows && x.num_cols == 1 ) {
if ( A.storage_type == Magma_CSR
|| A.storage_type == Magma_CSRL
|| A.storage_type == Magma_CSRU ) {
//printf("using CSR kernel for SpMV: ");
//magma_zgecsrmv( MagmaNoTrans, A.num_rows, A.num_cols, alpha,
// A.dval, A.drow, A.dcol, x.dval, beta, y.dval );
//printf("done.\n");
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);
cusparseZcsrmv( cusparseHandle,CUSPARSE_OPERATION_NON_TRANSPOSE,
A.num_rows, A.num_cols, A.nnz, &alpha, descr,
A.dval, A.drow, A.dcol, x.dval, &beta, y.dval );
cusparseDestroyMatDescr( descr );
cusparseDestroy( cusparseHandle );
}
else if ( A.storage_type == Magma_ELL ) {
//printf("using ELLPACKT kernel for SpMV: ");
magma_zgeelltmv( MagmaNoTrans, A.num_rows, A.num_cols,
A.max_nnz_row, alpha, A.dval, A.dcol, x.dval, beta,
y.dval, queue );
//printf("done.\n");
}
else if ( A.storage_type == Magma_ELLPACKT ) {
//printf("using ELL kernel for SpMV: ");
magma_zgeellmv( MagmaNoTrans, A.num_rows, A.num_cols,
A.max_nnz_row, alpha, A.dval, A.dcol, x.dval, beta,
y.dval, queue );
//printf("done.\n");
}
else if ( A.storage_type == Magma_ELLRT ) {
//printf("using ELLRT kernel for SpMV: ");
magma_zgeellrtmv( MagmaNoTrans, A.num_rows, A.num_cols,
A.max_nnz_row, alpha, A.dval, A.dcol, A.drow, x.dval,
beta, y.dval, A.alignment, A.blocksize, queue );
//printf("done.\n");
}
else if ( A.storage_type == Magma_SELLP ) {
//printf("using SELLP kernel for SpMV: ");
magma_zgesellpmv( MagmaNoTrans, A.num_rows, A.num_cols,
A.blocksize, A.numblocks, A.alignment,
alpha, A.dval, A.dcol, A.drow, x.dval, beta, y.dval, queue );
//printf("done.\n");
}
else if ( A.storage_type == Magma_DENSE ) {
//printf("using DENSE kernel for SpMV: ");
magmablas_zgemv( MagmaNoTrans, A.num_rows, A.num_cols, alpha,
A.dval, A.num_rows, x.dval, 1, beta, y.dval,
1 );
//printf("done.\n");
}
else if ( A.storage_type == Magma_SPMVFUNCTION ) {
//printf("using DENSE kernel for SpMV: ");
magma_zcustomspmv( alpha, x, beta, y, queue );
//printf("done.\n");
}
/* else if ( A.storage_type == Magma_BCSR ) {
//printf("using CUSPARSE BCSR kernel for SpMV: ");
// CUSPARSE context //
cusparseHandle_t cusparseHandle = 0;
cusparseStatus_t cusparseStatus;
cusparseStatus = cusparseCreate(&cusparseHandle);
cusparseSetStream( cusparseHandle, queue );
cusparseMatDescr_t descr = 0;
cusparseStatus = cusparseCreateMatDescr(&descr);
// end CUSPARSE context //
cusparseDirection_t dirA = CUSPARSE_DIRECTION_ROW;
int mb = (A.num_rows + A.blocksize-1)/A.blocksize;
int nb = (A.num_cols + A.blocksize-1)/A.blocksize;
cusparseZbsrmv( cusparseHandle, dirA,
CUSPARSE_OPERATION_NON_TRANSPOSE, mb, nb, A.numblocks,
&alpha, descr, A.dval, A.drow, A.dcol, A.blocksize, x.dval,
&beta, y.dval );
//printf("done.\n");
magmablasSetKernelStream( orig_queue );
return MAGMA_SUCCESS;
}*/
else {
printf("error: format not supported.\n");
magmablasSetKernelStream( orig_queue );
return MAGMA_ERR_NOT_SUPPORTED;
}
}
else if ( A.num_cols < x.num_rows || x.num_cols > 1 ) {
magma_int_t num_vecs = x.num_rows / A.num_cols * x.num_cols;
if ( A.storage_type == Magma_CSR ) {
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);
if ( x.major == MagmaColMajor) {
cusparseZcsrmm(cusparseHandle,
CUSPARSE_OPERATION_NON_TRANSPOSE,
A.num_rows, num_vecs, A.num_cols, A.nnz,
&alpha, descr, A.dval, A.drow, A.dcol,
x.dval, A.num_cols, &beta, y.dval, A.num_cols);
} else if ( x.major == MagmaRowMajor) {
cusparseZcsrmm2(cusparseHandle,
CUSPARSE_OPERATION_NON_TRANSPOSE,
CUSPARSE_OPERATION_TRANSPOSE,
A.num_rows, num_vecs, A.num_cols, A.nnz,
&alpha, descr, A.dval, A.drow, A.dcol,
x.dval, A.num_cols, &beta, y.dval, A.num_cols);
}
cusparseDestroyMatDescr( descr );
cusparseDestroy( cusparseHandle );
}
else if ( A.storage_type == Magma_ELL ) {
if ( x.major == MagmaColMajor) {
magma_zmgeelltmv( MagmaNoTrans, A.num_rows, A.num_cols,
num_vecs, A.max_nnz_row, alpha, A.dval, A.dcol, x.dval,
beta, y.dval, queue );
}
else if ( x.major == MagmaRowMajor) {
// transpose first to col major
magma_z_vector x2;
magma_zvtranspose( x, &x2, queue );
magma_zmgeellmv( MagmaNoTrans, A.num_rows, A.num_cols,
num_vecs, A.max_nnz_row, alpha, A.dval, A.dcol, x.dval,
beta, y.dval, queue );
magma_z_vfree(&x2, queue );
}
}
else if ( A.storage_type == Magma_ELLPACKT ) {
if ( x.major == MagmaColMajor) {
magma_zmgeellmv( MagmaNoTrans, A.num_rows, A.num_cols,
num_vecs, A.max_nnz_row, alpha, A.dval, A.dcol, x.dval,
beta, y.dval, queue );
}
else if ( x.major == MagmaRowMajor) {
// transpose first to col major
magma_z_vector x2;
magma_zvtranspose( x, &x2, queue );
magma_zmgeelltmv( MagmaNoTrans, A.num_rows, A.num_cols,
num_vecs, A.max_nnz_row, alpha, A.dval, A.dcol, x.dval,
beta, y.dval, queue );
magma_z_vfree(&x2, queue );
}
} else if ( A.storage_type == Magma_SELLP ) {
if ( x.major == MagmaRowMajor) {
magma_zmgesellpmv( MagmaNoTrans, A.num_rows, A.num_cols,
num_vecs, A.blocksize, A.numblocks, A.alignment,
alpha, A.dval, A.dcol, A.drow, x.dval, beta, y.dval, queue );
}
else if ( x.major == MagmaColMajor) {
// transpose first to row major
magma_z_vector x2;
magma_zvtranspose( x, &x2, queue );
magma_zmgesellpmv( MagmaNoTrans, A.num_rows, A.num_cols,
num_vecs, A.blocksize, A.numblocks, A.alignment,
alpha, A.dval, A.dcol, A.drow, x2.dval, beta, y.dval, queue );
magma_z_vfree(&x2, queue );
}
}/*
if ( A.storage_type == Magma_DENSE ) {
//printf("using DENSE kernel for SpMV: ");
magmablas_zmgemv( MagmaNoTrans, A.num_rows, A.num_cols,
num_vecs, alpha, A.dval, A.num_rows, x.dval, 1,
beta, y.dval, 1 );
//printf("done.\n");
magmablasSetKernelStream( orig_queue );
return MAGMA_SUCCESS;
}*/
else {
printf("error: format not supported.\n");
magmablasSetKernelStream( orig_queue );
return MAGMA_ERR_NOT_SUPPORTED;
}
}
}
// CPU case missing!
else {
printf("error: CPU not yet supported.\n");
magmablasSetKernelStream( orig_queue );
return MAGMA_ERR_NOT_SUPPORTED;
}
magmablasSetKernelStream( orig_queue );
return MAGMA_SUCCESS;
}
magma_int_t
magma_z_spmv( magmaDoubleComplex alpha, magma_z_sparse_matrix A,
magma_z_vector x, magmaDoubleComplex beta, magma_z_vector y )
{
if( A.memory_location != x.memory_location ||
x.memory_location != y.memory_location ){
printf("error: linear algebra objects are not located in same memory!\n");
printf("memory locations are: %d %d %d\n",
A.memory_location, x.memory_location, y.memory_location );
return MAGMA_ERR_INVALID_PTR;
}
// DEV case
if( A.memory_location == Magma_DEV ){
if( A.num_cols == x.num_rows ){
if( A.storage_type == Magma_CSR
|| A.storage_type == Magma_CSRL
|| A.storage_type == Magma_CSRU ){
//printf("using CSR kernel for SpMV: ");
//magma_zgecsrmv( MagmaNoTrans, A.num_rows, A.num_cols, alpha,
// A.val, A.row, A.col, x.val, beta, y.val );
//printf("done.\n");
cusparseHandle_t cusparseHandle = 0;
cusparseStatus_t cusparseStatus;
cusparseStatus = cusparseCreate(&cusparseHandle);
cusparseMatDescr_t descr = 0;
cusparseStatus = cusparseCreateMatDescr(&descr);
cusparseSetMatType(descr,CUSPARSE_MATRIX_TYPE_GENERAL);
cusparseSetMatIndexBase(descr,CUSPARSE_INDEX_BASE_ZERO);
cusparseZcsrmv( cusparseHandle,CUSPARSE_OPERATION_NON_TRANSPOSE,
A.num_rows, A.num_cols, A.nnz, &alpha, descr,
A.val, A.row, A.col, x.val, &beta, y.val );
cusparseDestroyMatDescr( descr );
cusparseDestroy( cusparseHandle );
return MAGMA_SUCCESS;
}
else if( A.storage_type == Magma_ELLPACK ){
//printf("using ELLPACK kernel for SpMV: ");
magma_zgeellmv( MagmaNoTrans, A.num_rows, A.num_cols,
A.max_nnz_row, alpha, A.val, A.col, x.val, beta, y.val );
//printf("done.\n");
return MAGMA_SUCCESS;
}
else if( A.storage_type == Magma_ELL ){
//printf("using ELL kernel for SpMV: ");
magma_zgeelltmv( MagmaNoTrans, A.num_rows, A.num_cols,
A.max_nnz_row, alpha, A.val, A.col, x.val, beta, y.val );
//printf("done.\n");
return MAGMA_SUCCESS;
}
else if( A.storage_type == Magma_ELLRT ){
//printf("using ELLRT kernel for SpMV: ");
magma_zgeellrtmv( MagmaNoTrans, A.num_rows, A.num_cols,
A.max_nnz_row, alpha, A.val, A.col, A.row, x.val,
beta, y.val, A.alignment, A.blocksize );
//printf("done.\n");
return MAGMA_SUCCESS;
}
else if( A.storage_type == Magma_SELLC ){
//printf("using SELLC kernel for SpMV: ");
magma_zgesellcmv( MagmaNoTrans, A.num_rows, A.num_cols,
A.blocksize, A.numblocks, A.alignment,
alpha, A.val, A.col, A.row, x.val, beta, y.val );
//printf("done.\n");
return MAGMA_SUCCESS;
}
else if( A.storage_type == Magma_SELLP ){
//printf("using SELLP kernel for SpMV: ");
magma_zgesellpmv( MagmaNoTrans, A.num_rows, A.num_cols,
A.blocksize, A.numblocks, A.alignment,
alpha, A.val, A.col, A.row, x.val, beta, y.val );
//printf("done.\n");
return MAGMA_SUCCESS;
}
else if( A.storage_type == Magma_DENSE ){
//printf("using DENSE kernel for SpMV: ");
magmablas_zgemv( MagmaNoTrans, A.num_rows, A.num_cols, alpha,
A.val, A.num_rows, x.val, 1, beta, y.val, 1 );
//printf("done.\n");
return MAGMA_SUCCESS;
}
/* else if( A.storage_type == Magma_BCSR ){
//printf("using CUSPARSE BCSR kernel for SpMV: ");
// CUSPARSE context //
cusparseHandle_t cusparseHandle = 0;
cusparseStatus_t cusparseStatus;
cusparseStatus = cusparseCreate(&cusparseHandle);
cusparseMatDescr_t descr = 0;
cusparseStatus = cusparseCreateMatDescr(&descr);
// end CUSPARSE context //
cusparseDirection_t dirA = CUSPARSE_DIRECTION_ROW;
int mb = (A.num_rows + A.blocksize-1)/A.blocksize;
int nb = (A.num_cols + A.blocksize-1)/A.blocksize;
cusparseZbsrmv( cusparseHandle, dirA,
CUSPARSE_OPERATION_NON_TRANSPOSE, mb, nb, A.numblocks,
&alpha, descr, A.val, A.row, A.col, A.blocksize, x.val,
&beta, y.val );
//printf("done.\n");
return MAGMA_SUCCESS;
}*/
else {
printf("error: format not supported.\n");
return MAGMA_ERR_NOT_SUPPORTED;
}
}
else if( A.num_cols < x.num_rows ){
magma_int_t num_vecs = x.num_rows / A.num_cols;
if( A.storage_type == Magma_CSR ){
//printf("using CSR kernel for SpMV: ");
magma_zmgecsrmv( MagmaNoTrans, A.num_rows, A.num_cols,
num_vecs, alpha, A.val, A.row, A.col, x.val, beta, y.val );
//printf("done.\n");
return MAGMA_SUCCESS;
}
else if( A.storage_type == Magma_ELLPACK ){
//printf("using ELLPACK kernel for SpMV: ");
magma_zmgeellmv( MagmaNoTrans, A.num_rows, A.num_cols,
num_vecs, A.max_nnz_row, alpha, A.val, A.col, x.val, beta, y.val );
//printf("done.\n");
return MAGMA_SUCCESS;
}
else if( A.storage_type == Magma_ELL ){
//printf("using ELL kernel for SpMV: ");
magma_zmgeelltmv( MagmaNoTrans, A.num_rows, A.num_cols,
num_vecs, A.max_nnz_row, alpha, A.val,
A.col, x.val, beta, y.val );
//printf("done.\n");
return MAGMA_SUCCESS;
}else if( A.storage_type == Magma_SELLP ){
//printf("using SELLP kernel for SpMV: ");
magma_zmgesellpmv( MagmaNoTrans, A.num_rows, A.num_cols,
num_vecs, A.blocksize, A.numblocks, A.alignment,
alpha, A.val, A.col, A.row, x.val, beta, y.val );
//printf("done.\n");
return MAGMA_SUCCESS;
}/*
if( A.storage_type == Magma_DENSE ){
//printf("using DENSE kernel for SpMV: ");
magmablas_zmgemv( MagmaNoTrans, A.num_rows, A.num_cols,
num_vecs, alpha, A.val, A.num_rows, x.val, 1,
beta, y.val, 1 );
//printf("done.\n");
return MAGMA_SUCCESS;
}*/
else {
printf("error: format not supported.\n");
return MAGMA_ERR_NOT_SUPPORTED;
}
}
}
// CPU case missing!
else{
printf("error: CPU not yet supported.\n");
return MAGMA_ERR_NOT_SUPPORTED;
}
return MAGMA_SUCCESS;
}