extern "C" magma_int_t
magma_ccumicgeneratesolverinfo(
magma_c_preconditioner *precond,
magma_queue_t queue )
{
magma_int_t info = 0;
cusparseHandle_t cusparseHandle=NULL;
cusparseMatDescr_t descrL=NULL;
cusparseMatDescr_t descrU=NULL;
// CUSPARSE context //
CHECK_CUSPARSE( cusparseCreate( &cusparseHandle ));
CHECK_CUSPARSE( cusparseSetStream( cusparseHandle, queue->cuda_stream() ));
CHECK_CUSPARSE( cusparseCreateMatDescr( &descrL ));
CHECK_CUSPARSE( cusparseSetMatType( descrL, CUSPARSE_MATRIX_TYPE_TRIANGULAR ));
CHECK_CUSPARSE( cusparseSetMatDiagType( descrL, CUSPARSE_DIAG_TYPE_NON_UNIT ));
CHECK_CUSPARSE( cusparseSetMatIndexBase( descrL, CUSPARSE_INDEX_BASE_ZERO ));
CHECK_CUSPARSE( cusparseSetMatFillMode( descrL, CUSPARSE_FILL_MODE_LOWER ));
CHECK_CUSPARSE( cusparseCreateSolveAnalysisInfo( &precond->cuinfoL ));
CHECK_CUSPARSE( cusparseCcsrsm_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 ));
CHECK_CUSPARSE( cusparseCreateMatDescr( &descrU ));
CHECK_CUSPARSE( cusparseSetMatType( descrU, CUSPARSE_MATRIX_TYPE_TRIANGULAR ));
CHECK_CUSPARSE( cusparseSetMatDiagType( descrU, CUSPARSE_DIAG_TYPE_NON_UNIT ));
CHECK_CUSPARSE( cusparseSetMatIndexBase( descrU, CUSPARSE_INDEX_BASE_ZERO ));
CHECK_CUSPARSE( cusparseSetMatFillMode( descrU, CUSPARSE_FILL_MODE_LOWER ));
CHECK_CUSPARSE( cusparseCreateSolveAnalysisInfo( &precond->cuinfoU ));
CHECK_CUSPARSE( cusparseCcsrsm_analysis( cusparseHandle,
CUSPARSE_OPERATION_TRANSPOSE, precond->M.num_rows,
precond->M.nnz, descrU,
precond->M.dval, precond->M.drow, precond->M.dcol, precond->cuinfoU ));
/*
// to enable also the block-asynchronous iteration for the triangular solves
CHECK( magma_cmtransfer( precond->M, &hA, Magma_DEV, Magma_CPU, queue ));
hA.storage_type = Magma_CSR;
CHECK( magma_ccsrsplit( 256, hA, &hD, &hR, queue ));
CHECK( magma_cmtransfer( hD, &precond->LD, Magma_CPU, Magma_DEV, queue ));
CHECK( magma_cmtransfer( hR, &precond->L, Magma_CPU, Magma_DEV, queue ));
magma_cmfree(&hD, queue );
magma_cmfree(&hR, queue );
CHECK( magma_c_cucsrtranspose( hA, &hAt, queue ));
CHECK( magma_ccsrsplit( 256, hAt, &hD, &hR, queue ));
CHECK( magma_cmtransfer( hD, &precond->UD, Magma_CPU, Magma_DEV, queue ));
CHECK( magma_cmtransfer( hR, &precond->U, Magma_CPU, Magma_DEV, queue ));
magma_cmfree(&hD, queue );
magma_cmfree(&hR, queue );
magma_cmfree(&hA, queue );
magma_cmfree(&hAt, queue );
*/
cleanup:
cusparseDestroyMatDescr( descrL );
cusparseDestroyMatDescr( descrU );
cusparseDestroy( cusparseHandle );
return info;
}
extern "C" magma_int_t
magma_ccumiccsetup(
magma_c_matrix A,
magma_c_preconditioner *precond,
magma_queue_t queue )
{
magma_int_t info = 0;
cusparseHandle_t cusparseHandle=NULL;
cusparseMatDescr_t descrA=NULL;
cusparseMatDescr_t descrL=NULL;
cusparseMatDescr_t descrU=NULL;
#if CUDA_VERSION >= 7000
csric02Info_t info_M=NULL;
void *pBuffer = NULL;
#endif
magma_c_matrix hA={Magma_CSR}, hACSR={Magma_CSR}, U={Magma_CSR};
CHECK( magma_cmtransfer( A, &hA, A.memory_location, Magma_CPU, queue ));
U.diagorder_type = Magma_VALUE;
CHECK( magma_cmconvert( hA, &hACSR, hA.storage_type, Magma_CSR, queue ));
// in case using fill-in
if( precond->levels > 0 ){
magma_c_matrix hAL={Magma_CSR}, hAUt={Magma_CSR};
CHECK( magma_csymbilu( &hACSR, precond->levels, &hAL, &hAUt, queue ));
magma_cmfree(&hAL, queue);
magma_cmfree(&hAUt, queue);
}
CHECK( magma_cmconvert( hACSR, &U, Magma_CSR, Magma_CSRL, queue ));
magma_cmfree( &hACSR, queue );
CHECK( magma_cmtransfer(U, &(precond->M), Magma_CPU, Magma_DEV, queue ));
// CUSPARSE context //
CHECK_CUSPARSE( cusparseCreate( &cusparseHandle ));
CHECK_CUSPARSE( cusparseSetStream( cusparseHandle, queue->cuda_stream() ));
CHECK_CUSPARSE( cusparseCreateMatDescr( &descrA ));
CHECK_CUSPARSE( cusparseCreateSolveAnalysisInfo( &(precond->cuinfo) ));
// use kernel to manually check for zeros n the diagonal
CHECK( magma_cdiagcheck( precond->M, queue ) );
#if CUDA_VERSION >= 7000
// this version has the bug fixed where a zero on the diagonal causes a crash
CHECK_CUSPARSE( cusparseCreateCsric02Info(&info_M) );
CHECK_CUSPARSE( cusparseSetMatType( descrA, CUSPARSE_MATRIX_TYPE_GENERAL ));
CHECK_CUSPARSE( cusparseSetMatIndexBase( descrA, CUSPARSE_INDEX_BASE_ZERO ));
int buffersize;
int structural_zero;
int numerical_zero;
CHECK_CUSPARSE(
cusparseCcsric02_bufferSize( cusparseHandle,
precond->M.num_rows, precond->M.nnz, descrA,
precond->M.dval, precond->M.drow, precond->M.dcol,
info_M,
&buffersize ) );
CHECK( magma_malloc((void**)&pBuffer, buffersize) );
CHECK_CUSPARSE( cusparseCcsric02_analysis( cusparseHandle,
precond->M.num_rows, precond->M.nnz, descrA,
precond->M.dval, precond->M.drow, precond->M.dcol,
info_M, CUSPARSE_SOLVE_POLICY_NO_LEVEL, pBuffer ));
CHECK_CUSPARSE( cusparseXcsric02_zeroPivot( cusparseHandle, info_M, &numerical_zero ) );
CHECK_CUSPARSE( cusparseXcsric02_zeroPivot( cusparseHandle, info_M, &structural_zero ) );
CHECK_CUSPARSE(
cusparseCcsric02( cusparseHandle,
precond->M.num_rows, precond->M.nnz, descrA,
precond->M.dval, precond->M.drow, precond->M.dcol,
info_M, CUSPARSE_SOLVE_POLICY_NO_LEVEL, pBuffer) );
#else
// this version contains the bug but is needed for backward compability
CHECK_CUSPARSE( cusparseSetMatType( descrA, CUSPARSE_MATRIX_TYPE_SYMMETRIC ));
CHECK_CUSPARSE( cusparseSetMatDiagType( descrA, CUSPARSE_DIAG_TYPE_NON_UNIT ));
CHECK_CUSPARSE( cusparseSetMatIndexBase( descrA, CUSPARSE_INDEX_BASE_ZERO ));
CHECK_CUSPARSE( cusparseSetMatFillMode( descrA, CUSPARSE_FILL_MODE_LOWER ));
CHECK_CUSPARSE( cusparseCcsrsm_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 ));
CHECK_CUSPARSE( cusparseCcsric0( cusparseHandle, CUSPARSE_OPERATION_NON_TRANSPOSE,
precond->M.num_rows, descrA,
precond->M.dval,
precond->M.drow,
precond->M.dcol,
precond->cuinfo ));
#endif
CHECK_CUSPARSE( cusparseCreateMatDescr( &descrL ));
CHECK_CUSPARSE( cusparseSetMatType( descrL, CUSPARSE_MATRIX_TYPE_TRIANGULAR ));
CHECK_CUSPARSE( cusparseSetMatDiagType( descrL, CUSPARSE_DIAG_TYPE_NON_UNIT ));
CHECK_CUSPARSE( cusparseSetMatIndexBase( descrL, CUSPARSE_INDEX_BASE_ZERO ));
CHECK_CUSPARSE( cusparseSetMatFillMode( descrL, CUSPARSE_FILL_MODE_LOWER ));
CHECK_CUSPARSE( cusparseCreateSolveAnalysisInfo( &precond->cuinfoL ));
CHECK_CUSPARSE( cusparseCcsrsm_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 ));
CHECK_CUSPARSE( cusparseCreateMatDescr( &descrU ));
CHECK_CUSPARSE( cusparseSetMatType( descrU, CUSPARSE_MATRIX_TYPE_TRIANGULAR ));
CHECK_CUSPARSE( cusparseSetMatDiagType( descrU, CUSPARSE_DIAG_TYPE_NON_UNIT ));
CHECK_CUSPARSE( cusparseSetMatIndexBase( descrU, CUSPARSE_INDEX_BASE_ZERO ));
CHECK_CUSPARSE( cusparseSetMatFillMode( descrU, CUSPARSE_FILL_MODE_LOWER ));
CHECK_CUSPARSE( cusparseCreateSolveAnalysisInfo( &precond->cuinfoU ));
CHECK_CUSPARSE( cusparseCcsrsm_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( precond->maxiter < 50 ){
//prepare for iterative solves
// copy the matrix to precond->L and (transposed) to precond->U
CHECK( magma_cmtransfer(precond->M, &(precond->L), Magma_DEV, Magma_DEV, queue ));
CHECK( magma_cmtranspose( precond->L, &(precond->U), queue ));
// extract the diagonal of L into precond->d
CHECK( magma_cjacobisetup_diagscal( precond->L, &precond->d, queue ));
CHECK( magma_cvinit( &precond->work1, Magma_DEV, hA.num_rows, 1, MAGMA_C_ZERO, queue ));
// extract the diagonal of U into precond->d2
CHECK( magma_cjacobisetup_diagscal( precond->U, &precond->d2, queue ));
CHECK( magma_cvinit( &precond->work2, Magma_DEV, hA.num_rows, 1, MAGMA_C_ZERO, queue ));
}
/*
// to enable also the block-asynchronous iteration for the triangular solves
CHECK( magma_cmtransfer( precond->M, &hA, Magma_DEV, Magma_CPU, queue ));
hA.storage_type = Magma_CSR;
magma_c_matrix hD, hR, hAt
CHECK( magma_ccsrsplit( 256, hA, &hD, &hR, queue ));
CHECK( magma_cmtransfer( hD, &precond->LD, Magma_CPU, Magma_DEV, queue ));
CHECK( magma_cmtransfer( hR, &precond->L, Magma_CPU, Magma_DEV, queue ));
magma_cmfree(&hD, queue );
magma_cmfree(&hR, queue );
CHECK( magma_c_cucsrtranspose( hA, &hAt, queue ));
CHECK( magma_ccsrsplit( 256, hAt, &hD, &hR, queue ));
CHECK( magma_cmtransfer( hD, &precond->UD, Magma_CPU, Magma_DEV, queue ));
CHECK( magma_cmtransfer( hR, &precond->U, Magma_CPU, Magma_DEV, queue ));
magma_cmfree(&hD, queue );
magma_cmfree(&hR, queue );
magma_cmfree(&hA, queue );
magma_cmfree(&hAt, queue );
*/
cleanup:
#if CUDA_VERSION >= 7000
magma_free( pBuffer );
cusparseDestroyCsric02Info( info_M );
#endif
cusparseDestroySolveAnalysisInfo( precond->cuinfo );
cusparseDestroyMatDescr( descrL );
cusparseDestroyMatDescr( descrU );
cusparseDestroyMatDescr( descrA );
cusparseDestroy( cusparseHandle );
magma_cmfree(&U, queue );
magma_cmfree(&hA, queue );
return info;
}
extern "C" magma_int_t
magma_ccumilugeneratesolverinfo(
magma_c_preconditioner *precond,
magma_queue_t queue )
{
magma_int_t info = 0;
cusparseHandle_t cusparseHandle=NULL;
cusparseMatDescr_t descrL=NULL;
cusparseMatDescr_t descrU=NULL;
magma_c_matrix hA={Magma_CSR}, hL={Magma_CSR}, hU={Magma_CSR};
if (precond->L.memory_location != Magma_DEV ){
CHECK( magma_cmtransfer( precond->M, &hA,
precond->M.memory_location, Magma_CPU, queue ));
hL.diagorder_type = Magma_UNITY;
CHECK( magma_cmconvert( hA, &hL , Magma_CSR, Magma_CSRL, queue ));
hU.diagorder_type = Magma_VALUE;
CHECK( magma_cmconvert( hA, &hU , Magma_CSR, Magma_CSRU, queue ));
CHECK( magma_cmtransfer( hL, &(precond->L), Magma_CPU, Magma_DEV, queue ));
CHECK( magma_cmtransfer( hU, &(precond->U), Magma_CPU, Magma_DEV, queue ));
magma_cmfree(&hA, queue );
magma_cmfree(&hL, queue );
magma_cmfree(&hU, queue );
}
// CUSPARSE context //
CHECK_CUSPARSE( cusparseCreate( &cusparseHandle ));
CHECK_CUSPARSE( cusparseSetStream( cusparseHandle, queue->cuda_stream() ));
CHECK_CUSPARSE( cusparseCreateMatDescr( &descrL ));
CHECK_CUSPARSE( cusparseSetMatType( descrL, CUSPARSE_MATRIX_TYPE_TRIANGULAR ));
CHECK_CUSPARSE( cusparseSetMatDiagType( descrL, CUSPARSE_DIAG_TYPE_UNIT ));
CHECK_CUSPARSE( cusparseSetMatIndexBase( descrL, CUSPARSE_INDEX_BASE_ZERO ));
CHECK_CUSPARSE( cusparseSetMatFillMode( descrL, CUSPARSE_FILL_MODE_LOWER ));
CHECK_CUSPARSE( cusparseCreateSolveAnalysisInfo( &precond->cuinfoL ));
CHECK_CUSPARSE( cusparseCcsrsm_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 ));
CHECK_CUSPARSE( cusparseCreateMatDescr( &descrU ));
CHECK_CUSPARSE( cusparseSetMatType( descrU, CUSPARSE_MATRIX_TYPE_TRIANGULAR ));
CHECK_CUSPARSE( cusparseSetMatDiagType( descrU, CUSPARSE_DIAG_TYPE_NON_UNIT ));
CHECK_CUSPARSE( cusparseSetMatIndexBase( descrU, CUSPARSE_INDEX_BASE_ZERO ));
CHECK_CUSPARSE( cusparseSetMatFillMode( descrU, CUSPARSE_FILL_MODE_UPPER ));
CHECK_CUSPARSE( cusparseCreateSolveAnalysisInfo( &precond->cuinfoU ));
CHECK_CUSPARSE( cusparseCcsrsm_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( precond->maxiter < 50 ){
//prepare for iterative solves
// extract the diagonal of L into precond->d
CHECK( magma_cjacobisetup_diagscal( precond->L, &precond->d, queue ));
CHECK( magma_cvinit( &precond->work1, Magma_DEV, precond->U.num_rows, 1, MAGMA_C_ZERO, queue ));
// extract the diagonal of U into precond->d2
CHECK( magma_cjacobisetup_diagscal( precond->U, &precond->d2, queue ));
CHECK( magma_cvinit( &precond->work2, Magma_DEV, precond->U.num_rows, 1, MAGMA_C_ZERO, queue ));
}
cleanup:
cusparseDestroyMatDescr( descrL );
cusparseDestroyMatDescr( descrU );
cusparseDestroy( cusparseHandle );
return info;
}
extern "C" magma_int_t
magma_ccumilusetup_transpose(
magma_c_matrix A,
magma_c_preconditioner *precond,
magma_queue_t queue )
{
magma_int_t info = 0;
magma_c_matrix Ah1={Magma_CSR}, Ah2={Magma_CSR};
cusparseHandle_t cusparseHandle=NULL;
cusparseMatDescr_t descrLT=NULL;
cusparseMatDescr_t descrUT=NULL;
// CUSPARSE context //
CHECK_CUSPARSE( cusparseCreate( &cusparseHandle ));
CHECK_CUSPARSE( cusparseSetStream( cusparseHandle, queue->cuda_stream() ));
// transpose the matrix
magma_cmtransfer( precond->L, &Ah1, Magma_DEV, Magma_CPU, queue );
magma_cmconvert( Ah1, &Ah2, A.storage_type, Magma_CSR, queue );
magma_cmfree(&Ah1, queue );
magma_cmtransposeconjugate( Ah2, &Ah1, queue );
magma_cmfree(&Ah2, queue );
Ah2.blocksize = A.blocksize;
Ah2.alignment = A.alignment;
magma_cmconvert( Ah1, &Ah2, Magma_CSR, A.storage_type, queue );
magma_cmfree(&Ah1, queue );
magma_cmtransfer( Ah2, &(precond->LT), Magma_CPU, Magma_DEV, queue );
magma_cmfree(&Ah2, queue );
magma_cmtransfer( precond->U, &Ah1, Magma_DEV, Magma_CPU, queue );
magma_cmconvert( Ah1, &Ah2, A.storage_type, Magma_CSR, queue );
magma_cmfree(&Ah1, queue );
magma_cmtransposeconjugate( Ah2, &Ah1, queue );
magma_cmfree(&Ah2, queue );
Ah2.blocksize = A.blocksize;
Ah2.alignment = A.alignment;
magma_cmconvert( Ah1, &Ah2, Magma_CSR, A.storage_type, queue );
magma_cmfree(&Ah1, queue );
magma_cmtransfer( Ah2, &(precond->UT), Magma_CPU, Magma_DEV, queue );
magma_cmfree(&Ah2, queue );
CHECK_CUSPARSE( cusparseCreateMatDescr( &descrLT ));
CHECK_CUSPARSE( cusparseSetMatType( descrLT, CUSPARSE_MATRIX_TYPE_TRIANGULAR ));
CHECK_CUSPARSE( cusparseSetMatDiagType( descrLT, CUSPARSE_DIAG_TYPE_UNIT ));
CHECK_CUSPARSE( cusparseSetMatIndexBase( descrLT, CUSPARSE_INDEX_BASE_ZERO ));
CHECK_CUSPARSE( cusparseSetMatFillMode( descrLT, CUSPARSE_FILL_MODE_UPPER ));
CHECK_CUSPARSE( cusparseCreateSolveAnalysisInfo( &precond->cuinfoLT ));
CHECK_CUSPARSE( cusparseCcsrsm_analysis( cusparseHandle,
CUSPARSE_OPERATION_NON_TRANSPOSE, precond->LT.num_rows,
precond->LT.nnz, descrLT,
precond->LT.dval, precond->LT.drow, precond->LT.dcol, precond->cuinfoLT ));
CHECK_CUSPARSE( cusparseCreateMatDescr( &descrUT ));
CHECK_CUSPARSE( cusparseSetMatType( descrUT, CUSPARSE_MATRIX_TYPE_TRIANGULAR ));
CHECK_CUSPARSE( cusparseSetMatDiagType( descrUT, CUSPARSE_DIAG_TYPE_NON_UNIT ));
CHECK_CUSPARSE( cusparseSetMatIndexBase( descrUT, CUSPARSE_INDEX_BASE_ZERO ));
CHECK_CUSPARSE( cusparseSetMatFillMode( descrUT, CUSPARSE_FILL_MODE_LOWER ));
CHECK_CUSPARSE( cusparseCreateSolveAnalysisInfo( &precond->cuinfoUT ));
CHECK_CUSPARSE( cusparseCcsrsm_analysis( cusparseHandle,
CUSPARSE_OPERATION_NON_TRANSPOSE, precond->UT.num_rows,
precond->UT.nnz, descrUT,
precond->UT.dval, precond->UT.drow, precond->UT.dcol, precond->cuinfoUT ));
cleanup:
cusparseDestroyMatDescr( descrLT );
cusparseDestroyMatDescr( descrUT );
cusparseDestroy( cusparseHandle );
magma_cmfree(&Ah1, queue );
magma_cmfree(&Ah2, queue );
return info;
}
extern "C" magma_int_t
magma_ccumiccsetup(
magma_c_matrix A,
magma_c_preconditioner *precond,
magma_queue_t queue )
{
magma_int_t info = 0;
cusparseHandle_t cusparseHandle=NULL;
cusparseMatDescr_t descrA=NULL;
cusparseMatDescr_t descrL=NULL;
cusparseMatDescr_t descrU=NULL;
magma_c_matrix hA={Magma_CSR}, hACSR={Magma_CSR}, U={Magma_CSR};
CHECK( magma_cmtransfer( A, &hA, A.memory_location, Magma_CPU, queue ));
U.diagorder_type = Magma_VALUE;
CHECK( magma_cmconvert( hA, &hACSR, hA.storage_type, Magma_CSR, queue ));
// in case using fill-in
if( precond->levels > 0 ){
magma_c_matrix hAL={Magma_CSR}, hAUt={Magma_CSR};
CHECK( magma_csymbilu( &hACSR, precond->levels, &hAL, &hAUt, queue ));
magma_cmfree(&hAL, queue);
magma_cmfree(&hAUt, queue);
}
CHECK( magma_cmconvert( hACSR, &U, Magma_CSR, Magma_CSRL, queue ));
magma_cmfree( &hACSR, queue );
CHECK( magma_cmtransfer(U, &(precond->M), Magma_CPU, Magma_DEV, queue ));
// CUSPARSE context //
CHECK_CUSPARSE( cusparseCreate( &cusparseHandle ));
CHECK_CUSPARSE( cusparseSetStream( cusparseHandle, queue ));
CHECK_CUSPARSE( cusparseCreateMatDescr( &descrA ));
CHECK_CUSPARSE( cusparseSetMatType( descrA, CUSPARSE_MATRIX_TYPE_SYMMETRIC ));
CHECK_CUSPARSE( cusparseSetMatDiagType( descrA, CUSPARSE_DIAG_TYPE_NON_UNIT ));
CHECK_CUSPARSE( cusparseSetMatIndexBase( descrA, CUSPARSE_INDEX_BASE_ZERO ));
CHECK_CUSPARSE( cusparseSetMatFillMode( descrA, CUSPARSE_FILL_MODE_LOWER ));
CHECK_CUSPARSE( cusparseCreateSolveAnalysisInfo( &(precond->cuinfo) ));
CHECK_CUSPARSE( cusparseCcsrsm_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 ));
CHECK_CUSPARSE( cusparseCcsric0( cusparseHandle, CUSPARSE_OPERATION_NON_TRANSPOSE,
precond->M.num_rows, descrA,
precond->M.dval,
precond->M.drow,
precond->M.dcol,
precond->cuinfo ));
CHECK_CUSPARSE( cusparseCreateMatDescr( &descrL ));
CHECK_CUSPARSE( cusparseSetMatType( descrL, CUSPARSE_MATRIX_TYPE_TRIANGULAR ));
CHECK_CUSPARSE( cusparseSetMatDiagType( descrL, CUSPARSE_DIAG_TYPE_NON_UNIT ));
CHECK_CUSPARSE( cusparseSetMatIndexBase( descrL, CUSPARSE_INDEX_BASE_ZERO ));
CHECK_CUSPARSE( cusparseSetMatFillMode( descrL, CUSPARSE_FILL_MODE_LOWER ));
CHECK_CUSPARSE( cusparseCreateSolveAnalysisInfo( &precond->cuinfoL ));
CHECK_CUSPARSE( cusparseCcsrsm_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 ));
CHECK_CUSPARSE( cusparseCreateMatDescr( &descrU ));
CHECK_CUSPARSE( cusparseSetMatType( descrU, CUSPARSE_MATRIX_TYPE_TRIANGULAR ));
CHECK_CUSPARSE( cusparseSetMatDiagType( descrU, CUSPARSE_DIAG_TYPE_NON_UNIT ));
CHECK_CUSPARSE( cusparseSetMatIndexBase( descrU, CUSPARSE_INDEX_BASE_ZERO ));
CHECK_CUSPARSE( cusparseSetMatFillMode( descrU, CUSPARSE_FILL_MODE_LOWER ));
CHECK_CUSPARSE( cusparseCreateSolveAnalysisInfo( &precond->cuinfoU ));
CHECK_CUSPARSE( cusparseCcsrsm_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( precond->maxiter < 50 ){
//prepare for iterative solves
// copy the matrix to precond->L and (transposed) to precond->U
CHECK( magma_cmtransfer(precond->M, &(precond->L), Magma_DEV, Magma_DEV, queue ));
CHECK( magma_cmtranspose( precond->L, &(precond->U), queue ));
// extract the diagonal of L into precond->d
CHECK( magma_cjacobisetup_diagscal( precond->L, &precond->d, queue ));
CHECK( magma_cvinit( &precond->work1, Magma_DEV, hA.num_rows, 1, MAGMA_C_ZERO, queue ));
// extract the diagonal of U into precond->d2
CHECK( magma_cjacobisetup_diagscal( precond->U, &precond->d2, queue ));
CHECK( magma_cvinit( &precond->work2, Magma_DEV, hA.num_rows, 1, MAGMA_C_ZERO, queue ));
}
/*
// to enable also the block-asynchronous iteration for the triangular solves
CHECK( magma_cmtransfer( precond->M, &hA, Magma_DEV, Magma_CPU, queue ));
hA.storage_type = Magma_CSR;
magma_c_matrix hD, hR, hAt
CHECK( magma_ccsrsplit( 256, hA, &hD, &hR, queue ));
CHECK( magma_cmtransfer( hD, &precond->LD, Magma_CPU, Magma_DEV, queue ));
CHECK( magma_cmtransfer( hR, &precond->L, Magma_CPU, Magma_DEV, queue ));
magma_cmfree(&hD, queue );
magma_cmfree(&hR, queue );
CHECK( magma_c_cucsrtranspose( hA, &hAt, queue ));
CHECK( magma_ccsrsplit( 256, hAt, &hD, &hR, queue ));
CHECK( magma_cmtransfer( hD, &precond->UD, Magma_CPU, Magma_DEV, queue ));
CHECK( magma_cmtransfer( hR, &precond->U, Magma_CPU, Magma_DEV, queue ));
magma_cmfree(&hD, queue );
magma_cmfree(&hR, queue );
magma_cmfree(&hA, queue );
magma_cmfree(&hAt, queue );
*/
cleanup:
cusparseDestroySolveAnalysisInfo( precond->cuinfo );
cusparseDestroyMatDescr( descrL );
cusparseDestroyMatDescr( descrU );
cusparseDestroyMatDescr( descrA );
cusparseDestroy( cusparseHandle );
magma_cmfree(&U, queue );
magma_cmfree(&hA, queue );
return info;
}
extern "C" magma_int_t
magma_ccumilusetup(
magma_c_matrix A,
magma_c_preconditioner *precond,
magma_queue_t queue )
{
magma_int_t info = 0;
cusparseHandle_t cusparseHandle=NULL;
cusparseMatDescr_t descrA=NULL;
cusparseMatDescr_t descrL=NULL;
cusparseMatDescr_t descrU=NULL;
//magma_cprint_matrix(A, queue );
// copy matrix into preconditioner parameter
magma_c_matrix hA={Magma_CSR}, hACSR={Magma_CSR};
magma_c_matrix hL={Magma_CSR}, hU={Magma_CSR};
CHECK( magma_cmtransfer( A, &hA, A.memory_location, Magma_CPU, queue ));
CHECK( magma_cmconvert( hA, &hACSR, hA.storage_type, Magma_CSR, queue ));
// in case using fill-in
if( precond->levels > 0 ){
magma_c_matrix hAL={Magma_CSR}, hAUt={Magma_CSR};
CHECK( magma_csymbilu( &hACSR, precond->levels, &hAL, &hAUt, queue ));
magma_cmfree(&hAL, queue);
magma_cmfree(&hAUt, queue);
}
CHECK( magma_cmtransfer(hACSR, &(precond->M), Magma_CPU, Magma_DEV, queue ));
magma_cmfree( &hA, queue );
magma_cmfree( &hACSR, queue );
// CUSPARSE context //
CHECK_CUSPARSE( cusparseCreate( &cusparseHandle ));
CHECK_CUSPARSE( cusparseSetStream( cusparseHandle, queue ));
CHECK_CUSPARSE( cusparseCreateMatDescr( &descrA ));
CHECK_CUSPARSE( cusparseSetMatType( descrA, CUSPARSE_MATRIX_TYPE_GENERAL ));
CHECK_CUSPARSE( cusparseSetMatDiagType( descrA, CUSPARSE_DIAG_TYPE_NON_UNIT ));
CHECK_CUSPARSE( cusparseSetMatIndexBase( descrA, CUSPARSE_INDEX_BASE_ZERO ));
CHECK_CUSPARSE( cusparseCreateSolveAnalysisInfo( &(precond->cuinfo) ));
CHECK_CUSPARSE( cusparseCcsrsm_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 ));
CHECK_CUSPARSE( cusparseCcsrilu0( cusparseHandle, CUSPARSE_OPERATION_NON_TRANSPOSE,
precond->M.num_rows, descrA,
precond->M.dval,
precond->M.drow,
precond->M.dcol,
precond->cuinfo ));
CHECK( magma_cmtransfer( precond->M, &hA, Magma_DEV, Magma_CPU, queue ));
hL.diagorder_type = Magma_UNITY;
CHECK( magma_cmconvert( hA, &hL , Magma_CSR, Magma_CSRL, queue ));
hU.diagorder_type = Magma_VALUE;
CHECK( magma_cmconvert( hA, &hU , Magma_CSR, Magma_CSRU, queue ));
CHECK( magma_cmtransfer( hL, &(precond->L), Magma_CPU, Magma_DEV, queue ));
CHECK( magma_cmtransfer( hU, &(precond->U), Magma_CPU, Magma_DEV, queue ));
CHECK_CUSPARSE( cusparseCreateMatDescr( &descrL ));
CHECK_CUSPARSE( cusparseSetMatType( descrL, CUSPARSE_MATRIX_TYPE_TRIANGULAR ));
CHECK_CUSPARSE( cusparseSetMatDiagType( descrL, CUSPARSE_DIAG_TYPE_UNIT ));
CHECK_CUSPARSE( cusparseSetMatIndexBase( descrL, CUSPARSE_INDEX_BASE_ZERO ));
CHECK_CUSPARSE( cusparseSetMatFillMode( descrL, CUSPARSE_FILL_MODE_LOWER ));
CHECK_CUSPARSE( cusparseCreateSolveAnalysisInfo( &precond->cuinfoL ));
CHECK_CUSPARSE( cusparseCcsrsm_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 ));
CHECK_CUSPARSE( cusparseCreateMatDescr( &descrU ));
CHECK_CUSPARSE( cusparseSetMatType( descrU, CUSPARSE_MATRIX_TYPE_TRIANGULAR ));
CHECK_CUSPARSE( cusparseSetMatDiagType( descrU, CUSPARSE_DIAG_TYPE_NON_UNIT ));
CHECK_CUSPARSE( cusparseSetMatIndexBase( descrU, CUSPARSE_INDEX_BASE_ZERO ));
CHECK_CUSPARSE( cusparseSetMatFillMode( descrU, CUSPARSE_FILL_MODE_UPPER ));
CHECK_CUSPARSE( cusparseCreateSolveAnalysisInfo( &precond->cuinfoU ));
CHECK_CUSPARSE( cusparseCcsrsm_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( precond->maxiter < 50 ){
//prepare for iterative solves
// extract the diagonal of L into precond->d
CHECK( magma_cjacobisetup_diagscal( precond->L, &precond->d, queue ));
CHECK( magma_cvinit( &precond->work1, Magma_DEV, hA.num_rows, 1, MAGMA_C_ZERO, queue ));
// extract the diagonal of U into precond->d2
CHECK( magma_cjacobisetup_diagscal( precond->U, &precond->d2, queue ));
CHECK( magma_cvinit( &precond->work2, Magma_DEV, hA.num_rows, 1, MAGMA_C_ZERO, queue ));
}
cleanup:
cusparseDestroySolveAnalysisInfo( precond->cuinfo );
cusparseDestroyMatDescr( descrA );
cusparseDestroyMatDescr( descrL );
cusparseDestroyMatDescr( descrU );
cusparseDestroy( cusparseHandle );
magma_cmfree( &hA, queue );
magma_cmfree( &hACSR, queue );
magma_cmfree(&hA, queue );
magma_cmfree(&hL, queue );
magma_cmfree(&hU, queue );
return info;
}