Teuchos::RCP<cusparseSolveAnalysisInfo_t> Kokkos::CUSPARSEdetails::createSolveAnalysisInfo()
{
cusparseSolveAnalysisInfo_t *info = new cusparseSolveAnalysisInfo_t;
cusparseStatus_t status = cusparseCreateSolveAnalysisInfo(info);
TEUCHOS_TEST_FOR_EXCEPTION(status == CUSPARSE_STATUS_ALLOC_FAILED, std::runtime_error,
"Kokkos::CUSPARSEdetails::createSolveAnalysisInfo(): the resources could not be allocated.")
TEUCHOS_TEST_FOR_EXCEPTION(status != CUSPARSE_STATUS_SUCCESS, std::runtime_error,
"Kokkos::CUSPARSEdetails::createSolveAnalysisInfo(): unspecified error.")
return Teuchos::rcpWithDealloc(info,CUSPARSESolveAnalysisDestroyer(),true);
}
extern "C" magma_int_t
magma_dcumicgeneratesolverinfo(
magma_d_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( 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 ));
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( 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 ));
/*
// to enable also the block-asynchronous iteration for the triangular solves
CHECK( magma_dmtransfer( precond->M, &hA, Magma_DEV, Magma_CPU, queue ));
hA.storage_type = Magma_CSR;
CHECK( magma_dcsrsplit( 256, hA, &hD, &hR, queue ));
CHECK( magma_dmtransfer( hD, &precond->LD, Magma_CPU, Magma_DEV, queue ));
CHECK( magma_dmtransfer( hR, &precond->L, Magma_CPU, Magma_DEV, queue ));
magma_dmfree(&hD, queue );
magma_dmfree(&hR, queue );
CHECK( magma_d_cucsrtranspose( hA, &hAt, queue ));
CHECK( magma_dcsrsplit( 256, hAt, &hD, &hR, queue ));
CHECK( magma_dmtransfer( hD, &precond->UD, Magma_CPU, Magma_DEV, queue ));
CHECK( magma_dmtransfer( hR, &precond->U, Magma_CPU, Magma_DEV, queue ));
magma_dmfree(&hD, queue );
magma_dmfree(&hR, queue );
magma_dmfree(&hA, queue );
magma_dmfree(&hAt, queue );
*/
cleanup:
cusparseDestroyMatDescr( descrL );
cusparseDestroyMatDescr( descrU );
cusparseDestroy( cusparseHandle );
return info;
}
extern "C" magma_int_t
magma_dcumiccsetup(
magma_d_matrix A,
magma_d_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_d_matrix hA={Magma_CSR}, hACSR={Magma_CSR}, U={Magma_CSR};
CHECK( magma_dmtransfer( A, &hA, A.memory_location, Magma_CPU, queue ));
U.diagorder_type = Magma_VALUE;
CHECK( magma_dmconvert( hA, &hACSR, hA.storage_type, Magma_CSR, queue ));
// in case using fill-in
if( precond->levels > 0 ){
magma_d_matrix hAL={Magma_CSR}, hAUt={Magma_CSR};
CHECK( magma_dsymbilu( &hACSR, precond->levels, &hAL, &hAUt, queue ));
magma_dmfree(&hAL, queue);
magma_dmfree(&hAUt, queue);
}
CHECK( magma_dmconvert( hACSR, &U, Magma_CSR, Magma_CSRL, queue ));
magma_dmfree( &hACSR, queue );
CHECK( magma_dmtransfer(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_ddiagcheck( 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(
cusparseDcsric02_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( cusparseDcsric02_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(
cusparseDcsric02( 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( 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 ));
CHECK_CUSPARSE( cusparseDcsric0( 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( 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 ));
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( 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( precond->maxiter < 50 ){
//prepare for iterative solves
// copy the matrix to precond->L and (transposed) to precond->U
CHECK( magma_dmtransfer(precond->M, &(precond->L), Magma_DEV, Magma_DEV, queue ));
CHECK( magma_dmtranspose( precond->L, &(precond->U), queue ));
// extract the diagonal of L into precond->d
CHECK( magma_djacobisetup_diagscal( precond->L, &precond->d, queue ));
CHECK( magma_dvinit( &precond->work1, Magma_DEV, hA.num_rows, 1, MAGMA_D_ZERO, queue ));
// extract the diagonal of U into precond->d2
CHECK( magma_djacobisetup_diagscal( precond->U, &precond->d2, queue ));
CHECK( magma_dvinit( &precond->work2, Magma_DEV, hA.num_rows, 1, MAGMA_D_ZERO, queue ));
}
/*
// to enable also the block-asynchronous iteration for the triangular solves
CHECK( magma_dmtransfer( precond->M, &hA, Magma_DEV, Magma_CPU, queue ));
hA.storage_type = Magma_CSR;
magma_d_matrix hD, hR, hAt
CHECK( magma_dcsrsplit( 256, hA, &hD, &hR, queue ));
CHECK( magma_dmtransfer( hD, &precond->LD, Magma_CPU, Magma_DEV, queue ));
CHECK( magma_dmtransfer( hR, &precond->L, Magma_CPU, Magma_DEV, queue ));
magma_dmfree(&hD, queue );
magma_dmfree(&hR, queue );
CHECK( magma_d_cucsrtranspose( hA, &hAt, queue ));
CHECK( magma_dcsrsplit( 256, hAt, &hD, &hR, queue ));
CHECK( magma_dmtransfer( hD, &precond->UD, Magma_CPU, Magma_DEV, queue ));
CHECK( magma_dmtransfer( hR, &precond->U, Magma_CPU, Magma_DEV, queue ));
magma_dmfree(&hD, queue );
magma_dmfree(&hR, queue );
magma_dmfree(&hA, queue );
magma_dmfree(&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_dmfree(&U, queue );
magma_dmfree(&hA, queue );
return info;
}
extern "C" magma_int_t
magma_dcumilugeneratesolverinfo(
magma_d_preconditioner *precond,
magma_queue_t queue )
{
magma_int_t info = 0;
cusparseHandle_t cusparseHandle=NULL;
cusparseMatDescr_t descrL=NULL;
cusparseMatDescr_t descrU=NULL;
magma_d_matrix hA={Magma_CSR}, hL={Magma_CSR}, hU={Magma_CSR};
if (precond->L.memory_location != Magma_DEV ){
CHECK( magma_dmtransfer( precond->M, &hA,
precond->M.memory_location, Magma_CPU, queue ));
hL.diagorder_type = Magma_UNITY;
CHECK( magma_dmconvert( hA, &hL , Magma_CSR, Magma_CSRL, queue ));
hU.diagorder_type = Magma_VALUE;
CHECK( magma_dmconvert( hA, &hU , Magma_CSR, Magma_CSRU, queue ));
CHECK( magma_dmtransfer( hL, &(precond->L), Magma_CPU, Magma_DEV, queue ));
CHECK( magma_dmtransfer( hU, &(precond->U), Magma_CPU, Magma_DEV, queue ));
magma_dmfree(&hA, queue );
magma_dmfree(&hL, queue );
magma_dmfree(&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( 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 ));
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( 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( precond->maxiter < 50 ){
//prepare for iterative solves
// extract the diagonal of L into precond->d
CHECK( magma_djacobisetup_diagscal( precond->L, &precond->d, queue ));
CHECK( magma_dvinit( &precond->work1, Magma_DEV, precond->U.num_rows, 1, MAGMA_D_ZERO, queue ));
// extract the diagonal of U into precond->d2
CHECK( magma_djacobisetup_diagscal( precond->U, &precond->d2, queue ));
CHECK( magma_dvinit( &precond->work2, Magma_DEV, precond->U.num_rows, 1, MAGMA_D_ZERO, queue ));
}
cleanup:
cusparseDestroyMatDescr( descrL );
cusparseDestroyMatDescr( descrU );
cusparseDestroy( cusparseHandle );
return info;
}
extern "C" magma_int_t
magma_dcumilusetup_transpose(
magma_d_matrix A,
magma_d_preconditioner *precond,
magma_queue_t queue )
{
magma_int_t info = 0;
magma_d_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_dmtransfer( precond->L, &Ah1, Magma_DEV, Magma_CPU, queue );
magma_dmconvert( Ah1, &Ah2, A.storage_type, Magma_CSR, queue );
magma_dmfree(&Ah1, queue );
magma_dmtransposeconjugate( Ah2, &Ah1, queue );
magma_dmfree(&Ah2, queue );
Ah2.blocksize = A.blocksize;
Ah2.alignment = A.alignment;
magma_dmconvert( Ah1, &Ah2, Magma_CSR, A.storage_type, queue );
magma_dmfree(&Ah1, queue );
magma_dmtransfer( Ah2, &(precond->LT), Magma_CPU, Magma_DEV, queue );
magma_dmfree(&Ah2, queue );
magma_dmtransfer( precond->U, &Ah1, Magma_DEV, Magma_CPU, queue );
magma_dmconvert( Ah1, &Ah2, A.storage_type, Magma_CSR, queue );
magma_dmfree(&Ah1, queue );
magma_dmtransposeconjugate( Ah2, &Ah1, queue );
magma_dmfree(&Ah2, queue );
Ah2.blocksize = A.blocksize;
Ah2.alignment = A.alignment;
magma_dmconvert( Ah1, &Ah2, Magma_CSR, A.storage_type, queue );
magma_dmfree(&Ah1, queue );
magma_dmtransfer( Ah2, &(precond->UT), Magma_CPU, Magma_DEV, queue );
magma_dmfree(&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( cusparseDcsrsm_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( cusparseDcsrsm_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_dmfree(&Ah1, queue );
magma_dmfree(&Ah2, queue );
return info;
}
int _tmain(int argc, _TCHAR* argv[])
{
int M = 0, N = 0, nz = 0, *I = NULL, *J = NULL;
cuDoubleComplex *val = NULL;
cuDoubleComplex *x, *y;
cuDoubleComplex *d_x, *d_y;
double duration, duration_setup;
std::clock_t setup_clock;
setup_clock = std::clock();
// This will pick the best possible CUDA capable device
cudaDeviceProp deviceProp;
int devID = findCudaDevice(argc, (const char **)argv);
if (devID < 0)
{
printf("no devices found...\n");
exit(EXIT_SUCCESS);
}
checkCudaErrors(cudaGetDeviceProperties(&deviceProp, devID));
// Statistics about the GPU device
printf("> GPU device has %d Multi-Processors, SM %d.%d compute capabilities\n\n",
deviceProp.multiProcessorCount, deviceProp.major, deviceProp.minor);
int version = (deviceProp.major * 0x10 + deviceProp.minor);
if (version < 0x11)
{
printf("Requires a minimum CUDA compute 1.1 capability\n");
cudaDeviceReset();
exit(EXIT_SUCCESS);
}
M = N = 8388608; //2 ^ 23
//M = N = 4194304; //2 ^ 22
//M = N = 2097152; //2 ^ 21
//M = N = 1048576; //2 ^ 20
//M = N = 524288; //2 ^ 19
nz = N * 8;
I = (int *)malloc(sizeof(int)*(N + 1));
J = (int *)malloc(sizeof(int)*nz);
val = (cuDoubleComplex *)malloc(sizeof(cuDoubleComplex)*nz);
genTridiag(I, J, val, N, nz);
x = (cuDoubleComplex*)malloc(sizeof(cuDoubleComplex)* N);
y = (cuDoubleComplex*)malloc(sizeof(cuDoubleComplex)* N);
//create an array for the answer array (Y) and set all of the answers to 0 for the test (could do random)
for (int i = 0; i < N; i++)
{
y[i] = make_cuDoubleComplex(0.0, 0.0);
}
//Get handle to the CUBLAS context
cublasHandle_t cublasHandle = 0;
cublasStatus_t cublasStatus;
cublasStatus = cublasCreate(&cublasHandle);
checkCudaErrors(cublasStatus);
//Get handle to the CUSPARSE context
cusparseHandle_t cusparseHandle = 0;
cusparseStatus_t cusparseStatus;
cusparseStatus = cusparseCreate(&cusparseHandle);
checkCudaErrors(cusparseStatus);
//Get handle to a CUSPARSE matrix descriptor
cusparseMatDescr_t descr = 0;
cusparseStatus = cusparseCreateMatDescr(&descr);
checkCudaErrors(cusparseStatus);
//Get handle to a matrix_solve_info object
cusparseSolveAnalysisInfo_t info = 0;
cusparseStatus = cusparseCreateSolveAnalysisInfo(&info);
checkCudaErrors(cusparseStatus);
cusparseSetMatType(descr, CUSPARSE_MATRIX_TYPE_GENERAL);
cusparseSetMatIndexBase(descr, CUSPARSE_INDEX_BASE_ZERO);
duration_setup = (std::clock() - setup_clock) / (double)CLOCKS_PER_SEC;
printf("setup_time: %f\r\n", duration_setup);
std::clock_t start;
start = std::clock();
checkCudaErrors(cudaMalloc((void **)&d_x, N*sizeof(float)));
checkCudaErrors(cudaMalloc((void **)&d_y, N*sizeof(float)));
cudaMemcpy(d_x, x, N*sizeof(float), cudaMemcpyHostToDevice);
cudaMemcpy(d_y, y, N*sizeof(float), cudaMemcpyHostToDevice);
//Analyze the matrix. The info variable is needed to perform additional operations on the matrix
cusparseStatus = cusparseZcsrsv_analysis(cusparseHandle, CUSPARSE_OPERATION_NON_TRANSPOSE, N, nz, descr, val, J, I, info);
//Uses infor gathered from the matrix to solve the matrix.
cusparseStatus = cusparseZcsrsv_solve(cusparseHandle, CUSPARSE_OPERATION_NON_TRANSPOSE, N, 0, descr, val, J, I, info, d_x, d_y);
//Get the result back from the device
cudaMemcpy(x, d_x, N*sizeof(float), cudaMemcpyDeviceToHost);
cudaMemcpy(y, d_y, N*sizeof(float), cudaMemcpyDeviceToHost);
duration = (std::clock() - start) / (double)CLOCKS_PER_SEC;
printf("time ellapsed: %f", duration);
//free up memory
cusparseDestroy(cusparseHandle);
cublasDestroy(cublasHandle);
free(I);
free(J);
free(val);
free(x);
cudaFree(d_x);
cudaDeviceReset();
//Wait for user input so they can see the results
char* s = (char*)malloc(sizeof(char) * 8);
scanf(s);
exit(0);
}
magma_int_t
magma_ccustomicsetup(
magma_c_matrix A,
magma_c_matrix b,
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};
char preconditionermatrix[255];
snprintf( preconditionermatrix, sizeof(preconditionermatrix),
"/Users/hanzt0114cl306/work/matrices/ani/ani7_crop_ichol.mtx" );
CHECK( magma_c_csr_mtx( &hA, preconditionermatrix , queue) );
// for CUSPARSE
CHECK( magma_cmtransfer( hA, &precond->M, Magma_CPU, Magma_DEV , queue ));
// 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 ));
// CUSPARSE context //
CHECK_CUSPARSE( cusparseCreate( &cusparseHandle ));
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( cusparseCcsrsv_analysis( cusparseHandle,
CUSPARSE_OPERATION_NON_TRANSPOSE, precond->M.num_rows,
precond->M.nnz, descrL,
precond->M.val, precond->M.row, precond->M.col, 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( cusparseCcsrsv_analysis( cusparseHandle,
CUSPARSE_OPERATION_TRANSPOSE, precond->M.num_rows,
precond->M.nnz, descrU,
precond->M.val, precond->M.row, precond->M.col, precond->cuinfoU ));
cleanup:
cusparseDestroy( cusparseHandle );
cusparseDestroyMatDescr( descrL );
cusparseDestroyMatDescr( descrU );
cusparseHandle=NULL;
descrL=NULL;
descrU=NULL;
magma_cmfree( &hA, queue );
return info;
}
/* Solve Ax=b using the conjugate gradient method a) without any preconditioning, b) using an Incomplete Cholesky preconditioner and c) using an ILU0 preconditioner. */
int main(int argc, char **argv)
{
const int max_iter = 1000;
int k, M = 0, N = 0, nz = 0, *I = NULL, *J = NULL;
int *d_col, *d_row;
int qatest = 0;
const float tol = 1e-12f;
float *x, *rhs;
float r0, r1, alpha, beta;
float *d_val, *d_x;
float *d_zm1, *d_zm2, *d_rm2;
float *d_r, *d_p, *d_omega, *d_y;
float *val = NULL;
float *d_valsILU0;
float *valsILU0;
float rsum, diff, err = 0.0;
float qaerr1, qaerr2 = 0.0;
float dot, numerator, denominator, nalpha;
const float floatone = 1.0;
const float floatzero = 0.0;
int nErrors = 0;
printf("conjugateGradientPrecond starting...\n");
/* QA testing mode */
if (checkCmdLineFlag(argc, (const char **)argv, "qatest"))
{
qatest = 1;
}
/* This will pick the best possible CUDA capable device */
cudaDeviceProp deviceProp;
int devID = findCudaDevice(argc, (const char **)argv);
printf("GPU selected Device ID = %d \n", devID);
if (devID < 0)
{
printf("Invalid GPU device %d selected, exiting...\n", devID);
exit(EXIT_SUCCESS);
}
checkCudaErrors(cudaGetDeviceProperties(&deviceProp, devID));
/* Statistics about the GPU device */
printf("> GPU device has %d Multi-Processors, SM %d.%d compute capabilities\n\n",
deviceProp.multiProcessorCount, deviceProp.major, deviceProp.minor);
int version = (deviceProp.major * 0x10 + deviceProp.minor);
if (version < 0x11)
{
printf("%s: requires a minimum CUDA compute 1.1 capability\n", sSDKname);
// cudaDeviceReset causes the driver to clean up all state. While
// not mandatory in normal operation, it is good practice. It is also
// needed to ensure correct operation when the application is being
// profiled. Calling cudaDeviceReset causes all profile data to be
// flushed before the application exits
cudaDeviceReset();
exit(EXIT_SUCCESS);
}
/* Generate a random tridiagonal symmetric matrix in CSR (Compressed Sparse Row) format */
M = N = 16384;
nz = 5*N-4*(int)sqrt((double)N);
I = (int *)malloc(sizeof(int)*(N+1)); // csr row pointers for matrix A
J = (int *)malloc(sizeof(int)*nz); // csr column indices for matrix A
val = (float *)malloc(sizeof(float)*nz); // csr values for matrix A
x = (float *)malloc(sizeof(float)*N);
rhs = (float *)malloc(sizeof(float)*N);
for (int i = 0; i < N; i++)
{
rhs[i] = 0.0; // Initialize RHS
x[i] = 0.0; // Initial approximation of solution
}
genLaplace(I, J, val, M, N, nz, rhs);
/* Create CUBLAS context */
cublasHandle_t cublasHandle = 0;
cublasStatus_t cublasStatus;
cublasStatus = cublasCreate(&cublasHandle);
checkCudaErrors(cublasStatus);
/* Create CUSPARSE context */
cusparseHandle_t cusparseHandle = 0;
cusparseStatus_t cusparseStatus;
cusparseStatus = cusparseCreate(&cusparseHandle);
checkCudaErrors(cusparseStatus);
/* Description of the A matrix*/
cusparseMatDescr_t descr = 0;
cusparseStatus = cusparseCreateMatDescr(&descr);
checkCudaErrors(cusparseStatus);
/* Define the properties of the matrix */
cusparseSetMatType(descr,CUSPARSE_MATRIX_TYPE_GENERAL);
cusparseSetMatIndexBase(descr,CUSPARSE_INDEX_BASE_ZERO);
/* Allocate required memory */
checkCudaErrors(cudaMalloc((void **)&d_col, nz*sizeof(int)));
checkCudaErrors(cudaMalloc((void **)&d_row, (N+1)*sizeof(int)));
checkCudaErrors(cudaMalloc((void **)&d_val, nz*sizeof(float)));
checkCudaErrors(cudaMalloc((void **)&d_x, N*sizeof(float)));
checkCudaErrors(cudaMalloc((void **)&d_y, N*sizeof(float)));
checkCudaErrors(cudaMalloc((void **)&d_r, N*sizeof(float)));
checkCudaErrors(cudaMalloc((void **)&d_p, N*sizeof(float)));
checkCudaErrors(cudaMalloc((void **)&d_omega, N*sizeof(float)));
cudaMemcpy(d_col, J, nz*sizeof(int), cudaMemcpyHostToDevice);
cudaMemcpy(d_row, I, (N+1)*sizeof(int), cudaMemcpyHostToDevice);
cudaMemcpy(d_val, val, nz*sizeof(float), cudaMemcpyHostToDevice);
cudaMemcpy(d_x, x, N*sizeof(float), cudaMemcpyHostToDevice);
cudaMemcpy(d_r, rhs, N*sizeof(float), cudaMemcpyHostToDevice);
/* Conjugate gradient without preconditioning.
------------------------------------------
Follows the description by Golub & Van Loan, "Matrix Computations 3rd ed.", Section 10.2.6 */
printf("Convergence of conjugate gradient without preconditioning: \n");
k = 0;
r0 = 0;
cublasSdot(cublasHandle, N, d_r, 1, d_r, 1, &r1);
while (r1 > tol*tol && k <= max_iter)
{
k++;
if (k == 1)
{
cublasScopy(cublasHandle, N, d_r, 1, d_p, 1);
}
else
{
beta = r1/r0;
cublasSscal(cublasHandle, N, &beta, d_p, 1);
cublasSaxpy(cublasHandle, N, &floatone, d_r, 1, d_p, 1) ;
}
cusparseScsrmv(cusparseHandle,CUSPARSE_OPERATION_NON_TRANSPOSE, N, N, nz, &floatone, descr, d_val, d_row, d_col, d_p, &floatzero, d_omega);
cublasSdot(cublasHandle, N, d_p, 1, d_omega, 1, &dot);
alpha = r1/dot;
cublasSaxpy(cublasHandle, N, &alpha, d_p, 1, d_x, 1);
nalpha = -alpha;
cublasSaxpy(cublasHandle, N, &nalpha, d_omega, 1, d_r, 1);
r0 = r1;
cublasSdot(cublasHandle, N, d_r, 1, d_r, 1, &r1);
}
printf(" iteration = %3d, residual = %e \n", k, sqrt(r1));
cudaMemcpy(x, d_x, N*sizeof(float), cudaMemcpyDeviceToHost);
/* check result */
err = 0.0;
for (int i = 0; i < N; i++)
{
rsum = 0.0;
for (int j = I[i]; j < I[i+1]; j++)
{
rsum += val[j]*x[J[j]];
}
diff = fabs(rsum - rhs[i]);
if (diff > err)
{
err = diff;
}
}
printf(" Convergence Test: %s \n", (k <= max_iter) ? "OK" : "FAIL");
nErrors += (k > max_iter) ? 1 : 0;
qaerr1 = err;
if (0)
{
// output result in matlab-style array
int n=(int)sqrt((double)N);
printf("a = [ ");
for (int iy=0; iy<n; iy++)
{
for (int ix=0; ix<n; ix++)
{
printf(" %f ", x[iy*n+ix]);
}
if (iy == n-1)
{
printf(" ]");
}
printf("\n");
}
}
/* Preconditioned Conjugate Gradient using ILU.
--------------------------------------------
Follows the description by Golub & Van Loan, "Matrix Computations 3rd ed.", Algorithm 10.3.1 */
printf("\nConvergence of conjugate gradient using incomplete LU preconditioning: \n");
int nzILU0 = 2*N-1;
valsILU0 = (float *) malloc(nz*sizeof(float));
checkCudaErrors(cudaMalloc((void **)&d_valsILU0, nz*sizeof(float)));
checkCudaErrors(cudaMalloc((void **)&d_zm1, (N)*sizeof(float)));
checkCudaErrors(cudaMalloc((void **)&d_zm2, (N)*sizeof(float)));
checkCudaErrors(cudaMalloc((void **)&d_rm2, (N)*sizeof(float)));
/* create the analysis info object for the A matrix */
cusparseSolveAnalysisInfo_t infoA = 0;
cusparseStatus = cusparseCreateSolveAnalysisInfo(&infoA);
checkCudaErrors(cusparseStatus);
/* Perform the analysis for the Non-Transpose case */
cusparseStatus = cusparseScsrsv_analysis(cusparseHandle, CUSPARSE_OPERATION_NON_TRANSPOSE,
N, nz, descr, d_val, d_row, d_col, infoA);
checkCudaErrors(cusparseStatus);
/* Copy A data to ILU0 vals as input*/
cudaMemcpy(d_valsILU0, d_val, nz*sizeof(float), cudaMemcpyDeviceToDevice);
/* generate the Incomplete LU factor H for the matrix A using cudsparseScsrilu0 */
cusparseStatus = cusparseScsrilu0(cusparseHandle, CUSPARSE_OPERATION_NON_TRANSPOSE, N, descr, d_valsILU0, d_row, d_col, infoA);
checkCudaErrors(cusparseStatus);
/* Create info objects for the ILU0 preconditioner */
cusparseSolveAnalysisInfo_t info_u;
cusparseCreateSolveAnalysisInfo(&info_u);
cusparseMatDescr_t descrL = 0;
cusparseStatus = cusparseCreateMatDescr(&descrL);
cusparseSetMatType(descrL,CUSPARSE_MATRIX_TYPE_GENERAL);
cusparseSetMatIndexBase(descrL,CUSPARSE_INDEX_BASE_ZERO);
cusparseSetMatFillMode(descrL, CUSPARSE_FILL_MODE_LOWER);
cusparseSetMatDiagType(descrL, CUSPARSE_DIAG_TYPE_UNIT);
cusparseMatDescr_t descrU = 0;
cusparseStatus = cusparseCreateMatDescr(&descrU);
cusparseSetMatType(descrU,CUSPARSE_MATRIX_TYPE_GENERAL);
cusparseSetMatIndexBase(descrU,CUSPARSE_INDEX_BASE_ZERO);
cusparseSetMatFillMode(descrU, CUSPARSE_FILL_MODE_UPPER);
cusparseSetMatDiagType(descrU, CUSPARSE_DIAG_TYPE_NON_UNIT);
cusparseStatus = cusparseScsrsv_analysis(cusparseHandle, CUSPARSE_OPERATION_NON_TRANSPOSE, N, nz, descrU, d_val, d_row, d_col, info_u);
/* reset the initial guess of the solution to zero */
for (int i = 0; i < N; i++)
{
x[i] = 0.0;
}
checkCudaErrors(cudaMemcpy(d_r, rhs, N*sizeof(float), cudaMemcpyHostToDevice));
checkCudaErrors(cudaMemcpy(d_x, x, N*sizeof(float), cudaMemcpyHostToDevice));
k = 0;
cublasSdot(cublasHandle, N, d_r, 1, d_r, 1, &r1);
while (r1 > tol*tol && k <= max_iter)
{
// Forward Solve, we can re-use infoA since the sparsity pattern of A matches that of L
cusparseStatus = cusparseScsrsv_solve(cusparseHandle, CUSPARSE_OPERATION_NON_TRANSPOSE, N, &floatone, descrL,
d_valsILU0, d_row, d_col, infoA, d_r, d_y);
checkCudaErrors(cusparseStatus);
// Back Substitution
cusparseStatus = cusparseScsrsv_solve(cusparseHandle, CUSPARSE_OPERATION_NON_TRANSPOSE, N, &floatone, descrU,
d_valsILU0, d_row, d_col, info_u, d_y, d_zm1);
checkCudaErrors(cusparseStatus);
k++;
if (k == 1)
{
cublasScopy(cublasHandle, N, d_zm1, 1, d_p, 1);
}
else
{
cublasSdot(cublasHandle, N, d_r, 1, d_zm1, 1, &numerator);
cublasSdot(cublasHandle, N, d_rm2, 1, d_zm2, 1, &denominator);
beta = numerator/denominator;
cublasSscal(cublasHandle, N, &beta, d_p, 1);
cublasSaxpy(cublasHandle, N, &floatone, d_zm1, 1, d_p, 1) ;
}
cusparseScsrmv(cusparseHandle,CUSPARSE_OPERATION_NON_TRANSPOSE, N, N, nzILU0, &floatone, descrU, d_val, d_row, d_col, d_p, &floatzero, d_omega);
cublasSdot(cublasHandle, N, d_r, 1, d_zm1, 1, &numerator);
cublasSdot(cublasHandle, N, d_p, 1, d_omega, 1, &denominator);
alpha = numerator / denominator;
cublasSaxpy(cublasHandle, N, &alpha, d_p, 1, d_x, 1);
cublasScopy(cublasHandle, N, d_r, 1, d_rm2, 1);
cublasScopy(cublasHandle, N, d_zm1, 1, d_zm2, 1);
nalpha = -alpha;
cublasSaxpy(cublasHandle, N, &nalpha, d_omega, 1, d_r, 1);
cublasSdot(cublasHandle, N, d_r, 1, d_r, 1, &r1);
}
printf(" iteration = %3d, residual = %e \n", k, sqrt(r1));
cudaMemcpy(x, d_x, N*sizeof(float), cudaMemcpyDeviceToHost);
/* check result */
err = 0.0;
for (int i = 0; i < N; i++)
{
rsum = 0.0;
for (int j = I[i]; j < I[i+1]; j++)
{
rsum += val[j]*x[J[j]];
}
diff = fabs(rsum - rhs[i]);
if (diff > err)
{
err = diff;
}
}
printf(" Convergence Test: %s \n", (k <= max_iter) ? "OK" : "FAIL");
nErrors += (k > max_iter) ? 1 : 0;
qaerr2 = err;
/* Destroy parameters */
cusparseDestroySolveAnalysisInfo(infoA);
cusparseDestroySolveAnalysisInfo(info_u);
/* Destroy contexts */
cusparseDestroy(cusparseHandle);
cublasDestroy(cublasHandle);
/* Free device memory */
free(I);
free(J);
free(val);
free(x);
free(rhs);
free(valsILU0);
cudaFree(d_col);
cudaFree(d_row);
cudaFree(d_val);
cudaFree(d_x);
cudaFree(d_y);
cudaFree(d_r);
cudaFree(d_p);
cudaFree(d_omega);
cudaFree(d_valsILU0);
cudaFree(d_zm1);
cudaFree(d_zm2);
cudaFree(d_rm2);
// cudaDeviceReset causes the driver to clean up all state. While
// not mandatory in normal operation, it is good practice. It is also
// needed to ensure correct operation when the application is being
// profiled. Calling cudaDeviceReset causes all profile data to be
// flushed before the application exits
cudaDeviceReset();
printf(" Test Summary:\n");
printf(" Counted total of %d errors\n", nErrors);
printf(" qaerr1 = %f qaerr2 = %f\n\n", fabs(qaerr1), fabs(qaerr2));
exit((nErrors == 0 &&fabs(qaerr1)<1e-5 && fabs(qaerr2) < 1e-5 ? EXIT_SUCCESS : EXIT_FAILURE));
}
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_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;
}
magma_int_t
magma_dcustomilusetup(
magma_d_matrix A,
magma_d_matrix b,
magma_d_preconditioner *precond,
magma_queue_t queue )
{
magma_int_t info = 0;
cusparseHandle_t cusparseHandle=NULL;
cusparseMatDescr_t descrL=NULL;
cusparseMatDescr_t descrU=NULL;
magma_d_matrix hA={Magma_CSR};
char preconditionermatrix[255];
// first L
snprintf( preconditionermatrix, sizeof(preconditionermatrix),
"precondL.mtx" );
CHECK( magma_d_csr_mtx( &hA, preconditionermatrix , queue) );
CHECK( magma_dmtransfer( hA, &precond->L, Magma_CPU, Magma_DEV , queue ));
// extract the diagonal of L into precond->d
CHECK( magma_djacobisetup_diagscal( precond->L, &precond->d, queue ));
CHECK( magma_dvinit( &precond->work1, Magma_DEV, hA.num_rows, 1, MAGMA_D_ZERO, queue ));
magma_dmfree( &hA, queue );
// now U
snprintf( preconditionermatrix, sizeof(preconditionermatrix),
"precondU.mtx" );
CHECK( magma_d_csr_mtx( &hA, preconditionermatrix , queue) );
CHECK( magma_dmtransfer( hA, &precond->U, Magma_CPU, Magma_DEV , queue ));
// extract the diagonal of U into precond->d2
CHECK( magma_djacobisetup_diagscal( precond->U, &precond->d2, queue ));
CHECK( magma_dvinit( &precond->work2, Magma_DEV, hA.num_rows, 1, MAGMA_D_ZERO, queue ));
// CUSPARSE context //
CHECK_CUSPARSE( cusparseCreate( &cusparseHandle ));
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( cusparseDcsrsv_analysis( cusparseHandle,
CUSPARSE_OPERATION_NON_TRANSPOSE, precond->L.num_rows,
precond->L.nnz, descrL,
precond->L.val, precond->L.row, precond->L.col, 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( cusparseDcsrsv_analysis( cusparseHandle,
CUSPARSE_OPERATION_NON_TRANSPOSE, precond->U.num_rows,
precond->U.nnz, descrU,
precond->U.val, precond->U.row, precond->U.col, precond->cuinfoU ));
cleanup:
cusparseDestroy( cusparseHandle );
cusparseDestroyMatDescr( descrL );
cusparseDestroyMatDescr( descrU );
cusparseHandle=NULL;
descrL=NULL;
descrU=NULL;
magma_dmfree( &hA, queue );
return info;
}
magma_int_t
magma_zailusetup( magma_z_sparse_matrix A, magma_z_preconditioner *precond ){
magma_z_sparse_matrix hAh, hA, hAL, hALCOO, hAU, hAUT, hAUCOO, dAL, dAU,
hL, hU, dL, dU, DL, RL, DU, RU;
// copy original matrix as CSRCOO to device
magma_z_mtransfer(A, &hAh, A.memory_location, Magma_CPU);
magma_z_mconvert( hAh, &hA, hAh.storage_type, Magma_CSR );
magma_z_mfree(&hAh);
// in case using fill-in
magma_zilustruct( &hA, precond->levels);
// need only lower triangular
hAL.diagorder_type == Magma_UNITY;
magma_z_mconvert( hA, &hAL, Magma_CSR, Magma_CSRL );
magma_z_mconvert( hAL, &hALCOO, Magma_CSR, Magma_CSRCOO );
magma_z_mtransfer( hALCOO, &dAL, Magma_CPU, Magma_DEV );
magma_z_mtransfer( hALCOO, &dAU, Magma_CPU, Magma_DEV );
// need only upper triangular
magma_z_mconvert( hA, &hAU, Magma_CSR, Magma_CSRU );
magma_z_cucsrtranspose( hAU, &hAUT );
magma_z_mconvert( hAUT, &hAUCOO, Magma_CSR, Magma_CSRCOO );
magma_z_mtransfer( hAUCOO, &dL, Magma_CPU, Magma_DEV );
magma_z_mtransfer( hAUCOO, &dU, Magma_CPU, Magma_DEV );
magma_z_mfree(&hALCOO);
magma_z_mfree(&hAL);
magma_z_mfree(&hAUCOO);
magma_z_mfree(&hAUT);
magma_z_mfree(&hAU);
for(int i=0; i<precond->sweeps; i++){
magma_zailu_csr_s( dAL, dAU, dL, dU );
}
magma_z_mtransfer( dL, &hL, Magma_DEV, Magma_CPU );
magma_z_mtransfer( dU, &hU, Magma_DEV, Magma_CPU );
magma_z_LUmergein( hL, hU, &hA);
magma_z_mtransfer( hA, &precond->M, Magma_CPU, Magma_DEV );
magma_z_mfree(&dL);
magma_z_mfree(&dU);
magma_z_mfree(&dAL);
magma_z_mfree(&dAU);
hAL.diagorder_type = Magma_UNITY;
magma_z_mconvert(hA, &hAL, Magma_CSR, Magma_CSRL);
hAL.storage_type = Magma_CSR;
magma_z_mconvert(hA, &hAU, Magma_CSR, Magma_CSRU);
hAU.storage_type = Magma_CSR;
magma_z_mfree(&hA);
magma_z_mfree(&hL);
magma_z_mfree(&hU);
magma_zcsrsplit( 256, hAL, &DL, &RL );
magma_zcsrsplit( 256, hAU, &DU, &RU );
magma_z_mtransfer( DL, &precond->LD, Magma_CPU, Magma_DEV );
magma_z_mtransfer( DU, &precond->UD, Magma_CPU, Magma_DEV );
// for cusparse uncomment this
magma_z_mtransfer( hAL, &precond->L, Magma_CPU, Magma_DEV );
magma_z_mtransfer( hAU, &precond->U, Magma_CPU, Magma_DEV );
// for ba-solve uncomment this
/*
if( RL.nnz != 0 )
magma_z_mtransfer( RL, &precond->L, Magma_CPU, Magma_DEV );
else{
precond->L.nnz = 0;
precond->L.val = NULL;
precond->L.col = NULL;
precond->L.row = NULL;
precond->L.blockinfo = NULL;
}
if( RU.nnz != 0 )
magma_z_mtransfer( RU, &precond->U, Magma_CPU, Magma_DEV );
else{
precond->U.nnz = 0;
precond->L.val = NULL;
precond->L.col = NULL;
precond->L.row = NULL;
precond->L.blockinfo = NULL;
}
*/
magma_z_mfree(&hAL);
magma_z_mfree(&hAU);
magma_z_mfree(&DL);
magma_z_mfree(&RL);
magma_z_mfree(&DU);
magma_z_mfree(&RU);
// CUSPARSE context //
cusparseHandle_t cusparseHandle;
cusparseStatus_t cusparseStatus;
cusparseStatus = cusparseCreate(&cusparseHandle);
if(cusparseStatus != 0) printf("error in Handle.\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_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 =
cusparseZcsrsv_analysis(cusparseHandle,
CUSPARSE_OPERATION_NON_TRANSPOSE, precond->L.num_rows,
precond->L.nnz, descrL,
precond->L.val, precond->L.row, precond->L.col, 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 =
cusparseZcsrsv_analysis(cusparseHandle,
CUSPARSE_OPERATION_NON_TRANSPOSE, precond->U.num_rows,
precond->U.nnz, descrU,
precond->U.val, precond->U.row, precond->U.col, precond->cuinfoU );
if(cusparseStatus != 0) printf("error in analysis.\n");
cusparseDestroyMatDescr( descrU );
cusparseDestroy( cusparseHandle );
return MAGMA_SUCCESS;
}
magma_int_t
magma_zaiccsetup( magma_z_sparse_matrix A, magma_z_preconditioner *precond ){
magma_z_sparse_matrix hAh, hA, hAL, hALCOO, dAL, hL, dL, DL, RL;
// copy original matrix as CSRCOO to device
magma_z_mtransfer(A, &hAh, A.memory_location, Magma_CPU);
magma_z_mconvert( hAh, &hA, hAh.storage_type, Magma_CSR );
magma_z_mfree(&hAh);
// in case using fill-in
magma_zilustruct( &hA, precond->levels);
magma_z_mconvert( hA, &hAL, Magma_CSR, Magma_CSRL );
magma_z_mconvert( hAL, &hALCOO, Magma_CSR, Magma_CSRCOO );
magma_z_mtransfer( hALCOO, &dAL, Magma_CPU, Magma_DEV );
magma_z_mtransfer( hALCOO, &dL, Magma_CPU, Magma_DEV );
magma_z_mfree(&hALCOO);
magma_z_mfree(&hAL);
magma_z_mfree(&hA);
for(int i=0; i<precond->sweeps; i++){
magma_zaic_csr_s( dAL, dL );
}
magma_z_mtransfer( dL, &hL, Magma_DEV, Magma_CPU );
magma_z_mfree(&dL);
magma_z_mfree(&dAL);
magma_z_mconvert(hL, &hAL, hL.storage_type, Magma_CSR);
// for CUSPARSE
magma_z_mtransfer( hAL, &precond->M, Magma_CPU, Magma_DEV );
magma_zcsrsplit( 256, hAL, &DL, &RL );
magma_z_mtransfer( DL, &precond->LD, Magma_CPU, Magma_DEV );
magma_z_mtransfer( RL, &precond->L, Magma_CPU, Magma_DEV );
magma_z_mfree(&hL);
magma_z_cucsrtranspose( hAL, &hL );
magma_zcsrsplit( 256, hL, &DL, &RL );
magma_z_mtransfer( DL, &precond->UD, Magma_CPU, Magma_DEV );
magma_z_mtransfer( RL, &precond->U, Magma_CPU, Magma_DEV );
magma_z_mfree(&hAL);
magma_z_mfree(&hL);
magma_z_mfree(&DL);
magma_z_mfree(&RL);
// CUSPARSE context //
cusparseHandle_t cusparseHandle;
cusparseStatus_t cusparseStatus;
cusparseStatus = cusparseCreate(&cusparseHandle);
if(cusparseStatus != 0) printf("error in Handle.\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 =
cusparseZcsrsv_analysis(cusparseHandle,
CUSPARSE_OPERATION_NON_TRANSPOSE, precond->M.num_rows,
precond->M.nnz, descrL,
precond->M.val, precond->M.row, precond->M.col, 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 =
cusparseZcsrsv_analysis(cusparseHandle,
CUSPARSE_OPERATION_TRANSPOSE, precond->M.num_rows,
precond->M.nnz, descrU,
precond->M.val, precond->M.row, precond->M.col, precond->cuinfoU );
if(cusparseStatus != 0) printf("error in analysis U.\n");
cusparseDestroyMatDescr( descrU );
cusparseDestroy( cusparseHandle );
return MAGMA_SUCCESS;
}
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;
}