Caffe::Caffe()
: cublas_handle_(NULL),cusparse_handle_(NULL),cusparse_descr_(NULL),curand_generator_(NULL),random_generator_(),mode_(Caffe::CPU), solver_count_(1), root_solver_(true){
// Try to create a cublas handler, and report an error if failed (but we will
// keep the program running as one might just want to run CPU code).
LOG(INFO)<<"caffe init.";
if (cublasCreate(&cublas_handle_) != CUBLAS_STATUS_SUCCESS) {
LOG(ERROR) << "Cannot create Cublas handle. Cublas won't be available.";
}
//add cusparse handler
if (cusparseCreate(&cusparse_handle_)!=CUSPARSE_STATUS_SUCCESS){
LOG(ERROR) << "cannot create Cusparse handle,Cusparse won't be available.";
}
if(cusparseCreateMatDescr(&cusparse_descr_)!=CUSPARSE_STATUS_SUCCESS){
LOG(ERROR) << "cannot create Cusparse descr,descr won't be available.";
}else{
cusparseSetMatType(cusparse_descr_,CUSPARSE_MATRIX_TYPE_GENERAL);
cusparseSetMatIndexBase(cusparse_descr_,CUSPARSE_INDEX_BASE_ZERO);
LOG(INFO)<<"init descr";
}
// Try to create a curand handler.
if (curandCreateGenerator(&curand_generator_, CURAND_RNG_PSEUDO_DEFAULT)
!= CURAND_STATUS_SUCCESS ||
curandSetPseudoRandomGeneratorSeed(curand_generator_, cluster_seedgen())
!= CURAND_STATUS_SUCCESS) {
LOG(ERROR) << "Cannot create Curand generator. Curand won't be available.";
}
LOG(INFO)<<"caffe finish";
}
CudaSparseSingleton()
{
cusparseCreate( & handle );
cusparseCreateMatDescr( & descra );
cusparseSetMatType( descra , CUSPARSE_MATRIX_TYPE_GENERAL );
cusparseSetMatIndexBase( descra , CUSPARSE_INDEX_BASE_ZERO );
}
// initialize CUDA
ssp_cuda *ssp_init_cuda() {
ssp_cuda *cudaHandle = (ssp_cuda*)malloc(sizeof(ssp_cuda));
if (!cudaHandle) {
fprintf(stderr,"ssp_init_cuda: cudaHandle memory allocation failed.\n");
return NULL;
}
cudaHandle->cusparse_handle = 0;
cudaHandle->cusparse_matDescr = 0;
cusparseStatus_t status = cusparseCreate(&cudaHandle->cusparse_handle);
if (status != CUSPARSE_STATUS_SUCCESS) {
ssp_finalize_cuda(cudaHandle);
fprintf(stderr,"ssp_init_cuda: cusparse initialization failed.\n");
return NULL;
}
status = cusparseCreateMatDescr(&cudaHandle->cusparse_matDescr);
if (status != CUSPARSE_STATUS_SUCCESS) {
ssp_finalize_cuda(cudaHandle);
fprintf(stderr,"ssp_init_cuda: cusparse matrix setup failed.\n");
return NULL;
}
cusparseSetMatType(cudaHandle->cusparse_matDescr,CUSPARSE_MATRIX_TYPE_GENERAL);
cusparseSetMatIndexBase(cudaHandle->cusparse_matDescr,CUSPARSE_INDEX_BASE_ZERO);
return cudaHandle;
}
xDense2Csr(StatisticalTimer& timer) : cusparseFunc(timer)
{
cusparseStatus_t err = cusparseCreateMatDescr(&descrA);
CUDA_V_THROW(err, "cusparseCreateMatDescr failed");
err = cusparseSetMatType(descrA, CUSPARSE_MATRIX_TYPE_GENERAL);
CUDA_V_THROW(err, "cusparseSetMatType failed");
err = cusparseSetMatIndexBase(descrA, CUSPARSE_INDEX_BASE_ZERO);
CUDA_V_THROW(err, "cusparseSetMatIndexBase failed");
n_rows = 0;
n_cols = 0;
n_vals = 0;
device_col_indices = nullptr;
device_row_offsets = nullptr;
device_values = nullptr;
device_A = nullptr;
nnzPerRow = nullptr;
devRowOffsets = nullptr;
devColIndices = nullptr;
devValues = nullptr;
}// end
void Caffe::SetDevice(const int device_id) {
int current_device;
CUDA_CHECK(cudaGetDevice(¤t_device));
if (current_device == device_id) {
return;
}
// The call to cudaSetDevice must come before any calls to Get, which
// may perform initialization using the GPU.
CUDA_CHECK(cudaSetDevice(device_id));
if (Get().cublas_handle_) CUBLAS_CHECK(cublasDestroy(Get().cublas_handle_));
if (Get().cusparse_descr_)CUSPARSE_CHECK(cusparseDestroyMatDescr(Get().cusparse_descr_));
if (Get().cusparse_handle_)CUSPARSE_CHECK(cusparseDestroy(Get().cusparse_handle_));
if (Get().curand_generator_) {
CURAND_CHECK(curandDestroyGenerator(Get().curand_generator_));
}
CUSPARSE_CHECK(cusparseCreate(&Get().cusparse_handle_));
CUSPARSE_CHECK(cusparseCreateMatDescr(&Get().cusparse_descr_));
// cusparseSetMatType(cusparse_descr_,CUSPARSE_MATRIX_TYPE_GENERAL);
// cusparseSetMatIndexBase(cusparse_descr_,CUSPARSE_INDEX_BASE_ZERO);
LOG(INFO)<<"set descr";
CUBLAS_CHECK(cublasCreate(&Get().cublas_handle_));
CURAND_CHECK(curandCreateGenerator(&Get().curand_generator_,
CURAND_RNG_PSEUDO_DEFAULT));
CURAND_CHECK(curandSetPseudoRandomGeneratorSeed(Get().curand_generator_,
cluster_seedgen()));
}
cuSparseHandleType(bool transposeA, bool transposeB){
cusparseStatus_t status;
status= cusparseCreate(&handle);
if (status != CUSPARSE_STATUS_SUCCESS) {
std::cerr << ("cusparseCreate ERROR") << std::endl;
return;
}
cusparseSetPointerMode(handle, CUSPARSE_POINTER_MODE_HOST);
if (transposeA){
transA = CUSPARSE_OPERATION_TRANSPOSE;
}
else {
transA = CUSPARSE_OPERATION_NON_TRANSPOSE;
}
if (transposeB){
transB = CUSPARSE_OPERATION_TRANSPOSE;
}
else {
transB = CUSPARSE_OPERATION_NON_TRANSPOSE;
}
status = cusparseCreateMatDescr(&a_descr);
if (status != CUSPARSE_STATUS_SUCCESS) {
std::cerr << "cusparseCreateMatDescr a_descr ERROR" << std::endl;
return;
}
cusparseSetMatType(a_descr,CUSPARSE_MATRIX_TYPE_GENERAL);
cusparseSetMatIndexBase(a_descr,CUSPARSE_INDEX_BASE_ZERO);
status = cusparseCreateMatDescr(&b_descr);
if (status != CUSPARSE_STATUS_SUCCESS) {
std::cerr << ("cusparseCreateMatDescr b_descr ERROR") << std::endl;
return;
}
cusparseSetMatType(b_descr,CUSPARSE_MATRIX_TYPE_GENERAL);
cusparseSetMatIndexBase(b_descr,CUSPARSE_INDEX_BASE_ZERO);
status = cusparseCreateMatDescr(&c_descr);
if (status != CUSPARSE_STATUS_SUCCESS) {
std::cerr << ("cusparseCreateMatDescr c_descr ERROR") << std::endl;
return;
}
cusparseSetMatType(c_descr,CUSPARSE_MATRIX_TYPE_GENERAL);
cusparseSetMatIndexBase(c_descr,CUSPARSE_INDEX_BASE_ZERO);
}
magma_int_t
magma_dapplycuicc_l( magma_d_vector b, magma_d_vector *x,
magma_d_preconditioner *precond ){
double one = MAGMA_D_MAKE( 1.0, 0.0);
// 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 =
cusparseSetMatFillMode(descrL,CUSPARSE_FILL_MODE_LOWER);
if(cusparseStatus != 0) printf("error in fillmode.\n");
cusparseStatus =
cusparseSetMatIndexBase(descrL,CUSPARSE_INDEX_BASE_ZERO);
if(cusparseStatus != 0) printf("error in IndexBase.\n");
// end CUSPARSE context //
cusparseStatus =
cusparseDcsrsv_solve( cusparseHandle,
CUSPARSE_OPERATION_NON_TRANSPOSE,
precond->M.num_rows, &one,
descrL,
precond->M.val,
precond->M.row,
precond->M.col,
precond->cuinfoL,
b.val,
x->val );
if(cusparseStatus != 0) printf("error in L triangular solve:%p.\n", precond->cuinfoL );
cusparseDestroyMatDescr( descrL );
cusparseDestroy( cusparseHandle );
magma_device_sync();
return MAGMA_SUCCESS;
}
void sparse_fully_connected_1x1_layer_tester_cuda::enqueue_forward_propagation(
cudaStream_t stream_id,
cuda_linear_buffer_device::ptr output_buffer,
const std::vector<cuda_linear_buffer_device::const_ptr>& schema_data,
const std::vector<cuda_linear_buffer_device::const_ptr>& data,
const std::vector<cuda_linear_buffer_device::const_ptr>& data_custom,
const std::vector<cuda_linear_buffer_device::const_ptr>& input_buffers,
const std::vector<cuda_linear_buffer_device::const_ptr>& persistent_working_data,
cuda_linear_buffer_device::ptr temporary_working_fixed_buffer,
cuda_linear_buffer_device::ptr temporary_working_per_entry_buffer,
unsigned int entry_count)
{
{
cusparse_safe_call(cusparseSetStream(cuda_config->get_cusparse_handle(), stream_id));
float alpha = 1.0F;
float beta = 0.0F;
cusparseMatDescr_t mat_descr;
cusparse_safe_call(cusparseCreateMatDescr(&mat_descr));
cusparse_safe_call(cusparseScsrmm(
cuda_config->get_cusparse_handle(),
CUSPARSE_OPERATION_NON_TRANSPOSE,
output_elem_count_per_entry,
entry_count,
input_elem_count_per_entry_list[0],
feature_map_connection_count,
&alpha,
mat_descr,
*data[0],
*data_custom[1],
*data_custom[0],
*input_buffers[0],
input_elem_count_per_entry_list[0],
&beta,
*output_buffer,
output_elem_count_per_entry));
}
// Add bias
{
cudnn_safe_call(cudnnSetStream(cuda_config->get_cudnn_handle(), stream_id));
cudnn_util::set_tensor_descriptor(
output_data_desc,
output_configuration_specific,
entry_count);
float alpha = 1.0F;
float beta = 1.0F;
cudnn_safe_call(cudnnAddTensor(
cuda_config->get_cudnn_handle(),
&alpha,
bias_desc,
*data[1],
&beta,
output_data_desc,
*output_buffer));
}
}
Teuchos::RCP<cusparseMatDescr_t> Kokkos::CUSPARSEdetails::createMatDescr()
{
cusparseMatDescr_t *desc = new cusparseMatDescr_t;
cusparseStatus_t status = cusparseCreateMatDescr(desc);
TEUCHOS_TEST_FOR_EXCEPTION(status == CUSPARSE_STATUS_ALLOC_FAILED, std::runtime_error,
"Kokkos::CUSPARSEdetails::createMatDescr(): the resources could not be allocated.")
TEUCHOS_TEST_FOR_EXCEPTION(status != CUSPARSE_STATUS_SUCCESS, std::runtime_error,
"Kokkos::CUSPARSEdetails::createMatDescr(): unspecified error.")
return Teuchos::rcpWithDealloc(desc,CUSPARSEMatDescDestroyer(),true);
}
xCsr2Dense( StatisticalTimer& timer, bool read_explicit_zeroes = true ): cusparseFunc( timer )
{
cusparseStatus_t err = cusparseCreateMatDescr( &descrA );
CUDA_V_THROW( err, "cusparseCreateMatDescr failed" );
err = cusparseSetMatType( descrA, CUSPARSE_MATRIX_TYPE_GENERAL );
CUDA_V_THROW( err, "cusparseSetMatType failed" );
err = cusparseSetMatIndexBase( descrA, CUSPARSE_INDEX_BASE_ZERO );
CUDA_V_THROW( err, "cusparseSetMatIndexBase failed" );
n_rows = 0;
n_cols = 0;
n_vals = 0;
explicit_zeroes = read_explicit_zeroes;
}
CudaSparseSingleton()
{
status = cusparseCreate(&handle);
if(status != CUSPARSE_STATUS_SUCCESS)
{
throw std::runtime_error( std::string("ERROR - CUSPARSE Library Initialization failed" ) );
}
status = cusparseCreateMatDescr(&descra);
if(status != CUSPARSE_STATUS_SUCCESS)
{
throw std::runtime_error( std::string("ERROR - CUSPARSE Library Matrix descriptor failed" ) );
}
cusparseSetMatType(descra , CUSPARSE_MATRIX_TYPE_GENERAL);
cusparseSetMatIndexBase(descra , CUSPARSE_INDEX_BASE_ZERO);
}
extern "C" magma_int_t
magma_dapplycumilu_r_transpose(
magma_d_matrix b,
magma_d_matrix *x,
magma_d_preconditioner *precond,
magma_queue_t queue )
{
magma_int_t info = 0;
cusparseHandle_t cusparseHandle=NULL;
cusparseMatDescr_t descrU=NULL;
double one = MAGMA_D_MAKE( 1.0, 0.0);
// CUSPARSE context //
CHECK_CUSPARSE( cusparseCreate( &cusparseHandle ));
CHECK_CUSPARSE( cusparseSetStream( cusparseHandle, queue->cuda_stream() ));
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( cusparseDcsrsm_solve( cusparseHandle,
CUSPARSE_OPERATION_NON_TRANSPOSE,
precond->UT.num_rows,
b.num_rows*b.num_cols/precond->UT.num_rows,
&one,
descrU,
precond->UT.dval,
precond->UT.drow,
precond->UT.dcol,
precond->cuinfoUT,
b.dval,
precond->UT.num_rows,
x->dval,
precond->UT.num_rows ));
cleanup:
cusparseDestroyMatDescr( descrU );
cusparseDestroy( cusparseHandle );
return info;
}
extern "C" magma_int_t
magma_capplycumicc_l(
magma_c_matrix b,
magma_c_matrix *x,
magma_c_preconditioner *precond,
magma_queue_t queue )
{
magma_int_t info = 0;
cusparseHandle_t cusparseHandle=NULL;
cusparseMatDescr_t descrL=NULL;
magmaFloatComplex one = MAGMA_C_MAKE( 1.0, 0.0);
// CUSPARSE context //
CHECK_CUSPARSE( cusparseCreate( &cusparseHandle ));
CHECK_CUSPARSE( cusparseSetStream( cusparseHandle, queue ));
CHECK_CUSPARSE( cusparseCreateMatDescr( &descrL ));
CHECK_CUSPARSE( cusparseSetMatType( descrL, CUSPARSE_MATRIX_TYPE_TRIANGULAR ));
CHECK_CUSPARSE( cusparseSetMatDiagType( descrL, CUSPARSE_DIAG_TYPE_NON_UNIT ));
CHECK_CUSPARSE( cusparseSetMatFillMode( descrL, CUSPARSE_FILL_MODE_LOWER ));
CHECK_CUSPARSE( cusparseSetMatIndexBase( descrL, CUSPARSE_INDEX_BASE_ZERO ));
CHECK_CUSPARSE( cusparseCcsrsm_solve( cusparseHandle,
CUSPARSE_OPERATION_NON_TRANSPOSE,
precond->M.num_rows,
b.num_rows*b.num_cols/precond->M.num_rows,
&one,
descrL,
precond->M.dval,
precond->M.drow,
precond->M.dcol,
precond->cuinfoL,
b.dval,
precond->M.num_rows,
x->dval,
precond->M.num_rows ));
magma_device_sync();
cleanup:
cusparseDestroyMatDescr( descrL );
cusparseDestroy( cusparseHandle );
return info;
}
sparse_matrix::sparse_matrix(sparse_matrix::descriptor_t descriptor,
int rows, int cols, int nonzeros,
const double* values, const int* col_ptr, const int* row_ind)
: descrA(), m(), n(), nnz(), csrValA(), csrRowPtrA(), csrColIndA()
{
// Create descriptor
assert(cusparseCreateMatDescr(&descrA) == cudaSuccess);
assert(cusparseSetMatIndexBase(descrA, CUSPARSE_INDEX_BASE_ZERO) == cudaSuccess);
// Set descriptor fields
switch (descriptor) {
case non_symmetric:
assert(cusparseSetMatDiagType(descrA, CUSPARSE_DIAG_TYPE_NON_UNIT) == cudaSuccess);
assert(cusparseSetMatType (descrA, CUSPARSE_MATRIX_TYPE_GENERAL) == cudaSuccess);
assert(cusparseSetMatFillMode(descrA, CUSPARSE_FILL_MODE_LOWER) == cudaSuccess); // doesn't matter which, presumably
break;
case symmetric_lower:
assert(cusparseSetMatDiagType(descrA, CUSPARSE_DIAG_TYPE_NON_UNIT) == cudaSuccess);
assert(cusparseSetMatType (descrA, CUSPARSE_MATRIX_TYPE_SYMMETRIC) == cudaSuccess);
assert(cusparseSetMatFillMode(descrA, CUSPARSE_FILL_MODE_UPPER) == cudaSuccess); // upper since we're coming with CSC and storing as CSR
break;
case symmetric_upper:
assert(cusparseSetMatDiagType(descrA, CUSPARSE_DIAG_TYPE_NON_UNIT) == cudaSuccess);
assert(cusparseSetMatType (descrA, CUSPARSE_MATRIX_TYPE_SYMMETRIC) == cudaSuccess);
assert(cusparseSetMatFillMode(descrA, CUSPARSE_FILL_MODE_LOWER) == cudaSuccess); // lower since we're coming with CSC and storing as CSR
break;
}
// Switch rows and cols becuase we're coming with CSC and storing as CSR
n = rows;
m = cols;
nnz = nonzeros;
// Allocate memory
assert(cudaMalloc(reinterpret_cast<void**>(&csrValA), nnz * sizeof(double)) == cudaSuccess);
assert(cudaMalloc(reinterpret_cast<void**>(&csrRowPtrA), (m+1) * sizeof(int)) == cudaSuccess);
assert(cudaMalloc(reinterpret_cast<void**>(&csrColIndA), nnz * sizeof(int)) == cudaSuccess);
// Copy values
assert(cudaMemcpy(csrValA, values, nnz * sizeof(double), cudaMemcpyHostToDevice) == cudaSuccess);
assert(cudaMemcpy(csrRowPtrA, col_ptr, (m+1) * sizeof(int), cudaMemcpyHostToDevice) == cudaSuccess);
assert(cudaMemcpy(csrColIndA, row_ind, nnz * sizeof(int), cudaMemcpyHostToDevice) == cudaSuccess);
}
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 TxMatrixOptimizationDataCU::ingestLocalMatrix(SparseMatrix& A) {
std::vector<local_int_t> i(A.localNumberOfRows + 1, 0);
// Slight overallocation for these arrays
std::vector<local_int_t> j;
j.reserve(A.localNumberOfNonzeros);
std::vector<double> a;
a.reserve(A.localNumberOfNonzeros);
scatterFromHalo.setNumRows(A.localNumberOfRows);
scatterFromHalo.setNumCols(A.localNumberOfColumns);
scatterFromHalo.clear();
// We're splitting the matrix into diagonal and off-diagonal block to
// enable overlapping of computation and communication.
i[0] = 0;
for (local_int_t m = 0; m < A.localNumberOfRows; ++m) {
local_int_t nonzerosInRow = 0;
for (local_int_t n = 0; n < A.nonzerosInRow[m]; ++n) {
local_int_t col = A.mtxIndL[m][n];
if (col < A.localNumberOfRows) {
j.push_back(col);
a.push_back(A.matrixValues[m][n]);
++nonzerosInRow;
} else {
scatterFromHalo.addEntry(m, col, A.matrixValues[m][n]);
}
}
i[m + 1] = i[m] + nonzerosInRow;
}
// Setup SpMV data on Device
cudaError_t err = cudaSuccess;
int* i_d;
err = cudaMalloc((void**)&i_d, i.size() * sizeof(i[0]));
CHKCUDAERR(err);
err = cudaMemcpy(i_d, &i[0], i.size() * sizeof(i[0]), cudaMemcpyHostToDevice);
CHKCUDAERR(err);
int* j_d;
err = cudaMalloc((void**)&j_d, j.size() * sizeof(j[0]));
CHKCUDAERR(err);
err = cudaMemcpy(j_d, &j[0], j.size() * sizeof(j[0]), cudaMemcpyHostToDevice);
CHKCUDAERR(err);
double* a_d;
err = cudaMalloc((void**)&a_d, a.size() * sizeof(a[0]));
CHKCUDAERR(err);
err = cudaMemcpy(a_d, &a[0], a.size() * sizeof(a[0]), cudaMemcpyHostToDevice);
CHKCUDAERR(err);
cusparseStatus_t cerr = CUSPARSE_STATUS_SUCCESS;
cerr = cusparseCreateMatDescr(&matDescr);
CHKCUSPARSEERR(cerr);
cerr = cusparseSetMatIndexBase(matDescr, CUSPARSE_INDEX_BASE_ZERO);
CHKCUSPARSEERR(cerr);
cerr = cusparseSetMatType(matDescr, CUSPARSE_MATRIX_TYPE_GENERAL);
CHKCUSPARSEERR(cerr);
cerr = cusparseCreateHybMat(&localMatrix);
CHKCUSPARSEERR(cerr);
cerr = cusparseDcsr2hyb(handle, A.localNumberOfRows, A.localNumberOfColumns,
matDescr, a_d, i_d, j_d, localMatrix, 27,
CUSPARSE_HYB_PARTITION_USER);
CHKCUSPARSEERR(cerr);
#ifndef HPCG_NOMPI
err = cudaMalloc((void**)&elementsToSend,
A.totalToBeSent * sizeof(*elementsToSend));
CHKCUDAERR(err);
err = cudaMemcpy(elementsToSend, A.elementsToSend,
A.totalToBeSent * sizeof(*elementsToSend),
cudaMemcpyHostToDevice);
CHKCUDAERR(err);
err = cudaMalloc((void**)&sendBuffer_d, A.totalToBeSent * sizeof(double));
CHKCUDAERR(err);
#endif
// Set up the GS data.
gelusStatus_t gerr = GELUS_STATUS_SUCCESS;
gelusSolveDescription_t solveDescr;
gerr = gelusCreateSolveDescr(&solveDescr);
CHKGELUSERR(gerr);
gerr = gelusSetSolveOperation(solveDescr, GELUS_OPERATION_NON_TRANSPOSE);
CHKGELUSERR(gerr);
gerr = gelusSetSolveFillMode(solveDescr, GELUS_FILL_MODE_FULL);
CHKGELUSERR(gerr);
gerr = gelusSetSolveStorageFormat(solveDescr, GELUS_STORAGE_FORMAT_HYB);
CHKGELUSERR(gerr);
gerr = gelusSetOptimizationLevel(solveDescr, GELUS_OPTIMIZATION_LEVEL_THREE);
CHKGELUSERR(gerr);
gerr = cugelusCreateSorIterationData(&gsContext);
CHKGELUSERR(gerr);
#ifdef HPCG_DEBUG
std::cout << A.localNumberOfRows << std::endl;
std::cout << A.localNumberOfColumns << std::endl;
std::cout << A.localNumberOfNonzeros << std::endl;
int myrank;
MPI_Comm_rank(MPI_COMM_WORLD, &myrank);
if (myrank == 0) {
dumpMatrix(std::cout, i, j, a);
}
#endif
gerr = cugelusDcsrsor_iteration_analysis(
A.localNumberOfRows, solveDescr, GELUS_SOR_SYMMETRIC, 1.0,
&i[0], &j[0], &a[0], gsContext);
gerr = gelusDestroySolveDescr(solveDescr);
CHKGELUSERR(gerr);
if (A.mgData) {
err =
cudaMalloc((void**)&f2c, A.mgData->rc->localLength * sizeof(local_int_t));
CHKCUDAERR(err);
err = cudaMemcpy(f2c, A.mgData->f2cOperator,
A.mgData->rc->localLength * sizeof(local_int_t),
cudaMemcpyHostToDevice);
CHKCUDAERR(err);
}
err = cudaMalloc((void**)&workvector, A.localNumberOfRows * sizeof(double));
CHKCUDAERR(err);
return (int)cerr | (int)gerr | (int)err;
}
/* 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_d_spmv(
double alpha,
magma_d_matrix A,
magma_d_matrix x,
double beta,
magma_d_matrix y,
magma_queue_t queue )
{
magma_int_t info = 0;
magma_d_matrix x2={Magma_CSR};
cusparseHandle_t cusparseHandle = 0;
cusparseMatDescr_t descr = 0;
// make sure RHS is a dense matrix
if ( x.storage_type != Magma_DENSE ) {
printf("error: only dense vectors are supported for SpMV.\n");
info = MAGMA_ERR_NOT_SUPPORTED;
goto cleanup;
}
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 );
info = MAGMA_ERR_INVALID_PTR;
goto cleanup;
}
// 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_CUCSR
|| A.storage_type == Magma_CSRL
|| A.storage_type == Magma_CSRU ) {
CHECK_CUSPARSE( cusparseCreate( &cusparseHandle ));
CHECK_CUSPARSE( cusparseSetStream( cusparseHandle, queue->cuda_stream() ));
CHECK_CUSPARSE( cusparseCreateMatDescr( &descr ));
CHECK_CUSPARSE( cusparseSetMatType( descr, CUSPARSE_MATRIX_TYPE_GENERAL ));
CHECK_CUSPARSE( cusparseSetMatIndexBase( descr, CUSPARSE_INDEX_BASE_ZERO ));
cusparseDcsrmv( 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 );
}
else if ( A.storage_type == Magma_ELL ) {
//printf("using ELLPACKT kernel for SpMV: ");
CHECK( magma_dgeelltmv( 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: ");
CHECK( magma_dgeellmv( 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: ");
CHECK( magma_dgeellrtmv( 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: ");
CHECK( magma_dgesellpmv( 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_dgemv( MagmaNoTrans, A.num_rows, A.num_cols, alpha,
A.dval, A.num_rows, x.dval, 1, beta, y.dval,
1, queue );
//printf("done.\n");
}
else if ( A.storage_type == Magma_SPMVFUNCTION ) {
//printf("using DENSE kernel for SpMV: ");
CHECK( magma_dcustomspmv( alpha, x, beta, y, queue ));
//printf("done.\n");
}
else if ( A.storage_type == Magma_BCSR ) {
//printf("using CUSPARSE BCSR kernel for SpMV: ");
// CUSPARSE context //
cusparseDirection_t dirA = CUSPARSE_DIRECTION_ROW;
int mb = magma_ceildiv( A.num_rows, A.blocksize );
int nb = magma_ceildiv( A.num_cols, A.blocksize );
CHECK_CUSPARSE( cusparseCreate( &cusparseHandle ));
CHECK_CUSPARSE( cusparseSetStream( cusparseHandle, queue->cuda_stream() ));
CHECK_CUSPARSE( cusparseCreateMatDescr( &descr ));
cusparseDbsrmv( 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 );
}
else {
printf("error: format not supported.\n");
info = 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 ) {
CHECK_CUSPARSE( cusparseCreate( &cusparseHandle ));
CHECK_CUSPARSE( cusparseSetStream( cusparseHandle, queue->cuda_stream() ));
CHECK_CUSPARSE( cusparseCreateMatDescr( &descr ));
CHECK_CUSPARSE( cusparseSetMatType( descr, CUSPARSE_MATRIX_TYPE_GENERAL ));
CHECK_CUSPARSE( cusparseSetMatIndexBase( descr, CUSPARSE_INDEX_BASE_ZERO ));
if ( x.major == MagmaColMajor) {
cusparseDcsrmm(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) {
/*cusparseDcsrmm2(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);
*/
}
} else if ( A.storage_type == Magma_SELLP ) {
if ( x.major == MagmaRowMajor) {
CHECK( magma_dmgesellpmv( 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
CHECK( magma_dvtranspose( x, &x2, queue ));
CHECK( magma_dmgesellpmv( 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 ));
}
}
/*if ( A.storage_type == Magma_DENSE ) {
//printf("using DENSE kernel for SpMV: ");
magmablas_dmgemv( 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");
}*/
else {
printf("error: format not supported.\n");
info = MAGMA_ERR_NOT_SUPPORTED;
}
}
}
// CPU case missing!
else {
printf("error: CPU not yet supported.\n");
info = MAGMA_ERR_NOT_SUPPORTED;
}
cleanup:
cusparseDestroyMatDescr( descr );
cusparseDestroy( cusparseHandle );
cusparseHandle = 0;
descr = 0;
magma_dmfree(&x2, queue );
return info;
}
/* ////////////////////////////////////////////////////////////////////////////
-- testing sparse matrix vector product
*/
int main( int argc, char** argv )
{
TESTING_INIT();
magma_queue_t queue;
magma_queue_create( /*devices[ opts->device ],*/ &queue );
magma_d_sparse_matrix hA, hA_SELLP, hA_ELL, dA, dA_SELLP, dA_ELL;
hA_SELLP.blocksize = 8;
hA_SELLP.alignment = 8;
real_Double_t start, end, res;
magma_int_t *pntre;
double c_one = MAGMA_D_MAKE(1.0, 0.0);
double c_zero = MAGMA_D_MAKE(0.0, 0.0);
magma_int_t i, j;
for( i = 1; i < argc; ++i ) {
if ( strcmp("--blocksize", argv[i]) == 0 ) {
hA_SELLP.blocksize = atoi( argv[++i] );
} else if ( strcmp("--alignment", argv[i]) == 0 ) {
hA_SELLP.alignment = atoi( argv[++i] );
} else
break;
}
printf( "\n# usage: ./run_dspmv"
" [ --blocksize %d --alignment %d (for SELLP) ]"
" matrices \n\n", (int) hA_SELLP.blocksize, (int) hA_SELLP.alignment );
while( i < argc ) {
if ( strcmp("LAPLACE2D", argv[i]) == 0 && i+1 < argc ) { // Laplace test
i++;
magma_int_t laplace_size = atoi( argv[i] );
magma_dm_5stencil( laplace_size, &hA, queue );
} else { // file-matrix test
magma_d_csr_mtx( &hA, argv[i], queue );
}
printf( "\n# matrix info: %d-by-%d with %d nonzeros\n\n",
(int) hA.num_rows,(int) hA.num_cols,(int) hA.nnz );
real_Double_t FLOPS = 2.0*hA.nnz/1e9;
magma_d_vector hx, hy, dx, dy, hrefvec, hcheck;
// init CPU vectors
magma_d_vinit( &hx, Magma_CPU, hA.num_rows, c_zero, queue );
magma_d_vinit( &hy, Magma_CPU, hA.num_rows, c_zero, queue );
// init DEV vectors
magma_d_vinit( &dx, Magma_DEV, hA.num_rows, c_one, queue );
magma_d_vinit( &dy, Magma_DEV, hA.num_rows, c_zero, queue );
#ifdef MAGMA_WITH_MKL
// calling MKL with CSR
pntre = (magma_int_t*)malloc( (hA.num_rows+1)*sizeof(magma_int_t) );
pntre[0] = 0;
for (j=0; j<hA.num_rows; j++ ) {
pntre[j] = hA.row[j+1];
}
MKL_INT num_rows = hA.num_rows;
MKL_INT num_cols = hA.num_cols;
MKL_INT nnz = hA.nnz;
MKL_INT *col;
TESTING_MALLOC_CPU( col, MKL_INT, nnz );
for( magma_int_t t=0; t < hA.nnz; ++t ) {
col[ t ] = hA.col[ t ];
}
MKL_INT *row;
TESTING_MALLOC_CPU( row, MKL_INT, num_rows );
for( magma_int_t t=0; t < hA.num_rows; ++t ) {
row[ t ] = hA.col[ t ];
}
start = magma_wtime();
for (j=0; j<10; j++ ) {
mkl_dcsrmv( "N", &num_rows, &num_cols,
MKL_ADDR(&c_one), "GFNC", MKL_ADDR(hA.val),
col, row, pntre,
MKL_ADDR(hx.val),
MKL_ADDR(&c_zero), MKL_ADDR(hy.val) );
}
end = magma_wtime();
printf( "\n > MKL : %.2e seconds %.2e GFLOP/s (CSR).\n",
(end-start)/10, FLOPS*10/(end-start) );
TESTING_FREE_CPU( row );
TESTING_FREE_CPU( col );
free(pntre);
#endif // MAGMA_WITH_MKL
// copy matrix to GPU
magma_d_mtransfer( hA, &dA, Magma_CPU, Magma_DEV, queue );
// SpMV on GPU (CSR) -- this is the reference!
start = magma_sync_wtime( queue );
for (j=0; j<10; j++)
magma_d_spmv( c_one, dA, dx, c_zero, dy, queue );
end = magma_sync_wtime( queue );
printf( " > MAGMA: %.2e seconds %.2e GFLOP/s (standard CSR).\n",
(end-start)/10, FLOPS*10/(end-start) );
magma_d_mfree(&dA, queue );
magma_d_vtransfer( dy, &hrefvec , Magma_DEV, Magma_CPU, queue );
// convert to ELL and copy to GPU
magma_d_mconvert( hA, &hA_ELL, Magma_CSR, Magma_ELL, queue );
magma_d_mtransfer( hA_ELL, &dA_ELL, Magma_CPU, Magma_DEV, queue );
magma_d_mfree(&hA_ELL, queue );
magma_d_vfree( &dy, queue );
magma_d_vinit( &dy, Magma_DEV, hA.num_rows, c_zero, queue );
// SpMV on GPU (ELL)
start = magma_sync_wtime( queue );
for (j=0; j<10; j++)
magma_d_spmv( c_one, dA_ELL, dx, c_zero, dy, queue );
end = magma_sync_wtime( queue );
printf( " > MAGMA: %.2e seconds %.2e GFLOP/s (standard ELL).\n",
(end-start)/10, FLOPS*10/(end-start) );
magma_d_mfree(&dA_ELL, queue );
magma_d_vtransfer( dy, &hcheck , Magma_DEV, Magma_CPU, queue );
res = 0.0;
for(magma_int_t k=0; k<hA.num_rows; k++ )
res=res + MAGMA_D_REAL(hcheck.val[k]) - MAGMA_D_REAL(hrefvec.val[k]);
if ( res < .000001 )
printf("# tester spmv ELL: ok\n");
else
printf("# tester spmv ELL: failed\n");
magma_d_vfree( &hcheck, queue );
// convert to SELLP and copy to GPU
magma_d_mconvert( hA, &hA_SELLP, Magma_CSR, Magma_SELLP, queue );
magma_d_mtransfer( hA_SELLP, &dA_SELLP, Magma_CPU, Magma_DEV, queue );
magma_d_mfree(&hA_SELLP, queue );
magma_d_vfree( &dy, queue );
magma_d_vinit( &dy, Magma_DEV, hA.num_rows, c_zero, queue );
// SpMV on GPU (SELLP)
start = magma_sync_wtime( queue );
for (j=0; j<10; j++)
magma_d_spmv( c_one, dA_SELLP, dx, c_zero, dy, queue );
end = magma_sync_wtime( queue );
printf( " > MAGMA: %.2e seconds %.2e GFLOP/s (SELLP).\n",
(end-start)/10, FLOPS*10/(end-start) );
magma_d_vtransfer( dy, &hcheck , Magma_DEV, Magma_CPU, queue );
res = 0.0;
for(magma_int_t k=0; k<hA.num_rows; k++ )
res=res + MAGMA_D_REAL(hcheck.val[k]) - MAGMA_D_REAL(hrefvec.val[k]);
printf("# |x-y|_F = %8.2e\n", res);
if ( res < .000001 )
printf("# tester spmv SELL-P: ok\n");
else
printf("# tester spmv SELL-P: failed\n");
magma_d_vfree( &hcheck, queue );
magma_d_mfree(&dA_SELLP, queue );
// SpMV on GPU (CUSPARSE - CSR)
// CUSPARSE context //
cusparseHandle_t cusparseHandle = 0;
cusparseStatus_t cusparseStatus;
cusparseStatus = cusparseCreate(&cusparseHandle);
cusparseSetStream( cusparseHandle, queue );
cusparseMatDescr_t descr = 0;
cusparseStatus = cusparseCreateMatDescr(&descr);
cusparseSetMatType(descr,CUSPARSE_MATRIX_TYPE_GENERAL);
cusparseSetMatIndexBase(descr,CUSPARSE_INDEX_BASE_ZERO);
double alpha = c_one;
double beta = c_zero;
magma_d_vfree( &dy, queue );
magma_d_vinit( &dy, Magma_DEV, hA.num_rows, c_zero, queue );
// copy matrix to GPU
magma_d_mtransfer( hA, &dA, Magma_CPU, Magma_DEV, queue );
start = magma_sync_wtime( queue );
for (j=0; j<10; j++)
cusparseStatus =
cusparseDcsrmv(cusparseHandle,CUSPARSE_OPERATION_NON_TRANSPOSE,
hA.num_rows, hA.num_cols, hA.nnz, &alpha, descr,
dA.dval, dA.drow, dA.dcol, dx.dval, &beta, dy.dval);
end = magma_sync_wtime( queue );
if (cusparseStatus != 0) printf("error in cuSPARSE CSR\n");
printf( " > CUSPARSE: %.2e seconds %.2e GFLOP/s (CSR).\n",
(end-start)/10, FLOPS*10/(end-start) );
cusparseMatDescr_t descrA;
cusparseStatus = cusparseCreateMatDescr(&descrA);
if (cusparseStatus != 0) printf("error\n");
cusparseHybMat_t hybA;
cusparseStatus = cusparseCreateHybMat( &hybA );
if (cusparseStatus != 0) printf("error\n");
magma_d_vtransfer( dy, &hcheck , Magma_DEV, Magma_CPU, queue );
res = 0.0;
for(magma_int_t k=0; k<hA.num_rows; k++ )
res=res + MAGMA_D_REAL(hcheck.val[k]) - MAGMA_D_REAL(hrefvec.val[k]);
printf("# |x-y|_F = %8.2e\n", res);
if ( res < .000001 )
printf("# tester spmv cuSPARSE CSR: ok\n");
else
printf("# tester spmv cuSPARSE CSR: failed\n");
magma_d_vfree( &hcheck, queue );
magma_d_vfree( &dy, queue );
magma_d_vinit( &dy, Magma_DEV, hA.num_rows, c_zero, queue );
cusparseDcsr2hyb(cusparseHandle, hA.num_rows, hA.num_cols,
descrA, dA.dval, dA.drow, dA.dcol,
hybA, 0, CUSPARSE_HYB_PARTITION_AUTO);
start = magma_sync_wtime( queue );
for (j=0; j<10; j++)
cusparseStatus =
cusparseDhybmv( cusparseHandle, CUSPARSE_OPERATION_NON_TRANSPOSE,
&alpha, descrA, hybA,
dx.dval, &beta, dy.dval);
end = magma_sync_wtime( queue );
if (cusparseStatus != 0) printf("error in cuSPARSE HYB\n");
printf( " > CUSPARSE: %.2e seconds %.2e GFLOP/s (HYB).\n",
(end-start)/10, FLOPS*10/(end-start) );
magma_d_vtransfer( dy, &hcheck , Magma_DEV, Magma_CPU, queue );
res = 0.0;
for(magma_int_t k=0; k<hA.num_rows; k++ )
res=res + MAGMA_D_REAL(hcheck.val[k]) - MAGMA_D_REAL(hrefvec.val[k]);
printf("# |x-y|_F = %8.2e\n", res);
if ( res < .000001 )
printf("# tester spmv cuSPARSE HYB: ok\n");
else
printf("# tester spmv cuSPARSE HYB: failed\n");
magma_d_vfree( &hcheck, queue );
cusparseDestroyMatDescr( descrA );
cusparseDestroyHybMat( hybA );
cusparseDestroy( cusparseHandle );
magma_d_mfree(&dA, queue );
printf("\n\n");
// free CPU memory
magma_d_mfree(&hA, queue );
magma_d_vfree(&hx, queue );
magma_d_vfree(&hy, queue );
magma_d_vfree(&hrefvec, queue );
// free GPU memory
magma_d_vfree(&dx, queue );
magma_d_vfree(&dy, queue );
i++;
}
magma_queue_destroy( queue );
TESTING_FINALIZE();
return 0;
}
int main(int argc, char* argv[])
{
const int bufsize = 512;
char buffer[bufsize];
int m,n,S;
double time_st,time_end,time_avg;
//omp_set_num_threads(2);
// printf("\n-----------------\nnumber of threads fired = %d\n-----------------\n",(int)omp_get_num_threads());
if(argc!=2)
{
cout<<"Insufficient arguments"<<endl;
return 1;
}
graph G;
cerr<<"Start reading ";
// time_st=dsecnd();
G.create_graph(argv[1]);
// time_end=dsecnd();
// time_avg = (time_end-time_st);
// cout<<"Success "<<endl;
// cerr<<"Reading time "<<time_avg<<endl;
cerr<<"Constructing Matrices ";
// time_st=dsecnd();
G.construct_MNA();
// time_end=dsecnd();
// time_avg = (time_end-time_st);
// cerr<<"Done "<<time_avg<<endl;
// G.construct_sparse_MNA();
m=G.node_array.size()-1;
n=G.voltage_edge_id.size();
cout<<endl;
cout<<"MATRIX STAT:"<<endl;
cout<<"Nonzero elements: "<<G.nonzero<<endl;
cout<<"Number of Rows: "<<m+n<<endl;
printf("\n Nonzero = %ld", G.nonzero);
printf("\n Rows = %d", m+n);
cout<<"MAT val: "<<endl;
int i,j;
// G.Mat_val[0] +=100;
/*
for(i=0;i<G.nonzero;i++)
cout<<" "<<G.Mat_val[i];
cout<<endl;
for(i=0;i<G.nonzero;i++)
cout<<" "<<G.columns[i];
cout<<endl;
for(i=0;i<m+n+1;i++)
cout<<" "<<G.rowIndex[i];
cout<<endl;
for(i=0;i<m+n;i++)
{
cout<<endl;
int startindex=G.rowIndex[i];
int endindex=G.rowIndex[i+1];
for(j=startindex;j<endindex;j++)
cout<<" "<<G.Mat_val[j];
cout<<endl;
}
*/
/* for (i=0;i<m+n+1;i++)
{
//cout<<endl;
if(G.rowIndex[i]==G.rowIndex[i+1])
break;
for(j=G.rowIndex[i];j<G.rowIndex[i+1];j++)
{
if(G.Mat_val[j]>10)
cout<<G.Mat_val[j]<<"\t";
}
//cout<<endl;
/*for(j=G.rowIndex[i];j<G.rowIndex[i+1];j++)
{
cout<<G.columns[j]<<"\t";
}
//cout<<endl;
}
cout<<endl;
*/
//printing the matrix
printf("\n Fine till here");
printf("\n");
// int* rowmIndex=(int*)calloc(m+1,sizeof(int));
printf("\n Fine till here");
printf("\n");
//int rowmIndex[5]={1,2,3,4,5};
/* for(i=0;i<m+1;i++)
{
rowmIndex[i]=G.rowIndex[i];
printf(" %d", rowmIndex[i]);
}
*/
cerr<<"Solving Equations "<<endl;
double r1, b, alpha, alpham1, beta, r0, a, na;
const double tol = 0.1;
const int max_iter = 1000000;
int *d_col, *d_row;
double *d_val, *d_x, dot;
double *d_r, *d_p, *d_Ax;
int k;
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);
cusparseMatDescr_t descr = 0;
cusparseStatus = cusparseCreateMatDescr(&descr);
checkCudaErrors(cusparseStatus);
cusparseSetMatType(descr,CUSPARSE_MATRIX_TYPE_GENERAL);
cusparseSetMatIndexBase(descr,CUSPARSE_INDEX_BASE_ZERO);
checkCudaErrors(cudaMalloc((void **)&d_col, G.nonzero*sizeof(int)));
checkCudaErrors(cudaMalloc((void **)&d_row, (m+n+1)*sizeof(int)));
checkCudaErrors(cudaMalloc((void **)&d_val, G.nonzero*sizeof(double)));
checkCudaErrors(cudaMalloc((void **)&d_x, (m+n)*sizeof(double)));
checkCudaErrors(cudaMalloc((void **)&d_r, (m+n)*sizeof(double)));
checkCudaErrors(cudaMalloc((void **)&d_p, (m+n)*sizeof(double)));
checkCudaErrors(cudaMalloc((void **)&d_Ax, (m+n)*sizeof(double)));
cudaMemcpy(d_col, G.columns, G.nonzero*sizeof(int), cudaMemcpyHostToDevice);
cudaMemcpy(d_row, G.rowIndex, (m+n+1)*sizeof(int), cudaMemcpyHostToDevice);
cudaMemcpy(d_val, G.Mat_val, G.nonzero*sizeof(double), cudaMemcpyHostToDevice);
cudaMemcpy(d_x, G.x, (m+n)*sizeof(double), cudaMemcpyHostToDevice);
cudaMemcpy(d_r, G.b, (m+n)*sizeof(double), cudaMemcpyHostToDevice);
alpha = 1.0;
alpham1 = -1.0;
beta = 0.0;
r0 = 0.;
printf("\n Data transferred\n");
cudaEvent_t start,stop;
cudaEventCreate(&start);
cudaEventCreate(&stop);
cudaEventRecord(start, 0);
cusparseDcsrmv(cusparseHandle,CUSPARSE_OPERATION_NON_TRANSPOSE, (m+n), (m+n), G.nonzero, &alpha, descr, d_val, d_row, d_col, d_x, &beta, d_Ax);
cublasDaxpy(cublasHandle, (m+n), &alpham1, d_Ax, 1, d_r, 1);
cublasStatus = cublasDdot(cublasHandle, (m+n), d_r, 1, d_r, 1, &r1);
k = 1;
while (r1 > tol && k <= max_iter)
{
if (k > 1)
{
b = r1 / r0;
cublasStatus = cublasDscal(cublasHandle, (m+n), &b, d_p, 1);
cublasStatus = cublasDaxpy(cublasHandle, (m+n), &alpha, d_r, 1, d_p, 1);
}
else
{
cublasStatus = cublasDcopy(cublasHandle, (m+n), d_r, 1, d_p, 1);
}
cusparseDcsrmv(cusparseHandle, CUSPARSE_OPERATION_NON_TRANSPOSE, (m+n), (m+n), G.nonzero, &alpha, descr, d_val, d_row, d_col, d_p, &beta, d_Ax);
cublasStatus = cublasDdot(cublasHandle, (m+n), d_p, 1, d_Ax, 1, &dot);
a = r1 / dot;
cublasStatus = cublasDaxpy(cublasHandle, (m+n), &a, d_p, 1, d_x, 1);
na = -a;
cublasStatus = cublasDaxpy(cublasHandle, (m+n), &na, d_Ax, 1, d_r, 1);
r0 = r1;
cublasStatus = cublasDdot(cublasHandle, (m+n), d_r, 1, d_r, 1, &r1);
// cudaThreadSynchronize();
// printf("iteration = %3d, residual = %e\n", k, sqrt(r1));
k++;
}
cudaEventRecord(stop, 0);
cudaEventSynchronize(stop);
float elapsedTime;
cudaEventElapsedTime(&elapsedTime, start, stop);
printf("Iterations = %3d\tTime : %.6f milli-seconds : \n", k, elapsedTime);
cudaMemcpy(G.x, d_x, (m+n)*sizeof(double), cudaMemcpyDeviceToHost);
/* printf("\n x = \n");
for(i=0;i<(m+n);i++)
{
printf("\n x[%d] = %.8f", i, G.x[i]);
}
*/ float rsum, diff, err = 0.0;
for (int i = 0; i < (m+n); i++)
{
rsum = 0.0;
for (int j = G.rowIndex[i]; j < G.rowIndex[i+1]; j++)
{
rsum += G.Mat_val[j]*G.x[G.columns[j]];
}
diff = fabs(rsum - G.b[i]);
if (diff > err)
{
err = diff;
}
}
cusparseDestroy(cusparseHandle);
cublasDestroy(cublasHandle);
/* free(I);
free(J);
free(val);
free(x);
free(rhs);
*/ cudaFree(d_col);
cudaFree(d_row);
cudaFree(d_val);
cudaFree(d_x);
cudaFree(d_r);
cudaFree(d_p);
cudaFree(d_Ax);
// 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("\n X is:\n");
for(i=0;i<(m+n);i++)
{
printf("\n X[%d] = %.4f", i, G.x[i]);
}
*/
printf("Test Summary: Error amount = %f\n", err);
exit((k <= max_iter) ? 0 : 1);
/*
culaSparseHandle handle;
if (culaSparseCreate(&handle) != culaSparseNoError)
{
// this should only fail under extreme conditions
std::cout << "fatal error: failed to create library handle!" << std::endl;
exit(EXIT_FAILURE);
}
StatusChecker sc(handle);
culaSparsePlan plan;
sc = culaSparseCreatePlan(handle, &plan);
culaSparseCsrOptions formatOpts;
culaSparseCsrOptionsInit(handle, &formatOpts);
//formatOpts.indexing=1;
sc = culaSparseSetDcsrData(handle, plan, &formatOpts, m+n, G.nonzero, &G.Mat_val[0], &G.rowIndex[0], &G.columns[0], &G.x[0], &G.b[0]);
printf("\n Fine till here");
printf("\n");
culaSparseConfig config;
sc = culaSparseConfigInit(handle, &config);
config.relativeTolerance = 1e-2;
config.maxIterations = 100000;
config.divergenceTolerance = 20;
culaSparseResult result;
/* sc = culaSparseSetHostPlatform(handle, plan, 0);
// set cg solver
sc = culaSparseSetCgSolver(handle, plan, 0);
printf("\n Fine till here");
printf("\n");
// perform solve (cg + no preconditioner on host)
culaSparseResult result;
sc = culaSparseExecutePlan(handle, plan, &config, &result);
sc = culaSparseGetResultString(handle, &result, buffer, bufsize);
std::cout << buffer << std::endl;
if (culaSparsePreinitializeCuda(handle) == culaSparseNoError)
{
// change to cuda accelerated platform
sc = culaSparseSetCudaPlatform(handle, plan, 0);
// perform solve (cg + ilu0 on cuda)
culaSparseGmresOptions solverOpts;
culaSparseStatus status = culaSparseGmresOptionsInit(handle, &solverOpts);
solverOpts.restart = 20;
sc = culaSparseSetCgSolver(handle, plan, 0);
// sc = culaSparseSetJacobiPreconditioner(handle, plan, 0);
// sc = culaSparseSetIlu0Preconditioner(handle, plan, 0);
sc = culaSparseExecutePlan(handle, plan, &config, &result);
sc = culaSparseGetResultString(handle, &result, buffer, bufsize);
std::cout << buffer << std::endl;
// change preconditioner to fainv
// this avoids data transfer by using data cached by the plan
// sc = culaSparseSetIlu0Preconditioner(handle, plan, 0);
// perform solve (cg + fainv on cuda)
// these timing results should indicate minimal overhead
/* sc = culaSparseExecutePlan(handle, plan, &config, &result);
sc = culaSparseGetResultString(handle, &result, buffer, bufsize);
std::cout << buffer << std::endl;
// change solver
// this avoids preconditioner regeneration by using data cached by the plan
sc = culaSparseSetBicgstabSolver(handle, plan, 0);
// perform solve (bicgstab + fainv on cuda)
// the timing results should indicate minimal overhead and preconditioner generation time
sc = culaSparseExecutePlan(handle, plan, &config, &result);
sc = culaSparseGetResultString(handle, &result, buffer, bufsize);
std::cout << buffer << std::endl;
sc = culaSparseSetGmresSolver(handle, plan, 0);
// perform solve (bicgstab + fainv on cuda)
// the timing results should indicate minimal overhead and preconditioner generation time
sc = culaSparseExecutePlan(handle, plan, &config, &result);
sc = culaSparseGetResultString(handle, &result, buffer, bufsize);
std::cout << buffer << std::endl;
}
else
{
std::cout << "alert: no cuda capable gpu found" << std::endl;
}
// cleanup plan
culaSparseDestroyPlan(plan);
// cleanup handle
culaSparseDestroy(handle);
FILE* myWriteFile;
myWriteFile=fopen("result.txt","w");
for (i = 0; i < n; i++)
{
fprintf(myWriteFile,"%1f\n",G.x[i]);
// printf ("\n x [%d] = % f", i, x[i]);
}
fprintf(myWriteFile,".end\n");
fclose(myWriteFile);
printf ("\n");
// time_st=dsecnd();
// solver(G.rowIndex,G.columns,G.Mat_val,G.b,G.x,m+n,G.nonzero);
// time_end=dsecnd();
// time_avg = (time_end-time_st);
// printf("Successfully Solved in : %.6f secs\n",time_avg);
cerr<<endl;
cerr<<"Fillup Graph ";
// time_st=dsecnd();
G.fillup_graph();
// time_end=dsecnd();
// time_avg = (time_end-time_st);
// cerr<<"Done "<<time_avg<<endl;
//G.output_graph_stdout();
cerr<<"Matching KCL ";
// time_st=dsecnd();
G.check_kcl();
// time_end=dsecnd();
// time_avg = (time_end-time_st);
// cerr<<"Done "<<time_avg<<endl;
/*for (int i=0;i<m+n;i++)
{
cout<<"M"<<i<<endl;
for (int j=0;j<m+n;j++)
cout<<" "<<j<<"#"<<M[i][j]<<endl;
}*/
}
/* ////////////////////////////////////////////////////////////////////////////
-- testing sparse matrix vector product
*/
int main( int argc, char** argv )
{
magma_int_t info = 0;
TESTING_CHECK( magma_init() );
magma_print_environment();
magma_queue_t queue=NULL;
magma_queue_create( 0, &queue );
magma_s_matrix hA={Magma_CSR}, hA_SELLP={Magma_CSR},
dA={Magma_CSR}, dA_SELLP={Magma_CSR};
magma_s_matrix hx={Magma_CSR}, hy={Magma_CSR}, dx={Magma_CSR},
dy={Magma_CSR}, hrefvec={Magma_CSR}, hcheck={Magma_CSR};
hA_SELLP.blocksize = 8;
hA_SELLP.alignment = 8;
real_Double_t start, end, res;
#ifdef MAGMA_WITH_MKL
magma_int_t *pntre=NULL;
#endif
cusparseHandle_t cusparseHandle = NULL;
cusparseMatDescr_t descr = NULL;
float c_one = MAGMA_S_MAKE(1.0, 0.0);
float c_zero = MAGMA_S_MAKE(0.0, 0.0);
float accuracy = 1e-10;
#define PRECISION_s
#if defined(PRECISION_c)
accuracy = 1e-4;
#endif
#if defined(PRECISION_s)
accuracy = 1e-4;
#endif
magma_int_t i, j;
for( i = 1; i < argc; ++i ) {
if ( strcmp("--blocksize", argv[i]) == 0 ) {
hA_SELLP.blocksize = atoi( argv[++i] );
} else if ( strcmp("--alignment", argv[i]) == 0 ) {
hA_SELLP.alignment = atoi( argv[++i] );
} else
break;
}
printf("\n# usage: ./run_sspmm"
" [ --blocksize %lld --alignment %lld (for SELLP) ] matrices\n\n",
(long long) hA_SELLP.blocksize, (long long) hA_SELLP.alignment );
while( i < argc ) {
if ( strcmp("LAPLACE2D", argv[i]) == 0 && i+1 < argc ) { // Laplace test
i++;
magma_int_t laplace_size = atoi( argv[i] );
TESTING_CHECK( magma_sm_5stencil( laplace_size, &hA, queue ));
} else { // file-matrix test
TESTING_CHECK( magma_s_csr_mtx( &hA, argv[i], queue ));
}
printf("%% matrix info: %lld-by-%lld with %lld nonzeros\n",
(long long) hA.num_rows, (long long) hA.num_cols, (long long) hA.nnz );
real_Double_t FLOPS = 2.0*hA.nnz/1e9;
// m - number of rows for the sparse matrix
// n - number of vectors to be multiplied in the SpMM product
magma_int_t m, n;
m = hA.num_rows;
n = 48;
// init CPU vectors
TESTING_CHECK( magma_svinit( &hx, Magma_CPU, m, n, c_one, queue ));
TESTING_CHECK( magma_svinit( &hy, Magma_CPU, m, n, c_zero, queue ));
// init DEV vectors
TESTING_CHECK( magma_svinit( &dx, Magma_DEV, m, n, c_one, queue ));
TESTING_CHECK( magma_svinit( &dy, Magma_DEV, m, n, c_zero, queue ));
// calling MKL with CSR
#ifdef MAGMA_WITH_MKL
TESTING_CHECK( magma_imalloc_cpu( &pntre, m + 1 ) );
pntre[0] = 0;
for (j=0; j < m; j++ ) {
pntre[j] = hA.row[j+1];
}
MKL_INT num_rows = hA.num_rows;
MKL_INT num_cols = hA.num_cols;
MKL_INT nnz = hA.nnz;
MKL_INT num_vecs = n;
MKL_INT *col;
TESTING_CHECK( magma_malloc_cpu( (void**) &col, nnz * sizeof(MKL_INT) ));
for( magma_int_t t=0; t < hA.nnz; ++t ) {
col[ t ] = hA.col[ t ];
}
MKL_INT *row;
TESTING_CHECK( magma_malloc_cpu( (void**) &row, num_rows * sizeof(MKL_INT) ));
for( magma_int_t t=0; t < hA.num_rows; ++t ) {
row[ t ] = hA.col[ t ];
}
// === Call MKL with consecutive SpMVs, using mkl_scsrmv ===
// warmp up
mkl_scsrmv( "N", &num_rows, &num_cols,
MKL_ADDR(&c_one), "GFNC", MKL_ADDR(hA.val), col, row, pntre,
MKL_ADDR(hx.val),
MKL_ADDR(&c_zero), MKL_ADDR(hy.val) );
start = magma_wtime();
for (j=0; j < 10; j++ ) {
mkl_scsrmv( "N", &num_rows, &num_cols,
MKL_ADDR(&c_one), "GFNC", MKL_ADDR(hA.val), col, row, pntre,
MKL_ADDR(hx.val),
MKL_ADDR(&c_zero), MKL_ADDR(hy.val) );
}
end = magma_wtime();
printf( "\n > MKL SpMVs : %.2e seconds %.2e GFLOP/s (CSR).\n",
(end-start)/10, FLOPS*10/(end-start) );
// === Call MKL with blocked SpMVs, using mkl_scsrmm ===
char transa = 'n';
MKL_INT ldb = n, ldc=n;
char matdescra[6] = {'g', 'l', 'n', 'c', 'x', 'x'};
// warm up
mkl_scsrmm( &transa, &num_rows, &num_vecs, &num_cols, MKL_ADDR(&c_one), matdescra,
MKL_ADDR(hA.val), col, row, pntre,
MKL_ADDR(hx.val), &ldb,
MKL_ADDR(&c_zero),
MKL_ADDR(hy.val), &ldc );
start = magma_wtime();
for (j=0; j < 10; j++ ) {
mkl_scsrmm( &transa, &num_rows, &num_vecs, &num_cols, MKL_ADDR(&c_one), matdescra,
MKL_ADDR(hA.val), col, row, pntre,
MKL_ADDR(hx.val), &ldb,
MKL_ADDR(&c_zero),
MKL_ADDR(hy.val), &ldc );
}
end = magma_wtime();
printf( "\n > MKL SpMM : %.2e seconds %.2e GFLOP/s (CSR).\n",
(end-start)/10, FLOPS*10.*n/(end-start) );
magma_free_cpu( row );
magma_free_cpu( col );
row = NULL;
col = NULL;
#endif // MAGMA_WITH_MKL
// copy matrix to GPU
TESTING_CHECK( magma_smtransfer( hA, &dA, Magma_CPU, Magma_DEV, queue ));
// SpMV on GPU (CSR)
start = magma_sync_wtime( queue );
for (j=0; j < 10; j++) {
TESTING_CHECK( magma_s_spmv( c_one, dA, dx, c_zero, dy, queue ));
}
end = magma_sync_wtime( queue );
printf( " > MAGMA: %.2e seconds %.2e GFLOP/s (standard CSR).\n",
(end-start)/10, FLOPS*10.*n/(end-start) );
TESTING_CHECK( magma_smtransfer( dy, &hrefvec , Magma_DEV, Magma_CPU, queue ));
magma_smfree(&dA, queue );
// convert to SELLP and copy to GPU
TESTING_CHECK( magma_smconvert( hA, &hA_SELLP, Magma_CSR, Magma_SELLP, queue ));
TESTING_CHECK( magma_smtransfer( hA_SELLP, &dA_SELLP, Magma_CPU, Magma_DEV, queue ));
magma_smfree(&hA_SELLP, queue );
magma_smfree( &dy, queue );
TESTING_CHECK( magma_svinit( &dy, Magma_DEV, dx.num_rows, dx.num_cols, c_zero, queue ));
// SpMV on GPU (SELLP)
start = magma_sync_wtime( queue );
for (j=0; j < 10; j++) {
TESTING_CHECK( magma_s_spmv( c_one, dA_SELLP, dx, c_zero, dy, queue ));
}
end = magma_sync_wtime( queue );
printf( " > MAGMA: %.2e seconds %.2e GFLOP/s (SELLP).\n",
(end-start)/10, FLOPS*10.*n/(end-start) );
TESTING_CHECK( magma_smtransfer( dy, &hcheck , Magma_DEV, Magma_CPU, queue ));
res = 0.0;
for(magma_int_t k=0; k < hA.num_rows; k++ ) {
res=res + MAGMA_S_REAL(hcheck.val[k]) - MAGMA_S_REAL(hrefvec.val[k]);
}
printf("%% |x-y|_F = %8.2e\n", res);
if ( res < accuracy )
printf("%% tester spmm SELL-P: ok\n");
else
printf("%% tester spmm SELL-P: failed\n");
magma_smfree( &hcheck, queue );
magma_smfree(&dA_SELLP, queue );
// SpMV on GPU (CUSPARSE - CSR)
// CUSPARSE context //
magma_smfree( &dy, queue );
TESTING_CHECK( magma_svinit( &dy, Magma_DEV, dx.num_rows, dx.num_cols, c_zero, queue ));
//#ifdef PRECISION_d
start = magma_sync_wtime( queue );
TESTING_CHECK( cusparseCreate( &cusparseHandle ));
TESTING_CHECK( cusparseSetStream( cusparseHandle, magma_queue_get_cuda_stream(queue) ));
TESTING_CHECK( cusparseCreateMatDescr( &descr ));
TESTING_CHECK( cusparseSetMatType( descr, CUSPARSE_MATRIX_TYPE_GENERAL ));
TESTING_CHECK( cusparseSetMatIndexBase( descr, CUSPARSE_INDEX_BASE_ZERO ));
float alpha = c_one;
float beta = c_zero;
// copy matrix to GPU
TESTING_CHECK( magma_smtransfer( hA, &dA, Magma_CPU, Magma_DEV, queue) );
for (j=0; j < 10; j++) {
cusparseScsrmm(cusparseHandle,
CUSPARSE_OPERATION_NON_TRANSPOSE,
dA.num_rows, n, dA.num_cols, dA.nnz,
&alpha, descr, dA.dval, dA.drow, dA.dcol,
dx.dval, dA.num_cols, &beta, dy.dval, dA.num_cols);
}
end = magma_sync_wtime( queue );
printf( " > CUSPARSE: %.2e seconds %.2e GFLOP/s (CSR).\n",
(end-start)/10, FLOPS*10*n/(end-start) );
TESTING_CHECK( magma_smtransfer( dy, &hcheck , Magma_DEV, Magma_CPU, queue ));
res = 0.0;
for(magma_int_t k=0; k < hA.num_rows; k++ ) {
res = res + MAGMA_S_REAL(hcheck.val[k]) - MAGMA_S_REAL(hrefvec.val[k]);
}
printf("%% |x-y|_F = %8.2e\n", res);
if ( res < accuracy )
printf("%% tester spmm cuSPARSE: ok\n");
else
printf("%% tester spmm cuSPARSE: failed\n");
magma_smfree( &hcheck, queue );
cusparseDestroyMatDescr( descr );
cusparseDestroy( cusparseHandle );
descr = NULL;
cusparseHandle = NULL;
//#endif
printf("\n\n");
// free CPU memory
magma_smfree( &hA, queue );
magma_smfree( &hx, queue );
magma_smfree( &hy, queue );
magma_smfree( &hrefvec, queue );
// free GPU memory
magma_smfree( &dx, queue );
magma_smfree( &dy, queue );
magma_smfree( &dA, queue);
#ifdef MAGMA_WITH_MKL
magma_free_cpu( pntre );
#endif
i++;
}
magma_queue_destroy( queue );
TESTING_CHECK( magma_finalize() );
return info;
}
extern "C" magma_int_t
magma_cmtransposeconjugate(
magma_c_matrix A,
magma_c_matrix *B,
magma_queue_t queue )
{
// for symmetric matrices: convert to csc using cusparse
magma_int_t info = 0;
cusparseHandle_t handle=NULL;
cusparseMatDescr_t descrA=NULL;
cusparseMatDescr_t descrB=NULL;
magma_c_matrix ACSR={Magma_CSR}, BCSR={Magma_CSR};
magma_c_matrix A_d={Magma_CSR}, B_d={Magma_CSR};
if( A.storage_type == Magma_CSR && A.memory_location == Magma_DEV ) {
// fill in information for B
B->storage_type = A.storage_type;
B->diagorder_type = A.diagorder_type;
B->memory_location = Magma_DEV;
B->num_rows = A.num_cols; // transposed
B->num_cols = A.num_rows; // transposed
B->nnz = A.nnz;
B->true_nnz = A.true_nnz;
if ( A.fill_mode == MagmaFull ) {
B->fill_mode = MagmaFull;
}
else if ( A.fill_mode == MagmaLower ) {
B->fill_mode = MagmaUpper;
}
else if ( A.fill_mode == MagmaUpper ) {
B->fill_mode = MagmaLower;
}
B->dval = NULL;
B->drow = NULL;
B->dcol = NULL;
// memory allocation
CHECK( magma_cmalloc( &B->dval, B->nnz ));
CHECK( magma_index_malloc( &B->drow, B->num_rows + 1 ));
CHECK( magma_index_malloc( &B->dcol, B->nnz ));
// CUSPARSE context //
CHECK_CUSPARSE( cusparseCreate( &handle ));
CHECK_CUSPARSE( cusparseSetStream( handle, queue->cuda_stream() ));
CHECK_CUSPARSE( cusparseCreateMatDescr( &descrA ));
CHECK_CUSPARSE( cusparseCreateMatDescr( &descrB ));
CHECK_CUSPARSE( cusparseSetMatType( descrA, CUSPARSE_MATRIX_TYPE_GENERAL ));
CHECK_CUSPARSE( cusparseSetMatType( descrB, CUSPARSE_MATRIX_TYPE_GENERAL ));
CHECK_CUSPARSE( cusparseSetMatIndexBase( descrA, CUSPARSE_INDEX_BASE_ZERO ));
CHECK_CUSPARSE( cusparseSetMatIndexBase( descrB, CUSPARSE_INDEX_BASE_ZERO ));
CHECK_CUSPARSE(
cusparseCcsr2csc( handle, A.num_rows, A.num_cols, A.nnz,
A.dval, A.drow, A.dcol, B->dval, B->dcol, B->drow,
CUSPARSE_ACTION_NUMERIC,
CUSPARSE_INDEX_BASE_ZERO) );
CHECK( magma_cmconjugate( B, queue ));
} else if ( A.memory_location == Magma_CPU ){
CHECK( magma_cmtransfer( A, &A_d, A.memory_location, Magma_DEV, queue ));
CHECK( magma_cmtransposeconjugate( A_d, &B_d, queue ));
CHECK( magma_cmtransfer( B_d, B, Magma_DEV, A.memory_location, queue ));
} else {
CHECK( magma_cmconvert( A, &ACSR, A.storage_type, Magma_CSR, queue ));
CHECK( magma_cmtransposeconjugate( ACSR, &BCSR, queue ));
CHECK( magma_cmconvert( BCSR, B, Magma_CSR, A.storage_type, queue ));
}
cleanup:
cusparseDestroyMatDescr( descrA );
cusparseDestroyMatDescr( descrB );
cusparseDestroy( handle );
magma_cmfree( &A_d, queue );
magma_cmfree( &B_d, queue );
magma_cmfree( &ACSR, queue );
magma_cmfree( &BCSR, queue );
if( info != 0 ){
magma_cmfree( B, queue );
}
return info;
}
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;
}
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;
}
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);
}
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;
}
int main(int argc, char **argv)
{
int N = 0, nz = 0, *I = NULL, *J = NULL;
float *val = NULL;
const float tol = 1e-5f;
const int max_iter = 10000;
float *x;
float *rhs;
float a, b, na, r0, r1;
float dot;
float *r, *p, *Ax;
int k;
float alpha, beta, alpham1;
printf("Starting [%s]...\n", sSDKname);
// This will pick the best possible CUDA capable device
cudaDeviceProp deviceProp;
int devID = findCudaDevice(argc, (const char **)argv);
checkCudaErrors(cudaGetDeviceProperties(&deviceProp, devID));
#if defined(__APPLE__) || defined(MACOSX)
fprintf(stderr, "Unified Memory not currently supported on OS X\n");
cudaDeviceReset();
exit(EXIT_WAIVED);
#endif
if (sizeof(void *) != 8)
{
fprintf(stderr, "Unified Memory requires compiling for a 64-bit system.\n");
cudaDeviceReset();
exit(EXIT_WAIVED);
}
if (((deviceProp.major << 4) + deviceProp.minor) < 0x30)
{
fprintf(stderr, "%s requires Compute Capability of SM 3.0 or higher to run.\nexiting...\n", argv[0]);
cudaDeviceReset();
exit(EXIT_WAIVED);
}
// 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);
/* Generate a random tridiagonal symmetric matrix in CSR format */
N = 1048576;
nz = (N-2)*3 + 4;
cudaMallocManaged((void **)&I, sizeof(int)*(N+1));
cudaMallocManaged((void **)&J, sizeof(int)*nz);
cudaMallocManaged((void **)&val, sizeof(float)*nz);
genTridiag(I, J, val, N, nz);
cudaMallocManaged((void **)&x, sizeof(float)*N);
cudaMallocManaged((void **)&rhs, sizeof(float)*N);
for (int i = 0; i < N; i++)
{
rhs[i] = 1.0;
x[i] = 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);
cusparseMatDescr_t descr = 0;
cusparseStatus = cusparseCreateMatDescr(&descr);
checkCudaErrors(cusparseStatus);
cusparseSetMatType(descr,CUSPARSE_MATRIX_TYPE_GENERAL);
cusparseSetMatIndexBase(descr,CUSPARSE_INDEX_BASE_ZERO);
// temp memory for CG
checkCudaErrors(cudaMallocManaged((void **)&r, N*sizeof(float)));
checkCudaErrors(cudaMallocManaged((void **)&p, N*sizeof(float)));
checkCudaErrors(cudaMallocManaged((void **)&Ax, N*sizeof(float)));
cudaDeviceSynchronize();
for (int i=0; i < N; i++)
{
r[i] = rhs[i];
}
alpha = 1.0;
alpham1 = -1.0;
beta = 0.0;
r0 = 0.;
cusparseScsrmv(cusparseHandle,CUSPARSE_OPERATION_NON_TRANSPOSE, N, N, nz, &alpha, descr, val, I, J, x, &beta, Ax);
cublasSaxpy(cublasHandle, N, &alpham1, Ax, 1, r, 1);
cublasStatus = cublasSdot(cublasHandle, N, r, 1, r, 1, &r1);
k = 1;
while (r1 > tol*tol && k <= max_iter)
{
if (k > 1)
{
b = r1 / r0;
cublasStatus = cublasSscal(cublasHandle, N, &b, p, 1);
cublasStatus = cublasSaxpy(cublasHandle, N, &alpha, r, 1, p, 1);
}
else
{
cublasStatus = cublasScopy(cublasHandle, N, r, 1, p, 1);
}
cusparseScsrmv(cusparseHandle, CUSPARSE_OPERATION_NON_TRANSPOSE, N, N, nz, &alpha, descr, val, I, J, p, &beta, Ax);
cublasStatus = cublasSdot(cublasHandle, N, p, 1, Ax, 1, &dot);
a = r1 / dot;
cublasStatus = cublasSaxpy(cublasHandle, N, &a, p, 1, x, 1);
na = -a;
cublasStatus = cublasSaxpy(cublasHandle, N, &na, Ax, 1, r, 1);
r0 = r1;
cublasStatus = cublasSdot(cublasHandle, N, r, 1, r, 1, &r1);
cudaThreadSynchronize();
printf("iteration = %3d, residual = %e\n", k, sqrt(r1));
k++;
}
printf("Final residual: %e\n",sqrt(r1));
fprintf(stdout,"&&&& uvm_cg test %s\n", (sqrt(r1) < tol) ? "PASSED" : "FAILED");
float rsum, diff, 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;
}
}
cusparseDestroy(cusparseHandle);
cublasDestroy(cublasHandle);
cudaFree(I);
cudaFree(J);
cudaFree(val);
cudaFree(x);
cudaFree(r);
cudaFree(p);
cudaFree(Ax);
cudaDeviceReset();
printf("Test Summary: Error amount = %f, result = %s\n", err, (k <= max_iter) ? "SUCCESS" : "FAILURE");
exit((k <= max_iter) ? EXIT_SUCCESS : EXIT_FAILURE);
}
extern "C" magma_int_t
magma_zcuspaxpy(
magmaDoubleComplex *alpha, magma_z_sparse_matrix A,
magmaDoubleComplex *beta, magma_z_sparse_matrix B,
magma_z_sparse_matrix *AB,
magma_queue_t queue )
{
if ( A.memory_location == Magma_DEV
&& B.memory_location == Magma_DEV
&& ( A.storage_type == Magma_CSR ||
A.storage_type == Magma_CSRCOO )
&& ( B.storage_type == Magma_CSR ||
B.storage_type == Magma_CSRCOO ) ) {
magma_z_sparse_matrix C;
C.num_rows = A.num_rows;
C.num_cols = A.num_cols;
C.storage_type = A.storage_type;
C.memory_location = A.memory_location;
magma_int_t stat_dev = 0;
C.val = NULL;
C.col = NULL;
C.row = NULL;
C.rowidx = NULL;
C.blockinfo = NULL;
C.diag = NULL;
C.dval = NULL;
C.dcol = NULL;
C.drow = NULL;
C.drowidx = NULL;
C.ddiag = NULL;
// CUSPARSE context //
cusparseHandle_t handle;
cusparseStatus_t cusparseStatus;
cusparseStatus = cusparseCreate(&handle);
cusparseSetStream( handle, queue );
if (cusparseStatus != 0) printf("error in Handle.\n");
cusparseMatDescr_t descrA;
cusparseMatDescr_t descrB;
cusparseMatDescr_t descrC;
cusparseStatus = cusparseCreateMatDescr(&descrA);
cusparseStatus = cusparseCreateMatDescr(&descrB);
cusparseStatus = cusparseCreateMatDescr(&descrC);
if (cusparseStatus != 0) printf("error in MatrDescr.\n");
cusparseStatus =
cusparseSetMatType(descrA,CUSPARSE_MATRIX_TYPE_GENERAL);
cusparseSetMatType(descrB,CUSPARSE_MATRIX_TYPE_GENERAL);
cusparseSetMatType(descrC,CUSPARSE_MATRIX_TYPE_GENERAL);
if (cusparseStatus != 0) printf("error in MatrType.\n");
cusparseStatus =
cusparseSetMatIndexBase(descrA,CUSPARSE_INDEX_BASE_ZERO);
cusparseSetMatIndexBase(descrB,CUSPARSE_INDEX_BASE_ZERO);
cusparseSetMatIndexBase(descrC,CUSPARSE_INDEX_BASE_ZERO);
if (cusparseStatus != 0) printf("error in IndexBase.\n");
// multiply A and B on the device
magma_int_t baseC;
// nnzTotalDevHostPtr points to host memory
magma_index_t *nnzTotalDevHostPtr = (magma_index_t*) &C.nnz;
cusparseSetPointerMode(handle, CUSPARSE_POINTER_MODE_HOST);
stat_dev += magma_index_malloc( &C.drow, (A.num_rows + 1) );
cusparseXcsrgeamNnz(handle,A.num_rows, A.num_cols,
descrA, A.nnz, A.drow, A.dcol,
descrB, B.nnz, B.drow, B.dcol,
descrC, C.row, nnzTotalDevHostPtr);
if (NULL != nnzTotalDevHostPtr) {
C.nnz = *nnzTotalDevHostPtr;
} else {
// workaround as nnz and base C are magma_int_t
magma_index_t base_t, nnz_t;
magma_index_getvector( 1, C.drow+C.num_rows, 1, &nnz_t, 1 );
magma_index_getvector( 1, C.drow, 1, &base_t, 1 );
C.nnz = (magma_int_t) nnz_t;
baseC = (magma_int_t) base_t;
C.nnz -= baseC;
}
stat_dev += magma_index_malloc( &C.dcol, C.nnz );
stat_dev += magma_zmalloc( &C.dval, C.nnz );
if( stat_dev != 0 ) {
magma_z_mfree( &C, queue );
return MAGMA_ERR_DEVICE_ALLOC;
}
cusparseZcsrgeam(handle, A.num_rows, A.num_cols,
alpha,
descrA, A.nnz,
A.dval, A.drow, A.dcol,
beta,
descrB, B.nnz,
B.dval, B.drow, B.dcol,
descrC,
C.dval, C.drow, C.dcol);
cusparseDestroyMatDescr( descrA );
cusparseDestroyMatDescr( descrB );
cusparseDestroyMatDescr( descrC );
cusparseDestroy( handle );
// end CUSPARSE context //
magma_z_mtransfer( C, AB, Magma_DEV, Magma_DEV, queue );
magma_z_mfree( &C, queue );
return MAGMA_SUCCESS;
}
else {
printf("error: CSRSPAXPY only supported on device and CSR format.\n");
return MAGMA_SUCCESS;
}
}
extern "C" magma_int_t
magma_dapplycumicc_r(
magma_d_vector b, magma_d_vector *x,
magma_d_preconditioner *precond,
magma_queue_t queue )
{
double one = MAGMA_D_MAKE( 1.0, 0.0);
// CUSPARSE context //
cusparseHandle_t cusparseHandle;
cusparseStatus_t cusparseStatus;
cusparseStatus = cusparseCreate(&cusparseHandle);
cusparseSetStream( cusparseHandle, queue );
if (cusparseStatus != 0) printf("error in Handle.\n");
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");
// end CUSPARSE context //
magma_int_t dofs = precond->M.num_rows;
cusparseStatus =
cusparseDcsrsm_solve( cusparseHandle,
CUSPARSE_OPERATION_TRANSPOSE,
precond->M.num_rows,
b.num_rows*b.num_cols/precond->M.num_rows,
&one,
descrU,
precond->M.dval,
precond->M.drow,
precond->M.dcol,
precond->cuinfoU,
b.dval,
precond->M.num_rows,
x->dval,
precond->M.num_rows);
if (cusparseStatus != 0)
printf("error in U triangular solve:%d.\n", precond->cuinfoU );
cusparseDestroyMatDescr( descrU );
cusparseDestroy( cusparseHandle );
magma_device_sync();
return MAGMA_SUCCESS;
}