void initialize_gpu_buffer()
{
cudaError_t err = cudaMemcpy(device_row_offsets, &row_offsets[0], row_offsets.size() * sizeof(int), cudaMemcpyHostToDevice);
CUDA_V_THROW(err, "cudaMalloc device_row_offsets");
err = cudaMemcpy(device_col_indices, &col_indices[0], col_indices.size() * sizeof(int), cudaMemcpyHostToDevice);
CUDA_V_THROW(err, "cudaMalloc device_col_indices");
err = cudaMemcpy(device_values, &values[0], values.size() * sizeof(T), cudaMemcpyHostToDevice);
CUDA_V_THROW(err, "cudaMalloc device_values");
err = cudaMemset(device_A, 0x0, n_rows * n_cols * sizeof(T));
CUDA_V_THROW(err, "cudaMalloc device_A");
// call csr2dense to get input in dense format
csr2dense_Function(true);
int nnzA;
// Compute number of nonzero elements per row
if (typeid(T) == typeid(float))
{
cuSparseStatus = cusparseSnnz(handle,
CUSPARSE_DIRECTION_ROW,
n_rows,
n_cols,
descrA,
reinterpret_cast< float*> (device_A),
n_rows,
nnzPerRow,
&nnzA);
CUDA_V_THROW(cuSparseStatus, "cusparseSnnz");
}
else if (typeid(T) == typeid(double))
{
cuSparseStatus = cusparseDnnz(handle,
CUSPARSE_DIRECTION_ROW,
n_rows,
n_cols,
descrA,
reinterpret_cast< double*> (device_A),
n_rows,
nnzPerRow,
&nnzA);
CUDA_V_THROW(cuSparseStatus, "cusparseDnnz");
}
else
{
// error
}
if (nnzA != n_vals)
{
// error
}
cudaDeviceSynchronize();
// Once we get input in dense format, no-loner input csr values are needed.
CUDA_V_THROW(cudaFree(device_values), "cudafree device_values");
CUDA_V_THROW(cudaFree(device_row_offsets), "cudafree device_row_offsets");
CUDA_V_THROW(cudaFree(device_col_indices), "cudafree device_col_indices");
}// end
cusparseStatus_t cusparseXnnz(cusparseDirection_t dirA, int m, int n, const cusparseMatDescr_t descrA,
const float *A, int lda, int *nnzPerRowColumn, int *nnzTotalDevHostPtr) {
return cusparseSnnz(g_context->cusparseHandle, dirA, m, n, descrA, A, lda, nnzPerRowColumn, nnzTotalDevHostPtr);
}