cusolverSpScsrcholBufferInfo(...) crashes due to "Access violation reading location"

Accelerated Computing GPU-Accelerated Libraries
1

I am a newbie to the CUDA. When I use cuSPARSE to solve a linear system, I adopt Cholesky factorization.
My program crashes when I use the function cusolverSpScsrcholBufferInfo(...) after cusolverSpXcsrcholAnalysis(...). Visual Studio shows “Access violation reading location”.
I doubt that I input the wrong parameters so I write a simple code with a small 3x3 matrix to test this function, but come with the same error.

My GPU device is GeForce GTX Titan Black, released in 2014.
CUDA version: 10.1
IDE: Visual Studio 2019

If I don’t offer enough information, please let me know.

Here is my code a.cu:

#include "cuda_runtime.h"
#include "device_launch_parameters.h"
#include <cusolverSp.h>
#include <cusolverSp_LOWLEVEL_PREVIEW.h> //cusolverSpXcsrcholAnalysis
#include <cusolverRf.h>
#include <stdio.h>
#include "cusolverSp.h"
#include <iostream>
using namespace std;

void cudaSafeCall(cudaError_t err, char* msg) {
    if (err != cudaSuccess) {
        printf("Error(%s)! number %d, %s\n", msg, err, cudaGetErrorString(err));
        exit(1);
    }
}

 void cusolverCheck(cusolverStatus_t st, const char* msg)
{
    //std::cout << st << std::endl;
    if (st != CUSOLVER_STATUS_SUCCESS)
    {
        std::cout << "[cusolver error( " << st << ")]: " << msg;
        throw std::exception();
    }
}

int main()
{
    cusolverSpHandle_t spHandle;
    cusolverSpCreate(&spHandle);
    cusparseMatDescr_t desc;
    cusparseCreateMatDescr(&desc);
    csrcholInfo_t m_info = nullptr;
    cusolverSpCreateCsrcholInfo(&m_info);
    int row = 3, nnz = 3;
    int* rowId, * colId;
    float* value;
    cudaMallocManaged(&rowId, sizeof(int) * (row + 1));
    cudaMallocManaged(&colId, sizeof(int) * nnz);
    cudaMallocManaged(&value, sizeof(float) * nnz);
    // [ 1 0 0 ]
    // [ 0 2 0 ]
    // [ 0 0 3 ]
    for (int i = 0; i < nnz; ++i)value[i] = i + 1;
    for (int i = 0; i < row + 1; ++i)rowId[i] = i;
    for (int i = 0; i < nnz; ++i)colId[i] = i;
    size_t* internalDataInBytes, * workspaceInBytes;
    cusolverSpXcsrcholAnalysis(spHandle, row, nnz, desc, rowId, colId, m_info);

    cusolverSpScsrcholBufferInfo(spHandle, row, nnz, desc, value, rowId, colId, m_info,
        internalDataInBytes, workspaceInBytes);
    cudaDeviceSynchronize();
    cudaSafeCall(cudaGetLastError(), "cusolverSpScsrcholBufferInfo Error!");

    printf("success!!\n");
2

Solved.
Change
size_t* internalDataInBytes, * workspaceInBytes; ... cusolverSpScsrcholBufferInfo(spHandle, row, nnz, desc, value, rowId, colId, m_info, internalDataInBytes, workspaceInBytes);
into
size_t internalDataInBytes, workspaceInBytes; ... cusolverSpScsrcholBufferInfo(spHandle, row, nnz, desc, value, rowId, colId, m_info, &internalDataInBytes, &workspaceInBytes);

1 Like