cuRandGenerateUniform Access violation

Accelerated Computing GPU-Accelerated Libraries
1

The following code is causing Access violation error when executing cuRandGenerateUniform call and I have no idea why, as it seems to be an example almost from the book:

        auto bufSize = count * sizeof( float ); 
        uint32_t seed = 24342;

        float *rand = nullptr;
        cudaMalloc( (void**) &rand, bufSize );

        curandGenerator_t prngGPU = nullptr;
        curandCreateGenerator(&prngGPU, CURAND_RNG_PSEUDO_MTGP32);
        curandSetStream(prngGPU, stream);
        curandSetPseudoRandomGeneratorSeed(prngGPU, seed);
        // this call is causing Access violation exception:
        curandGenerateUniform(prngGPU, rand, count);
2

My suggestion would be to use proper error checking throughout your code and run your code with a sanitizer such as cuda-memcheck or compute-sanitizer.

As a simple example, if your count value was too large, or you were otherwise out of GPU memory (perhaps because you are calling this in a loop and not properly freeing allocations), your cudaMalloc operation would fail.

I’m not certain that is the issue, but error checking may be a useful step. Your code seems to run fine for me, when I provide suitable missing values:

$ cat t1932.cu
#include <curand.h>
#include <cstdint>

int main(){
        cudaStream_t stream;
        cudaStreamCreate(&stream);
        size_t count = 1048576;
        auto bufSize = count * sizeof( float );
        uint32_t seed = 24342;

        float *rand = nullptr;
        cudaMalloc( (void**) &rand, bufSize );

        curandGenerator_t prngGPU = nullptr;
        curandCreateGenerator(&prngGPU, CURAND_RNG_PSEUDO_MTGP32);
        curandSetStream(prngGPU, stream);
        curandSetPseudoRandomGeneratorSeed(prngGPU, seed);
        // this call is causing Access violation exception:
        curandGenerateUniform(prngGPU, rand, count);
}
$ nvcc -o t1932 t1932.cu -lcurand
$ cuda-memcheck ./t1932
========= CUDA-MEMCHECK
========= ERROR SUMMARY: 0 errors
$

If you’re still having trouble after that, please provide a complete test case. The shortest but complete code that is necessary to see the issue, the operating system you are using, the CUDA version you are using, the GPU you are running on, and the compile command you use.

3

Thanks Robert. Yes, I am wrapping those calls in error check macros of our own, I just wanted to reduce our code to the barebones sample. I am not getting any errors on the way up to curandGenerateUniform.

Any thoughts where the access violation could come from?

Edit: Windows 10, same result on Quadro FX5200 and RTX 3070, cuda toolkit 10.4 and 11.5.

4

Quadro FX5200 isn’t supported by any recent CUDA toolkit.

I’d recommend trying my code verbatim. Did it produce an access violation? If not, we will need the things I asked for.

5

Sorry, I meant Quadro K5200.

It seems that the issue was with the supplied cuda stream. For some reason I thought that the curandSetStream would return an error if the stream was invalid.

Thanks.

6

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.