cudaErrorLaunchFailure when running MD5 5000 times

Question

I have a CUDA code like:

__global__ void test() 
{
    for (unsigned int loop = 1; loop < MAX_LOOP; loop++)
    {
        unsigned char hash[32];
        MD5Ctx ctx;

        md5_init(&ctx);
        md5_update(&ctx, "123", 3);
        md_final(&ctx, hash);
    }
}

Then I launch the kernel with only one thread like this:

cudaError_t status;
test<<<1, 1>>>();
cudaThreadSynchronize();
status = cudaGetLastError();

If I set MAX_LOOP to 3000, everything runs successfully. When I change MAX_LOOP to 5000 or even larger, cudaGetLastError will return back with the error: cudaErrorLaunchFailure(4).