CUDA 11.2 cudaGetDeviceCount bug

KlausT · December 17, 2020, 10:55pm

I just installed the CUDA Toolkit 11.2 on my Windows 10 system.
After recompiling some code I noticed that cudaGetDeviceCount always returns error code 1 (invalid argument). Older binaries are working just fine.

Is my installation broken, or can somebody reproduce it?
Here’s some sample code:

github.com

KlausT/CUDAinfo/blob/master/CUDAinfo/CUDAinfo.cpp

#include <cuda_runtime.h>
#include <iostream>

using namespace std;

int main()
{
	int device,deviceCount,version;
	cudaError_t error;
	cudaDeviceProp deviceProp;

	cout << "This program is using the Nvidia CUDA Toolkit " << CUDART_VERSION / 1000 << "." << (CUDART_VERSION % 1000) / 10 << endl << endl;
	error = cudaDriverGetVersion(&version);
	if (error != cudaSuccess)
	{
		cout << cudaGetErrorString(error) << endl;
		return -1;
	}

	int major = version / 1000;

This file has been truncated. show original

imkow · December 19, 2020, 10:26pm

dude, u should use 11.2 with latest NV driver but not the one for CUDA for WSL.