CUDA Error 101 with cudaMemPrefetchAsync Positioning on WSL2

[tinunkai]

I am running these code snippets on WSL using CUDA 12.1.
snip1.txt (1.5 KB)
snip2.txt (1.5 KB)
The only difference between snip1 and snip2 is the position of cudaMemPrefetchAsync. Running snip1 results in a CUDA error 101: “invalid device ordinal,” while snip2 runs without any issues.

What could be the mechanism behind this behavior?

// snip1
  cudaMemPrefetchAsync(noisy, bytes, deviceId);
  cudaMemPrefetchAsync(ising1, bytes, deviceId);
  cudaMemPrefetchAsync(ising2, bytes, deviceId);

  init_curand<<<numBlocks, threadsPerBlock>>>(states, time(NULL));
  checkError();

// snip2
  init_curand<<<numBlocks, threadsPerBlock>>>(states, time(NULL));
  checkError();

  cudaMemPrefetchAsync(noisy, bytes, deviceId);
  cudaMemPrefetchAsync(ising1, bytes, deviceId);
  cudaMemPrefetchAsync(ising2, bytes, deviceId);

[Robert_Crovella]

CUDA Error 101 with cudaMemPrefetchAsync Positioning

cudaMemPrefetchAsync is not supported on windows or WSL2. That is why you are getting the error in snip1. see here. In snip2 you get no errors because your error checking is broken. See here.

[tinunkai]

Thank you!