Multiple devices: Should you do cudaSetDevice() before every cuBLAS call? says:

However, it doesn’t say whether the thread should make sure that the current device ID is correctly set before every cuBLAS library call. Should it?

I find that if it doesn’t, cuBLAS returns error code 13, and works fine otherwise.

If there was an earlier discussion about this, I can’t find it.

Never mind (bug in my own code)