Manually set cudnn convolution algorithm

vision
21

Thank you!
Just for clarification: python example.py creates this log and fails with CUDNN_STATUS_NOT_SUPPORTED, right?

22

Yes! That’s correct.

BTW:

  • If C=512, it breaks at the cudnnGetConvolutionForwardWorkspaceSize function
Allocating Workspace
Error on line 87: CUDNN_STATUS_NOT_SUPPORTED
  • If C=256, it is able to compute and even allocate the memory, but it breaks at the cudnnConvolutionForward function.
Allocating Workspace
Workspace size: 6656884736 Bytes
Allocated: 6656884736 Bytes
Error on line 113: CUDNN_STATUS_BAD_PARAM
  • If C=128, it works just fine.

I am trying to get another machine to check if the issue is caused by not having enough GDDR. I have 8192MiB.

1 Like