I am using the libtorch C++ API. How can I programatically disable CUDNN for testing purposes?
I think this should work:
I am doing a quick test with
torch::conv2d on a simple RGB image, but whether i enable or disable CUDNN, the runtime seems to be the same.
torch::cuda::cudnn_is_available() is true
No CUDNN is faster, just needed to run more iterations and take an average
cudnn_is_available() would not return
False, if you disable the usage of cuDNN as in fact it’s still available.
You can profile the code via e.g. Nsight Systems or the PyTorch profiler to see the kernels.