「cudnn8.2」how to use conv3d kernel in cudnn with INT8?

I want to activate conv3d INT8 kernel in cudnn and I tried this way but failed:

hack pytorch backend API to CUDNN cudnnConvolutionForward()

To activate INT8 kernel I changed xDesc yDesc wDesc and convDesc like this

but when I run programme, I got an error :

RuntimeError: no valid convolution algorithms available in CuDNN

CUDA11.2 CuDNN8.2.0 GPU:RTX2080Ti/ RTX8000

So, Is there any support for INT8 Conv3d kernel in CuDNN? OR, How can I achieve my goal by other ways?

Thanks a lot!

Allright I know CUDNN8.2 is not support Conv3d INT8 with _NCHW layout and _NHWC layout

1 Like