How do I use cudnn convolutions with cudnn 8.0?

AI & Data Science Deep Learning (Training & Inference) cuDNN
1

Hello, I successfully installed CUDA 11.0 and cudnn 8.0.
I am trying to run a repo that uses the function: cudnnGetConvolutionForwardAlgorithm which looks deprecated. What can I substitute it with? Or do you suggest I downgrade my CUDA and cudnn versions? Thank you

1 Like
2

Please refer “Table 3. API functions and data types that were removed” in below link, cudnnGetConvolutionForwardAlgorithm is removed in cuDNN 8.0
https://docs.nvidia.com/deeplearning/sdk/cudnn-api/index.html#release-800-preview

You can use cudnnGetConvolutionForwardAlgorithm_v7 API Reference - NVIDIA Docs

For an exhaustive search for the fastest algorithm, use cudnnFindConvolutionForwardAlgorithm()

Thanks

3

Hello, thank you for the prompt reply. I am a beginner with this library. Is there a simple way to change/cast the parameters to fit the _v7 function, given the “original” function’s parameters?

4

v7 api gives you multiple choices that you can pick from, sorted by estimated speed.

If you refer to API, most of the params remains same.
https://docs.nvidia.com/deeplearning/sdk/cudnn-archived/cudnn_765/cudnn-api/index.html#cudnnGetConvolutionForwardAlgorithm

The output of cudnnGetConvolutionForwardAlgorithm_v7 is “cudnnConvolutionFwdAlgoPerf_t” which includes the “cudnnConvolutionFwdAlgo_t” as subset along with additional info.

You can also refer to the mnist sample where we have updated to use this v7 api.

Thanks

5

Hi, I had a followup question about this. I am also using CUDA 11.0 and CuDNN 8.0.

I notice that cudnnGetForwardAlgorithm() allows you to pass in a cudnnConvolutionFwdPreference_t and a memory limit. Previously, I had been able to pass CUDNN_CONVOLUTION_FWD_SPECIFY_WORKSPACE_LIMIT and a workspace limit of 0 to select a convolution algorithm that required no workspace.

Is there a way to select a no-workspace conv algorithm with the v7 api ( cudnnGetForwardAlgorithm_v7() ) ?