Pytorch FFT RuntimeError: FFT_INVALID_DESCRIPTOR

[jedcheng]

Hi all

I am working on a scientific solver based on pytorch. The software worked perfectly on Nvidia GPUs and Arm/x86 CPUs. 

However, there was an error related to the FFT function when I ran it on an Intel GPU. 

The error is down to the inverse real FFT call:

x = torch.rand(1000,1000,1,3).to(torch.device("xpu"))

torch.fft.irfftn(x)

which returns the error:

RuntimeError: FFT_INVALID_DESCRIPTOR

[AthiraM_Intel]

Hi,

Thank you for posting in Intel Communities.

Could you please share the below details:

1. OS and Hardware details

2. Torch and IPEX version

Thanks

[AthiraM_Intel]

Hi,

We have not heard back from you. Could you please give us an update?

Thanks

[jedcheng]

My apologies. I received the previous reply while having dinner and then completely forgot the issue.

I did open an issue on the intel GitHub repo.

https://github.com/intel/intel-extension-for-pytorch/issues/479

I ran it on the jupyter dev cloud.

It seems that I can't simply post the log generated by the environment script on the forum directly (it was flagged as html)

Please refer to this reply 

https://github.com/intel/intel-extension-for-pytorch/issues/479#issuecomment-1851202928

[AthiraM_Intel]

Hi,

Thank you for the response. We are checking on this internally.

Thanks

[asin · Accepted solution]

Hi Jed Cheng,

This issue is not reproducible on IPEX version 2.1. Please retry and let me know if the issue persists.

Thanks for confirming on the GitHub issue: https://github.com/intel/intel-extension-for-pytorch/issues/479