Sox transforms work on CUDA tensors?

[turian]

❓ Questions and Help

Do Sox transforms work on CUDA tensors? Or do they transfer the tensor to CPU and process there?

[mthrok]

Hi @turian

If what you mean by "Sox transforms" is the filtering functions implemented in torchaudio.functional, (found under the category of filtering in the official doc), then they work with CUDA. You can pass Tensors transferred to CUDA and filtering functions will perform the operation in CUDA. (however this does not necessarily faster than CPU, because torchaudio does not have efficient CUDA-specific implementation)

If you mean functions under torchaudio.sox_effects, then no. They only support CPU and functions will raise an error.

Either way, functions in torchaudio do not automatically transfer Tensors across devices. It's up to users to move Tensor to an appropriate device.

[turian]

Thanks, my question was about torchaudio.sox_effects. Perhaps this could be added to the docs, just to be clear?

[mthrok]

Hi @turian

Thanks for the suggestion. I agree. this can be better explained in the doc.
I filed issue #1456. If no one picks up, I will work on it before the next release.

[turian]

Thank you

[rien333]

Any chance if sox effects ever working with CUDA, or is that unrealistic?

[mthrok]

Hi @rien333

It's not impossible but it's a huge amount of work, and our engineering resource is very limited at the moment. So, yeah unfortunately it's unrealistic.

Right now, we have some filters re-implemented with PyTorch's generic Tensor operations. They are compatible with CUDA, but when tensors are moved to CUDA, operations become slower than CPU. Efficient filtering in CUDA requires implementations appropriate for CUDA style parallel computation.

[turian]

Thank you @mthrok for your wonderful work. Shinji speaks very highly of your project to me.

[mthrok]

@turian

Thanks for the such nice comments. I try to keep the bar high, but libraries cannot exist without users, so let's make something interesting together. :)