Support rotation on beta cuda

[mollyxu]

Add rotation support to BetaCudaDeviceInterface using torch::rot90

Parametrize over cpu and cuda:beta in test_rotation_applied_to_frames (note: since we are planning to move to beta cuda eventually, we've decided to currently not support this on cuda with ffmpeg backend)

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/meta-pytorch/torchcodec/1235

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

✅ No Failures

As of commit 298c9ef with merge base 2d1b5c6 ():
💚 Looks good so far! There are no failures yet. 💚

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[Dan-Flores]

The code structure makes it a bit hard to follow what preAllocatedOutputTensor is storing, perhaps we could change it to be more explicit when changes to the preallocated tensor are needed:

if rotation_ == Rotation::NONE:
  convertNV12FrameToRGB() # regular call
else:
  convertNV12FrameToRGB() # call without preallocated tensor
  applyRotation()

Ideally we could handle the rotation switch statement and call to torch::rot90 in a separate function.

There might be nuances to the logic I am missing that make this structure difficult, please double check me on this.

[NicolasHug]

Agreed it might help to have an applyRotation() helper.
I'm not sure about having convertNV12FrameToRGB() within if/else blocks, it might be tricky since there is also validatePreAllocatedTensorShape which depends on preAllocatedOutputTensor having the correct size (i.e. we'd still need the savedPreAllocatedOutputTensor hack above, I think).

[Dan-Flores]

I'm still a little hesitant on this approach - when the hack sets preAllocatedOutputTensor = std::nullopt, then validatePreAllocatedTensorShape does not actually do any validation. Excluding the validation call in the else block would be more explicit about that.

This is minor though, and could be addressed when we implement the less hacky solution which I believe will use the NPP functions?

[mollyxu]

Thanks for the discussion! I agree that it is clearer to be explicit in excluding the validation call and to reduce the number of variables

[NicolasHug]

Thanks @mollyxu , this LGTM, I'll let @Dan-Flores make another pass as well

[Dan-Flores]

Thanks @mollyxu!