ToDtype CV-CUDA Backend

[justincdavis]

Summary

Add the backend kernel for ToDtype transform using CV-CUDA

How to use

import cvcuda
import torchvision.transforms.v2.functional as F

cvc_tensor = cvcuda.Tensor((1, 224, 224, 3), cvcuda.Type.U8, cvcuda.TensorLayout.NHWC)
# Dispatches to F.to_dtype_cvcuda
cvc_fp32_tensor = F.to_dtype(cvc_tensor, torch.float32)

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/vision/9278

• 📄 Preview Python docs built from this PR

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

❗ 1 Active SEVs

There are 1 currently active SEVs. If your PR is affected, please view them below:

• Can't find 'action.yml', 'action.yaml' or 'Dockerfile' under '/home/ec2-user/actions-runner/_work/pytorch/pytorch/.github/actions/check-tpu'

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

[meta-cla]

Hi @justincdavis!

Thank you for your pull request and welcome to our community.

Action Required

In order to merge any pull request (code, docs, etc.), we require contributors to sign our Contributor License Agreement, and we don't seem to have one on file for you.

Process

In order for us to review and merge your suggested changes, please sign at https://code.facebook.com/cla. If you are contributing on behalf of someone else (eg your employer), the individual CLA may not be sufficient and your employer may need to sign the corporate CLA.

Once the CLA is signed, our tooling will perform checks and validations. Afterwards, the pull request will be tagged with CLA signed. The tagging process may take up to 1 hour after signing. Please give it that time before contacting us about it.

If you have received this in error or have any questions, please contact us at cla@meta.com. Thanks!

[AntoineSimoulin]

Hey @justincdavis, thanks a lot for the PR. I left some comments and questions as a first review. Let me know what you think!

[AntoineSimoulin]

@justincdavis could you complete the missing Contributor License Agreement (c.f. earlier comment from the meta-cla bot)?

[AntoineSimoulin]

Hey @justincdavis, thanks for addressing my first round of comments. I had another pass. Will it be possible to have another iteration on the PR based on my new comments? Thanks a lot for your time here!

[zy1git]

Hi,

This is just a light pass of the review. Let me know what you think.

[justincdavis]

Hi @zy1git thanks for the first pass! I have updated this PR to reflect the conventions of the flip PR, LMK what you think!

[NicolasHug]

Thanks a lot for the PR @justincdavis , I left a first pass

[NicolasHug]

Just a not that you should be able to remove these changes once #9305 lands.

[NicolasHug]

Thanks for adding this test!

[NicolasHug]

IIUC, this 255 tol is needed because in torch, when scale is False, we're doing a brutal .to(dtype) which is going to cause a lot of overflows, whereas in CVCUDA you either cap the result or always scale?

I'm hoping we can simplify this a bit, potentially by dropping support for uint8 <-> uint16 conversions when scale is False on CV-CUDA. I feel like that's not a really valid conversion to support anyway. The general idea is that for all transforms, we'll want the CVCUDA backend to have very close results to the existing tensor backend. A difference of 255 is too large.

BTW, we should be able to set atol to 0 or 1 when is_uint16_to_uint8 and scale is True?

[justincdavis]

I am looking into if we can reduce this more and if not I will update to drop support for uint16<->uint8 through cvcuda

[justincdavis]

@NicolasHug I made some changes to the atol calculations and dropped uint16->uint8 with scale=False from the CV-CUDA version. All atol values are <=1 now for all supported use cases. LMK if you want to see more changes/verification from this.