Port deform_conv2d (CUDA) to stable ABI.

[adabeyta]

Ports deform_conv2d's CUDA/ROCm kernels to the stable ABI. This is the last CUDA kernel, and the last operator, left in the legacy _C extension.

Notes

• Stable-ABI is missing aten::add.Tensor, so the bias add goes through subtract(self, other, -1.0). Follows the CPU (Port deform_conv2d (CPU) to stable ABI. #9610) and MPS (Port deform_conv2d (MPS) to stable ABI.  #9617) ports.

• Device setup (guard, current stream and max grid size for GET_BLOCKS) move off of at::cuda onto the stable ABI and cudaDeviceGetAttribute call is used, since the stable ABI has no getCurrentDeviceProperties().

Stable-ABI audit (torch-abi-audit)

Ran torch-abi-audit. deform_conv2d's CUDA op moves into the fully-STABLE _C_stable.so (unstable=0). The unstable _C.so count drops, 85 → 66: deform_conv2d was the last CUDA consumer of ATen tensor algebra.

    Package: torchvision
      Torch ABI:   UNSTABLE
      CPython ABI: n/a
      Bundled libs: 3
      -- bundled libs --
        [UNSTABLE] [not-abi3] _C.so            (stable_shim=0,  unstable=66)
        [STABLE  ] [not-abi3] _C_stable.so     (stable_shim=74, unstable=0)
        [STABLE  ] [not-abi3] image_stable.so  (stable_shim=78, unstable=0)

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/vision/9626

• 📄 Preview Python docs built from this PR

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

❌ 6 Pending, 1 Unclassified Failure

As of commit 203951e with merge base 2f411e7 ():

UNCLASSIFIED FAILURE - DrCI could not classify the following job because the workflow did not run on the merge base. The failure may be pre-existing on trunk or introduced by this PR:

• .github/workflows/pr-labels.yml (gh) (this job did not run on the merge base, so DrCI cannot tell whether the failure is pre-existing)

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

[NicolasHug]

Thanks @adabeyta !