[CUDA] RoiAlign for opset versions 16 and 22

[tianleiwu]

Support RoiAlign for opset versions 16 and 22

[github-actions]

You can commit the suggested changes from lintrunner.

[Copilot]

Pull request overview

Adds CUDA kernel support for RoiAlign across ONNX opset versions 16 and 22, including additional datatype coverage.

Changes:

• Register CUDA RoiAlign kernels for opset ranges 10–15, 16–21, and opset 22.
• Extend CUDA implementation to support MLFloat16 and BFloat16 (with accumulation types).
• Add CUDA tests covering Float16 (opset 16/22) and BFloat16 (opset 22), and update kernel documentation.

Reviewed changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated 4 comments.

Show a summary per file

File	Description
onnxruntime/test/providers/cpu/object_detection/roialign_test.cc	Adds CUDA EP tests for Float16/BFloat16 across opset 16 and 22.
onnxruntime/core/providers/cuda/object_detection/roialign_impl.cu	Improves numeric handling via accumulation types; enables half/BFloat16 specializations.
onnxruntime/core/providers/cuda/object_detection/roialign.cc	Adds versioned kernel registrations for opset 10–15, 16–21, and opset 22; enables MLFloat16/BFloat16 registrations.
onnxruntime/core/providers/cuda/cuda_execution_provider.cc	Adds kernel class declarations and registration entries for new RoiAlign variants/opsets.
docs/OperatorKernels.md	Updates documented opset/type support matrix for RoiAlign.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

You can also share your feedback on Copilot code review. Take the survey.

[yuslepukhin]

There is also an issue in the header which is a pre-existing issue but might as well be fixed now.

Default coordinate_transformation_mode for opset 16+:

roialign.h:129-131 — The attribute is only read inside the if (GetAttr(...).IsOK()) guard. If the attribute is absent (which is valid per the spec), the if body is skipped and half_pixel_ stays at its roialign.h:143. But the ONNX spec for opset 16+ says the default for coordinate_transformation_mode is "half_pixel", which should set half_pixel_ = true. So when the attribute is omitted, the kernel silently behaves as "output_half_pixel" instead of the spec-mandated "half_pixel".

Note: for opset 10 (which has no coordinate_transformation_mode attribute), the false default is correct — it matches the original opset 10 behavior. #Resolved

[yuslepukhin]

🕐

[Copilot]

Pull request overview

Adds CUDA Execution Provider support for RoiAlign across multiple ONNX operator versions, extending datatype coverage to FP16/BF16 where supported by the corresponding opset versions.

Changes:

• Register CUDA RoiAlign kernels for opset ranges 10–15 and 16–21, and opset 22 (including MLFloat16/BFloat16 where applicable).
• Update CUDA RoiAlign implementation to accumulate using a wider accumulation type for FP16/BF16 paths.
• Add CUDA-focused unit tests for opset 16 and 22 (FP16/BF16) and adjust default coordinate_transformation_mode behavior for opset 16+.

Reviewed changes

Copilot reviewed 6 out of 6 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
onnxruntime/test/providers/cpu/object_detection/roialign_test.cc	Adds CUDA EP tests for opset 16/22 with FP16/BF16 and additional coordinate mode/sampling coverage.
onnxruntime/core/providers/cuda/object_detection/roialign_impl.cu	Uses AccumulationType_t for safer FP16/BF16 accumulation; adds explicit instantiations for half/BFloat16.
onnxruntime/core/providers/cuda/object_detection/roialign.cc	Updates CUDA kernel registrations to be versioned for opsets 10–15 and 16–21, plus opset 22 typed registrations.
onnxruntime/core/providers/cuda/cuda_execution_provider.cc	Wires new RoiAlign kernel registrations into CUDA EP kernel registry and class declarations.
onnxruntime/core/providers/cpu/object_detection/roialign.h	Sets default half_pixel_ behavior for opset 16+ when the attribute is absent, aligning with spec defaulting.
docs/OperatorKernels.md	Updates generated kernel support table entries for CUDA RoiAlign across opset ranges and types.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[yuslepukhin]

roialign.h:129-131 — The attribute is only read inside the if (GetAttr(...).IsOK()) guard.

If the attribute is absent (which is valid per the spec), the if body is skipped and half_pixel_ stays at its roialign.h:143.

But the ONNX spec for opset 16+ says the default for coordinate_transformation_mode is "half_pixel", which should set half_pixel_ = true. So when the attribute is omitted, the kernel silently behaves as "output_half_pixel" instead of the spec-mandated "half_pixel".

[tianleiwu]

roialign.h:129-131 — The attribute is only read inside the if (GetAttr(...).IsOK()) guard.

If the attribute is absent (which is valid per the spec), the if body is skipped and half_pixel_ stays at its roialign.h:143.

But the ONNX spec for opset 16+ says the default for coordinate_transformation_mode is "half_pixel", which should set half_pixel_ = true. So when the attribute is omitted, the kernel silently behaves as "output_half_pixel" instead of the spec-mandated "half_pixel".

Current logic:

If there is coordinate_transformation_mode attribute
    set half_pixel_ according to whether the attribute value is "half_pixel". 
else
    use the default value according to spec. The default is "half_pixel" for opset 16+, and "output_half_pixel" for opset 10.
``` #Resolved

[yuslepukhin]

[yuslepukhin]