Intel GPU support (xpu)

[ssheorey]

• Include parity tests between CPU and GPU (CUDA or XPU) backends.
• Download and use the latest main-devel Open3D wheel instead of building from source.
• pointtransformer fully supports all 3 backends (cpu, cuda, xpu)

[Copilot]

Pull request overview

This PR adds Intel GPU (XPU/SYCL) support across the Open3D-ML PyTorch stack and introduces CPU↔accelerator parity tests to validate correctness across CPU, CUDA, and XPU backends. It also updates CI to test against prebuilt Open3D “main-devel” wheels instead of building Open3D from source.

Changes:

• Adds PyTorch CPU vs CUDA/XPU parity test helpers and a dedicated parity-focused torch model test suite.
• Extends device selection and custom-op gating to support xpu/sycl across pipelines, models, and utilities.
• Updates CI and dependency requirements to use prebuilt Open3D wheels and backend-specific PyTorch installs (CPU/CUDA/XPU).

Reviewed changes

Copilot reviewed 31 out of 31 changed files in this pull request and generated 4 comments.

Show a summary per file

File	Description
tests/torch_backend_parity.py	Adds shared CPU↔CUDA/XPU parity assertions for outputs and gradients.
tests/test_models_torch.py	New parity + smoke tests for registered PyTorch models across CPU/CUDA/XPU.
tests/test_models_tf.py	Removes legacy torch tests from TF test module; improves repo-root path handling.
tests/run_tests.sh	Runs each pytest test case in a separate process to avoid XPU/SYCL state issues.
scripts/run_pipeline.py	Expands device help/handling and adjusts distributed worker device string formatting.
requirements-torch.txt	Bumps CPU PyTorch/torchvision requirements.
requirements-torch-xpu.txt	Adds XPU-specific PyTorch/torchvision requirements.
requirements-torch-cuda.txt	Bumps CUDA PyTorch/torchvision requirements.
README.md	Documents installing XPU PyTorch requirements and links parity testing docs.
ml3d/utils/tensorboard_o3d.py	Adds helper to conditionally register Open3D TensorBoard plugin on GUI builds.
ml3d/utils/open3d_ops.py	Centralizes “PyTorch ops built” checks and a consistent error for missing ops.
ml3d/utils/builder.py	Extends device name conversion to include xpu/sycl.
ml3d/torch/utils/torch_utils.py	Adds default_training_device() to pick cuda/xpu/cpu consistently.
ml3d/torch/utils/roipool3d/roipool3d_utils.py	Switches custom-op import/guards to BUILD_PYTORCH_OPS-based gating (XPU-friendly).
ml3d/torch/utils/pointnet/pointnet2_utils.py	Switches custom-op import/guards to BUILD_PYTORCH_OPS-based gating (XPU-friendly).
ml3d/torch/utils/objdet_helper.py	Updates docstrings to explicitly include xpu device usage.
ml3d/torch/pipelines/semantic_segmentation.py	Uses conditional TensorBoard plugin registration; updates device defaulting to be accelerator-aware.
ml3d/torch/pipelines/object_detection.py	Uses conditional TensorBoard plugin registration; improves DDP device_ids handling.
ml3d/torch/pipelines/base_pipeline.py	Adds default device selection and xpu handling; adjusts distributed device setup logic.
ml3d/torch/models/sparseconvnet.py	Defaults model device based on available accelerator (cuda/xpu/cpu).
ml3d/torch/models/randlanet.py	Updates loss docstring to include xpu.
ml3d/torch/models/pvcnn.py	Defaults model device based on available accelerator; gates custom ops via BUILD_PYTORCH_OPS.
ml3d/torch/models/point_transformer.py	Fixes KNN device handling to preserve original device (cuda/xpu/cpu).
ml3d/torch/models/point_rcnn.py	Defaults model device based on available accelerator (cuda/xpu/cpu).
ml3d/torch/models/point_pillars.py	Defaults model device based on available accelerator (cuda/xpu/cpu).
ml3d/torch/models/base_model.py	Updates loss docstring to include xpu.
ml3d/torch/dataloaders/concat_batcher.py	Updates docstring to include xpu.
ml3d/metrics/init.py	Adds SYCL iou backend selection path.
docs/howtos.md	Adds documentation on CPU/CUDA/XPU parity testing and CI limitations.
ci/run_ci.sh	Reworks CI to download/install Open3D wheels per backend and run appropriate test subsets.
.github/workflows/ubuntu.yml	Updates GitHub Actions workflow to use the new CPU-wheel-based CI path.

Suppressed comments (2)

ml3d/torch/pipelines/semantic_segmentation.py:663

• ensure_tensorboard_plugin() can return False on headless Open3D builds (BUILD_GUI=OFF), but writer.add_3d(...) is still called unconditionally. That will raise AttributeError if the plugin couldn't be imported/registered. Consider guarding the add_3d call (e.g., if ensure_tensorboard_plugin(): ...) or skipping 3D summaries when the plugin is unavailable.

        for stage in self.summary:
            ensure_tensorboard_plugin()
            for key, summary_dict in self.summary[stage].items():
                label_to_names = summary_dict.pop('label_to_names', None)
                writer.add_3d('/'.join((stage, key)),
                              summary_dict,
                              epoch,
                              max_outputs=0,
                              label_to_names=label_to_names)

ml3d/torch/pipelines/object_detection.py:534

• Same issue as semantic_segmentation: ensure_tensorboard_plugin() return value is ignored, but writer.add_3d(...) is called unconditionally. On BUILD_GUI=OFF wheels this will likely fail at runtime when saving logs.

        for stage in self.summary.keys():
            ensure_tensorboard_plugin()
            for key, summary_dict in self.summary[stage].items():
                label_to_names = summary_dict.pop('label_to_names', None)
                writer.add_3d('/'.join((stage, key)),
                              summary_dict,
                              epoch,
                              max_outputs=0,
                              label_to_names=label_to_names)

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.