Arm backend: Support for aten.slice_scatter and slice_copy with non-unit step

[robell]

Arm backend: Add support for aten.slice_copy with non-unit step

• Decompose strided slice_copy into unit-step slice_copy plus optional
  right padding and view_copy reshapes
• Update SliceCopySupported check for the supported pattern
• Add non-unit-step slice tests

Arm backend: Add support for aten.slice_scatter.default

• Decompose edge.aten.slice_scatter into fast and general paths
• Fast path (step==1): slice_copy + cat on the updated dimension
• General path (step>1): arange + index_put with permute
• Fix slice_copy lowering to accept the 5-arg form

Change-Id: Ida60ee2f42c283d50c9e3185dca1f9ea2238cf83
Change-Id: Ic21e995bee8f722f6d03ec3cc3b04747186a8a9b
Co-authored-by: Rob Elliott Robert.Elliott@arm.com
Signed-off-by: Yufeng Shi yufeng.shi@arm.com
Signed-off-by: Rob Elliott Robert.Elliott@arm.com

cc @freddan80 @per @zingo @oscarandersson8218 @digantdesai

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/17413

• 📄 Preview Python docs built from this PR

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

❌ 5 New Failures, 6 Cancelled Jobs

As of commit 58faa08 with merge base b871398 ():

NEW FAILURES - The following jobs have failed:

• pull / test-samsung-models-linux / linux-job (gh)
  RuntimeError: Command docker exec -t 665f40d5db0ca37ee3700726bb5f656076678a9a6969c64e2aaec33a1070e94d /exec failed with exit code 1
• pull / test-samsung-quantmodels-linux / linux-job (gh)
  RuntimeError: Command docker exec -t 3cfd5ce5a3c63afb9c3461efdc76b364583213f6f1645de0eabc674c8635a3ee /exec failed with exit code 1
• pull / unittest / macos / macos-job (gh)
  export/tests/test_target_recipes.py::TestTargetRecipes::test_mv3_model
• trunk / test-arm-backend-ethos-u (test_pytest_ops_ethos_u55) / linux-job (gh)
  RuntimeError: Command docker exec -t d44a93a8351e24e17b47f2b76912bca0ef594ff40494b47e66f3c4410dc1c62c /exec failed with exit code 1
• trunk / test-mcu-cortex-m-backend / linux-job (gh)
  RuntimeError: Command docker exec -t e8be1c82248787e29b8361561e5fff3d63cb487014532924ad3694931062b0d0 /exec failed with exit code 1

CANCELLED JOBS - The following jobs were cancelled. Please retry:

• periodic / test-models-linux (buck2, mv3, portable, linux.2xlarge, 90) / linux-job (gh)
  ##[error]The operation was canceled.
• periodic / test-models-linux (buck2, mv3, xnnpack-quantization-delegation, linux.2xlarge, 90) / linux-job (gh)
  ##[error]The operation was canceled.
• periodic / test-models-linux (cmake, mv3, portable, linux.2xlarge, 90) / linux-job (gh)
  ##[error]The operation was canceled.
• periodic / test-models-linux (cmake, mv3, xnnpack-quantization-delegation, linux.2xlarge, 90) / linux-job (gh)
  ##[error]The operation was canceled.
• periodic / test-models-linux (cmake, vit, portable, linux.2xlarge, 90) / linux-job (gh)
  ##[error]The operation was canceled.
• periodic / test-models-linux (cmake, vit, xnnpack-quantization-delegation, linux.2xlarge, 90) / linux-job (gh)
  ##[error]The operation was canceled.

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

[Copilot]

Pull request overview

Adds Arm backend support for aten.slice_copy with non-unit step and aten.slice_scatter.default by introducing decomposition passes and expanding test coverage across TOSA/Ethos/VGF pipelines.

Changes:

• Add decomposition passes for strided slice_copy and for slice_scatter (fast path for step==1, general path for step>1).
• Update slice_copy lowering/support checks to handle the 5-arg form and validate dtype/profile constraints.
• Add new/expanded tests for non-unit-step slice and slice_scatter across backends.

Reviewed changes

Copilot reviewed 10 out of 10 changed files in this pull request and generated 6 comments.

Show a summary per file

File	Description
backends/arm/test/ops/test_slice_scatter.py	Adds slice_scatter tests for unit and non-unit step across TOSA/Ethos/VGF.
backends/arm/test/ops/test_slice.py	Adds non-unit-step slice tests across TOSA/Ethos/VGF.
backends/arm/test/models/stable_diffusion/test_CLIPTextModelWithProjection.py	Marks a known-failing TOSA FP test as xfail.
backends/arm/operators/op_slice.py	Updates slice_copy lowering to accept 5-arg form and enforce step==1 at lowering.
backends/arm/operator_support/tosa_profile_supported_op_lists.py	Adds aten.slice_scatter.default to supported ops list.
backends/arm/operator_support/slice_copy_support.py	Expands slice_copy support checks (arg count, positive step, dtype/profile validation).
backends/arm/_passes/decompose_strided_slice_copy_pass.py	Introduces decomposition for non-unit-step slice_copy into unit-step ops.
backends/arm/_passes/decompose_slice_scatter_pass.py	Introduces decomposition for slice_scatter into fast/general lowering forms.
backends/arm/_passes/arm_pass_manager.py	Wires new decomposition passes into TOSA and annotation pipelines.
backends/arm/_passes/init.py	Exports the newly added passes.

---

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

[Copilot]

arange is always emitted even when step == 1 (fast path), and the resulting index_positions is unused in that branch. If dead-code elimination doesn’t reliably remove it before partitioning/lowering, this can introduce an unnecessary (and potentially unsupported) op in the step==1 path. Move index_positions = arange(...) (and any dependent shape checks) inside the step > 1 branch; for step == 1, compute the expected slice length directly from start_i/end_i for validation.

Suggested change

	# index_positions: [W] where W = len(arange(start_i, end_i, step))
	index_positions = super().call_operator(
	arange_op,
	(start_i, end_i, step),
	{"dtype": torch.int32, "device": input_device},
	meta,
	updated=True,
	)
	src_val = src.data
	src_shape = src_val.shape
	index_shape = index_positions.data.shape
	# slice_shape is input_shape with dim_norm replaced by W
	# input_shape: [d0, ..., D, ..., d{r-1}]
	# -> slice_shape: [d0, ..., W, ..., d{r-1}]
	slice_shape = list(input_shape)
	slice_shape[dim_norm] = int(index_shape[0])
	src_val = src.data
	src_shape = src_val.shape
	if step == 1:
	# Fast path: contiguous slice, avoid materializing index_positions via arange.
	# slice_shape is input_shape with dim_norm replaced by W = end_i - start_i.
	# input_shape: [d0, ..., D, ..., d{r-1}]
	# -> slice_shape: [d0, ..., W, ..., d{r-1}]
	slice_shape = list(input_shape)
	slice_shape[dim_norm] = int(end_i - start_i)
	else:
	# index_positions: [W] where W = len(arange(start_i, end_i, step))
	index_positions = super().call_operator(
	arange_op,
	(start_i, end_i, step),
	{"dtype": torch.int32, "device": input_device},
	meta,
	updated=True,
	)
	index_shape = index_positions.data.shape
	# slice_shape is input_shape with dim_norm replaced by W
	# input_shape: [d0, ..., D, ..., d{r-1}]
	# -> slice_shape: [d0, ..., W, ..., d{r-1}]
	slice_shape = list(input_shape)
	slice_shape[dim_norm] = int(index_shape[0])

[Copilot]

arange is always emitted even when step == 1 (fast path), and the resulting index_positions is unused in that branch. If dead-code elimination doesn’t reliably remove it before partitioning/lowering, this can introduce an unnecessary (and potentially unsupported) op in the step==1 path. Move index_positions = arange(...) (and any dependent shape checks) inside the step > 1 branch; for step == 1, compute the expected slice length directly from start_i/end_i for validation.

[Copilot]

Avoid using assert for input/graph validation in production passes since assertions can be stripped with optimization flags and produce less actionable failures. Prefer raising NotImplementedError or RuntimeError with a message consistent with other pass validations (e.g., the way DecomposeSliceScatterPass handles step <= 0).

Suggested change

	assert step > 0, "slice_copy step must be positive"
	if step <= 0:
	raise RuntimeError(f"slice_copy step must be positive, got {step}")