Arm backend: Add aten.flip support via TOSA REVERSE

[vacu9708]

Summary

TOSA REVERSE reverses a single axis, while aten.flip can take several. A multi-axis flip is a chain of single-axis reversals, so:

• DecomposeFlipPass rewrites a multi-axis flip into single-axis flips (one dim passes through, zero dims is a no-op).
  • The quantizer annotates it as a shared-qspec op. The decompose runs after Q/DQ folding, so the chained flips inherit the shared quant params from the node metadata.
• FlipVisitor lowers each single-axis flip to one REVERSE (axis = dim % rank).

graph TD
    src["aten.flip(x, [d0, d1, d2])"]
    src -->|DecomposeFlipPass| F0

    subgraph flips["FX graph"]
        direction LR
        F0["aten.flip [d0]"] --> F1["aten.flip [d1]"] --> F2["aten.flip [d2]"]
    end

    subgraph reverses["Serialized TOSA"]
        direction LR
        R0["REVERSE axis=d0"] --> R1["REVERSE axis=d1"] --> R2["REVERSE axis=d2"]
    end

    F0 -.FlipVisitor.-> R0
    F1 -.FlipVisitor.-> R1
    F2 -.FlipVisitor.-> R2

Loading

Changes

I referred to existing ops for each new file: op_permute.py for op_flip.py, decompose_strided_slice_copy_pass for decompose_flip_pass.py, and test_permute.py for test_flip.py.

File	Change
_passes/decompose_flip_pass.py (new)	DecomposeFlipPass
operators/op_flip.py (new)	FlipVisitor
_passes/__init__.py	export DecomposeFlipPass
_passes/arm_pass_manager.py	register DecomposeFlipPass in the post-fold decompose stage
operators/__init__.py	register op_flip
operator_support/tosa_profile_supported_op_lists.py	add aten.flip.default to the INT and FP support lists
test/ops/test_flip.py (new)	Test cases

Testing

$ pytest backends/arm/test/ops/test_flip.py -q
33 passed, 14 skipped, 7 xfailed

$ lintrunner   # new and changed files
ok No lint issues.

Test	Dtype / target	Ranks / dims	What it verifies	Result
tosa_FP	FP32	1–4; single, negative, multi-axis	output matches torch.flip	pass
tosa_FP_reverse_count	FP32	1–4; single, negative, multi-axis	one TOSA REVERSE per flipped dim	pass
tosa_FP_bf16	bf16	rank 4; multi-axis	output matches torch.flip	pass
tosa_FP_fp8	fp8 e4m3 / e5m2	rank 4; multi-axis	output matches torch.flip	pass
bool_tosa_FP	bool	rank 2; single-axis	output matches torch.flip on a bool mask	pass
tosa_INT	INT8	1–4; single, negative, multi-axis	output matches torch.flip	pass
16a8w_tosa_INT	INT16 (16a8w)	1–4; single, negative, multi-axis	output matches torch.flip	pass
u55_INT_not_delegated	INT8 / U55	rank 4; single-axis	flip is rejected and runs on CPU (U55 has no REVERSE)	pass
u85_INT	INT8 / Ethos-U85	1–4; single, negative, multi-axis	runs on Ethos-U85	xfail locally (no FVP); runs in CI
vgf_no_quant, vgf_quant	FP & INT / VGF	1–4; single, negative, multi-axis	runs on the VGF backend	skip locally (no converter); runs in CI

cc @digantdesai @freddan80 @per @zingo @oscarandersson8218 @mansnils @Sebastian-Larsson @robell @rascani

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/20592

• 📄 Preview Python docs built from this PR

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

❌ 1 New Failure, 32 Pending

As of commit 6882925 with merge base d03dee5 ():

NEW FAILURE - The following job has failed:

• trunk / test-qnn-optimum-model (fp32, mobilevit_v1) / linux-job (gh)
  RuntimeError: Command docker exec -t 0276795c34f40508ff217236b3e045698278cdc129d0233a2ef0f0b244664322 /exec failed with exit code 92

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

[pytorch-bot]

Workflows were awaiting approval. CI has now been triggered for the ciflow labels on this PR.

[github-actions]

This PR needs a release notes: label

If your change should be included in the release notes (i.e. would users of this library care about this change?), please use a label starting with release notes:. This helps us keep track and include your important work in the next release notes.

To add a label, you can comment to pytorchbot, for example
@pytorchbot label "release notes: none"

For more information, see
https://github.com/pytorch/pytorch/wiki/PyTorch-AutoLabel-Bot#why-categorize-for-release-notes-and-how-does-it-work.

[zingo]

Would be nice with coverage for dims=[]

[zingo]

Thanks for the nice patch!
I started some extra testing that for some reason didn't get started automatically :).

[zingo]

@claude does the new files in this PR need to be added to any buck2 files?

[claude]

Claude finished @zingo's task in 1m 19s —— View job

---

Do the new files need buck2 (TARGETS) updates?

Short answer: only the test file. The two new source files are already covered by globs, but test/ops/test_flip.py must be added explicitly.

Details:

New file	buck2 target	Needs manual entry?
_passes/decompose_flip_pass.py	_passes/TARGETS → passes uses glob(["*.py"], ...)	❌ No — picked up automatically
operators/op_flip.py	operators/TARGETS → ops uses glob(["op_*.py", "ops_*.py"])	❌ No — matches op_*.py
test/ops/test_flip.py	test/targets.bzl → define_arm_tests uses an explicit list (ops/test_*.py are not globbed)	✅ Yes

The modified files (_passes/__init__.py, arm_pass_manager.py, operators/__init__.py, operator_support/tosa_profile_supported_op_lists.py) are all covered by existing globs, so no target changes there.

Required change

In backends/arm/test/targets.bzl, add test_flip.py to the operator test list (kept alphabetical-ish next to the other ops):

         "ops/test_exp.py",
         "ops/test_fft.py",
+        "ops/test_flip.py",

Note the passes tests are globbed (test_files += native.glob(["passes/test_*.py"])), so if there were a passes/test_decompose_flip*.py it would be automatic — but the op test list is explicit and needs this line.

Note

I made this change locally but couldn't push it — the branch's Meta CLA Check repository rule rejects pushes from this bot. Please add the one line above yourself (or let me know how you'd like to proceed).
· branch flip

[zingo]

Thanks this will be Ok to merge after the backends/arm/test/targets.bzl is updated (see comments) and extra nice if you also add a test case but that can be fixed later also :)

[zingo]

This should probably be added to backends/arm/test/targets.bzl see claude comment for more info :)

[zingo]

Super thanks!
I approved it but lets the tests run just to make sure then you can merge. If you can't let me know and Ill merge this for you.

[zingo]

All test passer the test-qnn-.. . Is unrelated and fails in other PRs also.

[vacu9708]

Thanks for the quick review and for kicking off the extra CI, @zingo!

Both review points have been addressed.

• Added ops/test_flip.py to backends/arm/test/targets.bzl.
• Added dims=[] coverage (test_flip_empty_dims_not_delegated). An empty flip is the identity, so the partitioner treats it as a no-op (_is_noop_flip) and keeps it on CPU instead of delegating.

I don't have merge access, can you please merge it if there are no more issues?