[zingo]

Summary

Only replace view_copy with a shared-storage view when its input is
contiguous. Retain mutation safety checks and copy channels-last
inputs in logical order in the portable kernel, keeping the existing
memcpy fast path for contiguous inputs.

Test plan

Cover layout and view-removal configurations, safe contiguous views,
and the Cortex-M patch-token regression.

Authored with assistance from OpenAI Codex.

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

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/24152

• 📄 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, 3 Cancelled Jobs, 6 Unrelated Failures

As of commit ffb9430 with merge base 05092c1 ():

NEW FAILURE - The following job has failed:

• trunk / test-multimodal-macos (gemma3-4b) / macos-job (gh)
  RuntimeError: Command bash /Users/runner/work/_temp/exec_script failed with exit code 1

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

• pull / test-qnn-python-imports-linux / linux-job (gh)
  Could not load credentials from any providers
• pull / test-static-llama-qnn-linux (stories_110m) / linux-job (gh)
  Could not load credentials from any providers
• pull / test-static-llama-qnn-linux (stories_260k_bc) / linux-job (gh)
  Could not load credentials from any providers

BROKEN TRUNK - The following jobs failed but were present on the merge base:

👉 Rebase onto the `viable/strict` branch to avoid these failures

• Build Presets / apple (ios-simulator) / build (gh) (trunk failure)
• Build Presets / apple (ios) / build (gh) (trunk failure)
• Build Presets / apple (llm) / build (gh) (trunk failure)
• Build Presets / apple (macos) / build (gh) (trunk failure)
• Build Presets / apple (profiling) / build (gh) (trunk failure)
• Build Presets / apple (pybind) / build (gh) (trunk failure)

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

[rascani]

I’d prefer to enforce this in view lowering instead of adding another correctness requirement behind _skip_dim_order. Whether view_copy can become memory.view depends on the input layout, regardless of which operator representation we use.

Concretely, I suggest:

1. Make layout compatibility a precondition of ReplaceViewCopyWithViewPass, alongside its mutation checks. Keep view_copy when aliasing would change element order.
2. Update portable view_copy_out to handle channels-last inputs by copying in logical element order, retaining the existing memcpy fast path for contiguous inputs. Keeping the copy alone is insufficient because the current kernel rejects those layouts.
3. Remove LegalizeViewCopyPass and its _skip_dim_order gate. The allocation-metadata cleanup can remain independently.

For validation, parameterize the channels-last flatten runtime test over both _skip_dim_order and remove_view_copy, and check exact output values in all four combinations. Also retain a contiguous-input check proving that safe views still avoid copying.

[zingo]

@rascani would this update match your idea, the new PR turned out quite differently, I should maybe started a new PR instead of force pushing to the same, sorry about that.