XNNPack: Extend to_copy support when skip_dim_order=True

[Erik-Lundell]

With the default XNNPACK test config, skip_dim_order=False rewrites aten._to_copy.default to dim_order_ops._to_dim_order_copy.default path is already supported through CloneDimOrderConfig.

Add a ToCopy partitioner config to partition the operator. _to_copy already had a node_visitor due to DimOrderOpsRevertPass creating _to_copy nodes for permutation. Expand this to also support dtype conversion using to_copy.

Extend RemoveRedundantCopy pass to handle to_copy
nodes with redundant dtypes.

cc @GregoryComer @digantdesai @cbilgin @JakeStevens @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/22075

• 📄 Preview Python docs built from this PR

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

✅ No Failures

As of commit 8f684de with merge base d7c5dd4 ():
💚 Looks good so far! There are no failures yet. 💚

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

[JakeStevens]

@claude please review

[claude]

Claude finished @JakeStevens's task in 7m 42s —— View job

---

Code Review

• Read changed files
• Analyze correctness & design
• Post review feedback

Nice work — the decomposition into TRANSPOSE/CAST/COPY is clean, the "make the source of truth the reshape pass, fall back to memory_format only when untagged" split is well reasoned, and the tests cover the new paths well. A few points below, ordered by importance.

---

1. ToCopyConfig permits dtype conversions the visitor can't faithfully serialize

ToCopyConfig.check_constraints (generic_node_configs.py:539) only validates the dim-order side. Any dtype pair is accepted and routed to the CAST → XNNConvert path. But NodeVisitor.get_serialized_dtype (node_visitor.py:266-278) only recognizes fp32/fp16/bf16 (plus quant types) — every other dtype silently falls back to xnn_datatype_fp32.

So an integer→float _to_copy (e.g. int32 → float32) would declare the input tensor as fp32 while its data is int32, and emit an fp32 → fp32 convert. For a constant this may be masked by buffer folding, but for a runtime integer activation this is a correctness bug.

Contrast with ToDimOrderCopyConfig (:507), which explicitly gates the dtype dimension. I'd recommend ToCopyConfig mirror that: reject changes_dtype conversions whose source/target aren't in the XNNPACK-supported float set (fp32/fp16/bf16), with a why(...) reason. Otherwise please confirm how integer sources are handled. Fix this →

2. Quantized redundant-copy path was not given the same dtype/shape guard

The new _same_tensor_type_and_shape / _preserves_tensor_type_and_shape guards were added to _try_remove_regular_redundant_to_copy (remove_redundant_copy_pass.py:49 and :71), which is correct. But _try_remove_quantized_redundant_to_copy (:83) still removes copies based purely on memory-format tags, with no dtype/shape check. If a dtype-changing _to_copy ever appears inside a dq → to_copy → q chain, it would be incorrectly elided. Probably out of scope for this PR, but worth either guarding for symmetry or adding a short comment noting the quantized path assumes dtype-preserving copies.

3. _append_output_tensor — redundant parameter

At the call site (op_to_copy.py:295-298), output_is_channels_last is passed both as the output_is_channels_last arg and again as convert_to_nhwc. They're always identical, so convert_to_nhwc is dead surface area on an already 12-parameter helper. Dropping it and using output_is_channels_last internally would simplify the signature. Minor, per the repo's "choose simpler, more concise" guidance.

---

Things I checked that look correct

• No circular import from the new top-level ChannelsLastTaggedReshapePass import in op_to_copy.py — node_visitor.py already imports it the same way.
• Intermediate-tensor dtype threading in the decomposition loop is consistent for both TRANSPOSE→CAST and CAST→TRANSPOSE orderings, and sort_decomposed_operations correctly runs the transpose on the smaller dtype.
• tag_node handling of None/preserve_format (inheriting the input's NCHW/NHWC tag) is sound, and the match / X | None / tuple[...] syntax is fine given requires-python >=3.10.
• Transpose num_dims is read back from the serialized tensor, so it stays consistent with define_tensor's NHWC permutation.
  · branch xnnpack-to-copy