[ExecuTorch][WebGPU] Add view_copy op (aten.view_copy.default)

[JulianCloudNTH]

Stack from ghstack (oldest at bottom):

• [ExecuTorch][WebGPU] cat op test suite (cases.py op-test framework) #20399
• [ExecuTorch][WebGPU] Add cat op + ValueList graph support (aten.cat.default) #20398
• [ExecuTorch][WebGPU] permute_copy op test suite (cases.py op-test framework) #20397
• [ExecuTorch][WebGPU] Add permute_copy + IntList graph support (aten.permute_copy.default) #20396
• [ExecuTorch][WebGPU] slice_copy op test suite (cases.py op-test framework) #20395
• [ExecuTorch][WebGPU] Add slice_copy op (aten.slice_copy.Tensor) #20394
• [ExecuTorch][WebGPU] squeeze_copy + unsqueeze_copy test suites (cases.py op-test framework) #20393
• [ExecuTorch][WebGPU] Add squeeze_copy + unsqueeze_copy (flat copies) #20392
• [ExecuTorch][WebGPU] sigmoid op test suite (cases.py op-test framework) #20391
• [ExecuTorch][WebGPU] Add sigmoid op (aten.sigmoid.default) #20390
• [ExecuTorch][WebGPU] select_copy op test suite (cases.py op-test framework) #20363
• [ExecuTorch][WebGPU] Add select_copy op (aten.select_copy.int) #20362
• [ExecuTorch][WebGPU] view_copy op test suite (cases.py op-test framework) #20361
• -> [ExecuTorch][WebGPU] Add view_copy op (aten.view_copy.default) #20360
• [ExecuTorch][WebGPU] mul op test suite (cases.py op-test framework) #20359

Adds aten.view_copy.default to the WebGPU delegate. A contiguous reshape on a dense row-major buffer backend is a flat copy output[i] = input[i], so the op is a single 1D-dispatch copy kernel.

Composition (single compute dispatch):

• runtime/ops/view_copy/view_copy.h — declares add_flat_copy(graph, in_id, out_id): fail-loud guards (both tensors, fp32, numel-preserving) + the view_copy.wgsl dispatch over compute_1d_workgroup_count(num_elements) with override wg_size; mirrors Vulkan add_view_copy_node.
• runtime/ops/view_copy/ViewCopy.cpp — reads args = [self, size, out], ignores the AOT-fixed size value-id (output shape comes from out_tensor.dims), calls add_flat_copy.
• runtime/ops/view_copy/view_copy.wgsl — guards idx >= num_elements, writes output[idx] = input[idx].
• add_flat_copy is factored into the header so the stacked squeeze/unsqueeze ops reuse it without a new kernel.
  @exported-using-ghexport

Differential Revision: D108793164

Differential Revision: D108793164

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/20360

• 📄 Preview Python docs built from this PR

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

❌ 2 New Failures, 3 Unrelated Failures

As of commit 867bd73 with merge base 0e65ba6 ():

NEW FAILURES - The following jobs have failed:

• pull / unittest / linux / linux-job (gh)
  RuntimeError: Command docker exec -t 5343bba6577171397a676a284384ed22d67f533f6fd73aea6a81dbbdceb57e96 /exec failed with exit code 1
• pull / unittest-editable / linux / linux-job (gh)
  RuntimeError: Command docker exec -t fb0359baa95e8f7195cec09bf28e4608b35bd93d7c7d99600a1348a25ddd94e4 /exec failed with exit code 1

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

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

• pull / test-llama-runner-qnn-linux (fp32, qnn_16a16w, qnn) / linux-job (gh) (trunk failure)
• pull / unittest / macos / macos-job (gh) (trunk failure)
  RuntimeError: Command bash /Users/ec2-user/runner/_work/_temp/exec_script failed with exit code 1
• pull / unittest-editable / macos / macos-job (gh) (trunk failure)
  RuntimeError: Command bash /Users/ec2-user/runner/_work/_temp/exec_script failed with exit code 1

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

[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.

[JulianCloudNTH]

@claude review

[claude]

Claude finished @JulianCloudNTH's task in 1m 14s —— View job

---

Code Review: WebGPU view_copy op

• Read changed files & sibling ops
• Check consistency with existing WebGPU op patterns
• Review correctness, resource handling, edge cases

---

Clean, well-scoped addition. The structure mirrors mul/BinaryOp.cpp faithfully (workgroup sizing, uniform creation, bind-group layout, release ordering), and factoring add_flat_copy into the header so stacked squeeze/unsqueeze ops can reuse the kernel is the right call. Resource lifetime matches the established convention — shader/BGL/pipeline-layout/uniform released after dispatch, while pipeline and bind group are retained by the graph. No leaks or double-frees.

A few observations, none blocking:

1. The "fp32 guard" is really a 4-byte-alignment guard — ViewCopy.cpp:44-49

if (in_tensor.nbytes % sizeof(float) != 0 || out_tensor.nbytes % sizeof(float) != 0)

This admits any dtype whose byte count is a multiple of 4 (e.g. an even-numel fp16 tensor), not just fp32. That happens to be safe here — a flat byte copy is dtype-agnostic and the in_numel == num_elements check below is effectively an nbytes-equal check — so the result is still correct. But the error string "expected fp32" overpromises. Compare with mul_impl (mul/BinaryOp.cpp:53-61), which asserts nbytes == numel * 4 against the dims-derived numel for a true dtype check. Since WebGPUTensor carries dims, you could tighten this to match mul's precision, or just soften the comment/message to "operand size not 4-byte aligned" to reflect what's actually enforced.

2. Contiguity assumption is implicit. The flat output[i] = input[i] is only correct when the input is dense row-major — a transposed/strided view would copy in the wrong order. This matches the buffer-backend design (and Vulkan's add_view_copy_node), and the PR body states it, but there's no runtime guard against a non-contiguous input. If non-contiguous inputs are structurally impossible at this point in the WebGPU delegate, a one-line comment in add_flat_copy noting that invariant would save a future reader the trip to the PR description.

3. args.at(args.size() - 1) for out — ViewCopy.cpp:154. Correct and robust against the AOT size value-id. Minor: a named local (const int out_id = args.at(args.size() - 1);) reads a touch clearer than the inline expression, but this is taste.

4. Generated header has unused constants — view_copy_wgsl.h:39-40 define kViewCopyWorkgroupSizeY/Z, which aren't referenced (only ...SizeX is). This is @generated, so presumably the codegen template emits all three uniformly — fine to leave, just flagging.

Nit: ViewCopyParams::_pad[3] pads the uniform to 16 bytes and the binding uses sizeof(ViewCopyParams) (16) while the WGSL Params is 4 bytes — that's valid (binding size ≥ struct size), just worth a one-word comment that the padding is for the 16-byte uniform alignment requirement (the existing comment at :24 covers this).

No correctness bugs found. The nbytes-based numel check correctly prevents the OOB input[] read, the idx >= num_elements shader guard is present, and the dispatch count goes through compute_1d_workgroup_count's limit check. LGTM with the optional polish above.