[ExecuTorch][WebGPU] Generate logical and arithmetic binary shader variants

[JCNTH]

Stack from ghstack (oldest at bottom):

• [ExecuTorch][WebGPU] Preserve Linear and Q4 embedding params across dynamic resize #21483
• -> [ExecuTorch][WebGPU] Generate logical and arithmetic binary shader variants #21451
• [ExecuTorch][WebGPU] Generate extrema and unary shader variants #21450
• [ExecuTorch][WebGPU] Generate typed to_copy shader variants #21449
• [ExecuTorch][WebGPU] Make WGSL generation rollback-safe #21448
• [ExecuTorch][WebGPU] Add Qwen3 K16 attention and fp16 KV cache #21136
• [ExecuTorch][WebGPU] Add HuggingFace rotate-half RoPE #21135
• [ExecuTorch][WebGPU] Repair dynamic SDPA routing, add attestation #21134
• [ExecuTorch][WebGPU] Select emdawnwebgpu for Emscripten builds #21133
• [ExecuTorch][WebGPU] Add Llama K16 online causal attention #21132
• [ExecuTorch][WebGPU] Replace broad QKV fusion with BK64 kernel #21131
• [ExecuTorch][WebGPU] Route exact prefill shapes to the BK64 GEMM #21130
• [ExecuTorch][WebGPU] Fuse the SwiGLU elementwise chain #21129
• [ExecuTorch][WebGPU] Coordinate dynamic dispatch routes #21128
• [ExecuTorch][WebGPU] Add certified scoped output suppression #21127

Generate byte-identical logical and arithmetic binary variants from shared WGSL families

Logical AND/OR and four arithmetic binary kernels duplicated shader skeletons and broadcast logic. This consolidates logical AND/OR behind one packed-Boolean family and minimum/pow/floor_divide/mul into the existing binary family, with a permanent mixed-rank broadcast contract.

Key changes:

• Generate logical AND/OR from one operator-token manifest.
• Generate minimum, pow, floor_divide, and mul beside the existing div/sub variants.
• Lock same-shape and mixed-rank expressions, exact payloads/workgroups, PTE delegation, and broadcast boundary cases.

No runtime C++ dispatch, bindings, pipeline construction, workgroups, or expanded shader payloads change. Four standalone WGSL inputs are removed, and future compatible variants require manifest entries instead of copied kernels. This follows the Vulkan binary-family pattern.

Co-authored-with: Claude Code.
@exported-using-ghexport

Differential Revision: D113979789

Differential Revision: D113979789

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/21451

• 📄 Preview Python docs built from this PR

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

⏳ No Failures, 16 Pending

As of commit d25c0c2 with merge base f4a21b8 ():
💚 Looks good so far! There are no failures yet. 💚

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.

[JCNTH]

v

[JCNTH]

@claude Double check for need of changes, modulariziation opportunity, etc

[claude]

Claude finished @JCNTH's task in 1m 58s —— View job

---

Review: WebGPU logical binary shader variants

• Read the full PR diff vs base
• Compare against Vulkan binary_op_buffer reference
• Assess correctness, modularization opportunities, drift risks
• Post review feedback

The shader consolidation is clean and correct. The generated logical_and_wgsl.h / logical_or_wgsl.h are byte-identical to before (only the @generated from provenance line changed), the $if OP == "&" branch keeps each comment accurate, and the new roundtrip test locks the template → expansion → embedded header → registry chain. The deterministic truth-table generators are a genuine improvement over the old random-seed masks — they now cover all four Boolean pairs (FF/FT/TF/TT) in byte order plus word-boundary shapes (words63/64/65), and test_logical_binary_case_contract pins the exact payloads. No correctness issues found.

The @claude ask was "double check for need of changes, modularization opportunity" — so, focusing on that:

Main modularization opportunity: the two handler .cpp files still duplicate ~130 lines

The PR unified the shaders but left logical_and/LogicalAnd.cpp and logical_or/LogicalOr.cpp as near-identical files. A diff shows they differ only in:

• the WGSL constant (kLogicalAndWGSL vs kLogicalOrWGSL) and workgroup constant,
• the op-name strings in error messages / dispatch labels ("logical_and" vs "logical_or"),
• the params struct name (LogicalAndParams vs LogicalOrParams — identical layout),
• the two WEBGPU_REGISTER_OP lines.

Everything else — validation, uniform setup, pipeline/bind-group construction, the resize hook — is byte-for-byte the same. This is exactly the drift risk the PR set out to eliminate at the shader layer, just moved down to the host layer.

The repo already has the pattern to fix this: boolean_op/BooleanOp.cpp consolidates 7 ops (eq/ne/le/ge/lt/gt.Scalar + logical_not) behind a single dispatch_bool_op(...) helper parameterized by (wgsl, wg_size_x, op_name), with thin per-op registration shims. Applying the same shape here — a single logical_binary/LogicalBinary.cpp with dispatch_logical_binary(graph, args, wgsl, wg_size_x, op_name) and two 1-line impls — would collapse the two files into one and put the handler in the same directory as the shader it dispatches (right now shaders live in logical_binary/ while handlers stay in logical_and/ + logical_or/, which is a bit split). This also mirrors Vulkan more faithfully, where binary_op_buffer is one shader family and one host dispatch path.

If you want this, I'm happy to implement it — it's a mechanical consolidation and the existing test_logical_binary_template_roundtrip_byte_identical already validates the shader/registry side; the handler hash assertions in that test would need their paths updated to the new single file.

Minor: test-fixture module ownership

LOGICAL_BINARY_CASES, logical_binary_gen_a, and logical_binary_gen_b now live in test/ops/test_logical_and.py and are imported by test_logical_or.py, test_bitwise.py, cases.py, and test_generator.py. Having the OR / bitwise / codegen tests reach into the AND test module for shared fixtures is slightly surprising. Consider a small shared helper (e.g. test/ops/logical_binary_fixtures.py) so the shared truth-table generators have a neutral home — parallels the runtime logical_binary/ consolidation. Low priority; purely organizational.

Nit

• test_logical_and.py _logical_binary_gen: the guard is numel % len(pattern) != 0 but the error string is hard-coded "...numel % 4 == 0". Fine today (len(pattern) == 4), just a latent inconsistency if the pattern length ever changes.

Net: no blocking issues — the shader-side work is solid and well-tested. The one worthwhile follow-up is folding the two duplicate handlers into a single dispatch_logical_binary à la BooleanOp.cpp; that completes the anti-drift goal end to end rather than stopping at the shader layer.