[ExecuTorch][WebGPU] SymInt arithmetic ops (add/sub/mul/floordiv) for dynamic shapes

[JulianCloudNTH]

Stack from ghstack (oldest at bottom):

• [ExecuTorch][WebGPU] 2D compute dispatch tests — prefill golden + fold unit test #20584
• [ExecuTorch][WebGPU] 2D compute dispatch — lift the 65535 per-dim cap (prefill path) #20583
• [ExecuTorch][WebGPU] Dynamic-shape integration test (allocate-at-max + per-op resize) #20582
• [ExecuTorch][WebGPU] Dynamic resize hook for slice_copy (dynamic start/end) #20581
• [ExecuTorch][WebGPU] Dynamic resize hook for SDPA (live seq-len S) #20580
• [ExecuTorch][WebGPU] Dynamic resize hook for view_copy #20579
• [ExecuTorch][WebGPU] Dynamic resize hooks for sigmoid and select_copy #20578
• [ExecuTorch][WebGPU] Dynamic resize hooks for add and mul #20577
• [ExecuTorch][WebGPU] Dynamic resize hook for linear_q4gsw #20576
• [ExecuTorch][WebGPU] Dynamic resize hooks for rms_norm, embedding, rope #20575
• [ExecuTorch][WebGPU] Dynamic tensor-shape resize engine core #20574
• -> [ExecuTorch][WebGPU] SymInt arithmetic ops (add/sub/mul/floordiv) for dynamic shapes #20573

Register scalar SymInt arithmetic so dynamic-shape graphs lower without an "unsupported op" failure.

Problem: A dynamic-shape exported program emits scalar SymInt arithmetic nodes (add/sub/mul/floordiv) to compute live sizes and positions (e.g. input_pos + S, seq_len // n). The WebGPU backend registered only et_vk.select_as_symint.default, so WebGPUGraph::build() threw unsupported op: add when loading any dynamic .pte.

Solution:

• Before: only select_as_symint produced a live SymInt; any arithmetic on it was unsupported.
• After: add/sub/mul/floordiv each recompute their output SymInt from the operands via a resize hook whenever a live operand changes.

Implementation:

• register_sym_binary reads each operand (live SymInt via read_symint, else a static Int), seeds the build-time value, and registers a resize hook on any live operand (set_symint on recompute).
• floordiv rounds toward negative infinity (Python semantics).
• Mirrors Vulkan backends/vulkan/runtime/graph/ops/impl/SymIntOps.cpp (add/sub/floordiv/mul).
• Registered under the bare targets add/sub/mul/floordiv — distinct registry keys from the tensor aten.add.Tensor/aten.mul.Tensor ops.

Constraints: An output that folded to a static Int is a no-op. No GPU kernel, no dispatch, no change to the static-shape path. sym_size.int is intentionally not added here (depends on the tensor-shape dim-source API in the following diff).

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

Differential Revision: D109906102

Differential Revision: D109906102

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/20573

• 📄 Preview Python docs built from this PR

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

⏳ No Failures, 2 Pending

As of commit fe4f84f with merge base db5f1ba ():
💚 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.

[JulianCloudNTH]

@claude review and check for any areas or opportunities for modularization

[claude]

Claude finished @JulianCloudNTH's task in 2m 6s —— View job

---

Review: WebGPU SymInt arithmetic ops

• Read changed file SelectAsSymint.cpp
• Compare against Vulkan reference SymIntOps.cpp
• Check WebGPUGraph API (read_symint/set_symint/add_resize_hook/propagate_resize)
• Analyze correctness & modularization opportunities

Overall this is a clean, well-scoped change. The resize-hook approach is a genuine improvement over the Vulkan reference it mirrors, and the diff does exactly what the description claims. Notes below, ordered by importance.

---

Modularization (the main ask)

1. register_sym_binary is the right call — and better than Vulkan. The Vulkan reference (backends/vulkan/runtime/graph/ops/impl/SymIntOps.cpp:53-166) copy-pastes the full _impl + resize_*_node + registration boilerplate four times. Collapsing all four ops into one helper parameterized by a std::function (SelectAsSymint.cpp:51-74) is the more maintainable design. Good instinct.

2. File/directory naming no longer matches contents. This is the clearest modularization opportunity. The file now contains general SymInt arithmetic (add/sub/mul/floordiv), but lives in ops/select_as_symint/SelectAsSymint.cpp — a directory named after one specific op. The repo convention is strictly one directory per op/concern (ops/add/, ops/mul/, ops/select/, …). A reader looking for where floordiv is registered won't find it under a select_as_symint path.

Consider extracting the arithmetic ops into a sibling ops/sym_int/SymIntOps.cpp (mirroring the Vulkan filename), leaving select_as_symint to own only select_as_symint. The shared helper read_scalar would move with them. This keeps the grep-to-op mapping the codebase relies on. Fix this → (if you'd rather keep the diff small and land the stack first, this is fine as a fast-follow — just flagging the convention.)

---

Correctness

3. Division by zero in floordiv is UB (SelectAsSymint.cpp:90-91). x / y and x % y with y == 0 are undefined behavior. The Vulkan reference has the same gap, so this isn't a regression, and a floordiv node with a zero divisor shouldn't arise from a valid export — but a one-line guard (throw, or return 0) would make it fail-loud like the rest of this file (read_symint, set_symint both throw on misuse) instead of crashing.

4. INT_MIN % y / INT_MIN / -1 overflow — theoretical only, since these values are tensor sizes/positions. Not worth handling; noting for completeness.

5. Self-referential operands register the hook twice. If a == b (e.g. x + x), both if branches at :68-72 fire and recompute is registered twice on the same SymInt. Harmless (recompute is idempotent, and set_symint no-ops when the value is unchanged), just slightly redundant. An if (a_is_symint) add_hook(a); else if (b_is_symint)… won't work since both may legitimately be distinct live operands — a b != a guard on the second branch would be the precise fix if you want it.

---

Minor / style

6. read_scalar silently assumes non-SymInt ⇒ Int (:43-48). If an operand were a Double (or any non-Int constant), get_int reads an unrelated flat array and returns a wrong value rather than erroring. Scalar arithmetic operands should only ever be Int/SymInt, so this is safe in practice; a comment or a type assert would document the assumption.

7. Chained arithmetic relies on registration order. propagate_resize (WebGPUGraph.cpp:148-158) iterates resize_hooks_ once in registration order; a hook that recomputes out marks out dirty so a later-registered hook keyed on out still fires in the same pass. This works only because ops are registered in topological order during build(). That's a reasonable invariant, but it's load-bearing and currently undocumented — worth a one-line comment on add_resize_hook or propagate_resize so a future single-pass refactor doesn't silently break multi-level SymInt chains (a + b, then that * c).

---

Verification

I could not build/run the WebGPU backend in this environment, so the above is static analysis only. Worth confirming with an actual dynamic-shape .pte that exercises a chained expression (e.g. (input_pos + S) // n) end-to-end, which would validate both the floor rounding (:88-95) and the chaining behavior in note 7.
· branch gh/JulianCloudNTH/65/head