perf(gpu): run DECODE (num_parts=1) DEEP/FRI on device

[ColoCarletti]

Extends the device-resident composition-parts path to num_parts==1 so the DECODE table's DEEP/FRI runs on GPU instead of host.

For a degree-1 composition, H is already the single part on the LDE coset — no decompose, no re-extension, just an interleaved→slab de-interleave (new comp_h_to_slabs_ext3 kernel, no NTT). The R2 commit, R3 OOD, R4 DEEP, FRI and openings already read the part count from the handle, so no downstream changes are needed; the host fallback is preserved.

Validation (ethrex 100tx, RTX 5090):

• Wall: 13.13s -> 12.18s (-7.25%, ABBA x3)
• Cross-verify OK; math-cuda + stark cuda suites green
• Unit parity: de-interleave bit-exact, evaluate_dev(keep) == host bit-for-bit

[ColoCarletti]

/bench-gpu

[ColoCarletti]

/ai-review

[github-actions]

GPU Benchmark (ABBA) — a9aead5753 vs main (14 pairs)

_{RTX 5090 · AMD Ryzen 9 9950X 16-Core Processor (32 threads) · Vast.ai datacenter @ $0.9347222222222221/hr · prover/cuda · ethrex real block, continuations · drift-free A/B/B/A}

=== ABBA paired result  (improvement: - = PR faster) ===
  pairs: 14   mean A (PR): 71.796s   mean B (base): 74.506s

  [parametric] paired-t   mean -3.55%   sd 4.69%   se 1.25%
               95% CI: [-6.26%, -0.85%]   (t df=13 = 2.16)
  [robust]     median -4.53%   Wilcoxon W+=14 W-=91  p(exact)=0.0134  (z=-2.39)

  --- server stability (this run; compare across servers) ---
  run-to-run jitter:    A CV 7.20%   B CV 6.93%        (lower = steadier)
  within-session drift: +6.97% over the run, 1st->2nd half +4.27%
    (jitter -> Tier-1 cached gate floor; drift -> whether the cached baseline can be trusted)

  VERDICT: REAL IMPROVEMENT - PR faster by ~3.55% (t-CI and Wilcoxon agree)

  raw pairs: /tmp/abba_run/pairs.csv

_{- = PR faster. Trust the verdict when paired-t and Wilcoxon agree.}

[github-actions]

Codex Code Review

No issues found in the PR changes.

[claude]

Review: perf(gpu): run DECODE (num_parts=1) DEEP/FRI on device

Reviewed the kernel, the math-cuda wrapper, the prover dispatch, and every downstream consumer of GpuLdeExt3 (build_comp_poly_tree_from_slabs_dev, try_barycentric_ext3_on_ext3_handle, deep_composition_ext3, gather_rows_ext3_on_device, download_ext3_columns, materialize_composition_parts_host). No Critical or High safety issues. The design holds up:

• comp_h_to_slabs_ext3 is a pure grid-strided transpose writing all 3 * num_rows u64, so the unsafe { stream.alloc } (no zero-init) is sound — the contrast with the zero-padded d=2 slab layout is correctly called out in the comment.
• Slab ordering out[k*num_rows + i] matches the handle contract (c*3 + k) * lde_size at m = 1, and matches decompose_d2_ext3's [h0.abc, h1.abc] layout.
• Cross-stream handoff mirrors the d=2 path: kernel enqueued on h.stream, PooledEvent recorded, consumers wait_ready_on device-side. GpuCompH::buf's stream-ordered free is ordered behind the kernel on the same stream.
• Every downstream consumer already reads the part count from handle.m / lde_composition_poly_evaluations.len(), so m = 1 needs no changes — confirmed, and num_parts == 1 already exercised the GPU comp-poly tree before this PR.
• The host fallbacks (downloaded_h → vec![h], and the else if number_of_parts == 1 arm) are value-equivalent to evaluator.evaluate(...).

Findings

Medium — device_only_for gate not extended; the want_host == false branch is dead. device_only_gate is still reached only when composition_poly_degree_bound(n) / n == 2 (prover.rs:1101), and host_trace_empty is set only when that gate held at R1. So for a num_parts == 1 table want_host is always true: the vec![Vec::new()] placeholder in try_deinterleave_comp_h_dev is unreachable, DECODE still does the full H D2H every proof, and the gate's comments ("the device-resident R2 path only exists for the d=2 quotient decomposition", prover.rs:1094 and :1101) are now wrong. Inline detail on the branch.

Low — gpu_comp_h_slabs_calls() is exported but never read. Every sibling counter is asserted in prover/tests/cuda_path_integration.rs; this one isn't, so a silent regression of the d=1 path to the host arm would go unnoticed. Add the assertion or drop the counter.

Low — comp_h_from_host_interleaved is test-only public API on math_cuda::constraint_interp. It's honestly documented as a parity helper and it does need to be pub for the integration test, so this is just a note that it widens the crate's surface for no production caller.

Low — the if number_of_parts == 1 { deinterleave } else { decompose_d2 } dispatch is duplicated verbatim in round_2_compute_composition_polynomial (prover.rs:1702) and xcheck_post_mortem (prover.rs:4299), as is the vec![host_h] vs decompose_and_extend_d2 split. A small helper would keep the two from drifting — the xcheck's whole value is being an apples-to-apples mirror of the producer.

Test coverage

comp_h_to_slabs.rs (bit-exact transpose over log 1..=14, raw non-canonical limbs) and the decode_shaped_program + evaluate_dev(keep=true) == host-drained H parity in gpu_constraint_interp.rs are both well targeted at the two things that could go wrong here. The gap is the integration level — nothing asserts the d=1 device path actually engages end to end (see the counter finding).

[github-actions]

AI Review

PR #946 · 7 changed files

Findings

No non-rejected structured findings were reported.

Reviewer Lanes

Lane	Model	Prompt	Status	Findings
glm	openrouter/z-ai/glm-5.2	general	success	0
kimi	openrouter/moonshotai/kimi-k2.7-code	general	success	0
minimax	minimax/MiniMax-M3	general	error: opencode failed (provider/auth/runtime error) and no findings were submitted	0
moonmath	zro/minimax-m3	general	error: opencode failed (provider/auth/runtime error) and no findings were submitted	0
nemotron	openrouter/nvidia/nemotron-3-ultra-550b-a55b	general	success	0

Native Codex and Claude reviews run separately and post their own comments. They are not included in this structured provenance report.

Raw lane outputs, candidates, final issues, and model metrics are uploaded as workflow artifacts.

[ColoCarletti]

/bench-gpu

[ColoCarletti]

/bench-gpu