perf(bb/msm): stream-walker accumulate — cache forward-pass dx (drop inverse-pass point reloads)

[AztecBot]

Goal

Make the memory-light stream-walker BN254 MSM accumulator faster — focused first on the ~28 ms accumulate kernel — without regressing its memory footprint or correctness. Approach: small, individually-validated changes, each its own commit. Base branch: stream-walker-impl.

Change 1 — cache forward-pass dx, drop inverse-pass point reloads (ba_stream_walker)

The batched-inversion inner loop has three phases per iteration over the S=8 slots: forward prefix product, one safegcd inversion, inverse pass + backward affine-add peel.

The inverse pass recomputed each slot's dx = (p_rx − p_lx) purely to advance the running inverse (inv = inv·dx_k). That recompute re-loaded both x-coordinates through the l0_index → point_x double indirection and re-ran fr_sub_f8 for every active slot, every iteration — although the forward pass already computed exactly those dx one phase earlier.

This caches the S forward-pass dx in a private array and reuses them.

Removed per iteration (S=8): (S−1)·(2 load_pt_x + 1 fr_sub_f8) = up to 14 storage-buffer point reads + 7 field subtractions, traded for S private dx slots. Arithmetic is bit-for-bit identical. The l0_index → point_x chain is a dependent (latency-bound) load on mobile GPUs, so cutting it from the inner loop is the target.

Memory budget unchanged. dx_slot is per-thread private/register state — not workgroup-shared (the 16 KB Mali / 32 KB Apple·Adreno pref_scratch cap is untouched) and not a device buffer (≤100 MB budget untouched). Confirmed: pool GPU bytes are identical baseline vs optimized on every device measured.

Change 0 — COI-free harness (dev page)

Local box has no GPU and the bb WASM-MT oracle is unavailable (needs cross-origin isolation + a built barretenberg-threads.wasm). Added two autorun modes that need neither:

• msm-noble-check — cross-checks WebGPU vs the in-process noble (JS bigint) reference MSM. Correctness gate under headless SwiftShader.
• msm-bench-gpu — times the WebGPU MSM, reporting the per-pass GPU breakdown (incl. the stream_walker accumulate kernel via timestamp-query) + pool GPU bytes. No WASM, runs on any WebGPU device behind BrowserStack.

Lowered dev LOGN_MIN 10 → 8 so logn=8 is reachable.

Correctness — VERIFIED ✅

SwiftShader (software WebGPU, headless Chromium), WebGPU walker vs noble reference, both baseline and optimized:

logn	n	baseline	optimized
8	256	agree ✅	agree ✅
10	1024	agree ✅	agree ✅

BrowserStack timings — real devices, logn=17, 5 reps (msm-bench-gpu)

Apple M2 · macOS Sequoia · Chrome 148 — per-kernel A/B ✅

metric	baseline	optimized (dx-cache)	Δ
stream_walker (accumulate) median	67.57 ms	66.72 ms	−0.85 ms (−1.3%)
stream_walker mean (5 reps)	67.19 ms	66.69 ms	−0.50 ms (−0.7%)
wall median	88.6 ms	87.9 ms	−0.7 ms
pool GPU bytes	71,712,496	71,712,496	0 (memory-neutral)

Small but consistent — every optimized rep landed below baseline's max, with lower variance (opt 66.2–67.0 ms vs baseline 66.1–68.0 ms). Modest on Apple's bandwidth-rich TBDR GPU, where the removed point-loads were largely L1/L2 cache hits and the safegcd inversion dominates the kernel.

Adreno (Snapdragon 8 Elite) · Galaxy S25 Ultra · Android 15 · Chrome 145 — wall A/B

metric	baseline	optimized	Δ
wall median	523.4 ms	516.4 ms	−7.0 ms (−1.3%)
wall mean (5 reps)	527.2 ms	514.8 ms	−12.3 ms (−2.3%)
pool GPU bytes	71,712,496	71,712,496	0 (memory-neutral)

Same direction as Apple, slightly larger on this bandwidth-limited mobile GPU. Caveat: Android Chrome does not expose the timestamp-query feature, so per-pass isolation is unavailable — only wall time, whose ~±5% rep-to-rep variance is too coarse to tightly resolve a single-kernel change. The consistent mean/median direction is suggestive, not conclusive.

Mali (Pixel 9 Pro XL) — not separately A/B'd

Same Android-Chrome timestamp-query limitation (no per-kernel isolation), and the device was occupied by other agents throughout the window (the shared 2-seat BrowserStack pool). Memory-neutrality holds by construction (no device buffer added). A wall A/B can be run later when a Mali seat frees.

Summary

A small, safe, memory-neutral first step on the accumulate kernel: ~1–2% faster on both Apple and Adreno, identical pool bytes, correctness GREEN at logn=8/10. The win is bounded on Apple because the safegcd inversion (~47% of the kernel) dominates and Apple's caches hide the redundant reads; the larger relative move on Adreno fits the "dependent-load on a bandwidth-limited GPU" thesis. Next levers (separate, individually-benchmarked commits): a forward-pass dead-store removal, idle-anchor point hoisting, and a GPU partials-reduction path.

Commits

1. test(bb/msm): COI-free noble cross-check autorun + dev floor → logn=8
2. perf(bb/msm): walker — cache forward-pass dx, drop inverse-pass point reloads
3. test(bb/msm): COI-free GPU-only bench autorun (msm-bench-gpu)