feat(bb/msm): benchmark-gated WebGPU MSM harness — SwiftShader correctness + Adreno baseline + memory attribution

[AztecBot]

WebGPU MSM optimization — measured foundation (WIP)

Working branch off stream-walker-impl. Goal: a memory- and time-optimal BN254 WebGPU MSM for laptop + mobile GPUs (Apple TBDR, Adreno, Mali), within the 32 KB / 16 KB workgroup-shared-memory and ≤100 MB total-memory constraints up to n = 2²⁰.

This session was benchmark-gated: get a real WebGPU correctness path + a real device perf/memory path working first, record baselines, and only then change the algorithm. No unmeasured claims — so this PR ships the validated harness + the baseline measurements + the analysis that should drive the next change, rather than a speculative rewrite.

1. Harness enablement (the blocker this box had)

The box has no GPU. Nothing here could validate WebGPU before. Fixed:

• SwiftShader headless correctness (dev/msm-webgpu/scripts/swiftshader-crosscheck.mjs). The pre-existing test-msm-swiftshader.mjs couldn't get a WebGPU adapter on Linux. Working incantation (Playwright Chromium + SwiftShader Vulkan ICD):

  VK_ICD_FILENAMES=.../vk_swiftshader_icd.json
  --enable-unsafe-webgpu --enable-features=WebGPU,Vulkan
  --use-vulkan=swiftshader --use-webgpu-adapter=swiftshader
  

• ?ref=noble cross-check mode. The threaded-WASM oracle in the harness (barretenberg.wasm.gz) is a 213-byte placeholder here, so GPU-vs-WASM is unavailable. The new mode cross-checks the WebGPU result against the in-page @noble/curves reference at the actual logN and skips WASM — valid GPU-less validation. LOGN_MIN 10 → 8 so 2⁸ is genuinely testable (was clamped to 2¹⁰).
• GPU-only bench mode (?gpu_only=1, auto for the msm-bench autorun): skips the unavailable WASM path so timing runs don't depend on it; Run/Run×5 enable on SRS-ready.
• Per-buffer GPU-memory attribution (MsmV2Pool.statsBreakdown()): turns one opaque "working-set" number into a labeled per-buffer map, surfaced in the [mem] log and bench results.
• Bench wall-time fix: the Run handler clears $log on click, so the bench rep loop always parsed wallMs = 0. Now the dispatch→readback wall is read from __lastGpuWallMs. This is the trustworthy timing on GPUs with broken timestamp-queries (see Adreno finding below).

2. Correctness — GREEN (stream-walker path, this branch)

SwiftShader (software Vulkan), WebGPU cross-checked against noble:

logN	n	WebGPU vs noble
8	256	✅ agree
10	1024	✅ agree

(SwiftShader wall times are software-rasterizer numbers — correctness only, never perf.)

3. Baseline — real device (Samsung Galaxy S25 Ultra · Adreno, Chrome 145 Mobile)

Stream-walker accumulate path, n = 2¹⁷:

metric	value	notes
GPU wall (dispatch→readback)	~491 ms	the reliable number ([gpu] returned in)
per-pass timestamp sum	~25 000 ms	garbage — Adreno timestamp-queries are unreliable; do not use per-phase deltas on Adreno
GPU working set (excl. SRS)	60.4 MiB	shared scratch + instance buffers
SRS pool	8.0 MiB	n·64 B
total GPU buffers	68.4 MiB	under the 100 MB budget at 2¹⁷

4. Memory attribution (key finding)

statsBreakdown() shows the working set is not the "~9 MB memory-light" figure the plan implies — it's ~60 MiB at 2¹⁷:

• walkerPartials ≈ 8.5 MiB is fixed at every n — it's soaBuf(2 · STREAM_T · S) with STREAM_T = 8192 hardcoded (msm_v2.ts:1614), independent of n. At small n it's massively over-provisioned (it is the entire scratch at logN=10); at 2¹⁷ it's ~14% of scratch.
• The rest (~52 MiB at 2¹⁷) is the n/B-scaled Pippenger buffers: bucket sums, scalar decomposition outputs, reduction buffers. These scale toward the 100 MB ceiling by 2²⁰ and are the real budget pressure.
• The legacy V2 pair-tree / old stream-accum buffers are already 4-byte stubs (plan §14 step 9) — no easy dead-buffer win remains.

5. Where the next change should go (analysis, not yet implemented)

Ordered by leverage; each must clear SwiftShader correctness + a BrowserStack number before landing:

1. GLV endomorphism — the standout memory and time lever. φ(P)=(βx,y) gives s = s₁+λs₂ with 128-bit halves ⇒ effective windows T ≈ 9 vs 17 at c=15 ⇒ roughly halves both the bucket-accumulation work and the n·T- and B-scaled accumulation-column memory. Was removed from the bb.js port; re-wiring on the warm SRS path is greenfield but mathematically small. (Details: src/msm_webgpu/MSM_DESIGN_ANALYSIS.md §6.4.)
2. Right-size STREAM_T for small n — frees most of the fixed 8.5 MiB walkerPartials below ~2¹⁴. Neutral at 2¹⁷⁺. Low risk but needs the indirect-dispatch + taskCuts/walkerNodes sizing to track it.
3. GPU partials-reduction kernel + warp-coalesced task layout (plan's deferred wins).

6. Blocked / honest status

• Only the Adreno baseline was captured this session. BrowserStack has 2 seats shared across ~10 agents; both were occupied when I went for a second device, and the rule is to wait, never preempt. Apple + Mali baselines are pending a free seat — the harness + breakdown are ready to capture them in one run each (?autorun=msm-bench&logn=17).
• No algorithm change is included precisely because I can't yet validate one's perf/memory at scale on this contended path, and the brief forbids unmeasured claims.

Reproduce locally (no GPU):

cd barretenberg/ts && yarn dev:msm-webgpu   # :5173
VK_ICD_FILENAMES=/opt/ms-playwright/chromium-1148/chrome-linux/vk_swiftshader_icd.json \
CHROMIUM_PATH=/opt/ms-playwright/chromium-1148/chrome-linux/chrome \
  node dev/msm-webgpu/scripts/swiftshader-crosscheck.mjs 10   # → "WebGPU and noble agree"