perf(bb/msm): arch-aware stream-walker TPB (128 on ≥32 KB devices, 64 on Mali)

[AztecBot]

Summary

The memory-light stream-walker accumulator hardcoded WALKER_TPB = 64 in msm_v2.ts, sizing its var<workgroup> pref_scratch to 16 KB so it always fits Mali Bifrost's 16 KB workgroup-storage cap. On Apple (TBDR) and Adreno — which grant 32 KB — that leaves half the workgroup-memory budget and a slice of occupancy unused, and is part of why the walker trails the V2 pair-tree near logn≈17.

This wires the walker's TPB to the device's granted maxComputeWorkgroupStorageSize (already requested at the adapter max in gpu.ts):

• ≥ 32 KB (Apple / Adreno) → TPB = 128 (pref_scratch = 128·S·2·16 B = 32 KB at S=8)
• otherwise (Mali Bifrost, 16 KB) → TPB = 64 (16 KB) — unchanged

The constants STREAM_WALKER_TPB / STREAM_WALKER_TPB_MALI_FALLBACK already lived in ba_stream_plan.ts for exactly this ("Mali targets … must drop this to 64 at adapter init") but were never read; this connects them.

Changes

• msm_v2.ts — arch-aware WALKER_TPB from device.limits.maxComputeWorkgroupStorageSize; new walkerTpb MsmConfig override.
• dev/msm-webgpu/main.ts — ?walkertpb= URL knob so the bench can A/B 64 vs 128 on one device.

Why this is correctness-safe (structural argument)

In ba_stream_walker.template.wgsl the walker has no cross-thread sharing and no workgroup barriers (KNOB 1): pref_scratch is sized TPB * S * 2 and each thread touches only pref_scratch[local_id*S*2 .. +S*2); partial slots are 2*(t*S+k)+{0,1} (disjoint per thread); accumulators are private registers. partition_task and the walker are generated from the same WALKER_TPB, so nwg = ceil(num_active/TPB) stays consistent. Changing TPB therefore changes only (a) how the fixed NUM_THREADS=8192 threads pack into workgroups and (b) the symbolic pref_scratch length — never the per-thread computation → bit-identical result across TPB by construction.

Memory

Workgroup-shared only; device buffers unchanged → still within the ≤100 MB budget up to n=2²⁰. 32 KB is exactly the Apple/Adreno cap and 2× Mali's, so the 64 fallback is required there (a 32 KB request fails device validation on Mali).

Validation status — full transparency

Type-check: no type errors from this change (the dev tsconfig's only complaints are pre-existing TS6059 rootDir structural notes affecting every src/ import).

SwiftShader local cross-check (logn 8/10): attempted, BLOCKED by the environment. The WebGPU harness cannot run in this container: the only available browser (Playwright's pre-installed /opt/ms-playwright/chromium-1148, launched via executablePath) does not expose navigator.gpu under any configuration tried — --enable-unsafe-webgpu with --use-vulkan=swiftshader / --use-angle=swiftshader, in both headless=old and headless=new under xvfb. navigator.gpu is undefined in every case (no working Dawn↔SwiftShader WebGPU backend in this build on a GPU-less host). yarn install, @noble/curves, and playwright-core all work; only the WebGPU adapter is missing. Correctness therefore rests on the structural argument above pending real-GPU confirmation.

BrowserStack device numbers: attempted, BLOCKED by seat contention. The shared BrowserStack pool (2 seats across ~10 concurrent agents) was at 2/2 for essentially the entire session (confirmed across ~90 min of escalating 4→10→30 min backoffs). When a seat momentarily freed, staging the cloudflared tunnel and completing the MCP worker handshake before another agent re-claimed the seat did not succeed within the worker-id window. The driver + tunnel are fully working (run-browserstack.mjs, cloudflared installed). One worker created during an attempt was pointed at a stale tunnel URL by mistake and was immediately deleted (browserstack_workers now returns [] — no leaked seats).

To complete the measurements (exact repro, when a seat is free)

cd barretenberg/ts
PATH="/tmp/bin:$PATH" node dev/msm-webgpu/scripts/run-browserstack.mjs \
  --target macos --page index --autorun msm-cross-check --n 10   # correctness
# then ?autorun=msm-bench and A/B ?walkertpb=64 vs 128 on macos / s25-ultra / pixel-9-pro-xl

No device numbers are reported here because none were obtained — I will not publish unverified figures.

Base: stream-walker-impl.