feat(bb/msm): adaptive bucket-accumulate — coop-walker G=1 on Adreno (1.67–2.05× over stream-walker), walker elsewhere

[AztecBot]

Summary

A single adaptive codebase, proven fastest on real Apple, Adreno, and Mali GPUs:

• Adreno (Qualcomm) → coop accumulator with inversion-granularity G=1: measured 1.67–2.05× faster than the stream-walker on a Galaxy S25 Ultra across logN 12/14/16.
• Mali, Apple, and everything else → the stream-walker, which is fastest there (coop regresses).

accum:'auto' (the default) reads the adapter vendor/architecture and selects per device. The choice is data-driven, not a blanket "mobile → coop" — Mali measured opposite to Adreno, so coop is selected only on Adreno.

The earlier "mobile is BLOCKED" conclusion was wrong

The previous revision claimed BrowserStack's real-Android target served a non-WebGPU "stock browser" and that no Adreno/Mali number was obtainable. That was a blind-diagnostic artifact: the cross-check autorun emitted no early heartbeat and the full-size SRS download stalled over the tunnel, so the page looked dead.

Two harness fixes settle it:

• ?autorun=env-probe — posts UA + navigator.gpu + adapter info before any SRS/WebGPU/WASM work.
• ?cfg=<base64> — BrowserStack's /5/worker truncates the device URL at the first &; encoding the whole query as one base64 param survives it.

Result: BrowserStack's "Android Browser" on a Galaxy S25 Ultra is Chrome 145 with full WebGPU on a real Adreno (vendor:qualcomm, architecture:adreno-8xx); the Pixel 9 Pro XL is Chrome 145 on Mali (vendor:arm, architecture:valhall). Both fully measurable.

Real-hardware A/B (BrowserStack, min ms over 8 reps, drift-controlled single page load)

coop G=1 speedup vs stream-walker:

logN	Apple M2 (Chrome 148)	Adreno · S25 Ultra (Chrome 145)	Mali · Pixel 9 Pro XL (Chrome 145)
12	1.16×	1.67×	1.31×
14	0.61×	2.05×	0.87×
16	0.52×	1.85×	0.79×

Full inversion-granularity sweep (speedup vs walker; G = threads sharing one batched inversion):

Adreno · Galaxy S25 Ultra — coop wins, monotonic in G:

logN	walker (ms)	g1	g8	g16	g32
12	52.2	1.67×	1.51×	1.35×	1.14×
14	129.9	2.05×	1.56×	1.13×	0.99×
16	274.9	1.85×	1.49×	1.28×	0.96×

(g64 — the old coop default scan — is worst here and triggered a device-lost at logN≥14.)

Mali · Pixel 9 Pro XL — coop wins only at logN 12, regresses at 14/16:

logN	walker (ms)	g1	g8	g16	g32
12	103.3	1.31×	0.77×	0.59×	0.41×
14	173.7	0.87×	0.45×	0.33×	0.20×
16	352.1	0.79×	0.39×	0.33×	0.18×

Apple M2 — walker wins at the sizes that matter:

logN	walker (ms)	g1	g8	g16	g32	g64
12	15.1	1.16×	0.85×	0.64×	0.44×	1.02×
14	23.4	0.61×	0.41×	0.30×	0.19×	0.54×
16	49.7	0.52×	0.36×	0.26×	0.17×	0.46×

(Apple g64 0.54×/0.46× at 14/16 matches this PR's prior Apple table — same methodology.)

Why G=1, and why Adreno only

coop sets slots-per-thread=1 and shares the affine-add batch inversion across the workgroup. The inversion-granularity knob G controls how many threads share one safegcd:

• G=1 — each thread inverts its own dx: no workgroup memory, no in-loop barriers, one accumulator/thread ⇒ maximal occupancy. But it does ~S× more safegcd inversions than the walker's S-wide batch (no amortisation).
• G=TPB (the old default scan) — one inversion/workgroup, but 2·log2(TPB) barriers/round + a large workgroup-memory footprint.

Adreno is memory-/occupancy-bound enough that G=1's huge occupancy gain hides the extra inversions → a real 1.67–2.05× win. Mali does not hide them (it has fewer resident-workgroup headroom / different inversion throughput), so G=1 wins only at tiny logN 12 and loses at 14/16. Apple (cache-rich, not occupancy-starved) loses outright above logN 12. Hence auto ships coop only on Adreno/Qualcomm; everything else stays on the walker. This avoids a one-axis Mali/Apple regression while capturing the Adreno win.

Correctness

GPU vs @noble/curves, headless SwiftShader: GREEN for both walker and coop, logN 8 and 10 (multiple configs incl. accum:'auto'). All perf runs are cross-checked.

What changed

• env-probe autorun + base64 cfg param (truncation-proof BrowserStack config).
• get_device: stash adapter info under device.__adapterInfo (Chrome's device.adapterInfo is a read-only getter — the prior selector read an always-undefined device.adapterInfo and could never recognise a mobile GPU; this was a latent bug that made the old "adaptive" path inert).
• resolveAccum: auto → coop G=1 on Adreno/Qualcomm (measured win), walker elsewhere; returns the per-device default granularity. Explicit accum/coopG always honoured.

Honest status

• ✅ coop-walker kernel + granularity knob G, correct (noble/SwiftShader), drop-in.
• ✅ Real-HW proven: coop G=1 is 1.67–2.05× over the walker on Adreno (S25, all of logN 12/14/16).
• ✅ Mali + Apple measured → coop regresses at logN≥14 → auto correctly keeps them on the walker (no regression).
• ✅ Adaptive selector verified against real adapter strings (Adreno qualcomm/adreno-8xx, Mali arm/valhall, Apple, SwiftShader).
• Caveats: one device per GPU family; the workgroup-scan mode (G=TPB) is unstable on Adreno (device-lost at logN≥14) so it is never auto-selected; at logN 12 coop-G=1 also edges out the walker on Mali/Apple, but auto conservatively keeps them on the walker to avoid the larger-n regression.