perf(bb/msm): streamed-Yuval karat-stream montmul knob

[AztecBot]

Summary

Adds a MsmConfig.streamedYuval switch that renders the Karatsuba+Yuval montgomery_product body with reduce iters 0..9 inlined between the lo and hi/cr half-products. After lo closes, t[0..9] are final (lo writes t[k]+= for k=0..18; hi and cr only touch t[10..38]), so iters 0..9 can run there and t[0..9] die before hi/cr's 27-schoolbook + ~15-limb live set lands — roughly 10 fewer accumulator GPRs alive through the high-pressure half-products.

Off by default. Threaded through ShaderManager via a new optional streamed_yuval ctor param; only the production MSM ShaderManager in MsmV2.create flips it. Same arithmetic and same template — multiply_body carries the inlined iters and the mustache yuval_iters section starts at i=10 instead of 0.

Measurements

Workload	Device	Baseline	Streamed	Δ
Field-mul microbench (n=2^18 × k=100)	S25 Ultra / Chrome 148 / Adreno 750	48.3 ms	47.1 ms	−2.5 %
Full MSM n=2^16, c=13 (5 reps median)	S25 Ultra / Chrome 148 / Adreno 750	65.4 ms	60.7 ms	−7.2 %
Full MSM n=2^16, c=13 (5 reps min)	S25 Ultra / Chrome 148 / Adreno 750	58.0 ms	59.0 ms	flat (noise)
Full MSM n=2^16, c=13 (3 reps median)	M2 / Sequoia / Chrome 148 / Metal	~40.0 ms	40.4 ms	flat
Adreno shader compile (all MSM pipelines)	S25 Ultra	5712 ms	5554 ms	−2.8 %

Adreno is the only target where the register-pressure win materialises — Metal is comfortably under the GPR cliff already, so the flat M2 result is expected.

Correctness

Validated bit-exact against the host BigInt reference at 64 random scalar/point pairs on both Metal (M2) and Adreno (S25) paths, both via the field-mul microbench (bench-field-mul.html) and via the dev-page MSM cross-check. Identical x coordinate, identical y.

Implementation notes

• shader_manager.ts: new optional 5th ctor arg streamed_yuval = false; passed through to renderKaratYuvalMont({ streamed: streamed_yuval }). The renderer now accepts { streamed?: boolean } and, when streamed, emits 10 inline { t_mask, carry; t[i+1..i+20] += ... } blocks after the lo half-product. The mustache template's yuval_iters section is rendered with yuvalStart = 10 so the remaining iters (10..18), standard reduce, and final drain are unchanged.
• msm_v2.ts: MsmConfig.streamedYuval field; threaded to the single MsmV2.create ShaderManager instantiation that compiles the MSM pipelines (the MsmV2Pool shader manager only compiles the point-conversion shader and doesn't reference mont_product_src, so I left that call site alone).

Why on Adreno specifically

The S25's Adreno 750 has a tight GPR-per-wave budget; the karat (grouped Karatsuba) body sits in a regime where the cr block's ~67-GPR live set lives right at the edge of half-occupancy. Killing the 10 t[0..9] slots before cr runs drops the peak by ~10 GPRs, giving the register allocator slack the scheduler can use. M2's Metal driver allocates ~2× the registers per wave at this size and isn't binding on this kernel — the streamed body is a no-op there, which is exactly what the numbers show.

How to A/B locally

Production callers pick up the knob via MsmConfig.streamedYuval = true. The bench page wiring + autorun=msm-bench harness used to produce these numbers lives on zw/msm-webgpu-experiments-v2 ahead of this commit; I can land it as a separate follow-up PR if you want the bench autorun + base64-cfg URL scheme in tree too.

Test plan

• Land on zw/msm-webgpu-experiments-v2; rebench n=2^18 / n=2^20 on S25 before flipping the production default
• Spot-check Metal/Intel paths show no regression (expected: flat, M2 was already saturated)
• If the n=2^18+ numbers also hold, flip MsmConfig.streamedYuval default to true in MsmV2.create (separate PR)

---

Created by claudebox · group: slackbot

[AztecBot]

Automatically closing this stale claudebox draft PR (no updates for 5+ days). Re-open if still needed.