perf(bb/msm): cache dx + idx in ba_fused_super pref_scratch

[AztecBot]

Summary

Adds a fusedSuperOpt knob (default off) to the ba_fused_super_bench kernel — the bin-packed pair-tree bucket-accumulate that runs for every level of the MSM, with level 0 being the largest single dispatch in the pipeline.

The optimisation has two parts:

1. Cache dx in pref_scratch. Forward stores dx[k] alongside prefix[k] (slot grows 8 → 16 u32, PREF_STRIDE 2 → 4). The inverse pass then loads dx_back from the cache instead of re-running 2 × load_active_x + 1 fr_sub.
2. Pre-load chunk_plan indices into a function-local array<u32, S> at kernel entry; forward, inverse, and backward all reuse them instead of re-reading the index buffer.

Both are gated by one Mustache flag (super_opt) so the original code path is byte-identical when the knob is off. Wired through MsmConfig.fusedSuperOpt, the host prefScratchBuf / estimateMem sizing, the shader manager, and a ?superopt=1 URL knob on the dev bench page.

Measurements (n=2^16, c=13, S=8)

Platform	fused baseline	fused super_opt	Δ
macOS Sequoia / M-class	31.3 ms	25.0 ms	−20 %
Samsung S25 Ultra (Adreno 830)	32.3 ms	32.3 ms	~0 %

The macOS win lands across the higher levels (l1+), where the inverse-pass dx recomputation and chunk_plan re-reads are a meaningful fraction of per-thread time. Level 0 itself is unchanged on macOS (~12 ms before/after) — the pool-indexed loads aren't the bottleneck there either.

On Adreno 830 fused_l0 is compute-bound on the safegcd inverse + montgomery products: removing the indirect storage loads doesn't move the needle, and the doubled pref_scratch write traffic exactly cancels the read savings. Default is left off so the macOS-specific savings stay opt-in until a per-device default lands. Enable with ?superopt=1 or config.fusedSuperOpt = true.

Why not split the inverse loop too?

A third optimisation — splitting k==0 out of the inverse loop into an epilogue so the per-iter divergent branch goes away — was implemented but reverted: Metal's WGSL compiler dropped the last inv = mont(inv, dx_back) update when it became loop-tail-with-only-out-of-loop-use, producing a wrong inv_dx[0]. Keeping the if/else branch (where the update sits inside the else block) is treated as load-bearing by the compiler. The if/else is uniform across the wave at runtime so the cost is negligible.

Test plan

• Open ?superopt=0 and ?superopt=1 on the dev bench page (macOS + S25), verify result x matches across runs
• Confirm fused median drops ~20 % on macOS and doesn't regress on S25
• Run the existing ?autorun=msm-cross-check flow on both platforms with ?superopt=1 to confirm the MSM result still cross-checks against the multi-threaded WASM path

---

Created by claudebox · group: slackbot