[flyq]

Summary

Ports PR #3 of suryuhh/salt-optimizations (commit 35436e1): the variable-base MSM of proof verification, eight-wide on AVX-512 IFMA.

multi_scalar_mul runs a Pippenger whose bucket accumulation and reduction use vector lanes when the host has AVX-512 IFMA and the MSM has at least 8,192 points. Smaller MSMs and other hosts keep the scalar path, so only large witnesses change.

Stacked on the batch-decode PR (liquan/perf/batch-witness-decode). Nothing under salt/src changes.

Changes on top of the external commit

• One cutoff for every thread count. The external dispatch used 8,192 points in a multi-thread rayon pool and 128 in a one-thread pool, which also sent the IPA's fixed 274-point MSM to the vector path. It is 8,192 everywhere now, so the path depends on input size and CPU alone. The single-thread cutoff has not been measured here and can be revisited with numbers.
• The bucket reduction starts at the highest non-empty bucket. It walked all 2^c - 1 buckets per group of windows with two full additions each. A signed digit's magnitude is at most 2^(c-1), so the upper half is never written and the walk added the identity to the identity there. Not measured.

Testing

Passing on this head, on an Apple Silicon machine:

• cargo fmt --all -- --check; cargo clippy --all-targets -- -D warnings, native and --target x86_64-unknown-linux-gnu
• cargo test; with test-bucket-resize (CI's settings); the random stress test; --no-default-features with and without test-bucket-resize
• cargo check -p salt --target riscv64imac-unknown-none-elf --no-default-features
• cargo check --all-targets --target i686-unknown-linux-gnu
• Suppression hygiene. The mutation gates have nothing to do: no salt/src change.

Not run:

• Any AVX-512 path, including the reduction change above. No such CPU or emulator on the machine used; the module was only compiled and linted for x86_64. lane_msm_matches_ark_msm checks the result against arkworks for every window size and runs in the Test Suite (AVX-512 IFMA) job.
• Benchmarks. bench_lane_msm (ignored test) gives ns per point on a host with IFMA.

Merging

Please keep this trailer in the squash message:

Co-authored-by: suryuhh <suryuhh@users.noreply.github.com>

🤖 Generated with Claude Code

[github-actions]

Mutation testing - PASS

Nothing to test: no viable mutants were generated.

[github-actions]

Mutation testing - PASS

Nothing to test: no viable mutants were generated.

[codecov]

Codecov Report

❌ Patch coverage is 5.33141% with 657 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
banderwagon/src/ifma_msm.rs	0.49%	599 Missing and 1 partial ⚠️
banderwagon/src/ifma.rs	0.00%	47 Missing ⚠️
banderwagon/src/element.rs	77.27%	9 Missing and 1 partial ⚠️

Files with missing lines	Coverage Δ
banderwagon/src/lib.rs	70.00% <ø> (ø)
banderwagon/src/element.rs	79.72% <77.27%> (-12.28%)	⬇️
banderwagon/src/ifma.rs	1.54% <0.00%> (-89.59%)	⬇️
banderwagon/src/ifma_msm.rs	0.49% <0.49%> (ø)

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[github-actions]

Performance Benchmark Comparison

Compared 5 benchmark(s) against the latest main baseline.

Detailed Comparison

Benchmark	Baseline Throughput (Kelem/s)	New Throughput (Kelem/s)	Change
update 10000 KVs/1 threads	83.52	81.66	-2.23%
update 10000 KVs/2 threads	160.32	158.24	-1.30%
update 10000 KVs/4 threads	299.94	294.70	-1.75%
update 10000 KVs/8 threads	490.69	477.45	-2.70%
update 10000 KVs/16 threads	617.42	596.16	-3.44%