logup parallelism and twiddle deduplication#518

Closed

diegokingston wants to merge 18 commits into

mainfrom

investigate/prover-performance

Commits on Apr 21, 2026

docs: add prover parallelism improvement plan (6 tasks)
diegokingston
committed
perf: parallelize LogUp fingerprint computation with rayon
diegokingston
committed
perf: parallelize table_contribution sum with rayon reduce
diegokingston
committed
perf: chunked parallel batch inverse for LogUp fingerprints
diegokingston
committed
revert: undo LogUp parallelization (caused +3.9% regression due to nested Rayon over-subscription)
diegokingston
committed
perf: fused chunk-local LogUp processing with parallel prefix sum
diegokingston
committed
perf: optimize FRI fold arithmetic and leaf construction
diegokingston
committed
perf: eliminate per-row heap allocation in Merkle leaf hashing
diegokingston
committed
revert: undo FRI fold twiddle precomputation (caused regression, needs isolated benchmarking)
diegokingston
committed
revert: undo per-row buffer reuse in Merkle hashing (allocation was not the bottleneck, Keccak dominates)
diegokingston
committed
fix: extract fingerprint computation closure to reduce cfg duplication in chunked LogUp
diegokingston
committed
perf: skip Rayon for small Merkle trees (< 1024 nodes)
diegokingston
committed
perf: deduplicate Domain + LdeTwiddles across tables by (trace_length, blowup)
diegokingston
committed
chore: remove outdated parallelism plan (approach changed)
diegokingston
committed
fix: resolve clippy warnings (div_ceil, loop variable, complex type)
diegokingston
committed
style: cargo fmt --all
diegokingston
committed

Commits on Apr 22, 2026

fix: update run_debug_checks to accept Arc<Domain> and Arc<LdeTwiddles>
diegokingston
committed

Commits on Apr 23, 2026

Merge branch 'main' into investigate/prover-performance
diegokingston
authored