perf(prover-ray): 2x faster FRI commit by gbotrel · Pull Request #3749 · LFDT-Lineth/lineth-monorepo

gbotrel · 2026-08-07T16:33:07Z

FRI commit is now faster than native Plonky3 (AVX-512, Rayon, fat LTO) in every measured configuration. Roots bit-identical.

shape, rate 2	before	after	plonky3
2^20 x 16	77.9 ms	31.8 ms	65.7 ms
2^16 x 256	17.1 ms	12.0 ms	25.0 ms
2^12 x 4096	17.9 ms	10.9 ms	18.5 ms
10 GiB (2^22 x 640)	2.67 s	1.40 s	1.76 s

96 cores, cold one-shot processes, medians of 5.

What:

Parallel Merkle levels + gnark-crypto's 16-lane state-seeded Poseidon2 compression (gnark-crypto#871, pinned commit 000c30b45f17).
Field-specific COBRA bit-reversal selection over gnark-crypto's generic helpers.
Rate-2 LDE: first half of the bit-reversed codeword is bitrev(input); second half is one half-size coset FFT.
Leaves hash directly into tree node storage; codewords slab-allocated per size.
Opt-in WithConsumeWitness + GOMEMLIMIT: 10 GiB peak RSS 31.6 -> ~23 GiB.

Checklist

I wrote new tests for my new core changes.
Focused tests, purego tests, formatting, and golangci-lint pass locally.
If this change is deployed to any environment (including Devnet), E2E test coverage exists or is included in this PR.
No breaking API changes.

Copilot

Pull request overview

This PR accelerates prover-ray’s KoalaBear FRI commitment path (encode + Merkle) via SIMD-batched Poseidon2 internal-node hashing, cache-friendly bit-reversal, a rate-2 LDE fast path, and reduced cold-start allocation overhead, while keeping commitment roots bit-identical to the scalar reference implementation.

Changes:

Add a 16-lane Poseidon2 compression-chain kernel (AVX-512 with purego fallback) and use it to batch hash internal Merkle tree levels in parallel.
Introduce COBRA-tiled bit-reversal for 4-byte field elements and a rate-2 encoder fast path (EncodeInto / EncodeExtInto) that avoids a full-size FFT.
Reduce allocation overhead via slab-allocated codeword buffers and add optional witness consumption (WithConsumeWitness) plus expanded benchmark tooling/docs.

Reviewed changes

Copilot reviewed 16 out of 16 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
prover-ray/crypto/koalabear/poseidon2/poseidon2_batch.go	Adds public batched Poseidon2 compression-chain API + generic reference implementation.
prover-ray/crypto/koalabear/poseidon2/poseidon2_batch_test.go	Tests batched kernel vs scalar compression chains (including generic fallback).
prover-ray/crypto/koalabear/poseidon2/poseidon2_batch_noasm.go	Purego/!amd64 fallback wiring to the generic implementation.
prover-ray/crypto/koalabear/poseidon2/poseidon2_batch_amd64.s	AVX-512 assembly kernel for 16-lane state-seeded Poseidon2 chain compression.
prover-ray/crypto/koalabear/poseidon2/poseidon2_batch_amd64.go	amd64 dispatcher + AVX-512 feature gate and glue symbols for the asm kernel.
prover-ray/crypto/koalabear/fri/tree.go	Refactors tree construction and batches internal level hashing (with parallelism threshold).
prover-ray/crypto/koalabear/fri/tree_test.go	Adds scalar-reference equivalence tests covering aux/non-aux and parallel paths.
prover-ray/crypto/koalabear/fri/reedsolomon.go	Adds rate-2 coset fast path and `EncodeInto`/`EncodeExtInto` to support slab allocation.
prover-ray/crypto/koalabear/fri/reedsolomon_fastpath_test.go	Verifies rate-2 fast path matches the generic path for base/ext columns.
prover-ray/crypto/koalabear/fri/PERFORMANCE.md	Documents the optimization session, benchmark protocol, and remaining work.
prover-ray/crypto/koalabear/fri/pcs.go	Extends PCS.Commit to accept commit options.
prover-ray/crypto/koalabear/fri/fri.go	Zero-copy tree bottom for ext-only trees (fills tree nodes directly in parallel).
prover-ray/crypto/koalabear/fri/commitment.go	Adds `CommitOption`/`WithConsumeWitness`, slab codeword allocation, and zero-copy Merkleize bottom.
prover-ray/crypto/koalabear/fri/bitreverse.go	Adds COBRA-tiled bit reversal + out-of-place bit-reversed copy helper.
prover-ray/crypto/koalabear/fri/bitreverse_test.go	Correctness tests and a benchmark for the bit-reversal implementations.
prover-ray/crypto/koalabear/fri/benchmarks/2026-08-07/summary.csv	Captures baseline benchmark summary data.
prover-ray/crypto/koalabear/fri/benchmarks/2026-08-07/samples.csv	Captures baseline benchmark sample data.
prover-ray/crypto/koalabear/fri/benchmarks/2026-08-07/README.md	Baseline benchmark description and reproduction notes.
prover-ray/crypto/koalabear/fri/benchmarks/2026-08-07/plonky3-commit-shapes.rs	Plonky3 harness used for baseline measurements.
prover-ray/crypto/koalabear/fri/benchmarks/2026-08-07/plonky3-cargo.patch	Notes required Plonky3 dev-deps for consistent Rayon enablement.
prover-ray/crypto/koalabear/fri/benchmarks/2026-08-07/memory.csv	Baseline RSS + user/system timing captures.
prover-ray/crypto/koalabear/fri/benchmarks/2026-08-07/environment.txt	Baseline environment/toolchain manifest.
prover-ray/crypto/koalabear/fri/benchmarks/2026-08-07-parallel-tree/summary.csv	Intermediate benchmark summary for parallel-tree stage.
prover-ray/crypto/koalabear/fri/benchmarks/2026-08-07-parallel-tree/samples.csv	Intermediate benchmark samples for parallel-tree stage.
prover-ray/crypto/koalabear/fri/benchmarks/2026-08-07-parallel-tree/README.md	Intermediate stage benchmark write-up and reproduction steps.
prover-ray/crypto/koalabear/fri/benchmarks/2026-08-07-parallel-tree/prover-ray.patch	Patch snapshot used for intermediate measurements.
prover-ray/crypto/koalabear/fri/benchmarks/2026-08-07-parallel-tree/plonky3-commit-shapes.rs	Plonky3 harness used for intermediate measurements.
prover-ray/crypto/koalabear/fri/benchmarks/2026-08-07-parallel-tree/plonky3-cargo.patch	Plonky3 Cargo changes for intermediate comparison.
prover-ray/crypto/koalabear/fri/benchmarks/2026-08-07-parallel-tree/memory.csv	Intermediate RSS captures.
prover-ray/crypto/koalabear/fri/benchmarks/2026-08-07-parallel-tree/environment.txt	Intermediate environment/toolchain manifest.
prover-ray/crypto/koalabear/fri/benchmarks/2026-08-07-optimized/summary.csv	Optimized benchmark summary data.
prover-ray/crypto/koalabear/fri/benchmarks/2026-08-07-optimized/samples.csv	Optimized benchmark sample data + root markers.
prover-ray/crypto/koalabear/fri/benchmarks/2026-08-07-optimized/README.md	Optimized benchmark description and protocol notes.
prover-ray/crypto/koalabear/fri/benchmarks/2026-08-07-optimized/merkle-only.csv	Merkle-only and encode-only diagnostics for optimized stage.
prover-ray/crypto/koalabear/fri/benchmarks/2026-08-07-optimized/environment.txt	Optimized environment/toolchain manifest.
prover-ray/crypto/koalabear/fri/bench/main.go	Bench harness: adds phase selection, JSON output, consume-input mode, and CPU profiling.
prover-ray/crypto/koalabear/fri/bench_test.go	Adds commit/encode/merkle shape benchmarks and configuration via env var.

YaoJGalteland

have you compared the performance across different FRI rates, e.g. 2, 4, 8, 16?

I remember we used rate 16 for Vortex before. Is that also a plausible parameter for FRI, or is the practical range more constrained here?

AVX-512 kernel derived from gnark-crypto's permutation16x16xN_columns with the Merkle-Damgard chain state loaded from memory instead of zeroed, so one call computes 16 direct C(left,right) compressions or C(C(left,right),aux) chains. Bit-identical to scalar Compress (lane-by-lane test, purego fallback included). Candidate for upstreaming to gnark-crypto. Signed-off-by: Gautam Botrel <gautam.botrel@gmail.com>

Full-commit medians vs native AVX-512+Rayon Plonky3, 96 cores, cold one-shot (was: pre-branch serial-tree baseline): 2^20x16 rate 2: 77.9 ms -> 31.8 ms (P3: 65.7 ms) 2^16x256 rate 2: 17.1 ms -> 12.0 ms (P3: 25.0 ms) 2^12x4096 rate 2: 17.9 ms -> 10.9 ms (P3: 18.5 ms) 10 GiB rate 2: 2.67 s -> 1.40 s (P3: 1.76 s) - Parallelize internal Merkle tree levels; hash them 16 nodes per call with the batched state-seeded Poseidon2 kernel, writing parents in place (scalar hashNode was 55% of the tall-shape commit). - COBRA cache-friendly bit-reversal for 4-byte elements; gnark's utils.BitReverse is always naive for sizeof<8 and burned 43% of the 4 GiB commit CPU. - Rate-2 LDE fast path: the first half of the bit-reversed codeword is exactly bitrev(input), so encode is one bit-reversed copy plus a half-size coset FFT (DIT inverse + DIF-on-coset). No full-size FFT, no standalone coefficient bit-reversal. - Hash bottom Merkle leaves directly into tree node storage (removes a leaf-array alloc+copy the size of the encoded bottom table). - Slab-allocate codewords per size bucket (EncodeInto): per-column large-object allocation under 96 threads spent 25 s in the kernel page-fault path on cold commits (372 ms -> 17.5 ms on rate-4 wide). - Opt-in WithConsumeWitness to release plaintext columns during encode; with GOMEMLIMIT this cuts 10 GiB commit peak RSS from 31.6 to ~23 GiB. Commitment roots are bit-identical throughout: deterministic root markers unchanged, fast path tested against the generic path, tree levels against a scalar reference. Signed-off-by: Gautam Botrel <gautam.botrel@gmail.com>

PERFORMANCE.md: current results vs native Plonky3, what changed and why, the cold-allocation measurement trap, remaining bottlenecks and next experiments, benchmark protocol, correctness gates. benchmarks/: raw samples, medians, environment manifests for the generic-P3 baseline, the parallel-tree baseline, and the optimized run. Signed-off-by: Gautam Botrel <gautam.botrel@gmail.com>

Signed-off-by: Gautam Botrel <gautam.botrel@gmail.com>

gbotrel · 2026-08-11T00:52:20Z

have you compared the performance across different FRI rates, e.g. 2, 4, 8, 16?

I remember we used rate 16 for Vortex before. Is that also a plausible parameter for FRI, or is the practical range more constrained here?

On AMD EPYC 9R45, for the fixed 64 MiB tall/balanced/wide workloads, rate 2 measured 31.8/12.0/10.9 ms and rate 4 measured 44.8/20.7/19.2 ms. I haven’t measured rates 8 or 16. We can tune the rate later once we benchmark representative production workflows

AlexandreBelling · 2026-08-11T12:54:22Z

+	gutils "github.com/consensys/gnark-crypto/utils"
+)
+
+func bitReverse[T any](v []T) {


Nit: seems a bit superfluous to have private wrappers for theses.

so this is the result of a round trip of experiments ; following @YaoJGalteland review I reverted some optimizations, I will need these when we tune them depending on the size and target architecture 👍

AlexandreBelling · 2026-08-11T12:56:16Z

You have conflicts. Let me know when they are fixed. Otherwise, I do not see any blocker.

…kle-tree Signed-off-by: Gautam Botrel <gautam.botrel@gmail.com> # Conflicts: # prover-ray/crypto/koalabear/fri/tree.go # prover-ray/go.mod # prover-ray/go.sum

gbotrel · 2026-08-11T16:43:31Z

tests are failing independently from this PR changes , otherwise ready to merge

Copilot AI lite review requested due to automatic review settings August 7, 2026 16:33

Copilot started reviewing on behalf of gbotrel August 7, 2026 16:33 View session

Copilot AI reviewed Aug 7, 2026

View reviewed changes

Comment thread prover-ray/crypto/koalabear/fri/commitment.go

gbotrel commented Aug 7, 2026

View reviewed changes

Comment thread prover-ray/crypto/koalabear/fri/commitment.go Outdated

gbotrel requested review from AlexandreBelling and YaoJGalteland August 7, 2026 17:20

YaoJGalteland reviewed Aug 10, 2026

View reviewed changes

Comment thread prover-ray/crypto/koalabear/fri/fri.go

Comment thread prover-ray/crypto/koalabear/fri/fri.go

Comment thread prover-ray/crypto/koalabear/fri/bitreverse.go Outdated

Comment thread prover-ray/crypto/koalabear/fri/bitreverse.go Outdated

gbotrel added 10 commits August 11, 2026 00:48

style: remove benchmark results

7c6acce

Signed-off-by: Gautam Botrel <gautam.botrel@gmail.com>

style: remove doc file

cc93548

Signed-off-by: Gautam Botrel <gautam.botrel@gmail.com>

refactor(prover-ray): use upstream FRI primitives

13c6263

Signed-off-by: Gautam Botrel <gautam.botrel@gmail.com>

build(prover-ray): repin gnark-crypto

1bfd7a2

Signed-off-by: Gautam Botrel <gautam.botrel@gmail.com>

style: remove verbose tests

15152a7

Signed-off-by: Gautam Botrel <gautam.botrel@gmail.com>

chore: update to latest gnark-crypto

5be8773

Signed-off-by: Gautam Botrel <gautam.botrel@gmail.com>

fix(prover-ray): address FRI review feedback

2a57992

Signed-off-by: Gautam Botrel <gautam.botrel@gmail.com>

gbotrel force-pushed the perf/fri-parallel-merkle-tree branch from 94acd38 to 2a57992 Compare August 11, 2026 00:48

gbotrel requested a review from YaoJGalteland August 11, 2026 00:53

AlexandreBelling previously approved these changes Aug 11, 2026

View reviewed changes

AlexandreBelling reviewed Aug 11, 2026

View reviewed changes

Merge remote-tracking branch 'origin/main' into perf/fri-parallel-mer…

b1bdee0

…kle-tree Signed-off-by: Gautam Botrel <gautam.botrel@gmail.com> # Conflicts: # prover-ray/crypto/koalabear/fri/tree.go # prover-ray/go.mod # prover-ray/go.sum

gbotrel dismissed AlexandreBelling’s stale review via b1bdee0 August 11, 2026 15:51

gbotrel enabled auto-merge (squash) August 11, 2026 15:52

YaoJGalteland approved these changes Aug 11, 2026

View reviewed changes

AlexandreBelling approved these changes Aug 12, 2026

View reviewed changes

gbotrel merged commit ae9326e into main Aug 12, 2026
38 of 39 checks passed

gbotrel deleted the perf/fri-parallel-merkle-tree branch August 12, 2026 01:30

Navigation Menu

Uh oh!

perf(prover-ray): 2x faster FRI commit - #3749

perf(prover-ray): 2x faster FRI commit#3749
gbotrel merged 11 commits into
mainfrom
perf/fri-parallel-merkle-tree

gbotrel commented Aug 7, 2026 •

edited

Loading

Uh oh!

Copilot AI left a comment

Uh oh!

Uh oh!

Uh oh!

YaoJGalteland left a comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

gbotrel commented Aug 11, 2026

Uh oh!

AlexandreBelling Aug 11, 2026

Uh oh!

gbotrel Aug 11, 2026

Uh oh!

AlexandreBelling commented Aug 11, 2026

Uh oh!

gbotrel commented Aug 11, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

Uh oh!

Conversation

gbotrel commented Aug 7, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Checklist

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Uh oh!

Uh oh!

YaoJGalteland left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

gbotrel commented Aug 11, 2026

Uh oh!

AlexandreBelling Aug 11, 2026

Choose a reason for hiding this comment

Uh oh!

gbotrel Aug 11, 2026

Choose a reason for hiding this comment

Uh oh!

AlexandreBelling commented Aug 11, 2026

Uh oh!

gbotrel commented Aug 11, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

gbotrel commented Aug 7, 2026 •

edited

Loading