Add optional CUDA-accelerated PLONK prover for BLS12-381 - #1792

Open

seunlanlege wants to merge 23 commits into

Consensys-Incorporated:masterfrom

polytope-labs:gpu-base

Commits on Jun 27, 2026

backend/plonk/bls12-381: device-resident GPU rho-loop for the quotient numerator (icicle/CUDA)
seunlanlege
committed
a218ac4
Add CUDA-accelerated PLONK backend: registration hooks + internal/gpu cgo bindings
seunlanlege
committed
a48f6dc
Device-resident GPU PLONK prover (p2 layer); build against upstream gnark-crypto
seunlanlege
committed
4444d17
Remove dead code: cuda hook backend, gpu.go FFT wrappers, prove cruft, test-only p2 toolkit
seunlanlege
committed
0b18fdd
Overlap SRS point-conversion + NTT warmup with the CPU witness solve
seunlanlege
committed
df4c308
Fold quotient shards on-device from the resident quotient (no host fold, no hFolded upload)
seunlanlege
committed
52243a7
Cache the circuit-fixed trace on the proving key and clone it per-proof
seunlanlege
committed
18cd433
Cache canonical base Qk on the proving key to skip the per-proof linearized iFFT
seunlanlege
committed
9c4b8c2
Overlap the linearized S1/S2 evaluations and parallelize the qcp accumulation
seunlanlege
committed
c8ba3d8
Add device opening primitives + resident batch-open (eval/fold/divide/commit on device handles)
seunlanlege
committed
e730c89
Run batchOpening evals + gamma-fold on device (upload path), verified shadow-match
seunlanlege
committed
b424725
Keep wires resident through the openings: device l/r/o(zeta) + resident batchOpening, skip the wire download
seunlanlege
committed
715a92a
Feed linearized dS3 from the resident canonical S3 so S1/S2/S3 stay resident (skip their download)
seunlanlege
committed
13c3f95
Feed linearized Ql/Qr/Qm/Qo from resident canonical buffers; skip all selector downloads
seunlanlege
committed
342b411
Recycle per-proof trace clones (pool + cloneInto); batch the openings/linearized poly evals
seunlanlege
committed
cdbbf87
Generate the quotient's per-coset scale table on-device (gpu_vec_powers + bitreverse)
seunlanlege
committed
4a0e645

Commits on Jun 28, 2026

Commits on Jul 6, 2026

Merge branch 'Consensys:master' into gpu-base
seunlanlege
authored
f329a0b
go mod tidy: drop orphaned gnark-crypto go.sum entry
seunlanlege
committed
5be470e

Navigation Menu

Uh oh!

Add optional CUDA-accelerated PLONK prover for BLS12-381 - #1792

Add optional CUDA-accelerated PLONK prover for BLS12-381#1792
seunlanlege wants to merge 23 commits into
Consensys-Incorporated:masterfrom
polytope-labs:gpu-base

Commits on Jun 27, 2026

backend/plonk/bls12-381: device-resident GPU rho-loop for the quotient numerator (icicle/CUDA)

Add CUDA-accelerated PLONK backend: registration hooks + internal/gpu cgo bindings

Device-resident GPU PLONK prover (p2 layer); build against upstream gnark-crypto

Remove dead code: cuda hook backend, gpu.go FFT wrappers, prove cruft, test-only p2 toolkit

Overlap SRS point-conversion + NTT warmup with the CPU witness solve

Fold quotient shards on-device from the resident quotient (no host fold, no hFolded upload)

Cache the circuit-fixed trace on the proving key and clone it per-proof

Cache canonical base Qk on the proving key to skip the per-proof linearized iFFT

Overlap the linearized S1/S2 evaluations and parallelize the qcp accumulation

Add device opening primitives + resident batch-open (eval/fold/divide/commit on device handles)

Run batchOpening evals + gamma-fold on device (upload path), verified shadow-match

Keep wires resident through the openings: device l/r/o(zeta) + resident batchOpening, skip the wire download

Feed linearized dS3 from the resident canonical S3 so S1/S2/S3 stay resident (skip their download)

Feed linearized Ql/Qr/Qm/Qo from resident canonical buffers; skip all selector downloads

Recycle per-proof trace clones (pool + cloneInto); batch the openings/linearized poly evals

Generate the quotient's per-coset scale table on-device (gpu_vec_powers + bitreverse)

Commits on Jun 28, 2026

Remove GPU prover debug/tracing scaffolding (timings, traces, shadow checks, dev toggles)

Pre-merge cleanup: remove dead code + cuda-path CPU fallbacks (GPU-only); lazy-init via SetDevice

Strict GPU-or-die: cuda device errors panic instead of falling back to CPU

Pre-merge fixes: restore SetRandom error check; abort on resident-openings structural miss

Merge pull request #1 from polytope-labs/gpu-plonk-prover

Commits on Jul 6, 2026

Merge branch 'Consensys:master' into gpu-base

go mod tidy: drop orphaned gnark-crypto go.sum entry