Cuda Exp 11#575

Draft

MauroToscano wants to merge 38 commits into

mainfrom

cuda/exp-11-merge

Commits on May 2, 2026

feat: GPU-accelerated coset LDE (batched, Goldilocks base field)

Mauro Toscano
authored and
MauroToscano
committed
perf(cuda): rayon-parallel host pack + median-of-10 microbench

Mauro Toscano
authored and
MauroToscano
committed
perf(cuda): ext3 aux-trace LDE via componentwise decomposition

Mauro Toscano
authored and
MauroToscano
committed
perf(cuda): GPU ext3 extend_half_to_lde path (dormant until caller scales up)

Mauro Toscano
authored and
MauroToscano
committed
perf(cuda): GPU ext3 LDE for Round 4 DEEP-poly extension

Mauro Toscano
authored and
MauroToscano
committed
perf(cuda): GPU ext3 evaluate-on-coset for R2 composition parts

Mauro Toscano
authored and
MauroToscano
committed
docs(math-cuda): update NOTES.md with final speedup numbers

Mauro Toscano
authored and
MauroToscano
committed
perf(cuda): GPU Keccak-256 Merkle leaf hashing for main-trace commit

Mauro Toscano
authored and
MauroToscano
committed
perf(cuda): GPU Keccak-256 Merkle commit for aux trace (ext3)

Mauro Toscano
authored and
MauroToscano
committed
docs(math-cuda): update NOTES with post-C1 state and path to 2x

Mauro Toscano
authored and
MauroToscano
committed
feat(cuda): barycentric OOD kernels + ext3 arithmetic building blocks

Lambda Dev
authored and
MauroToscano
committed
feat(cuda): GPU Merkle inner-tree kernel + parity tests

Lambda Dev
authored and
MauroToscano
committed
perf(cuda): fuse LDE + leaf-hash + Merkle tree into one on-device pipeline

Lambda Dev
authored and
MauroToscano
committed
docs(math-cuda): add fib 2M/4M timings after fused tree build

Lambda Dev
authored and
MauroToscano
committed
perf(cuda): GPU Merkle tree for R2 commit_composition_poly

Lambda Dev
authored and
MauroToscano
committed
feat(cuda): FRI layer Merkle tree kernel + parity (infra, unwired)

Lambda Dev
authored and
MauroToscano
committed
docs(math-cuda): update NOTES with post-R2-commit state + next-unlock analysis

Lambda Dev
authored and
MauroToscano
committed
perf(cuda): GPU-resident LDE handles + GPU R4 deep-composition

Lambda Dev
authored and
MauroToscano
committed
docs(math-cuda): NOTES — post-item-4 numbers and hook table

Lambda Dev
authored and
MauroToscano
committed
perf(cuda): R3 OOD barycentric reads LDE from device handles

Lambda Dev
authored and
MauroToscano
committed
perf(cuda): skip CPU trace-slab extraction when GPU R3 OOD handles it

Lambda Dev
authored and
MauroToscano
committed
feat(cuda): FRI fold + twiddle-update kernels (infra, unwired)

Lambda Dev
authored and
MauroToscano
committed
perf(cuda): memcpy + parallel pack of inv_h/inv_t for GPU R4 deep

Lambda Dev
authored and
MauroToscano
committed
docs: update NOTES to 1.52× baseline

Lambda Dev
authored and
MauroToscano
committed
perf(cuda): FRI commit phase fully device-resident

Lambda Dev
authored and
MauroToscano
committed
docs(math-cuda): NOTES — post-FRI-on-device state (1.52×, FRI table entry)

Lambda Dev
authored and
MauroToscano
committed
bench(cuda): add 15-trial fib_1M bench; NOTES at 1.57× (tighter mean)

Lambda Dev
authored and
MauroToscano
committed
perf(cuda): keep composition-parts LDE on device when R2 GPU path fires

Lambda Dev
authored and
MauroToscano
committed
docs(math-cuda): tier-3 analysis — why nothing shipped on this branch

Lambda Dev
authored and
MauroToscano
committed
docs(math-cuda): nsys profile of fib_1M — GPU is not the bottleneck

Lambda Dev
authored and
MauroToscano
committed
perf(cuda): parallel Montgomery batch inverse + R4 DEEP denoms on GPU

Lambda Dev
authored and
MauroToscano
committed
perf(cuda): LogUp aux-build on GPU — full port of compute_logup_batched_term_column

Mauro Toscano
authored and
MauroToscano
committed
docs(cuda): exp-8 checkpoint — scale bench + Zisk/pil2-proofman analysis

Mauro Toscano
authored and
MauroToscano
committed
perf(cuda): exp-9 — serialize GPU LogUp dispatch across tables

Mauro Toscano
authored and
MauroToscano
committed
docs(cuda): exp-11 checkpoint — device-resident main trace design

Mauro Toscano
authored and
MauroToscano
committed
Merge origin/main into cuda/exp-11-device-trace

Mauro Toscano
authored and
MauroToscano
committed
build(math-cuda): pin cudarc to cuda-12080 for forward-compat

Mauro Toscano
authored and
MauroToscano
committed
chore: drop experiment artifacts and profile logs from PR
MauroToscano
committed