feat: GPU trace gen and sppark NTT backend - #81

Merged

samuelburnham merged 34 commits into

mainfrom

sb/trace-sharding-gpu

Oct 2, 2026

Commits on Oct 2, 2026

cuda: hash single-chunk BLAKE3 rows with one thread per message
samuelburnham
committed
ac980a8
cuda: fuse wide NTTs at any height and trim LDE memory passes
samuelburnham
committed
38f7cc0
cuda: per-device prover binding, host pools per device, GPU trace witnesses
samuelburnham
committed
3c87a59
cuda: remove the Merkle tree cache, keep per-construction device admission
samuelburnham
committed
9869234
cuda: hash height groups wider than 32 KiB per row on the host
samuelburnham
committed
9b172ed
cuda: share Goldilocks arithmetic with generated trace kernels
samuelburnham
committed
39d59cb
cuda: admit lookup graphs only for resident LDEs that kept their trace
samuelburnham
committed
d4d09c7
cuda: let trace generators keep device state, released after the lookup job
samuelburnham
committed
89747d3
cuda: spans for LDE shapes, commit sources and the FRI opening phases
samuelburnham
committed
2ec82fa
cuda: cache the FRI coset per process; double-buffer staged uploads
samuelburnham
committed
a1fcbf5
cuda: collect lightweight prover operation metrics
samuelburnham
committed
82c6c9b
cuda: sppark's Goldilocks NTT as an optional comparison backend
samuelburnham
committed
1691727
cuda: build sppark's runtime without exceptions or threads
samuelburnham
committed
fa64c28
cuda: the sppark fork's dev branch carries the runtime mode
samuelburnham
committed
aa714fd
cuda: pin sppark at the fork's dev branch, 6c5d826
samuelburnham
committed
d964550
cuda: sppark adapter resolves the device by ordinal and fences before releasing its stream
samuelburnham
committed
44ea38c
cuda: resident coset LDE through sppark behind a runtime switch
samuelburnham
committed
9fef499
cuda: admit the sppark panel scratch, dispatch by height, count shapes per backend
samuelburnham
committed
62cee20
cuda: every transform of a proof through sppark above the height threshold
samuelburnham
committed
24ef583
cuda: the sppark adapter parses its settings without strtoul
samuelburnham
committed
a15a979
cuda: sppark dispatch by whole shape, admitted everywhere, labelled in spans
samuelburnham
committed
4739cb7
cuda: sppark panels transform their columns in batched launch sequences
samuelburnham
committed
9c51794
cuda: sppark height threshold at 2^18, short shapes in the resident bench
samuelburnham
committed
0ac94c1
cuda: sppark panel glue as tiled transposes, the fused expansion measured
samuelburnham
committed
e7be72c
README: the sppark transform backend and its settings
samuelburnham
committed
1326c65
README: AIUR_METRICS is a path for Ix, any value for the counters
samuelburnham
committed
fde18f8
build: pin sppark at the fork's batched entry and patch record
samuelburnham
committed
41fd44c
cuda: one dispatch decision per transform, tables built for that backend only
samuelburnham
committed
ce77de2
cuda: the sppark panels' scratch comes from the stream's memory pool
samuelburnham
committed
b6104b6
Use sppark for all CUDA transforms
samuelburnham
committed
9c62a10
Enable device snapshots with event-only tracing filters
samuelburnham
committed
bc864c1
Validate transform plans against CUDA buffer dimensions
samuelburnham
committed
7eae079
Format src/cuda/mmcs.rs with rustfmt
samuelburnham
committed
6516490
Drop the unused metrics layer, debug knobs and experiment docs
samuelburnham
committed
c9532e3