feat: GPU trace gen and sppark NTT backend

Cargo.toml

@@ -5,6 +5,7 @@ edition = "2024"
 authors = ["Argument Engineering <engineering@argument.xyz>"]
 license = "MIT OR Apache-2.0"
 rust-version = "1.98"
+links = "multi_stark_cuda"
 
 # See more keys and their definitions at https://doc.rust-lang.org/cargo/reference/manifest.html
 
@@ -32,6 +33,14 @@ p3-merkle-tree = { git = "https://github.com/Plonky3/Plonky3", rev = "3152b14a89
 p3-symmetric = { git = "https://github.com/Plonky3/Plonky3", rev = "3152b14a89067c83775a8076cc262ffc48a1fd7c" }
 p3-util = { git = "https://github.com/Plonky3/Plonky3", rev = "3152b14a89067c83775a8076cc262ffc48a1fd7c" }
 
+# The CUDA transform engine. The fork supplies column batching, borrowed
+# streams and a runtime compatible with the Lean executable's C++ linkage.
+[dependencies.sppark]
+git = "https://github.com/argumentcomputer/sppark"
+rev = "e10e107673aa22861f0f8b9758fc62169ab919ae"
+optional = true
+features = ["cuda"]
+
 [dev-dependencies]
 criterion = "0.5"
 p3-baby-bear = { git = "https://github.com/Plonky3/Plonky3", rev = "3152b14a89067c83775a8076cc262ffc48a1fd7c" }
@@ -45,10 +54,9 @@ harness = false
 
 [features]
 parallel = ["p3-maybe-rayon/parallel"]
-# Use the first-party CUDA Goldilocks DFT/LDE backend. Enabling this feature
-# requires a CUDA toolkit at build time and an NVIDIA GPU at runtime; the
-# default CPU build never invokes nvcc or links the CUDA runtime.
-cuda = ["dep:itertools", "dep:rayon"]
+# CUDA requires the toolkit at build time and an NVIDIA GPU at runtime.
+# CPU builds neither compile sppark nor link CUDA.
+cuda = ["dep:itertools", "dep:rayon", "dep:sppark"]
 
 # Similar to `release`, but preserves debug info
 [profile.dev-ci]

README.md

@@ -17,7 +17,7 @@ lookup arguments for shared state.
   a batteries-included Goldilocks/Blake3 instantiation is provided
 - **Serialization** — `Proof::to_bytes` / `Proof::from_bytes` via bincode
 - **Parallel proving** — opt-in via the `parallel` feature flag
-- **CUDA transforms** — opt-in first-party Goldilocks DFT/LDE backend via the
+- **CUDA transforms** — opt-in sppark Goldilocks DFT/LDE backend via the
   `cuda` feature; normal builds remain independent of CUDA
 
 ## Reference configuration
@@ -80,7 +80,7 @@ are enabled by default via `.cargo/config.toml`.
 ## CUDA acceleration
 
 The optional `cuda` feature routes the production Goldilocks configuration's
-PCS and quotient transforms through first-party CUDA kernels. CUDA and CPU
+PCS and quotient transforms through sppark. CUDA and CPU
 proofs generated by the same revision are byte-identical and use the same CPU
 verifier. Canonical field serialization changes newly generated CPU proof bytes
 relative to revisions before this backend; previously generated proofs remain
@@ -104,6 +104,29 @@ hardware and downstream Ix measurements live in
 See [cuda/README.md](cuda/README.md) for architecture, build settings, the NVIDIA
 correctness harness, platform limitations, and benchmark commands.
 
+### sppark transforms
+
+Every CUDA build uses [sppark](https://github.com/argumentcomputer/sppark)'s
+Goldilocks NTT, pinned to the fork's `dev` branch. Its borrowed streams keep
+panel allocation, transforms and freeing on the caller's stream. Tiny generic
+host matrices retain the CPU DFT. CPU builds do not depend on sppark or CUDA.
+
+An immutable per-device plan fixes each transform's panel size, launch groups
+and coset powers. Admission and execution share the plan. Dimensions must fit
+the compiled domain (currently 2^28 rows), checked byte arithmetic and at least
+one column within the panel budget; invalid shapes fail before allocation.
+
+`MULTI_STARK_SPPARK_PANEL_BYTES` caps the transform scratch, 4 GiB by default;
+a budget smaller than one column is a configuration error. Columns are
+transformed in launch groups sized to the device's L2 cache. Both are captured
+when the device DFT is constructed. The fork's `SPPARK_NO_CXX_RUNTIME` mode
+aborts on CUDA errors with diagnostics.
+
+```sh
+cargo test --release --features parallel,cuda --lib cuda::sppark::tests::
+MULTI_STARK_CUDA_BENCH_SHAPES="20,533,2" cargo run --release --features parallel,cuda --example cuda_resident_lde_bench
+```
+
 ## License
 
 MIT or Apache-2.0