feat: add experimental cuda-oxide fGN backend

Author: dancixx

Cargo.toml

@@ -10,6 +10,10 @@ members = [
     "stochastic-rs-viz",
     "stochastic-rs-py",
 ]
+# Device-only cuda-oxide kernel crate: its own `[workspace]` root (built
+# standalone with the cuda-oxide codegen backend to emit NVVM IR), excluded so
+# the umbrella workspace does not claim it as a member.
+exclude = ["stochastic-rs-stochastic/fgn-oxide-kernels"]
 
 [workspace.package]
 version = "2.3.0"
@@ -90,6 +94,9 @@ cubecl = { version = "0.9.0", default-features = false }
 cubecl-cuda = "0.9.0"
 cubecl-wgpu = "0.9.0"
 cudarc = { version = "0.19.2", features = ["cuda-12080", "cuda-version-from-build-system"] }
+cuda-core = { git = "https://github.com/NVlabs/cuda-oxide.git", rev = "4a56e4220aab8ce5d085a411e7f806cebb647d14" }
+cuda-device = { git = "https://github.com/NVlabs/cuda-oxide.git", rev = "4a56e4220aab8ce5d085a411e7f806cebb647d14" }
+cuda-host = { git = "https://github.com/NVlabs/cuda-oxide.git", rev = "4a56e4220aab8ce5d085a411e7f806cebb647d14" }
 gpu-fft = "1.1.1"
 metal = "0.33.0"
 
@@ -140,6 +147,9 @@ cubecl = { workspace = true, optional = true, default-features = false }
 cubecl-cuda = { workspace = true, optional = true }
 cubecl-wgpu = { workspace = true, optional = true }
 cudarc = { workspace = true, optional = true, features = [ "cuda-12080", "cuda-version-from-build-system", ] }
+cuda-core = { workspace = true, optional = true }
+cuda-device = { workspace = true, optional = true }
+cuda-host = { workspace = true, optional = true }
 flate2 = { workspace = true }
 gpu-fft = { workspace = true, optional = true }
 gauss-quad = { workspace = true }
@@ -221,6 +231,11 @@ name = "fgn_cuda_native"
 harness = false
 required-features = ["cuda-native"]
 
+[[bench]]
+name = "fgn_cuda_oxide"
+harness = false
+required-features = ["cuda-oxide-experimental"]
+
 [[bench]]
 name = "fgn_cuda_compare"
 harness = false
@@ -328,6 +343,12 @@ required-features = ["cuda-native"]
 [features]
 ai = ["dep:stochastic-rs-ai", "stochastic-rs-ai/quant"]
 cuda-native = ["dep:cudarc", "cudarc/cufft", "stochastic-rs-stochastic/cuda-native"]
+cuda-oxide-experimental = [
+    "dep:cuda-core",
+    "dep:cuda-device",
+    "dep:cuda-host",
+    "stochastic-rs-stochastic/cuda-oxide-experimental",
+]
 default = []
 # Experimental: opt in to the dual-stream RNG (`SimdRngDual` /
 # `SimdNormalDual`). ~5–11% speedup on Ziggurat-based Normal/Exp bulk

benches/fgn_cuda_oxide.rs

@@ -0,0 +1,54 @@
+use std::hint::black_box;
+use std::time::Duration;
+
+use criterion::BenchmarkId;
+use criterion::Criterion;
+use criterion::criterion_group;
+use criterion::criterion_main;
+use stochastic_rs::simd_rng::Unseeded;
+use stochastic_rs::stochastic::noise::fgn::Fgn;
+use stochastic_rs::traits::ProcessExt;
+
+fn bench_fgn_cuda_oxide(c: &mut Criterion) {
+  let mut group = c.benchmark_group("FGN_cuda_oxide_experimental");
+  group.measurement_time(Duration::from_secs(3));
+  group.warm_up_time(Duration::from_millis(700));
+  group.sample_size(30);
+
+  let cases = [
+    (4096usize, 32usize),
+    (16384usize, 128usize),
+    (65536usize, 128usize),
+  ];
+
+  for &(n, m) in &cases {
+    let label = format!("n={n},m={m}");
+    let fgn = Fgn::new(0.7f32, n, None, Unseeded);
+    let _ = fgn
+      .sample_cuda_oxide_with_module(m, "fgn_cuda_oxide")
+      .expect("cuda-oxide warmup should succeed");
+
+    group.bench_with_input(BenchmarkId::new("cpu/sample_par", &label), &m, |b, &m| {
+      b.iter(|| black_box(fgn.sample_par(m)));
+    });
+
+    group.bench_with_input(
+      BenchmarkId::new("cuda_oxide/sample", &label),
+      &m,
+      |b, &m| {
+        b.iter(|| {
+          black_box(
+            fgn
+              .sample_cuda_oxide_with_module(m, "fgn_cuda_oxide")
+              .expect("cuda-oxide sampling should succeed"),
+          )
+        });
+      },
+    );
+  }
+
+  group.finish();
+}
+
+criterion_group!(benches, bench_fgn_cuda_oxide);
+criterion_main!(benches);

stochastic-rs-stochastic/Cargo.toml

@@ -26,6 +26,12 @@ gpu = ["dep:cubecl", "dep:gpu-fft"]
 gpu-cuda = ["gpu", "dep:cubecl-cuda"]
 gpu-wgpu = ["gpu", "dep:cubecl-wgpu"]
 cuda-native = ["dep:cudarc", "cudarc/cufft"]
+cuda-oxide-experimental = [
+    "dep:cuda-core",
+    "dep:cuda-device",
+    "dep:cuda-host",
+    "dep:fgn-oxide-kernels",
+]
 metal = ["dep:metal"]
 accelerate = []
 python = ["dep:pyo3", "dep:numpy", "stochastic-rs-core/python", "stochastic-rs-distributions/python"]
@@ -64,6 +70,14 @@ cubecl = { workspace = true, optional = true, default-features = false }
 cubecl-cuda = { workspace = true, optional = true }
 cubecl-wgpu = { workspace = true, optional = true }
 cudarc = { workspace = true, optional = true, features = ["cuda-12080", "cuda-version-from-build-system"] }
+cuda-core = { workspace = true, optional = true }
+cuda-device = { workspace = true, optional = true }
+cuda-host = { workspace = true, optional = true }
+# Device-only kernel crate (its own `[workspace]`); provides the `#[kernel]`
+# markers the host `cuda_launch!` calls resolve against. Built as host stubs by
+# a normal `cargo build`; its NVVM IR is pre-generated and embedded — see
+# `noise/fgn/cuda_oxide.rs`.
+fgn-oxide-kernels = { path = "fgn-oxide-kernels", optional = true }
 gpu-fft = { workspace = true, optional = true }
 metal = { workspace = true, optional = true }
 pyo3 = { workspace = true, features = ["extension-module"], optional = true }

stochastic-rs-stochastic/fgn-oxide-kernels/Cargo.toml

@@ -0,0 +1,34 @@
+# Device-only kernel crate for the experimental cuda-oxide fGN backend.
+#
+# This crate contains ONLY `#[kernel]` device code. It is compiled two ways:
+#   1. Normally (plain `cargo build`) it expands to host-side kernel markers
+#      that `stochastic-rs-stochastic`'s `cuda_launch!` calls resolve against.
+#   2. Once, by the maintainer, with the cuda-oxide rustc codegen backend
+#      (`RUSTFLAGS="-Z codegen-backend=…"` + `CUDA_OXIDE_PTX_DIR`), it emits
+#      `fgn_oxide_kernels.ptx`. That PTX is committed and `include_str!`-ed into
+#      the lib so downstream users run with a plain `cargo build` — no
+#      `cargo oxide` / precompile step on their side.
+#
+# Standalone `[workspace]` so the cuda-oxide device build does not pull in the
+# whole stochastic-rs workspace (host-only code can't be lowered to PTX).
+[package]
+name = "fgn-oxide-kernels"
+version = "0.1.0"
+edition = "2024"
+license = "MIT"
+description = "Device-side fGN kernels for the experimental cuda-oxide backend."
+
+# Not a workspace root: the umbrella `Cargo.toml` lists this path under
+# `[workspace].exclude`, so it builds standalone for cuda-oxide PTX generation
+# yet is usable as a path dependency without a "multiple workspace roots" error.
+
+[lib]
+doctest = false
+
+[dependencies]
+# The `#[kernel]` macro expands to code referencing all three cuda-oxide crates
+# (device intrinsics + host-side registration markers), so all are required even
+# though this crate is device-only.
+cuda-device = { git = "https://github.com/NVlabs/cuda-oxide.git", rev = "4a56e4220aab8ce5d085a411e7f806cebb647d14" }
+cuda-host = { git = "https://github.com/NVlabs/cuda-oxide.git", rev = "4a56e4220aab8ce5d085a411e7f806cebb647d14" }
+cuda-core = { git = "https://github.com/NVlabs/cuda-oxide.git", rev = "4a56e4220aab8ce5d085a411e7f806cebb647d14" }

stochastic-rs-stochastic/fgn-oxide-kernels/rust-toolchain.toml

@@ -0,0 +1,3 @@
+[toolchain]
+channel = "nightly-2026-04-03"
+components = ["rust-src", "rustc-dev", "llvm-tools"]