NVIDIA/cuda-rust

CUDA SIMT kernels can be written natively in pure Rust -- no DSLs, no foreign language bindings -- and made available to the broader Rust community. The project is in an early stage (alpha) and under active development: you should expect bugs, incomplete features, and API breakage as we work to improve it. That said, we hope you'll try it in your own work and help shape its direction by sharing feedback on your experience.

Please see CONTRIBUTING.md if you're interested in contributing to the project.

The above example defines a generic #[kernel] function map that accepts any Fn(T) -> T closure. #[cuda_module] embeds the generated device artifact into the host binary and generates a typed module.map::<f32, _>(...) launch method. The closure move |x| x * factor is captured, scalarized, and passed as kernel parameters automatically. LaunchConfig is intentionally raw data, so using it to launch a kernel is unsafe: the caller must prove that its dimensions and resources match the kernel. Kernels with #[launch_contract(...)] instead use a checked PreparedLaunch through the safe generated method.

For composable async GPU work, stream: disappears, {kernel}_async returns a lazy DeviceOperation, and execution happens when you call .sync() or .await.

See the async_mlp example for the full async setup. The host runtime (cuda-core, cuda-async) is shared with cutile-rs and published from NVlabs/cutile-rs; the cuda-oxide SIMT surface lives under its simt modules.