baracuda-runtime

Safe Rust wrappers for the CUDA Runtime API.

The Runtime API is "higher level" than the Driver API: contexts are implicit (each device has a primary context the runtime manages), kernels are typically linked at build time, and most operations dispatch to the current thread's current device. baracuda-runtime mirrors the Driver-side types where it makes sense (Device, Stream, Event, DeviceBuffer) and provides Runtime-API-specific facilities (memory pools, graph capture from the runtime side, set-device per thread).

use baracuda_runtime::{set_device, DeviceBuffer};

# fn demo() -> Result<(), Box<dyn std::error::Error>> {
set_device(0)?;
let host = vec![1.0f32, 2.0, 3.0];
let dev = DeviceBuffer::from_slice(&host)?;
let mut back = vec![0.0f32; 3];
dev.copy_to_host(&mut back)?;
assert_eq!(host, back);
# Ok(()) }

Coverage

Mirror of baracuda-driver for everything the Runtime API exposes:

• Device + set_device — enumeration, attributes, current-device control per thread.
• Memory — typed DeviceBuffer<T>, slices, async copies, pinned host memory, managed memory, memory pools, multicast.
• Stream + Event — ordered async work + sync/timing.
• Graph — capture and replay; cross-API interop with baracuda-driver::Graph via from_raw / as_raw.
• Launch — LaunchBuilder (typed, lifetime-checked) for typed kernel invocations.
• External memory + semaphores, IPC, VMM, query helpers.

Driver / Runtime interop

Enable the driver-interop feature (forwarded from the umbrella crate's driver-runtime-interop) to get zero-cost conversion between baracuda_driver::Stream ↔ baracuda_runtime::Stream, baracuda_driver::Event ↔ baracuda_runtime::Event, etc. Same handle under the hood — the conversion just changes the Rust type.

Dynamic loading

Same model as the rest of baracuda: cargo build needs no CUDA installed; the resulting binary opens libcudart.so / cudart64_*.dll at runtime via the loader in baracuda-cuda-sys.

Part of the baracuda workspace.

License

Dual MIT / Apache-2.0.