baracuda-cuda-sys

Raw FFI bindings + dynamic loader for the NVIDIA CUDA Driver and Runtime APIs.

This is the foundational -sys crate of the baracuda workspace. It covers libcuda (Driver API: cu* symbols) and libcudart (Runtime API: cuda* symbols) in a single crate, with a versioned symbol resolver that uses cuGetProcAddress_v2 to pick the right ABI when CUDA exposes multiple versions of the same call (e.g. cuCtxCreate_v3 vs cuCtxCreate_v2).

cargo build succeeds with no CUDA installed: symbols resolve at runtime via libloading, and the resulting binary opens whatever libcuda.so.1 / nvcuda.dll / libcudart.so the host has.

Most users want baracuda-driver or baracuda-runtime instead — those crates expose RAII handles, typed memory, lifetime-checked slices, and a kernel-launch builder over the same symbols.

What's exposed

• Driver API (cu_*): contexts, modules, kernels, memory, streams, events, graphs, textures, surfaces, tensor maps, multicast, external memory + semaphores, IPC, peer access, graphics interop, profiler, green contexts.
• Runtime API (cuda_*): the same surface area where it overlaps, plus runtime-only symbols (memory pools, set-device, runtime graph capture).
• Type definitions for every CU* / cuda* enum, struct, and handle.

Version policy

CUDA floor: 11.4. Symbols added after 11.4 are bound but resolve to None at load time on older drivers — the safe-wrapper crates surface that as Feature::Unavailable rather than panicking.

Part of the baracuda workspace.

License

Dual MIT / Apache-2.0.