baracuda — idiomatic Rust wrappers for the NVIDIA CUDA stack.

Umbrella crate. Re-exports individual safe-API crates behind Cargo features. All features are opt-in; defaults enable just driver + runtime so downstream consumers pay for only what they use.

Feature matrix

Feature	Re-exports
`driver`	[`driver`] — CUDA Driver API
`runtime`	[`runtime`] — CUDA Runtime API
`nvrtc`	[`nvrtc`] — runtime C++→PTX compiler
`nvjitlink`	[`nvjitlink`] — CUDA 12+ JIT linker
`cublas`	[`cublas`] — BLAS
`curand`	[`curand`] — RNG
`cufft`	[`cufft`] — FFT
`cusparse`	[`cusparse`] — sparse linear algebra
`cusolver`	[`cusolver`] — dense/sparse solvers
`cudnn`	[`cudnn`] — deep-learning primitives
`nccl`	[`nccl`] — multi-GPU collectives
`npp`	[`npp`] — performance primitives
`nvjpeg`	[`nvjpeg`] — GPU JPEG codec
`nvcomp`	[`nvcomp`] — GPU compression (scaffolding)
`cvcuda`	[`cvcuda`] — CV-CUDA (scaffolding)
`nvml`	[`nvml`] — driver-bundled GPU monitoring
`cufile`	[`cufile`] — GPUDirect Storage (Linux only)

Bundles: math = cuBLAS + cuRAND + cuFFT + cuSPARSE + cuSOLVER; imaging = NPP + nvJPEG + CV-CUDA; ml = driver + runtime + nvrtc + nvjitlink + math + cuDNN + NCCL; full = ml + imaging + nvcomp + nvml + cufile.

Quickstart

[dependencies]
baracuda = "0.1"

use baracuda::driver::{Context, Device, DeviceBuffer};
# fn main() -> baracuda::driver::Result<()> {
let device = Device::get(0)?;
let ctx = Context::new(&device)?;
let data = DeviceBuffer::from_slice(&ctx, &[1.0f32, 2.0, 3.0])?;
# let _ = data;
# Ok(())
# }

Shared vocabulary

[types] (always available, no feature flag) exposes Half, BFloat16, Complex32, Complex64, DeviceRepr, CudaVersion, etc. — the types shared across every safe crate.

baracuda 0.0.1-alpha.2

Feature matrix

Quickstart

Shared vocabulary