baracuda-types

Shared type vocabulary for the baracuda CUDA stack.

This crate contains only pure-data types — no I/O, no libloading, no runtime machinery. It's intended to be a cheap dependency for any crate (including user-authored CUDA wrappers) that wants to speak baracuda's type vocabulary without pulling in the loader infrastructure.

What's here

• Half, BFloat16, Complex32, Complex64 — first-class scalar types that match CUDA's __half, __nv_bfloat16, cuFloatComplex, cuDoubleComplex ABI-for-ABI.
• DeviceRepr — marker trait for types with a stable, ABI-compatible layout. Implemented for every primitive numeric type, fixed-size arrays of DeviceRepr elements, and tuples up to arity 12. Derive it on your own #[repr(C)] types via #[derive(DeviceRepr)] (with the derive feature).
• KernelArg — blanket-impl'd for &T where T: DeviceRepr, so any DeviceRepr value can be passed by reference to a CUDA kernel launch builder.
• CudaVersion — (major, minor) pair with parse_version_string.
• Feature — runtime feature-availability tag used by the safe wrapper crates to gate APIs added after CUDA 11.4.
• CudaStatus — unifying error trait that every per-library error type implements, so application-level error types can absorb all of them.

Optional integrations (feature flags)

Example

use baracuda_types::{DeviceRepr, Half};

#[repr(C)]
#[derive(Copy, Clone, baracuda_types::DeviceRepr)]
struct MyKernelParams {
    scale: f32,
    half_offset: Half,
    flags: u32,
}

Part of the baracuda workspace.

License

Dual MIT / Apache-2.0.