Skip to main content

Crate baracuda_types

Crate baracuda_types 

Source
Expand description

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.

§Modules

Re-exports§

pub use device_repr::DeviceRepr;
pub use host_slice::HostSlice;
pub use kernel_arg::KernelArg;
pub use numeric::BFloat16;
pub use numeric::Complex32;
pub use numeric::Complex64;
pub use numeric::Half;
pub use status::CudaStatus;
pub use stream_mode::StreamMode;
pub use version::supports;
pub use version::CudaVersion;
pub use version::Feature;
pub use zero_bits::ValidAsZeroBits;

Modules§

device_repr
The DeviceRepr marker trait: types with a stable, ABI-compatible layout suitable for being laid out in GPU-visible memory.
external_impls
Third-party type integrations.
host_slice
HostSlice — generic abstraction over “things that are a slice of T”.
kernel_arg
The KernelArg trait: a uniform way to hand values to cuLaunchKernel / cudaLaunchKernel, which expect void** — an array of pointers to each argument.
numeric
Numeric types used across the CUDA stack.
status
A uniform interface every NVIDIA-library status/return-code enum implements.
stream_mode
The default-stream semantics switch.
version
CUDA version identifier + feature-gating enum.
zero_bits
ValidAsZeroBits — safety marker for zero-initializable types.