baracuda-build

Build-script helpers shared by every baracuda-*-sys crate.

This is a [build-dependencies] crate. It does not link against CUDA itself and contains no runtime code — it just gives build.rs files a common, well-tested place to detect a CUDA toolkit, find nvcc, parse versions, and (optionally) configure bindgen for CUDA headers.

Quick start

// build.rs
fn main() {
    baracuda_build::emit_rerun_hints();

    if let Some(install) = baracuda_build::detect_cuda() {
        baracuda_build::emit_version_cfg(&install);
        // install.root, install.include, install.lib, install.nvcc, install.version
        // are now available for downstream cc / bindgen invocations.
    }
}

What's here

• detect_cuda() — returns a CudaInstall (toolkit root, include/ path, library path, parsed cuda.h version, and nvcc path if found) by walking $CUDA_PATH / $CUDA_HOME / $CUDA_ROOT / $CUDA_TOOLKIT_ROOT_DIR plus standard locations on Linux/Windows.
• find_nvcc() — locates nvcc via $NVCC, the detected install, or $PATH walk.
• parse_nvcc_version(stdout) — parses the release X.Y line from nvcc --version output.
• emit_rerun_hints() — emits cargo:rerun-if-env-changed= for every env var the detector consults.
• emit_version_cfg(&install) — emits cargo:rustc-cfg=cuda_<major>_<minor>
  • cuda_<major> so downstream crates can #[cfg]-gate version-dependent code.
• find_library(&install, stem) — locates libcuda.so.X / cublas64_X.dll / etc. for static-link scenarios.
• bindgen_builder(&install) (feature bindgen) — preconfigured bindgen::Builder with the right clang_args for CUDA headers (-D__CUDACC__, -D__host__=, etc.), enum style, layout/derive flags baracuda relies on.

Used by

Every baracuda-*-sys crate. Also re-exported via baracuda-forge for toolkit detection inside the kernel-build pipeline.

Part of the baracuda workspace.

License

Dual MIT / Apache-2.0.