baracuda-cutlass-sys

Header acquisition for NVIDIA CUTLASS as a baracuda workspace dependency. Use this crate as a [build-dependencies] or [dependencies] entry to make CUTLASS headers available to other build.rs files (e.g., baracuda-forge) and to downstream cc / cxx / bindgen invocations.

What it does

• Sparse-checks-out CUTLASS headers from GitHub at build time.
• Caches under ~/.baracuda-cutlass-sys/git/checkouts/ (file-locked for concurrent-build safety, derived from the cudaforge fetch logic).
• Emits Cargo metadata (cargo:include, cargo:root, cargo:rustc-env=BARACUDA_CUTLASS_INCLUDE_DIR) so consumers find the headers without hand-rolling path discovery.
• Exposes INCLUDE_DIR at the Rust level for runtime access.

Version selection

The cutlass-2-11 feature is the supported escape hatch for users who want to keep baracuda's CUDA-11.4 floor across the CUTLASS path. CUTLASS 4.x requires CUDA 12+.

Usage

# Cargo.toml

[build-dependencies]

baracuda-cutlass-sys = "0.0.1-alpha.5"

// build.rs
fn main() {
    // The DEP_CUTLASS_INCLUDE env var is set by baracuda-cutlass-sys.
    let include = std::env::var("DEP_CUTLASS_INCLUDE")
        .expect("baracuda-cutlass-sys should set DEP_CUTLASS_INCLUDE");
    println!("cargo:rustc-link-search=native={include}");
    // Or feed `include` into cc::Build::include() / bindgen::Builder.
}

When pairing with baracuda-forge, the wiring is automatic — forge picks up the DEP_CUTLASS_INCLUDE set by this crate and uses it for with_cutlass, skipping forge's own fetch.

License

Dual MIT / Apache-2.0 (matches the workspace and CUTLASS-adjacent ecosystem). The CUTLASS source itself remains under NVIDIA's BSD-3-Clause; this crate fetches it at build time but does not redistribute it.

See NOTICE for attribution to upstream cudaforge for the sparse-checkout fetch logic.