Installation

Minimal:

[dependencies]
numkong = "8"

With host-side parallel helpers:

[dependencies]
numkong = { version = "8", features = ["parallel", "std"] }

§ Compilation and Backend Selection

The crate builds the C library through CMake, which probes the compiler for every backend of the target architecture, so the build machine needs CMake 3.28 or newer. All supported backends are compiled into a single binary and selected at runtime. Setting NUMKONG_LIBRARY_DIR to the exact directory holding a numkong_static archive CMake already built skips the build and links that archive. For Visual Studio build trees, include the configuration directory, such as build_msvc2022_x64/Release. Cargo-managed builds use CMake’s install target and link from its installed lib directory.

The two CPU features are std, which enables standard library support, and parallel, which adds host-side orchestration via ForkUnion and implies std.

The cuda and rocm features add the NVIDIA and AMD GPU kernels, and may be combined. They need nvcc or hipcc on the build machine, and the binary then links libcudart or libamdhip64 from the toolkit CMake found.

[dependencies]
numkong = { version = "8", features = ["cuda"] }

Backend selection follows the target architecture. ARM gets NEON, SVE, and SME, with SME available on Linux, FreeBSD, and macOS. x86-64 gets Haswell (AVX2), Skylake/Ice Lake/Sapphire Rapids AVX-512 variants, and AMX on Linux and Windows only. RISC-V gets RVV backends on Linux and FreeBSD. WebAssembly gets the one SIMD capability the Rust target declares, as the next section shows. A backend the compiler cannot build is left out, and NUMKONG_MARCH=native tunes the scaffolding for the build host, as described in CONTRIBUTING.md.

§ WebAssembly

An engine validates a WebAssembly module whole, so the crate compiles one SIMD capability, the one the Rust target declares: v128relaxed under -C target-feature=+relaxed-simd, v128 under +simd128, and the serial kernels alone without either, for engines that lack SIMD. A target declaring neither fails the build, rather than quietly falling back to the serial kernels. build.rs passes that capability to CMake as NUMKONG_MARCH, beside the toolchain the target needs:

RUSTFLAGS="-C target-feature=+relaxed-simd" cargo build --target wasm32-wasip1