ashvardanian/NumKong

NumKong is a portable mixed-precision math, linear algebra, and retrieval library featuring over 2,000 SIMD kernels for x86, Arm, RISC-V, LoongArch, Power, and WebAssembly. It covers 15+ numeric types from 4-bit integers and 6-bit floats to 128-bit complex numbers, and is validated against 118-bit extended-precision baselines. The library leverages rare algebraic transforms with both 1D and 2D registers like AMX and SME, and is co-designed with Tensor abstractions in C++, Python, Rust, JavaScript, GoLang, and Swift. Unlike typical BLAS libraries, NumKong promotes to wider accumulators to prevent overflow and achieves a 5–100x smaller binary footprint while delivering high throughput (e.g., 14.8 gso/s for i8 dot products on Intel Sapphire Rapids).