ashvardanian/NumKong

SIMD-accelerated distances, dot products, matrix ops, geospatial & geometric kernels for 16 numeric types β€” from 6-bit floats to 64-bit complex β€” across x86, Arm, RISC-V, and WASM, with bindings for Python, Rust, C, C++, Swift, JS, and Go πŸ“

What it solves

NumKong is a portable, mixed-precision math and linear algebra library designed to eliminate "quiet overflow" and numerical instability common in standard BLAS libraries. It provides high-performance SIMD kernels that promote results to wider accumulators (e.g., Int8 $\to$ Int32) to ensure results stay in range, while maintaining a significantly smaller binary footprint than alternatives like PyTorch or NumPy.

How it works

The library implements over 2,000 SIMD kernels across 30+ backends for x86, Arm, RISC-V, LoongArch, Power, and WebAssembly. It leverages hardware-specific features like AMX and SME and supports 15+ numeric types, ranging from 4-bit integers to 128-bit complex numbers.

Key design choices include:

  • Avoiding loop unrolling: Relying on modern CPU hardware to handle out-of-order execution and register renaming to keep the binary size small and avoid instruction-cache pressure.
  • Thread-agnostic design: It does not manage its own thread pools, instead exposing row-range parameters so developers can integrate it into any parallelism model (e.g., Python's concurrent.futures) without thread oversubscription.
  • Memory independence: It avoids hidden internal allocations, making it compatible with arbitrary allocators and alignment requirements.

Who it’s for

It is built for developers needing high-performance, numerically stable linear algebra across multiple languages (C++, Python, Rust, JS, Go, Swift) and platforms, particularly those working in robotics, edge AI, or real-time systems where deterministic latency and small binary sizes are critical.

Highlights

  • Mixed-Precision Support: Handles 17 numeric types, including rare formats like 6-bit floats.
  • Lighter Footprint: Binary size is 5-100x smaller than other BLAS-like alternatives.
  • Broad Hardware Support: Optimized for x86, Arm, RISC-V, LoongArch, Power, and WebAssembly.
  • Numerical Validation: Every kernel is validated against 118-bit extended-precision baselines.
  • Multi-language Bindings: Native support for six major programming languages.

κ΄€λ ¨

  • ν”„λ‘œμ νŠΈ
    ggml-org/ggml

    κ³ νš¨μœ¨μ„±κ³Ό ν¬λ‘œμŠ€ν”Œλž«νΌ ν˜Έν™˜μ„±μ„ κ³ λ €ν•œ, μ˜μ‘΄μ„± 없이 κ°€λ²Όμš΄ C/C++ ν…μ„œ 라이브러리둜, 기계 ν•™μŠ΅μ„ μœ„ν•œ 섀계.

  • ν”„λ‘œμ νŠΈ
    kokkos/kokkos-kernels

    Kokkos 병렬 ν”„λ‘œκ·Έλž˜λ° λͺ¨λΈ 기반의 μ„ ν˜• λŒ€μˆ˜ 및 κ·Έλž˜ν”„ 연산을 μœ„ν•œ μ„±λŠ₯ 포터블 C++ μˆ˜ν•™ 컀널 λͺ¨μŒ.

  • ν”„λ‘œμ νŠΈ
    mratsim/Arraymancer

    Nim용 κ³ μ„±λŠ₯ n차원 ν…μ„œ 라이브러리둜, CPU, CUDA, OpenCL λ°±μ—”λ“œμ—μ„œ κ³Όν•™ 계산, λ¨Έμ‹ λŸ¬λ‹ 및 λ”₯λŸ¬λ‹μ„ μœ„ν•œ 도ꡬλ₯Ό μ œκ³΅ν•©λ‹ˆλ‹€.

  • ν”„λ‘œμ νŠΈ
    OpenMathLib/OpenBLAS

    λ‹€μ–‘ν•œ CPU μ•„ν‚€ν…μ²˜μ— 걸쳐 κ³ μ„±λŠ₯ μ„ ν˜• λŒ€μˆ˜ 연산을 μ œκ³΅ν•˜λŠ” μ΅œμ ν™”λœ κΈ°λ³Έ μ„ ν˜• λŒ€μˆ˜ ν•˜μœ„ ν”„λ‘œκ·Έλž¨(BLAS) λΌμ΄λΈŒλŸ¬λ¦¬μž…λ‹ˆλ‹€.

  • ν”„λ‘œμ νŠΈ
    ROCm/composable_kernel

    λ‹€μ–‘ν•œ GPU 및 CPU μ•„ν‚€ν…μ²˜ 간에 ν¬ν„°λΈ”ν•œ μ„±λŠ₯ μ€‘μ‹¬μ˜ κΈ°κ³„ν•™μŠ΅ 컀널을 μž‘μ„±ν•˜κΈ° μœ„ν•œ ν”„λ‘œκ·Έλž˜λ° λͺ¨λΈμ„ μ œκ³΅ν•˜λŠ” λΌμ΄λΈŒλŸ¬λ¦¬μž…λ‹ˆλ‹€.