stillwater-sc/universal

Large collection of number systems providing custom arithmetic for mixed-precision algorithm development and optimization for AI, Machine Learning, Computer Vision, Signal Processing, CAE, EDA, control, optimization, estimation, and approximation.

What it solves

Universal is a C++ template library designed to replace native integer and floating-point types with custom arithmetic types. This allows developers to optimize algorithms for specific precision and dynamic range requirements, which is critical for improving performance and energy efficiency in embedded systems and Deep Learning applications where standard IEEE-754 floating-point may be overkill or inefficient.

How it works

It provides a header-only library of plug-in replacement types that can be used in parameterized computational kernels. Instead of using double or float, developers can use specialized formats like half-precision (half), bfloat16, or fixed-point types. The library includes fast implementations of formats that lack universal hardware support across different architectures (x86, ARM, RISC-V, GPUs).

Who it’s for

The library is intended for developers working in AI, Digital Signal Processing (DSP), High-Performance Computing (HPC), and High-Frequency Trading (HFT) who need to experiment with or implement mixed-precision algorithms.

Highlights

  • Wide Variety of Formats: Supports quarter, half, and quad precision, as well as NVIDIA TensorFloat and Google's bfloat16.
  • Custom Number Systems: Includes support for static and elastic integers, decimals, fixed-points, rationals, linear floats, tapered floats, logarithmic, and adaptive-precision types.
  • Hardware Agnostic: Provides software implementations for formats that are not natively supported by all hardware.
  • Low Friction: Header-only design allows for easy integration into existing C++ projects via template parameterization.

Related

  • Project
    mlpack/ensmallen

    A high-quality C++ header-only library providing a wide range of non-linear numerical optimizers, including gradient-based and gradient-free methods.

  • Project
    softwareQinc/qpp

    A modern C++ general-purpose quantum computing library for simulating arbitrary quantum processes with high performance and low dependencies.

  • Project
    uxlfoundation/oneDAL

    A C++ and DPC++ library that provides hardware-accelerated machine learning routines for tabular data across CPUs, GPUs, and distributed systems.

  • Project
    eyalroz/cuda-api-wrappers

    A header-only C++ library providing thin, modern wrappers for NVIDIA's CUDA Runtime, Driver, NVRTC, and NVTX APIs to simplify GPU programming and resource management.

  • Project
    mlpack/mlpack

    A fast, header-only C++ machine learning library that provides a wide array of ML methods and functions with bindings for Python, Julia, Go, and R.