AMM.c

AMM = Approximated Matrix Multiplication

Motivation

• Dependency-Free
• Tiny
• Multi-Platform
• Simple API (callable via FFI)
• Fast
• The best platform for experiments
• Supporting Lowert Bits (2 ~ 8bits integers)

Project Status

• FP32 Maddness
• (Experimental) INT2 ~ INT8 Encoder Training

Project Design

• In the future, better approximation algorithms may emerge. When that happens, I want to be able to add implementations with as little thought as possible. Even if there are commonalities between methods, I will not perform any abstraction. Instead, I will provide multiple entry points.

Supported Architectures

• arm64
• x64 (AVX2, AVX512)
• CUDA
• Metal

Build

Requirements: Build, cmake > 3.14, Ninja

$ mkdir build && cd build
$ cmake ..
$ cmake --build .

Testing

• Requirements: cunit

TODO

• Implement Maddness
• Build FFI and .dylib
• Test on CI (FFI Call from Common Lisp or Python)
• Benchmark on CI