Real-XNOR-popcount-GEMM-Linear-PyTorch-Extension-CPU_GPU_C++_CUDA_Python

This repository provides the real 1-bit XNOR GEMM (GEneral Matrix Multiplication) PyTorch extension for research purpose. It may helps for those who are doing research or project related to binary neural networks (1-bit quantization).

Introduction

XNOR GEMM is a fast and efficient GEMM for binary matrices multiplication (all elements are +1 or -1).

This implementation provides

(1) Both CPU and CUDA XNOR GEMM implementation of PyTorch extensions.
(2) The implementation of training a simple binary neural network.

So, if you want to save your pytorch model and make inference using real 1-bit numbers (XNOR GEMM), it may helps.

Code structure

Real-XNOR-popcount-GEMM-Linear-PyTorch-Extension-CPU_GPU_C++_CUDA_Python
  │ 
  ├── cpu
  │    ├── setup.py: setup modules
  │    ├── test.py: test modules
  │    └── xnor_cpu.cpp: C++ implementation of XNOR GEMM
  ├── cuda
  │    ├── setup.py: setup modules
  │    ├── test.py: test modules
  │    ├── xnor_cuda.cpp:  Pytorch C++ interface of CUDA XNOR GEMM
  │    └── xnor_kernel.cu: CUDA implementation of XNOR operations 
  │
  ├── binarized_modules.py: Python (PyTorch) implementation of XNOR GEMM
  └── main.py: The binary (1-bit) MLP model that uses XNOR GEMM

For CPU/CUDA implementation, it includes

1. The C++/CUDA implementation of XNOR GEMM
2. A simple python test file to check if XNOR GEMM works well
3. A setup file

The repo also provide a simple binary MLP for test the XNOR Linear layer.

binarized_modules.py provides the PyTorch implementation of the XNOR Linear layer. main.py provides a simple MLP model for testing the XNOR Linear modules.

Dependencies

Ubuntu 18.04 LTS / Mac Os Catalina/ Windows 10 or later
C++/CUDA
Python >= 3.6
PyTorch >= 1.4.0 (pytorch >= 1.9.0 are recommended)

Tested envs

1. Ubuntu 18.04 LTS, Python 3.6, PyTorch 1.4.0, Both CPU & CUDA
2. Ubuntu 18.04 LTS, Python 3.8, Pytorch 1.7.0, Both CPU & CUDA
3. MacOS Ventura 13.3.1(a), Python 3.9, PyTorch 2.0.1, Intel x86-64 CPU

How to Use

1. Simulated quantization-aware training (simulated 1-bit uantization)

python main.py

2. Setup the custom C++/CUDA PyTorch extensions

Setup the XNOR GEMM custom operations. For CPU example, run

cd cpu
pip install .  (or python setup.py install (deprecated))
python test.py

You would see the correct output of XNOR GEMM :), and then I think you would know how to use this code

3. Real XNOR GEMM inference (1-bit inference)

You first need to uncommnet some code (e.g., the xnor_cpu and xnor_cuda in the binarized_modules.py)

cd cuda
pip install .  (or python setup.py install (deprecated))
python test.py
cd ..
python main.py

Contact

Email: [email protected]

Donations

I am interested in deep learning, and your support becomes my main motivations. If this repo helps, you can give me a star, or buy me a coffee.