Block Floating Point (BFP) and DSConv with GPU support

August 17, 2020 ยท View on GitHub

This repo implements BFP and DSConv in cuda kernels to be used with PyTorch

Requirements

  • Python >= 3.6
  • PyTorch >= 1.0
  • CUB == 1.8

Build

  1. Download CUB and put it in /home/your_username/libs/ (or the file indicated at NUQ/BlackBox/Quantization/src/setup.py:22)
  2. cd /path/to/NUQ/BlackBox/Quantization/src/ then python build_ext --inplace.