In this work we implement in C++ multi-term adders that interleave the maximum exponent calculation and fraction accumulation, for floating point addition that is used in dot-product units. We evaluate the area and power by performing reduction that occur in matrix multiplications from the inference of the BERT model using input from the GLUE dataset. More specifically, we run the pytorch model from huggingface and extracted inter layer results for the different tests included in GLUE dataset to use as inputs in main.cc.
The core framework of this project is the Fast-Float4HLS library, publicly available on github.
This repository is organized as follows:
.
├── src
│ ├── file_io.h
│ ├── main_incl.h
│ ├── main.cc
│ ├── max_lzc.h
│ ├── multi_term_adders.h
│ └── norm_and_round.h
│
├── utils
│ ├── pack.py
│ └── run.py
│
├── .gitignore
├── LICENSE
├── README.md
└── setup.sh./src/This directory contains the C++ implementation of multi-term adders../utils/This directory contains Python utility scripts../setup.shA bash script to fetch all required dependencies.
- Add synthesis scripts.
- Python scripts for automatically loading and extracting BERT input on GLUE.
- Modify normalization and round logic to return a FP32 float.
- Integrade multi-term online adders onto dot-product units.
TODO
Currently active: Kosmas Alexandridis and Giorgos Dimitrakopoulos
Online-fp-add is licensed with the MIT License. You are completely free to re-distribute your work derived from Online-fp-add