GPT-LS: Generative Pre-Trained Transformer with Off-line Reinforcement Learning for Logic Synthesis

Overview

GPT-LS is a new algorithm developed to optimize logic synthesis (LS) in electronic design automation (EDA). LS is a process that transforms a high-level circuit description into a gate-level netlist, typically utilizing a unified heuristic algorithm to optimize various combinational circuits. The GPT-LS model uses decision transformer (DT), a form of offline reinforcement learning, to generate a primitive sequence (PS) that achieves design goals in a shorter time compared to traditional machine learning-based approaches. GPT-LS has been trained on a large-scale logic synthesis dataset and has achieved results that match those of previous state-of-the-art (SOTA) methods in a significantly shorter time.

Environment

We recommend using venv or Anaconda environment to install pre-requisites packages for running our framework and models. We list down the packages which we used on our side for experimentations. We recommend installing the packages using requirements.txt file provided in our repository.

• cudatoolkit = 10.1
• numpy >= 1.20.1
• pandas >= 1.2.2
• pickleshare >= 0.7.5
• python >=3.9
• pytorch = 1.8.1
• scikit-learn = 0.24.1
• torch-geometric=1.7.0
• tqdm >= 4.56
• seaborn >= 0.11.1
• networkx >= 2.5
• joblib >= 1.1.0

Here are few resources to install the packages (if not using requirements.txt)

• Pytorch
• Torch-geometric
• Networkx
• berkeley-ABC
• EPFL-benchmark

Run Example

/bin/bash auto_run_DT.sh

Cite Us

@misc{chenyang2023gptls,
      title={GPT-LS: Generative Pre-Trained Transformer with Off-line Reinforcement Learning for Logic Synthesis}, 
      author={Chenyang Lv, Ziling Wei, Weikang Qian, Junjie Ye, Chang Feng, Zhezhi He},
      booktitle={40th International Conference on Computer Design (ICCD)},
      year={2023},
      organization={IEEE}
}