FGNN2

This is the official open source repository for the paper

Functionality matters in netlist representation learning, ACM/IEEE Design Automation Conference (DAC), 2022.

Ziyi Wang, Chen Bai, Zhuolun He, Guangliang Zhang, Qiang Xu, Tsung-Yi Ho, Yu Huang, Bei Yu

Citation:

@inproceedings{wang-dac2022-fgnn,
    title={Functionality matters in netlist representation learning},
    author={Wang, Ziyi and Bai, Chen and He, Zhuolun and Zhang, Guangliang and Xu, Qiang and Ho, Tsung-Yi and Yu, Bei and Huang, Yu},
    booktitle=ACM/IEEE Design Automation Conference,
    pages={61--66},
    year={2022}
}

and FGNN2: a Powerful Pre-training Framework for Learning the Logic Functionality of Circuits, IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems (TCAD), 2024.

Ziyi Wang, Chen Bai, Zhuolun He, Guangliang Zhang, Qiang Xu, Tsung-Yi Ho, Yu Huang, Bei Yu

Citation:

@article{wang-tcad24-fgnn2,
    title={FGNN2: a Powerful Pre-training Framework for Learning the Logic Functionality of Circuits},
    author={Wang, Ziyi and Bai, Chen and He, Zhuolun and Zhang, Guangliang and Xu, Qiang and Ho, Tsung-Yi and Yu, Bei and Huang, Yu},
    journal=IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems,
    year={2024}
}

Setup

The experiments are conducted on Linux, with Python version 3.7.13, PyTorch version 1.13.0, and Dgl version 0.8.1.

Pretraining Dataset

Please refer to our Pretraining Dataset repo for download instructions and documentation.

Pretrain

First, download the pretaining dataset and extract it to <repo root>/datasets/pretrain_data/.

You can also choose to generate the pretraining dataset using the rawdata provided. You can first extract the rawdata to <repo root>/rawdata/pretrain_data/, go to the folder <repo root>/src and run the following command for i = 4...7

python pretrain_dataparser.py --rawdata_path ../rawdata/pretrain_data --datapath ../datasets/pretrain_data/ --num_input i

After you get the pretraining dataset, go to the folder <repo root>/src and run the following command:

python pretrain.py  --flag_inv --weighted --datapath ../datasets/pretrain_data --checkpoint ../checkpoints/example

We have also provided our pretrained weight, which can be downloaded here.

Downstream task

Here we show an example application of our pretrained FGNN2 model on the netlist classification task. Suppose that the pretrained weight is put under <repo root>/checkpoints/example/weight.pth.

Firstly, download the global task dataset here and extract it to <repo root>/datasets/global_data/.

go to the folder <repo root>/src and run the following command:

python train_global.py --datapath ../dataset/global_data/ --batch_size 128 --nlabels 4  --ratio 0.5 --flag_inv --pre_train --start_point `../checkpoints/example/weight.pth`  --checkpoint ../checkpoints/global/example

Code structure

FunctionConv.py: our customized GNN model implementation.

options.py: the hyperparameters.

pretrain.py: the model pretraining logic.

pretrain_dataparser.py: the data reader and pre-processor for pertaining.

train_local.py: the local downstream task training/fine-tuning and testing logic.

train_global.py: the global downstream task training/fine-tuning and testing logic.

truthvalue_parser.py: the parser for truthtable circuits.

ZeayW / FGNN2