LOSTIN: Logic Optimization via Spatio-Temporal Information with Hybrid Graph Models

Benchmark

All considered designs come from EPFL logic synthesis benchmark suite

• 9 designs from the arithmetic benchmarks
• 2 designs from the random/control benchmarks

Data Preprocessing

From verilog to graph

verilog2graph parses verilog designs into graphs.

Specifically,

• verilog_cleanser.py removes the symbols incompatible to our parser from the original verilog files.
• verilog2graph.py converts the processed verilog files into graphs, which are saved in json files.

Ground truth (i.e. label) generation

dataset-generation generates the ground truth based on the ABC logic synthesis tool.

Specifically,

• flow_generation.ipynb generates random synthesis flows with different lengths.
  • flow_10.csv, flow_15.csv, flow_20.csv, flow_25.csv are the previously generated synthesis flows.
• run_abc_syn.py invokes ABC to produce area and delay, as the ground truth.
  • The previously generated ground truth files are saved in dataset-ground-truth.

Baselines

• CNN-based model
  • Implemented based on Developing synthesis flows without human knowledge
  • Data preprocessing: cnn_data_gen.py
  • How to run:

    python train_cnn.py

• LSTM-based model
  • Requirements: torchtext 0.6.0
  • Implemented based on Decision Making in Synthesis cross Technologies using LSTMs and Transfer Learning
  • Data preprocessing: data_preprocessing_for_lstm.ipynb
  • How to run:

    python LSTM_area.py

  • Note that the dataset files should be unzipped before running the code.

Hybrid GNN models

• GNN with a supernode
  • Requirements: Pytorch Geometric
  • How to run:

    python main_gnn.py

• GNN + LSTM
  • Requirements: Pytorch Geometric, and torchtext 0.6.0
  • How to run:

    python main_gnn_customized_area.py

  • Note that the dataset files related to LSTM should be unzipped before running the code.

Contact and Citation

• If there is any question, please shoot an email to [email protected]
• If you find LOSTIN useful, please cite our paper:

  @inproceedings{wu2022lostin,
     title={LOSTIN: Logic Optimization via Spatio-Temporal Information with Hybrid Graph Models},
     author={Wu, Nan and Lee, Jiwon and Xie, Yuan and Hao, Cong},
     booktitle={Proceedings of the 33rd IEEE International Conference on Application-specific Systems, Architectures and Processors},
     year={2022},
     organization={IEEE}
  }