DeepGate

Extract features from graph-based netlist.

About

PDF Link: here
Authors: Min Li, Sadaf Khan, Qiang Xu
Lab: CURE Lab of CUHK

Innovative Designs

Represent a netlist as an and-inverter directed acyclic graph.
Use logic-simulated probabilities (the probability of being logic ‘1’) as the supervision metrics.
Propose a novel GNN including attention mechanisms and reversed propagation layers.
Add direct edges between the fan-out node and the reconvergence node (skip connection) to solve the reconvergence problem.

Diagram

example image

Diagram of the whole training process of DeepGate

Reference

@inproceedings{li2022deepgate,
  title={Deepgate: Learning neural representations of logic gates},
  author={Li, Min and Khan, Sadaf and Shi, Zhengyuan and Wang, Naixing and Yu, Huang and Xu, Qiang},
  booktitle={Proceedings of the 59th ACM/IEEE Design Automation Conference},
  pages={667--672},
  year={2022}
}