R-GCN Based Function Inference for Gate-level Netlist

Abstract

: Graph neural networks are a type of deep-learning model for classification of graph domains. To infer arithmetic functions in a netlist, we applied relational graph convolutional networks (R-GCN), which can directly treat relations between nodes and edges. However, because original R-GCN supports only for node level labeling, it cannot be directly used to infer set of functions in a netlist. In this paper, by considering the distribution of labels for each node, we show a R-GCN based function inference method and data augmentation technique for netlist having multi- ple functions. According to our result, 91.4% accuracy is obtained from 1,000 training data, thus demonstrating that R-GCN-based methods can be e ff ective for graphs with multiple functions.