Hardware Acceleration for GCNs via Bidirectional Fusion

Derived from the fusion of graph traversal and neural networks, graph convolutional neural networks (GCNs) have achieved state-of-the-art performance in graph learning. However, the hybrid execution pattern, caused by the opposite characteristics of graph traversal based phase and neural network based transformation phase, poses huge challenges to the efficient execution of traditional architectures. Although GCN accelerators have emerged to address these challenges, they fail to harvest both bidirectional execution and inter-phase fusion opportunities exposed by the alternate execution phases in GCNs. Previous works either concentrate on a single execution direction or exchange the execution order of phases without inter-phase fusion, hence failing to further improve performance and efficiency. Therefore, we propose a novel hardware unit named BiFusion, which can be easily applied to existing GCN accelerators with hybrid architecture in order to harvest both of the above opportunities. BiFusion enables dynamic direction selection and inter-phase fusion, and helps significantly reduce the amounts of data access and computation. Experiments show that integrating the BiFusion unit helps the state-of-the-art GCN accelerator achieve 2x speedup on average.