Multi-View Learning for Parallelism Discovery of Sequential Programs

Identifying suitable parallelizable regions in sequential programs is a crucial task for performance optimizations. Traditional methods like static and dynamic analysis have flaws like insufficient accuracy or high overhead runtime. Recent studies are more interested in applying machine learning techniques to this topic. The crux of parallelism discovery with machine learning is to generate meaningful code representations. One promising route is to exploit the dependence graph through Graph Neural Networks (GNNS). In this paper, a novel multi-view framework is proposed to automatically detect potential parallelism opportunities. Sequential programs are first represented by program execution graphs encompassing both semantic and structural information. Then two independent views are defined: namely, a structural pattern view and a node feature view. In the structural view, local graph structural patterns are captured via random anonymous walks and then fed into a Graph Convolutional Network (GCN). The node features, both dynamic and static, are fed into another GCN in the node feature view. In addition, a multi-view model is designed to unify the node features and the structural features for parallelism detection. Our approach achieves comparable state-of-the-art performance on parallel region classification with an accuracy up to 92.6% when evaluated with popular parallel computing benchmarks.