Predicting Cross Architecture Performance of Source Codes using Graph Neural Networks

Abstract

In this creative component, we explore the application of Graph Neural Networks (GNNs) to predict the cross-architecture performance of source code, focusing specifically on intermediate representation-based features. The goal of this work is to identify patterns that can help optimize the performance of programs when executed on different architectures, such as CPUs and GPUs. By leveraging GNNs, we aim to capture the relationships and structures inherent in source code, which are often challenging for traditional performance prediction models. Our experiments show promising results in predicting performance variations and highlight potential opportunities for program parallelization based on GNN-derived insights.