GRAFT: GNN-based Adaptive Framework for Efficient CGRA Mapping

Coarse-grained reconfigurable architecture (CGRA) is an emerging computing architecture that provides a trade-off between energy efficiency and flexibility. Although the CGRA mapping problem has been explored for many years, efficiently mapping complex loop kernels to large-scale CGRAs is still challenging. In this paper, we present GRAFT, GRaph neural network (GNN) Adaptive Framework for efficienT CGRA Mapping, that can be tuned automatically to generate high-quality mapping for various CGRAs under a graph-based architecture two-level intermediate representation (IR). The low-level IR binds computing resources with routing resources annotated with architectural information, while the high-level IR provides a concise way to describe hardware modules and elastic interconnections among them. GRAFT consists of two independent components, 1) GNN-based placement distribution prediction and 2) a progressive graph-based mapper. We abstract the architecture in graph and use GNN to analyze the graph attributes and identify the subgraph relationships between data flow graphs (DFGs) and the architecture graph, providing a panoramic view for the agile mapper. In addition, we enhance the graph-based mapping algorithm that introduces a local cache and pruning strategies and takes into account the impact of DFG’s placement distribution on the architecture. Experiments show that GRAFT achieves a 64.5× faster mapping speed with better performance on large-scale CGRA compared to the state-of-the-art GNN framework LISA. Moreover, GRAFT can attain 1.5× higher resource utilization with a 122× mapping speedup over Morpher by architecture tuning.