Data Flow Graph Partitioning Method for CGRA Temporal Mapping Based on Bayesian Optimization

Coarse-Grained Reconfigurable Arrays (CGRAs) are attracting more and more attention for their high flexibility and energy efficiency. Due to the limited resources, mapping large data flow graphs (DFGs) that represent application kernels onto a CGRA is difficult, for which partitioning is employed. However, existing partitioning methods in the CGRA domain are unable to solve large kernels. In this work, we propose BOPart, an efficient DFG partitioning method based on Bayesian optimization. This enables the mapping of large DFGs that surpass the capacity of the target CGRA. Moreover, we design a graph coarsening method to reduce the complexity of the partitioning problem, which further improves the performance and convergence of BOPart. BOPart can handle benchmarks with up to 333 operations, surpassing the capability of state-of-the-art temporal mapping and partitioning method, which can only handle benchmarks with up to 94 operations.

Introduction

Coarse-Grained Reconfigurable Architectures (CGRAs) attract increasing interest in the domain-specific architectures (DSAs) domain due to their high energy efficiency and post-silicon programmability. CGRAs are widely applied in emerging fields, such as near-data computing, object inference, and neural networks.