Optimizing Data Reuse for CGRA Mapping Using Polyhedral-based Loop Transformations
Liao Huang, Dajiang Liu · 2023
Coarse-Grained Reconfigurable Arrays (CGRA) can provide high energy efficiency while keeping moderate flexibility. With flexible connections, modern CGRAs are allowed to construct register chains on demand such that data reuse could be achieved. However, existing works put little effort into loop transformations for better data reuse. Therefore, this paper proposes an efficient loop transformation approach considering data reuse for the overall performance. Using reduced polyhedral formulation and Dynamical Programming (DP) based searching, loop structures could be thoroughly and efficiently explored for optimized solutions. The experimental results show that our approach can achieve 1.11-1.15 × speedup compared to the state-of-the-art approach.