Comparing Branch Predictors for Distributed-Controlled Coarse-Grained Reconfigurable Arrays

Abstract

Coarse-Grained Reconfigurable Array (CGRA) is a kind of spatial architecture that can achieve high parallelism. However, the control flow between instructions limits the parallelism and introduces pipeline stalls, significantly degrading the performance of the applications with intensive control flows. So, branch prediction is indispensable. Since CGRAs are composed of many processing elements (PEs) and the pipeline in a PE is not as deep as that in traditional processors, the branch predictors need to be reappraised. In this paper, we propose a hybrid branch predictor for CGRAs by exploiting the advantages of two complementary branch predictors. The result shows that the hybrid predictor performs best among all the predictors in the experiment. The hybrid predictor improves prediction accuracy by 4.48% compared with the bimodal predictor which is used in a latest CGRA related research.