A Reconfigurable Branch Predictor for Spatial Computing Architectures

Abstract

Branch predictors are widely used in general purpose processors to deal with control flows. However, control flows are spatially dispersed in spatial computing architectures and traditional branch predictors are not so much effective. This paper proposes a novel branch predictor that can be configured into different prediction modes according to the control flow characteristics. Experiment results show that the proposed predictor exceeds all the traditional predictors in terms of accuracy on a coarsegrained reconfigurable array platform. It improves the prediction accuracy by 5.09% compared to a state-of-the-art technique in the cost of slight area increment.