Configuration Cache Management for Coarse-Grained Reconfigurable Architecture with Multi-Array

Abstract

Coarse-Grained Reconfigurable Architectures (CGRAs) can achieve both high performance and flexibility, and CGRAs with multi-array are used to meet the increasing performance requirement of multimedia applications. Meanwhile, the context size also becomes quite large, so many CGRAs use a configuration cache to reduce reconfiguration overhead. However, with high power consumption, configuration cache management is still a challenge. This paper first analyzes context features of media algorithms, and introduces the base hardware architecture. Then a configuration cache management technique is proposed to implement H.264 video decoding on the base architecture. It includes a novel configuration cache structure and a configuration cache replacement algorithm based on Context Sequence Prefetching & Priority (CSPP). The experimental results show that the proposed approach can drastically improve system performance and reduce power consumption. The average configuration cache hit rate of CSPP is 96.83%, the speedup ranges from 64% to 109%, and our approach can support H.264 1080p@30fps decoding at a 200MHz working frequency.