针对粗粒度可重构架构的非完美循环映射方法

An Imperfect Loop Mapping Method for Course-Grained Reconfigurable Architecture

针对粗粒度可重构架构,提出了一个解决非完美循环映射问题的方法.该方法从最外层到最内层循环,依次进行循环分裂,生成多个完美循环并为每个完美循环生成虚拟配置包,然后使用配置包合并技术,最后将合并后的配置包映射到可重构阵列上.该方法兼顾可重构阵列的效率和重构次数,较现有的双流水映射提高了24.2%的PE利用率,减少了61.7%的重构次数.

This paper proposes an imperfect loop mapping method for course-grained reconfigurable architecture. Our method iteratively adopts the loop fission, from the outermost level to innermost level, to generate perfect loops. Then the method generates virtual configuration package for each perfect loop, and the technique of configuration package combination is used. At last, the combined configuration package is mapped on the reconfigurable array. This method pays attention to both the efficiency and reconfiguration frequency of reconfigurable processor. Compared to existing dual-pipelining method, our method generates 24. 2% more PE utilization rate and reduces 61.7% reconfiguration frequency.