关键循环到粗粒度可重构体系结构的存储感知映射

针对已有工作面向粗粒度可重构结构(CGRA)研究循环映射的不足,提出一种新颖的存储感知的关键循环映射方法 MALP。该方法定义RCP_CGRA体系结构模型并阐述关键循环到CGRA的映射问题,通过引入结合数组分簇的多面体数据域划分方法进行循环存储分析,根据分析结果,结合体系结构资源约束实现了循环的有效映射。实验结果表明,与已有的方法相比,MALP方法能够快速分析存储需求并有效降低循环映射的资源占用率,提高数据吞吐量,进一步提升了CGRA上循环映射的性能。

Critical loop memory-aware mapping onto coarse-grained reconfigurable architecture

Targeting at the mapping of key loops onto CGRA (Coarse-Grained Reconfigurable Architectures), this research proposes a novel approach called memory-aware kernel loop pipelining mapping (MALP). The RCP_CGRA model and the critical loop mapping formulation were shown first. Based on polyhedral model, then the array clustering and data domain partition were presented. An analysis of the critical loop storage requirement was described. Based on this analysis result, the MALP provided an efficient way for loop mapping under the resource constraints of CGRA. Experiment results show that MALP can improve the data throughput rate while costing less resource. MALP makes the loop mapping on CGRA more efficient.