An Imperfect Loop Mapping Method for Course-Grained Reconfigurable Architecture

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.