Rotated parallel mapping: A novel approach for mapping data parallel applications on CGRAs

Abstract

In this paper we present a new way of mapping data-parallel applications on coarse-grained reconfigurable architectures (CGRAs) to increase their performance. Traditional mapping approaches aim to map an application to a minimum number of contexts. In this work we gave up this idea. We propose to use the temporal domain with multiple contexts, as the preferred mapping domain. The benefit of this approach is that enough free resources are made accessible for a parallel execution of a datapath, which enables a higher utilization of a CGRA's resources, and thus a performance increase can be achieved. To show the validity of the proposed method, the speedup of various applications is evaluated using both, theoretical and experimental studies. The results show a performance improvement of up to 122% when compared to traditional application mapping techniques.