Loop tiling for reconfigurable accelerators

Abstract

In this paper, we focus on system level-optimizations for automatic parallelization of nested loop on Reconfigurable Accelerators. Specifically, as off-chip bandwidth plays a major role in total performances for such implementations, we propose some partitioning techniques based on loop tiling which can take advantage of the hierarchically structured RA memory systems.