SCALING KERNEL SPEEDUP TO APPLICATION-LEVEL PERFORMANCE WITH CGRAS

Abstract

While accelerators often generate impressive speedup at the kernel level, the speedup often do not scale to the application-level performance improvement due to several reasons. In this paper we identify key factors impacting the application-level performance of CGRA (Coarse-Grained Reconfigurable Architecture) accelerators using stream programs as the target application. As a practical remedy, we also propose a low-cost architecture extension focusing on the nested loops appearing very frequently in stream programs. We also present detailed application-level performance evaluation for the full StreamIt benchmark applications, which suggests that CGRAs can realistically accelerate stream applications by 3.6∼4.0 times on average, compared to software-only execution on a typical mobile processor.