A compiler directed approach to hiding configuration latency in chameleon processors

Abstract

The Chameleon CS2112 chip is the industry’s first reconfigurable communication processor. To attain high performance, the reconfiguration latency must be effectively tolerated in such a processor. In this paper, we present a compiler directed approach to hiding the configuration loading latency. We integrate multithreading, instruction scheduling, register allocation, and prefetching techniques to tolerate the configuration loading latency. Furthermore, loading configuration is overlapped with communication to further enhance performance. By running some kernel programs on a cycle-accurate simulator, we showed that the chip performance is significantly improved by leveraging such compiler and multithreading techniques. © 2000 Springer-Verlag Berlin Heidelberg.