Temporal partitioning to amortize reconfiguration overhead for dynamically reconfigurable architectures

Abstract

In these days, many dynamically reconfigurable architectures have been introduced to fill the gap between ASICs and software-programmed processors such as GPPs and DSPs. These reconfigurable architectures have shown to achieve higher performance compared to software-programmed processors. However, the current reconfigurable architectures suffer from a significant reconfiguration overhead so that it severely degrade the overall performance of reconfigurable architectures. In order to deal with this reconfigurable verhead, there had been intensive research on hardware and software techniques such as configuration compression, configuration prefetching, configuration sharing, multi-context switching, and partial reconfiguration. Among these techniques, many commercial and academic reconfigurable architectures provide the partial reconfiguration. When compiling applications written in high-level languages for these reconfigurable architectures with partial reconfiguration, compilers or synthesis tools should split into the temporal partitions using temporal partitioning. And then, each temporal partition is executed on these reconfigurable architectures through three stages: fetch, configuration, and computation. Therefore, it is important to split applications overlapping of the execution stages on different temporal partitions such that the reconfiguration overheads are minimized. In this dissertation, the temporal partitioning methodology are introduced to cope with partial reconfiguration. The proposed temporal partitioning methodology splits the application into several temporal partitions such that the reconfiguration overhead of each temporal partition can be overlapped with a computation time of the previous temporal partition. Therefore, the proposed temporal partitioning are able to amortize the reconfiguration overhead at high-level synthesis phase or compilation time. Also, temporal partitions inside a loop can be locally iterated by using a intra-...