Configuration Optimization Research And Implementation For Coarse-grained Reconfigurable Processor

The dynamic runtime reconfiguration characteristic of reconfigurableprocessor makes its design process more flexible. However, with morehardware resource and complicated interconnection integrated in the chip,the configuration overhead has increased quickly. At the same time,applications need to be split into small parts because they are always largerthan reconfigurable system hardware resource. Thus, longer configurationtime and frequent reconfiguration make configuration overhead impedefurther performance speed up.This paper proposes a configuration optimization method forcoarse-grained reconfigurable processor REmus to decrease itsconfiguration overhead. The configuration optimization includes hardwareand software two parts. Hardware part exploits the reuse amongconfiguration contexts. Software part consider the reuse potential whenchoosing the partition scheme, which exposes more optimization space tohardware part.In the hardware level, we propose configuration contexts reuse anddifference reconfiguration methods, which target the repeatability andsimilarity features among configuration contexts. These methods increasesystem efficiency by avoiding unnecessary configuration wordstransmission and simpler configuration process. The result fromconfiguration by hand shows that it achieves14%average or35%at mostin performance improvement.In the software level, we propose a combining of configurationoptimization and temporal partition method. The original temporalpartition algorithm is certain. We propose a new probabilistic temporal partition algorithm based on the multi-objectives particle swarmoptimization. When divides the DFG, this algorithm takes the reuse featureamong configurations into consideration and regards it as a part ofoptimize goal.We choose various benchmarks from H264and mediabench.Experimental results demonstrate that comparing with the originaltemporal partition without configuration optimization, new temporalpartition based on multi-objectives particle swarm optimization andconfiguration optimization can achieve40.58%decrease in executiondelay,6.76%decrease in communication overhead and14.83%increase inresource efficiency on average. Comparing with traditional optimizingalgorithms, MOPSO has a more balanced optimization for variousobjectives and obtains a better optimization result.