Research Of Hardware/Software Partitioning Methodology Based On Multi-version Implementation Exploration

With the development of embedded system, software and hardware have been deeplycombined. Traditional design method can’t satisfy with the increasing complex designrequirements. In order to overcome the insufficiency of traditional design method, software/hardware co-design method has been proposed and improved. The decision to map functionsinto hardware or software is referred to as hardware/software partitioning, which is one of thecritical techniques in software/hardware co-design and has a significant impact on the systemperformance.Existing hardware/software partitioning approaches typically consider only a singleimplementation of each function region, called binary partitioning, overlooking the importantpoint that a region may have hundreds of potential hardware implementations, which is calledmulti-version implementation exploration. This paper proposed the improved geneticalgorithm to solve the multi-version implementation in the hardware/software partitioning.The dissertation studied the multi-version implementation problem in thehardware/software partitioning and compared a local search algorithm called BUB withgenetic algorithm. The experiment showed that genetic algorithm performed better. Toimprove the local searching ability of genetic algorithm, this paper replaced the mutationoperator with reinforcement learning. So that in mutation step, chromes would select aadaptive action to get a better result. In the experiment, the improved algorithm performedbetter and more stably.