一种有效的面向多目标软硬件划分的遗传算法  

Effective Multi-objective Genetic Algorithm for Hardware-software Partitioning

作　　者：罗莉[1] 夏军[1] 何鸿君[1] 刘瀚[2] 

机构地区：[1]国防科技大学计算机学院,长沙410073 [2]电子工程学院,合肥230037

出　　处：《计算机科学》2010年第12期275-279,共5页Computer Science

基　　金：863国家专项基金项目(2008AA01A202)资助

摘　　要：软硬件划分是软硬件协同设计的关键技术之一,划分结果对最终的设计方案有非常重要的影响。软硬件划分根据优化目标的数量,可分为单目标划分和多目标划分。多目标划分问题是一个NP-hard问题,一般不存在传统意义上的"最优解",而是存在一组互不支配的Pareto最优解。遗传算法因其具有并行、群体搜索的特点而非常适于求解多目标优化问题。通过抽象描述将一个实际SOC设计问题转化为多目标软硬件划分问题,采用遗传算法便可获得最优设计方案。为克服过早收敛及加快搜索速度,改进了适应度函数的定义,通过自适应参数调整,加入惩罚函数的适应度定义,提高了进化速度,从而有效地获得了Pareto最优解集。在实际问题的应用中,多目标软硬件划分遗传算法是能有效求取平衡系统成本、硬件面积、功耗和时间特性的最优化方案。Hardware/Software partitioning is one of the critical step in Hardware/Software Codesign flow,and has very important influence on the final design.In terms of the number of optimizing objects,it can be classified as single object partitioning and multi objects partitioning.Multi objects partitioning is NP-hard problem,a multi objective partitioning algorithm usually gets irrelevant Pareto results,no traditional optimum results.Genetic algorithm benfits to solve multi objective partitioning for its parallel colony research.Fitness function was investigated and a redefined fitness function was proposed,which adopts self-adaptive parameter and penalization function to escape from the premature convergence and improve evolution speed.The practical experiment results demonstrate that this algorithm is more efficient to ba-lance all the parameters to optimize multi system objects under some constraints,for instance,executing time,cost,hardware area power,etc.

关 键 词：软硬件划分 遗传算法 多目标划分 适应度函数 

分 类 号：TP302[自动化与计算机技术—计算机系统结构]