多处理部件并行优化方法研究  

Research on Optimization Method of Multiple Processing Element Parallelization

作　　者：李钊[1] 郑红[1] 

机构地区：[1]北京航空航天大学自动化科学与电气工程学院,北京100191

出　　处：《计算机工程》2014年第9期305-311,共7页Computer Engineering

基　　金：国家自然科学基金资助项目(60543006);博士点基金资助项目(201003259);光电信息重点实验室基金资助项目(9140C150105100C1502)

摘　　要：针对多处理单元(PE)并行优化中运行时间和资源消耗随PE数量变化而增加的问题,分析多PE并行中运行时间和资源消耗随PE数量的变化规律,建立基于运行时间和资源消耗的优化目标函数,并从理论上证明优化目标函数最小值的存在性和唯一性,提出基于运行时间与资源消耗的多PE并行优化方法。该优化方法可在最小资源消耗的情况下实现运行时间的最优化。利用灰度共生矩阵和单精度浮点矩阵乘法的多PE优化方法进行验证。实验结果表明,多PE并行的优化方法实现了运行时间和资源消耗的优化,在运行时间上该方法比已有方法最高快6.79倍,在运行时间和资源消耗的综合对比上该方法最高为已有方法的3.3倍,能够实现基于运行时间和资源消耗的优化。The changing of run time and resource consumption with the number of the Processing Element（PE）is contrary. The rules of run time and resource consumption with the number of PE are analyzed. And the variation trend for resource consumption and run time with the number of PE is got. The optimization objective function based on run time and resource consumption is established. The existence and uniqueness of the minimum for optimization objective function are proved. The multi-PE optimization method based on run time and resource consumption is proposed. This method can realize the run time optimization with the least resource consumption. In order to validate the method,the optimal design of the calculation of the gray level co-occurrence matrix and single float matrix multiplication are proposed. Experimental results indicate that the runtime of gray level co-occurrence matrix is at most6.79 times than the old method. The integrated result about runtime and area consumption is3.3times than the old method. The optimization of runtime and area consumption is implemented.

关 键 词：多处理单元并行 优化方法 运行时间 资源消耗 灰度共生矩阵 单精度浮点矩阵乘法 

分 类 号：TP316[自动化与计算机技术—计算机软件与理论]