支持CPU+GPU协同计算的C源程序预处理划分策略  

Pre-processing partition strategies for C source program to support CPU+GPU coordination computing

作　　者：姚鼎[1,2] 曾国荪[1,2] 丁春玲[3] 

机构地区：[1]同济大学计算机科学及技术系,上海200092 [2]国家高性能计算机工程技术中心同济分中心,上海200092 [3]同济大学化学系,上海200092

出　　处：《计算机应用》2013年第A02期67-72,76,共7页journal of Computer Applications

基　　金：国家863计划项目(2009AA012201);国家自然科学基金资助项目(61272107;61202173;61103068);上海市优秀学科带头人计划项目(10XD1404400);教育部网络时代的科技论文快速共享专项(20110740001)

摘　　要：GPU技术的兴起为构建高性能异构协调计算平台提供了一个新的有效途径,针对如何在CPU+GPU并存的环境下,设计高效的应用程序以更好地发挥GPU技术,开展支持CPU+GPU协同计算的C源程序预处理划分策略研究。从计算任务的C语言源程序出发,分析程序的内存访问、运算密度、流程控制结构、数据并行性等各种特征,同时考虑CPU和GPU的特点和差异,以此作为关键依据,对C源程序进行预处理,给出划分和标识程序段适合CPU或GPU执行的策略和算法。此外,以传统典型程序为测试用例开展实验。实验结果表明,策略和算法具有一定的有效性,在一定程度上提升了计算的效率。The emerging of GPU technology provides a new effective way for building high-performance heterogeneous computing platform. In order to design efficient application programs with GPU technology under CPU ＋ GPU environment, this paper studied the pre-processing partition strategies for C source programs to support CPU ＋ GPU coordination computing. By analyzing characteristics of the program in memory access, arithmetic density, control flow structure and data parallelism, and considering the features and difference of CPU and GPU, strategies and algorithms which partition and label the program were presented. In addition, experiments were conducted by using some typical programs. The result shows these proposed strategies and algorithms are effective and improve the computing efficiency.

关 键 词：协调计算 CPU+GPU C源程序 划分策略 

分 类 号：TP338[自动化与计算机技术—计算机系统结构]