摘要
介绍了ＨＰＦ编译系统ｐ＿ＨＰＦ中的通信实现策略及其相关算法．综合数组的分布方式、对准、ＦＯＲＡＬＬ语句的空间索引变量范围以及数组元素的下标表达式，将程序语句中所隐含的通信划分为３种情形，即无通信、ＳＨＩＦＴ通信和其它通信（ＲＥＭＡＰ通信），指出了这种划分的意义，给出相应的通信检测算法，描述了具体的实现细节；详细讨论了ＳＨＩＦＴ通信以及利用阴影区技术对它进行的优化，对于ＲＥＭＡＰ通信的典型形式及其优化也进行了讨论．
In this paper, the implementation strategy of communication in a HPF compiling system, p _HPF, and related algorithms are introduced. The hierarchy architecture of the compiling system is presented. Thinking about the way of alignment and distribution of a array, the range of a index variable in a FORALL statement, and the subscript expression of the element of a array, the communication requirements in an array assignment are classified into three categories: no communication, SHIFT communication, and other type of communication(REMAP communication). The meaning of the classification and the algorithms for detecting them are given, some details of the implementation are described. Meanwhile, SHIFT communication is discussed in details, including the definition of SHIFT communication, the detecting algorithms and the optimization using “ghost area(shadow)” technique. Two examples of HPF programs are presented to show ghost area technique. The typical cases of REMAP communication are listed and discussed, and several optimization strategies in carrying out REMAP communication are described.
出处
《计算机学报》
EI
CSCD
北大核心
1999年第5期486-495,共10页
Chinese Journal of Computers
基金
国家"攀登计划"
八六三高技术研究发展计划基金
关键词
HPF语言
群通信
通信实现策略
编译器
算法
Data parallelism,HPF language,collective communication,SHIFT communication,REMAP communication.
