OpenMP中隐式数据并行编译策略  

Compiler strategy for implicit data parallelism in OpenMP

作　　者：蒋弘山[1] 田金兰[1] 张素琴[1] 胡长军[1] 

机构地区：[1]清华大学计算机科学与技术系,北京100084

出　　处：《清华大学学报（自然科学版）》2004年第1期54-57,共4页Journal of Tsinghua University(Science and Technology)

基　　金：国家自然科学基金资助项目(60083004;60173004);Intel公司国际合作计划项目

摘　　要：为了实现OpenMPFortranAPI2.0中通过WORKSHARE语句提供的对隐式数据并行的支持,提出了计算划分原则、计算划分算法以及实现方案。还给出了针对编译中产生的冗余同步以及相容的DO循环等情况的优化技术。将实现方案与EFC(Intel产品编译器)进行了对比测试。结果表明,该并行编译策略有效地实现了对隐式数据并行的支持。所述优化技术能够提高并行程序执行效率:消除冗余同步对于负载不平衡的情况所提高的效率与各线程最小总等待时间成正比;合并循环体能提高7%左右的效率。Data division principles and algorithms were developed to implement implicit data parallelism in OpenMP Fortran API 2.0, as indicated by the WORKSHARE directive. Compiler optimization techniques were developed for redundant synchronization and consistent DO loops. Tests showed that this parallel compilation strategy more effectively supports implicitly data parallelism than the Intel compiles EFC and that the optimization techniques speed up parallel program execution. With unbalanced workloads, the efficiency improvement is proportional to the minimum waiting time of all threads. In one example with a combination of loops, the efficiency was enhanced by about 7%.

关 键 词：OPENMP 隐式数据 并行编译 软件编译程序 

分 类 号：TP314[自动化与计算机技术—计算机软件与理论]