循环携带反依赖的MPI自动并行化研究

传统的面向MPI的自动并行化使用的依赖测试技术,只能确定代码中是否含有循环携带依赖,而不对循环携带依赖的类型进行判定。经研究发现,当循环携带的是反依赖时,代码仍然可以在一定条件下实现面向MPI的自动并行化。基于依赖测试方法和数据流信息,通过创建合理的依赖数据副本,提出了一种循环携带反依赖的MPI自动并行方法。实验结果表明,所提出的方法能够有效识别存在循环携带反依赖的并行循环,将其作为后端生成MPI代码的依据可有效提高MPI程序的效率。

Loop-carried Anti-dependence MPI Auto-parallelization Research

Traditional MPI auto-parallclization dependence testing methods can only detect whether there arc loop-car- ried dependences, but not their types. It was proved that auto-parallclization with loop-carried anti dependence can be a- chieved under certain conditions. 13y creating reasonable copies of dependence data, a MPI auto-parallclization method with loop-carried anti-dependence was proposed based on the dependence testing methods and the data flow informa- tion. I}he experimental results show that the proposed method can effectively recognize the parallel loops with loop-car- ried anti-dependence. Using the results of the method to generate MPI codes can efficiently improve the efficiency of MPI programs.