一种面向众核处理器的嵌套循环多维并行识别方法*

现有并行识别方法用于众核处理器时存在一定不足,当选择的循环并行维迭代数较少时可能导致严重地负载不均衡。针对这一问题,提出了一种面向众核处理器的多维并行识别方法,在现有并行识别方法无法做到较好的负载均衡时,选择嵌套循环的多个维进行并行,将多个并行维的迭代空间合并后再做任务划分,减少负载不均衡对程序并行效率的影响。此方法已在课题组开发的自动并行化系统中进行了实现,实际应用过程中能够提升一些应用程序在众核处理器上并行执行的效率。

A multi-dimensional parallelism recognition method of nested loop for many-core processors

There were some shortcomings in the existing parallelism recognition methods for the many-core processors. It could lead to serious load imbalance when the selected loop parallel dimension iteration number was small. To solve this problem, a multi-dimensional parallel recognition method is proposed for many-core processor. When it is difficult for the existing recognition methods to reach a better load balancing, the paper takes a multi-dimensional parallel approach to the nested loops, and a task partition scheme after multi-dimensional parallel iteration space merging, to reduce the impact of load imbalance on parallel efficiency of the program. It has been implemented in the automatic parallelization system developed by the research group, which can improve the parallel execution efficiency of some applications on many-core processor.