基于多面体表示的向量化收益评估方法

摘要

Loop transformations can bring program speedups, but after vectorization, program performance probably cannot benefit as expected. In order to solve the problem which is not done sufficiently by researchers, combining with the features of the targeted architecture, this paper implements a benefit evaluation framework in Open64 based on polyhedral representation. The benefit evaluation model can efficiently analyze cost of various loop transform scheme so that the best vectorization can be chosen as the ultimate vectorization selection. SPEC test set is done including swim and other four programs. Results show that the polyhedral cost model can accurately predict program speedups compared with real time tests and thus avoid blind optimization.%循环变换可提高程序性能,但对其向量化后可能会导致代码性能损失,并不一定会得到预期性能提升.针对该问题,结合目标体系结构特征,在Open64中实现一个基于多面体表示指导循环变换的向最化收益评估模型.该模型可以有效分析各种循环变换方案的代价,选择向量化收益最大的方案组合作为最终的向量化方案.对SPEC测试集的swim等5个程序进行测试,结果表明,收益评估结果与实测向量化加速比相近,可避免盲目优化.