面向向量化的局部数据重组

Vectorization-oriented Local Data Regrouping

目前,利用微处理器的多媒体扩展对非多媒体程序的向量化已成为提高程序性能的一个重要手段.然而,和多媒体程序相比,非多媒体程序存在大量的非连续和非对齐的数据引用方式,严重影响程序的向量化发掘和向量化性能.提出一种新的向量化方法—基于局部数据重组的向量化技术(.通过改变局部数据的布局,将循环中不连续的数据引用变为连续的数据引用,进而完成对循环的向量化;并对数据引用作对齐分析和对齐优化,从而提高程序的向量化性能.以SPEC CPU2000浮点测试集为例,该方法不仅可以向量化对于ICC编译器无法向量化的程序,而且对这些程序都有很好的性能提升,在当前的测试环境下某些程序性能最高可提高241.6%.

At present, non-multimedia program vectorization with multimedia extension has become an important way to improve the program performance. However, compared to multimedia program, there are a large amount of non-adjacent and non-alignment data references in the non-multimedia program, which seriously impede the program vectorized and decrease program vectorization performance. The present paper proposes a novel vetorization method based on data regrouping locally. This method changes non-adjacent data reference to adjacent data reference by data regrouping locally so as to vectorize regrouped loop, and makes alignment analysis and alignment optimization so as to improve program vectorization performance. For SPEC CPU2000 float point test sets, the proposed method can vectorize some program which can＇ t be vectorized by other compilers such as ICC, and highly improves the performance of some SPEC CPU2000 floating point program hy up to 241.6%.