面向SIMD向量化的循环优化技术研究  

Research on Loop Optimization Techniques Oriented to SIMD Vectorization

作　　者：高伟[1] 徐金龙[1] 孙回回[1] 李敏[1] 

机构地区：[1]信息工程大学,河南郑州450001

出　　处：《信息工程大学学报》2016年第4期496-503,共8页Journal of Information Engineering University

基　　金：国家科技重大专项资助项目(2009ZX01036-001-001-00102)

摘　　要：随着SIMD扩展技术的不断发展,自动向量化逐步成为程序优化的重点。然而,大部分程序本身不能直接被向量化,需要通过循环变换来发掘潜在的并行性。首先以Open64中的循环变换为例概括介绍了传统的循环优化方式,然后从多面体表示、循环变换建模方法和代价模型等3个方面具体介绍了多面体模型指导的循环变换技术,最后对两种循环优化方式的优缺点进行了对比分析。With the development of SIMD extension, automatic vectorization gradually becomes the focus of program optimization. Nevertheless, a plurality of programs fail to be vectorized directly, which need loop transformation to exploit potential parallelism. First, this study introduces loop transformations in Open64 as an example of traditional loop optimization. Then polyhedron model guided loop transformation is described in detail from three perspectives： polyhedron presentation, loop transformation modeling and cost model. Finally, the advantages and disadvantages of these two kinds of loop optimization techniques are analyzed contrastively.

关 键 词：自动向量化 循环优化 多面体模型 代价模型 

分 类 号：TP332[自动化与计算机技术—计算机系统结构]