面向自动向量化的结构体优化  

Structure Optimization for Automatic Vectorization

作　　者：于海宁[1] 韩林[1] 李鹏远[1] 

机构地区：[1]解放军信息工程大学数学工程与先进计算国家重点实验室,郑州450001

出　　处：《计算机科学》2016年第2期210-215,共6页Computer Science

基　　金："核高基"国家科技重大专项(2009ZX01036)资助

摘　　要：结构体广泛应用在科学计算等应用程序中,向量化结构体数组存在的非连续和非对齐访存会严重影响程序的向量化效果。为减少结构体数组SIMD向量化过程中的非连续和非对齐数据访问,提出了基于域访问亲和度与域数据类型相结合的结构体拆分模型,以消除域存储间的内存"间隙";同时利用结构体数组到二维数组的地址映射方式来满足结构体数组向量化时的访存连续和对齐要求,以降低Cache的失效率,从而提升应用程序性能。在自动向量化系统SW-VEC上,选取gcc-vec、spec2000和spec2006标准测试集中部分相关的测试用例,测试结果表明:与相应的串行程序相比,采用该方法后,测试用例程序性能加速比提高了8%以上。Sturcture is used more extensively to promote the performance of application program such as scientific computing.The noncontinuity and the nonalignment of its non-array memory address have a dramatic influence on the efficiency of program＇s vectorization.To reduce the access to these addresses during the SIMD＇s vectorization,this paper applied a structure peeling model based on the structure which combines field access affinity with type to eliminate the＂clearance＂of memory between field storage,and proposed an address conversion method of structure array one by one mapping to the two dimensional array to meet the request of the continuity and the alignment of its non-array memory address,further reducing the failure rate of Cache,so as to improve application performance.By using the test suites of gcc-vec,spec2000 and spec2006,the experimental results on the compiler of automatic vector show that using the method,the performance of optimized programs can be improved by more than eight percent.

关 键 词：访问亲和度 结构体拆分 地址映射 SIMD向量化 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]