向量并行度指导的循环SIMD向量化方法

发布时间：2018-02-16 06:23

本文关键词： SIMD扩展部件向量并行度 Loop-aware 循环展开　出处：《软件学报》2017年04期 　论文类型：期刊论文

【摘要】：SIMD扩展部件是集成到通用处理器中的加速部件,旨在发掘多媒体和科学计算等领域程序的数据级并行.当前,两种基本的向量发掘方法分别是发掘迭代间并行的Loop-based方法和发掘迭代内并行的SLP方法.Loopaware方法是对SLP方法的改进,其思想是:首先,通过循环展开将迭代间并行转换为迭代内并行,使循环体内的同构语句条数足够多;再利用SLP方法进行向量发掘.但当循环展开不合法或者并行度低于向量化因子时,Loop-aware方法无法实现程序向量并行性的发掘.因此提出了向量并行度指导的循环向量化方法,依据迭代间并行度、迭代内并行度和向量化因子构建循环向量化方法选择方案,同时提出了不充分向量化方法发掘并行度低于向量化因子的循环向量并行性,最后,依据向量并行度对生成的向量循环进行展开.经过标准测试集测试,向量并行度指导的循环SIMD向量化方法比Loop-aware方法的识别率提升了107.5%,性能提升了12.1%.
[Abstract]:The SIMD extension is an accelerator integrated into a general-purpose processor designed to explore data-level parallelism in fields such as multimedia and scientific computing. The two basic vector mining methods are to discover the parallel Loop-based method between iterations and the SLP method. Loopaware method is the improvement of the SLP method. The main ideas are as follows: first of all, The parallelism between iterations is transformed into intra-iteration parallelism by loop expansion, so that the number of isomorphic statements in the loop is sufficient. Then the SLP method is used for vector discovery. However, when the cyclic expansion is illegal or the degree of parallelism is lower than the vectorization factor, Loop-aware cannot realize the discovery of program vector parallelism. Therefore, a cyclic vectorization method guided by vector parallelism is proposed. According to the degree of parallelism between iterations, the degree of parallelism within iterations and the vectorization factor, the selection scheme of cyclic vectorization method is constructed. At the same time, an inadequate vectorization method is proposed to discover the parallelism of cyclic vectors with less parallelism than the vectorization factor. Finally, The vector cycle is expanded according to the degree of vector parallelism. Through the standard test set test, the recognition rate of the vector parallelism directed cyclic SIMD vectorization method is 107.5% higher than that of the Loop-aware method, and the performance is improved by 12.1%.
【作者单位】：数学工程与先进计算国家重点实验室(解放军信息工程大学);防空兵指挥学院;
【基金】：“核高基”国家科技重大专项(2009ZX01036)~~
【分类号】：TP332

【参考文献】