基于有向图可达性的SLP向量化识别方法

发布时间：2019-03-04 09:03

【摘要】：SLP(superword level parallelism)是一种实现SIMD(single instruction multiple data)向量化的方法,当前的主流向量化编译器都实现了这种向量化方法.然而,当前算法在进行SLP向量化时,对应用程序中可向量化循环的分析过程过于保守,导致其识别SLP向量化的能力不足.为了提升该能力,本文提出了一种基于有向图可达性的SLP向量化识别方法.首先,基于数组依赖图构建包含数组和语句依赖信息的有向图,使同一条语句内的所有数组节点都在一个强连通分量内,并对强连通分量之间的依赖边进行剪枝;其次,分析不同强连通分量节点之间的可达性,根据节点的可达性获得识别SLP向量化所需的所有依赖信息,从而确定语句中的循环是否可以进行SLP向量化.将该方法在Open64-5.0编译器中实现后,SLP向量化效果得到大幅提升.对gcc-vect测试集中程序的实测结果表明,优化后的Open64-5.0编译器识别SLP向量化循环的能力优于GCC4.9,与Intel ICC14.0相当,生成的向量化代码性能优于当前最优算法.
[Abstract]:SLP (superword level parallelism) is a method to implement SIMD (single instruction multiple data) vectorization, which is implemented by current mainstream vectorization compilers. However, when SLP vectorization is carried out in current algorithms, the analysis process of vectorization cycles in applications is too conservative, which leads to insufficient ability to identify SLP vectorization. In order to improve this capability, a SLP vectorization method based on directed graph reachability is proposed in this paper. Firstly, a directed graph containing information of array and statement dependency is constructed based on array dependency graph, so that all array nodes in the same statement are within a strongly connected component, and the dependency edges between strongly connected components are pruned. Secondly, the reachability between nodes with different strongly connected components is analyzed, and all the dependent information needed to identify SLP vectorization is obtained according to the reachability of nodes, so as to determine whether the loop in the statement can be vectorized by SLP. After the implementation of this method in the Open64-5.0 compiler, the SLP vectorization effect is greatly improved. The experimental results of gcc-vect test set show that the optimized Open64-5.0 compiler has better ability to identify SLP vectorization cycles than GCC4.9, and Intel ICC14.0, and the performance of generated vectorized codes is better than that of current optimal algorithms.
【作者单位】：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;
【基金】：“核高基”国家科技重大专项(批准号:2009ZX01036-001-001-2) 数学工程与先进计算国家重点实验室开放课题(批准号:2013A11)资助项目
【分类号】：TP314

【相似文献】