一种改进的控制流SIMD向量化方法

发布时间：2018-08-15 13:36

【摘要】：SIMD扩展部件是近年来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程序的数据级并行.控制依赖给发掘程序中的数据级并行带来了阻碍,当前,无论基于loop-based还是SLP的控制流向量化方法都需要if转换,而没有考虑循环内蕴含的向量并行度,导致生成的向量代码效率较低.此外,不精确的代价模型指导控制流向量化,同样导致生成的向量代码效率较低.为此,提出了改进的控制流SIMD向量化方法.首先,提出了含有控制依赖的循环分布算法,分离循环的可向量化部分和不可向量化部分,同时考虑分布时数据的局部性;其次,提出了一种直接向量化控制流的方法,该方法考虑了基本块间的向量重用;最后,利用精确的代价模型指导超字选择指令和超字条件分支指令的生成.实验结果表明:与现有的控制流向量化方法相比,改进方法生成的向量代码性能提高了24%.
[Abstract]:The SIMD extension is an accelerator integrated into a general purpose processor in recent years, which aims to explore the data-level parallelism of programs such as multimedia and scientific computing. Control dependence is a hindrance to data-level parallelism in mining programs. At present, both loop-based and SLP control flow direction quantization methods need if transformation without considering the vector parallelism in the loop. The resulting vector code is inefficient. In addition, the inexact cost model guides the quantization of the control flow, which also leads to the low efficiency of the generated vector code. Therefore, an improved control flow SIMD vectorization method is proposed. First of all, a cyclic distribution algorithm with control dependency is proposed to separate the quantifiable part of the loop from the non-vectorized part, and the locality of the data is considered at the same time. Secondly, a method of direct vectorization control flow is proposed. This method takes into account the vector reuse between the basic blocks. Finally, the exact cost model is used to guide the generation of superword selection instructions and superword conditional branch instructions. The experimental results show that the performance of vector code generated by the improved method is improved by 24% compared with the existing methods.
【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);
【基金】：“核高基”国家科技重大专项(2009ZX01036)~~
【分类号】：TP314;TP332

【相似文献】