面向多簇架构DSP的树匹配向量化算法

SIMD Algorithm Based on Tree Matching for Multi-cluster and VLIW DSP

BWDSP是针对高性能计算设计的一款新型的处理器,采用多簇超长指令字体系结构和SIMD架构,有丰富的指令集.为充分利用BWDSP提供的向量化资源,迫切需要提出一种向量化算法.本文在open64基础上研究并实现了面向多簇超长指令字(VLIW)DSP的SIMD编译优化算法.算法基于OPEN64的中间语言WHIRL,能够充分地利用BWDSP丰富的硬件资源和向量化指令.最终实验结果表明,对于能够合成双字和单字的循环程序,该优化算法能够平均取得6倍和4倍的加速比.

BWDSP is a new type of processor designed for high performance computing, using multi-cluster VLIW structure and SIMD architecture, including a rich instruction set. In order to make full use of the resources of BWDSP, a SIMD algorithm is to be proposed. In this paper, an algorithm for DSP SIMD compiler optimization based on open64 infrastructure is studied and implemented. This algorithm is based on WHIRL intermediate language of Open64 and can make full use of rich hardware resources and vector instruction set. The experimental result shows that the vectorization algorithm achieves 6 times performance improvement for double-word vectorization and 4 times performance for single-word vectorization on average.