一种面向SIMD扩展部件的向量化统一架构

Unified Vectorization Framework for SIMD Extensions

随着多媒体应用的普及和高性能计算的需求,越来越多的处理器集成了SIMD扩展。为了针对不同SIMD扩展部件自动生成高效的向量化代码,设计了一套虚拟向量指令集,在此基础上构建了一种面向SIMD扩展部件的向量化统一架构。将输入程序通过向量识别等阶段转变为虚拟向量指令的中间表示,而后通过向量长度解虚拟化和指令集解虚拟化,将其转变为特定SIMD部件的向量指令集。在申威1600、DSP和Alpha上的实验结果表明:统一架构能够针对3种平台自动变换出高效的向量化代码,在DSP上的加速比要明显优于其它两种平台。

With the popularization of multimedia applications and the requirement of high performance computing,more and more processors are integrated with SIMD extensions. A virtual simd instruction set was designed for generating efficient simd code for different SIMD extension units, and a unification framework facing isomeric SIMD extension units was built. Program input can be transformed into intermediate representation of virtual simd instruction and then translated into simd instruction set of specific SIMD extensions through vector length and instruction set devirtualization solutions. The experimental results on SW1600, DSP and Alpha show that efficient simd code in three platforms is generated automatically on this framework and DSP has the better speedup than the other two platform.