一种针对SIMD架构的指令调度优化方法  

Method for SIMD Instruction Scheduling Optimization

作　　者：王飞[1] 王俊[1] 李中升[1] 

机构地区：[1]江南计算技术研究所,无锡214000

出　　处：《高性能计算技术》2014年第5期5-8,共4页

摘　　要：现代超标量RISC处理器大多具有多条可并发执行的指令流水线，目前流行的多核处理器更是将SIMD（Single—Instruction Multi-Data）作为一种提高性能的通用扩展。SIMD扩展流水线独立于标量流水线，并可与标量流水线并发执行。本文针对SIMD体系结构的特点，提出了一种编译器指导下的指令变换和调度优化策略，该方法将一定比例的SIMD指令转换为标量指令，并将混合的标量和SIMD指令分别调度到标量流水线和SIMD流水线并发执行，从而充分开发了程序的指令级并行。目前已在Open64开源编译器上实现了该优化，实验表明优化后可明显提高系统的指令级并行性。Multiple instruction pipelines that can execute parallel are supported by most modem RISC processors, especially in multicore processors. This SIMD pipeline is independent with scalar pipeline and can execute with scalar pipeline in parallel. This paper presents a new compiler optimization method with regard to instruction transformation and scheduling. The technique transform a proportional SIMD operation to scalar operation, and then schedule the mixed instruction to scalar and SIMD pipeline, which exploits a good use of machine＇s parallelism. Implement the technique in Open64 compiler, and the experiment results show a satisfying speedup.

关 键 词：SIMD流水线 循环展开 指令调度 指令级并行 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]