面向BW104x软流水框架①.PDFVIP

计 算 机 系 统 应 用 2016 年 第25 卷 第 10 期 ① 面向 BW104x 软流水框架 洪立涛, 郑启龙 ( 中国科学与技术大学 计算机科学与技术学院, 合肥 230039) 摘 要: 现代高性能数字信号处理器大多数采用超长指令字体系结构, 通过在同一时钟周期发射多条指令以便 获得更高的运算性能来发掘目标机器指令级别并行性. 介绍了 BW104x 目标体系特征, BWDSP104X 是一款针对 高性能计算领域设计的处理器, 采用 16 发射、单指令流, 多数据流架构. 为了充分利用多簇及簇内硬件资源, 基 于open64 编译基础设施提出了后端软流水优化, 其中包括循环选择, 资源依赖数据依赖计算, 采用经典的模调度 方法进行软流水调度, 为解决不同迭代变量冲突引入模变量拓展模块. 实验结果证明流水后性能相对流水前有 了很好的提升. 关键词: 编译器; 软流水; 迭代间隔; 模调度; 模变量拓展; 代码生成 Software Pipelining Framework for BW104x HONG Li-Tao, ZHENG Qi-Long (School of Computer Science and Technology, University of Science and Technology of China, Hefei 230039, China) Abstract: The digital signal processor (Digital Signal Processing, DSP) is widely used in the field of signal processing, digital communication. The majority of modern high-performance DSP use long instruction word architecture, by exploiting instruction-level parallelism to launch multiple instructions at the same clock cycle out for a higher level of calculating performance. The article describes target system characteristics on BWDSP104x, BWDSP104x is designed in the light of high performance computing and processor, uses 16 launch, single instruction stream and multiple data stream architecture.in order to make full use of multi-cluster hardware resources, this paper proposes the back-end optimization about software pipelining based on the open-source compiler named Open64. Including the early stage of cycle options, resource constraints and precedence constraints computing, the classic Module-Scheduling algorithm is used in SWP-Scheduling, module variable expansion is for the conflict of registers using in different iteration. The experimental results show that program has better performance aft