| 本文已被：浏览 4875次 下载 6532次 |
码上扫一扫！ |
|
|
| 避免模调度中cache代价的优化方法 |
|
刘利1, 李文龙2, 郭振宇1, 李胜梅1, 汤志忠1
|
|
1.清华大学,计算机科学与技术系,北京,100084;2.Intel中国研究中心,编译组,北京,100080
|
|
| 摘要: |
| 软件流水能够加快循环的执行速度.模调度是一种被广泛采用的软件流水的启发式.为了改善存储系统,cache使用了分级机制,但这也带来了额外的存储延迟-cache代价.证明了模调度可能导致cache代价,并提出了一种可以避免模调度的cache代价的PCPMS(prevent cache penalty in modulo scheduling)算法.实验结果表明,PCPMS能够避免模调度中的cache代价,提高程序性能. |
| 关键词: 软件流水 模调度 存储优化 cache代价 |
| DOI： |
| 分类号: |
| 基金项目:Supported by the National Natural Science Foundation of China under Grant No.60573100(国家自然科学基金) |
|
| Optimization to Prevent Cache Penalty in Modulo Scheduling |
|
LIU Li,LI Wen-Long,GUO Zhen-Yu,LI Sheng-Mei,TANG Zhi-Zhong
|
| Abstract: |
| Software pipelining can speedup the loop execution. Modulo scheduling is a widely used heuristic for software pipelining. Cache hierarchies can improve memory systems, but probably all processor implementations introduce additional delays (cache penalty). This paper demonstrates possible cache penalty due to modulo scheduling, and presents an algorithm named Prevent Cache Penalty in Modulo Scheduling (PCPMS), which can prevent cache penalty due to modulo scheduling. Experimental results show that PCPMS can prevent cache penalty and improve the performance of programs. |
| Key words: software pipeline modulo scheduling memory optimization cache penalty |
