• 中国计算机学会会刊
  • 中国科技核心期刊
  • 中文核心期刊

计算机工程与科学

• 论文 • 上一篇    下一篇

以访存为中心的阵列众核处理器核心流水线设计

张昆,郑方,谢向辉   

  1. (数学工程与先进计算国家重点实验室,江苏 无锡 214125)
  • 收稿日期:2016-09-27 修回日期:2016-12-08 出版日期:2017-12-25 发布日期:2017-12-25
  • 基金资助:

    国家863计划(2015AA01A301);国家自然科学基金(91430214)

A load-centric core pipeline design in
array many-core processors
 

ZHANG Kun,ZHENG Fang,XIE Xiang-hui   

  1. (State Key Laboratory of Mathematical Engineering and Advanced Computing,Wuxi 214125,China)
  • Received:2016-09-27 Revised:2016-12-08 Online:2017-12-25 Published:2017-12-25

摘要:

传统的流水线设计是以转移指令为中心的,大量逻辑资源被用于提高处理器转移预测的能力,以保证向流水线发射和执行部件提供充足的指令流。在阵列众核处理器中提出了一种以访存为中心的核心流水线设计。通过提高访存装载指令在流水线中的执行优先级,以及访存装载指令的预测执行机制,可以有效减少顺序流水线因访存延迟所带来的停顿,提高流水线性能和能效比。测试结果表明,以4 KB容量的装载指令访存地址表为例,访存为中心的流水线设计可以带来8.6%的流水线性能提升和7%的流水线能效比提高。

关键词: 众核处理器, 核心流水线, 访存优化, 阵列众核

Abstract:

Traditional processor pipeline is a branch-instruction-centric design where a large number of chip resources are used to improve the prediction accuracy of branches. We present a load-centric core pipeline design in array many-core processors. In the load-centric pipeline, the load instruction has higher priority to be issued and executed. Besides, we also propose a prediction mechanism to generate the load instruction’s source address in advance. The load-centric design decreases the stall latency of load instructions and therefore improves the pipeline’s performance and energy efficiency. Experimental results show that equipped with a 4KB size prediction table, the load-centric design can improve the pipeline performance and energy efficiency by 8.6% and 7% respectively
 

Key words: many-core processor, core pipeline, optimization of memory accesses, array many-core processors