• 中国计算机学会会刊
  • 中国科技核心期刊
  • 中文核心期刊

J4 ›› 2016, Vol. 38 ›› Issue (01): 52-56.

• 论文 • 上一篇    下一篇

一种高效的DMA核间同步传输方法

田玉恒,马胜,鲁建壮,杨柳   

  1. (国防科学技术大学计算机学院,湖南 长沙 410073)
  • 收稿日期:2014-12-11 修回日期:2015-03-25 出版日期:2016-01-25 发布日期:2016-01-25
  • 基金资助:

    国家自然科学基金(61303065);教育部博士点基金(20134307120028);湖南省自然科学基金(JC130602);国防科学技术大学科研计划项目(JC130602)

An efficient intercore synchronization
transmission method for DMA 

TIAN Yuheng,MA Sheng,LU Jianzhuang,YANG Liu   

  1. (College of Computer,National University of Defense Technology,Changsha 410073,China)
  • Received:2014-12-11 Revised:2015-03-25 Online:2016-01-25 Published:2016-01-25

摘要:

对HPL基准程序的高速处理亟需一种能满足算法需求的高效DMA传输方式,同时DDR访问效率也直接决定了算法的运行速度。GEMM算法的操作占整个HPL基准程序操作的90%,为了提高该算法的处理速度,结合DDR访问特点,提出了一种基于核间同步的点对点DMA传输模式。实验结果表明,由于访存效率的提升,该传输模式相比于普通传输模式平均获得了25674%左右数据传输效率提升,明显减少了数据搬运时间开销,有效满足了HPL算法的处理需求。

关键词: DMA, 核间同步, 多核DSP, 传输方法

Abstract:

The highspeed processing for HPL benchmark programs urgently needs to create a DMA transmission way to improve the processing speed of the algorithm as much as possible. Meanwhile, the processing speed of the GEMM algorithm is also determined by the DDR's memory access efficiency. However, the GEMM algorithm operation accounts for 90% in that of the HPL benchmark programs. Thus in order to improve the processing speed of this algorithm, combining the characteristics of the DMA transmission mode, we elaborate a pointtopoint transmission mode design scheme based on intercore synchronization. The actual measurements show that, compared with general transmission modes, the transmission speed and the memory access efficiency of the  DMA is improved by 256.74%, which greatly reduces the time overhead and meets the processing need of the HPL algorithm.

Key words: DMA;inter-core synchronization;multicore DSP;transmission method