通用圖形處理器（GPGPU）可大大提升規(guī)則應用的計算性能。然而,很多應用中存在非規(guī)則內存訪問模式,大大限制了GPU的性能優(yōu)勢。近年來,一些研究提出解決方案來移除靜態(tài)非規(guī)則內存訪問。然而,利用軟件消除動態(tài)非規(guī)則內存訪問仍然面臨嚴峻挑戰(zhàn)。本文提出一種純軟件解決方案用于消除動態(tài)非規(guī)則內存訪問,尤其是間接內存訪問,無需硬件擴展和離線分析。提出數據重組和索引重定向以減少內存訪問次數,從而提高GPU內核性能。為提高數據重組效率,卸載重組數據操作至GPU以降低開銷并傳輸數據。通過并發(fā)執(zhí)行數據重組和數據處理內核的統(tǒng)一計算設備架構（CUDA）流,可降低數據重組開銷。完成這些優(yōu)化后,相比于CUSPARSE基準測試,使用該方法GPU內核的內存數據傳輸減少了16.7%–50%;同時,NVIDIA Tesla P4 GPU上的內核性能提高了9.64%–34.9%。 

【文章來源】：Frontiers of Information Technology & Electronic Engineering. 2020,21(09)EISCICSCD

【文章頁數】：18 頁

【文章目錄】：
1 Introduction
2 Related work
3 System analysis and design
    3.1 Analysis of memory access pattern
    3.2 Data reordering with CPU and GPU
    3.3 System design
4 Irregularity elimination
    4.1 Data reordering
    4.2 Index redirection
5 Overhead optimization
    5.1 Overlapping remapping with computa-tion
    5.2 Cache
6 Experiments and evaluation
    6.1 Experimental setup
    6.2 Benefits of data reordering on GPU
        6.2.1 Reduction of the number of memory transac-tions
        6.2.2 Data reordering varying from CPU to GPU
    6.3 Performance optimization with overlap and cache
        6.3.1 Evaluation of the CG program
        6.3.2 Evaluation of the SP program
        6.3.3 Evaluation of the MD program
    6.4 Overview of results
7 Conclusions



本文編號：3305444