HXDSP异构计算框架的设计与优化  

Design and Optimization of HXDSP Heterogeneous Computing Framework

作　　者：宁成明 蔡恒雨 郑启龙[1,2] 耿锐 NING Cheng-ming;CAI Heng-yu;ZHENG Qi-long;GENG Rui(School of Computer Science,University of Science and Technology of China,Hefei 230026,China;Anhui Province Key Laboratory of High Performance Computing,University of Science and Technology of China,Hefei 230026,China;Anhui Siliepoch Technology Co.,Ltd,Hefei 230088,China)

机构地区：[1]中国科学技术大学计算机科学与技术学院,合肥230026 [2]中国科学技术大学高性能计算安徽省重点实验室,合肥230026 [3]安徽芯纪元科技有限公司,合肥230088

出　　处：《小型微型计算机系统》2022年第1期179-185,共7页Journal of Chinese Computer Systems

基　　金：国家核高基重大专项项目(2012ZX01034-001-001)资助。

摘　　要：由于CPU在处理海量数据时所面临的性能瓶颈,使得基于异构系统的异构并行计算成为并行计算领域的研究热点之一.HXDSP是中国电子科技集团第三十八所自主研制的DSP芯片,能满足多种高性能计算领域的需求.为了充分发挥HXDSP的计算能力以及多HXDSP设备并行计算的能力,本文设计了基于OpenCL的HXDSP异构计算框架.本文根据计算任务的特性设计了不同的OpenCL设备映射方式,使得HXDSP异构计算框架能适应不同的应用场景,充分利用HXDSP的计算资源并节省能耗.本文在HXDSP异构计算框架设计完成的基础上针对HXDSP异构计算应用程序进行优化,包括数据传输优化和数据访问优化.最后本文通过实验分析了HXDSP异构计算框架的程序性能,验证了HXDSP异构计算框架的可行性以及有效性.Due to the performance bottleneck that CPU faces when processing massive amounts of data, heterogeneous parallel computing based on heterogeneous systems has become one of the research hotspots in the field of parallel computing.HXDSP is a DSP chip independently developed by China Electronics Technology Group 38,which can meet the needs of a variety of high-performance computing fields.In order to give full play to the computing power of HXDSP and the parallel computing power of multiple HXDSP devices, this paper designs a heterogeneous computing framework for HXDSP based on OpenCL.In this paper, different OpenCL device mapping methods are designed according to the characteristics of computing tasks, so that the HXDSP heterogeneous computing framework can adapt to different application scenarios, make full use of HXDSP computing resources and save energy consumption.Based on the design of the HXDSP heterogeneous computing framework, this article optimizes the HXDSP heterogeneous computing applications, including data transmission optimization and data access optimization.Finally, this paper analyzes the program performance of the HXDSP heterogeneous computing framework through experiments, and verifies the feasibility and effectiveness of the HXDSP heterogeneous computing framework.

关 键 词：HXDSP 异构计算 OPENCL 程序性能优化 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]