基于OpenCL的MUSER CLEAN算法研究与实现

发布时间：2018-03-22 11:09

本文选题：太阳　切入点：射电辐射　出处：《天文学报》2017年02期 　论文类型：期刊论文

【摘要】：天文软件开发中迫切需要在单机环境下进行高性能数据处理工作,但由于机器配置不同,采用传统的多线程、CUDA(Compute Unified Device Architecture)+GPU(Graphic Processing Unit)等方式都存在明显的局限,不利于天文软件的快速移植和无缝运行.对明安图频谱射电日像仪(MingantU SpEctral Radioheliograph,MUSER)数据处理系统开发中所采用的OpenCL(Open Computing Language)技术进行介绍,并基于OpenCL实现H銉gbom CLEAN算法.整体工作通过Python语言和PyOpenCL扩展包实现并行洁化处理.实验结果表明:基于OpenCL实现的CLEAN算法与基于CUDA实现的CLEAN算法具有大致相当的运行效率,同时也可以无需修改代码直接实现纯CPU(Central Processing Unit)环境下的高性能数据处理,解决了对CUDA+GPU环境依赖的问题,在保证MUSER数据处理系统洁化过程性能的基础上,提高了系统对硬件平台的适用性.该工作验证了OpenCL在科学数据处理中的可用性,可以预见:由于OpenCL所具有的异构环境下高性能计算特性,OpenCL将是未来天文高性能软件开发的首选技术.
[Abstract]:In the development of astronomical software, there is an urgent need for high performance data processing in a single computer environment. However, due to the different configuration of the machine, there are obvious limitations in the traditional methods of GPU(Graphic Processing unit, such as CUDAT computer Unified Device Architecture). This paper introduces the OpenCL(Open Computing language technology used in the development of MingantU SpEctral Radioheliographer data processing system, which is not conducive to the rapid transplantation and seamless operation of astronomical software. The algorithm of H gbom CLEAN is implemented based on OpenCL. The parallel and clean processing is realized by Python language and PyOpenCL extension packet. The experimental results show that the CLEAN algorithm based on OpenCL and CLEAN algorithm based on CUDA have approximately equal running efficiency. At the same time, the high performance data processing in pure CPU(Central Processing unit environment can be realized directly without modifying code. The problem of dependence on CUDA GPU environment is solved, and on the basis of ensuring the clean process performance of MUSER data processing system, It improves the applicability of the system to hardware platform. This work verifies the usability of OpenCL in scientific data processing. It can be predicted that OpenCL will be the first choice technology of astronomical high performance software development in the future because of its high performance computing characteristics in heterogeneous environment.
【作者单位】：昆明理工大学云南省计算机技术应用重点实验室;中国科学院云南天文台;云南省信息技术发展中心;
【基金】：中国科学院-国家自然科学基金委员会天文联合基金项目(U1231205、U1531132)资助
【分类号】：P161

【相似文献】