基于嵌入式CPU+GPU异构平台的遥感图像滤波加速  

Acceleration of Remote Sensing Image Filtering Based on Embedded CPU+GPU Heterogeneous Platform

作　　者：谭鹏源 薛长斌 周莉 TAN Pengyuan;XUE Changbin;ZHOU Li(National Space Science Center,Chinese Academy of Sciences,Beijing 100190;University of Chinese Academy of Sciences,Beijing 100049)

机构地区：[1]中国科学院国家空间科学中心,北京100190 [2]中国科学院大学,北京100049

出　　处：《空间科学学报》2024年第1期95-102,共8页Chinese Journal of Space Science

基　　金：中国科学院国防科技重点实验室基金项目资助(CXJJ-20S017)。

摘　　要：针对遥感图像在轨实时处理提出一种基于嵌入式CPU+GPU异构平台的遥感图像滤波加速设计方法.以加速拉普拉斯滤波为例,利用GPU的并行计算特点,通过数据划分及数据映射的方法对算法进行并行设计;利用GPU的向量单元和缓存等硬件资源,通过采取向量化和向量重组以及工作组调优方法进一步提高了算法的运行速度.在嵌入式开发板上验证了加速设计的可行性和高效性.实验结果表明,相比于单CPU的串行实现,在增加GPU并行处理后的拉普拉斯滤波获得了4.08~16.92倍的加速比.进一步利用GPU硬件资源优化性能后,加速比可达15.38~56.41倍.A method is proposed for accelerating remote sensing image filtering in real-time using an embedded CPU+GPU heterogeneous platform for satellite-based image processing.the algorithm was initially parallelized through data division and mapping,leveraging the parallel computing capabilities of the GPU.Subsequently,hardware resources like the vector unit and cache of the GPU were employed to enhance algorithm speed through vectorization,vector permutation,and workgroup tuning.The feasibility and efficiency of this accelerated design were validated on an embedded development board.The experiments demonstrate a speedup ranging from 4.08 to 16.92 times when incorporating GPU parallel processing,compared to the serial implementation on a single CPU.Further optimization using GPU hardware resources can push the speedup to 15.38 to 56.41 times.

关 键 词：嵌入式GPU 遥感图像滤波 OPENCL 向量化 向量重组 

分 类 号：V19[航空宇航科学与技术—人机与环境工程] TP391[自动化与计算机技术—计算机应用技术]