基于GPU粗细粒度和混合精度的SAR后向投影算法的并行加速研究  

Research on Parallel Acceleration Processing Technology of SAR Back Projection Algorithm Based on Two Granularities and Mixing Precision of GPU

作　　者：田卫明[1,3] 刘富强 谢鑫 王长军 王健[2] 邓云开 TIAN Weiming;LIU Fuqiang;XIE Xin;WANG Changjun;WANG Jian;DENG Yunkai(Radar Research Lab,School of Information and Electronics,Beijing Institute of Technology,Beijing 100081,China;Beijing Building Research Institute Co.,Ltd.of CSCEC,Beijing 100076,China;Chongqing Innovation Center,Beijing Institute of Technology,Chongqing 401135,China;Key Laboratory of Electronics and Information Technology in Satellite Navigation,Beijing Institute of Technology,Ministry of Education,Beijing 100081,China)

机构地区：[1]北京理工大学信息与电子学院雷达技术研究所,北京100081 [2]北京中建建筑科学研究院有限公司,北京100076 [3]北京理工大学重庆创新中心,重庆401135 [4]北京理工大学卫星导航电子信息技术教育部重点实验室,北京100081

出　　处：《信号处理》2023年第12期2213-2224,共12页Journal of Signal Processing

基　　金：国家重点研发计划(2021YFC3001903);中建股份科技研发课题(CSCEC-2022-Z-4);国家自然科学基金(62101036);北京理工大学青年教师学术启动计划。

摘　　要：SAR(Synthelic Aperture Radar,合成孔径雷达)后向投影成像算法计算量大,严重影响SAR信息获取的时效性。GPU(Graphics Processing Unit,图像处理单元)具有强大的浮点计算能力及高度并行的架构,在处理可并行任务中具有显著优势。该文基于GPU的粗细粒度和混合精度,针对SAR后向投影成像算法提出了一种并行加速技术方案。该方案基于异步流技术压缩了回波传输与脉冲压缩的耗时,解决了脉冲压缩效率受限于回波大小的问题;基于GPU共享内存,解决了矩阵转置过程中GPU内存不足的问题,并提高了矩阵转置效率;基于GPU粗细粒度,实现了后向投影算法在线程和线程块两层粒度上的加速并行,并基于混合精度的数据处理方法进一步提高了后向投影成像算法效率,提高了GPU计算资源利用率。通过对实测数据处理分析,验证了所提的GPU并行加速处理方案的正确性和加速性能。在相同实验条件下,较于CPU平台,双精度和混合精度的处理方法均获得了较大加速比,显著提升了成像效率。Synthetic Aperture Radar(SAR)backward projection imaging algorithm requires a large amount of computation,which affects the timeliness of SAR information acquisition seriously.Graphics processing unit(GPU)has strong floating point computing capability and highly parallel architecture,which has significant advantages in processing parallel tasks.This paper proposes a parallel acceleration technique for SAR backward projection imaging algorithm based on coarse grained and mixing accuracy of GPU.The scheme is based on asynchronous stream technology to reduce the time consumption of echo transmission and pulse compression and solve the problem that the pulse compression efficiency limited the size of radar echo.The efficient transpose of matrix is realized by using the GPU shared memory,which could solve the problem of insufficient GPU memory.The two-level granularity parallel optimization of the backward projection imaging algorithm is realized through managing the blocks and threads.And the data processing method based on mixed precision further improve the efficiency of the backward projection imaging algorithm and resource utility.By analyzing the measured data,the correctness and acceleration performance of the proposed GPU parallel acceleration processing scheme are verified.Compared with the CPU platform under the same conditions,the methods of double precision and mixed precision based on GPU both obtain obvious acceleration radio,which significantly improves the imaging efficiency.

关 键 词：合成孔径雷达 后向投影 图像处理单元 并行优化 混合精度 

分 类 号：TN958[电子电信—信号与信息处理]