基于GPU的多模式SAR实时成像算法研究

合成孔径雷达（SAR）成像处理需要较大的计算量，在基于中央处理器（CPU ）平台开发的SAR成像系统上处理一般需要消耗很长时间，无法满足实时成像的要求。借助于计算统一设备架构（CUDA ）编程模型，基于图形处理器（GPU ）提出了一种适用于多模式SAR的实时成像方案。该方案通过数据分段处理技术解决了计算设备GPU显存容量不足的问题；通过分析成像处理任务的并行度，利用异步执行流处理技术减少数据处理对数据交互的等待时间；通过优化GPU内存访问机制并使用特殊函数单元（SFU ）减少计算时钟周期。同时，该方案能够支持多GPU设备的并行处理，充分利用了GPU设备的计算资源。在NVIDIA GT 740M和INTEL Q6600上的实验结果表明，该方案与传统的基于CPU的单线程SAR成像技术相比，有了近150倍的速度提升，大大提高了SAR成像处理的计算效率，具有很好的工程应用前景。

Research on real time imaging algorithm for multi mode SAR with GPU

The SAR imaging system is time‐consuming ,which is developed based on central processing unit (CPU ) , because of the huge computation of SAR image processing ,making real‐time imaging impossible .Based on Computed Unified Device Architecture ,This study proposes a new plan for real‐time SAR imaging operated on graphic processing unit (GPU) which works for multiple working‐mode SAR .This new proposal makes the plan work with small GPU memory by dividing SAR data into pieces ,uses the asynchronous execution of multi‐stream technology to cut down the time wasting in data transfer ,optimizes memory management to get less access time of memory and uses light‐weight function of special function unit (SFU) to reduce clock cycle .All computational resources are totally exploited because this plan is suitable for multi‐GPU .It has been shown by an experiment on an NVIDIA GT740M and INTEL Q6600 that the proposed plan has an acceleration ratio of nearly 150 .The proposed real‐time SAR imaging system improves the imaging efficiency greatly and has a better application future .