基于NVIDIA GPU的机载SAR实时成像处理算法CUDA设计与实现

摘要

合成孔径雷达(SAR)成像处理的运算量较大，在基于中央处理器(Central Processing Unit, CPU)的工作站或服务器上一般需要耗费较长的时间，无法满足实时性要求。借助于通用并行计算架构(CUDA)编程架构，该文提出一种基于图形处理器(GPU)的SAR成像处理算法实现方案。该方案解决了GPU显存不足以容纳一景SAR数据时数据处理环节与内存/显存间数据传输环节的并行化问题，并能够支持多GPU设备的并行处理，充分利用了GPU设备的计算资源。在NVIDIA K20C和INTEL E5645上的测试表明，与传统基于GPU的SAR成像处理算法相比，该方案能够达到数十倍的速度提升，显著降低了处理设备的功耗，提高了处理设备的便携性，能够达到每秒约36兆采样点的实时处理速度。%Synthetic Aperture Radar (SAR) image processing requires a considerable amount of computational resources. Traditionally, this task runs on a workstation or a server based on Central Processing Units (CPUs) and is rather time-consuming, making real-time processing of SAR data impossible. Based on Compute Unified Device Architecture (CUDA) technology, a new plan for a SAR imaging algorithm operated on an NVIDIA Graphic Processing Unit (GPU) is proposed. The new proposal makes it possible for the data processing procedure and the CPU/GPU data exchange to execute concurrently, especially when the size of SAR data exceeds the total GPU global memory size. A multi-GPU is suitably supported by the new proposal, and all computational resources are fully exploited. It has been shown by an experiment on an NVIDIA K20C and INTEL E5645 that the proposed solution accelerates SAR data processing by tens of times. Consequently, a GPU based SAR processing system that embeds the proposed solution is much more efficient and portable, thereby making it qualified to be a real-time SAR data processing system. Experiments showed that SAR data can be processed in real-time at a rate of 36 megapixels per second by a K20C when the new solution is implemented.