Research On GPU Parallel Computing Technology Of SAR Imaging

The synthetic aperture radar owns huge practical value in military and civil fields. SAR system generates huge volume of data, and the imaging algorithm compute complexly, thus it acquire higher processing speed. The effectivity of traditional CPU clusters has become quite low, meanwhile with high cost. General-purpose computing on GPU has become an effective means on efficient imaging of SAR system as its outstanding floating point calculations performance and high speed bandwidth. Current research of SAR imaging process based on GPU is not enough in-depth, mainly concentrated in specific algorithm implementation on single GPU. This thesis designs and implements SAR imaging system architecture of large-scale SAR data on GPU clusters. The concrete work includes the following several aspects.First, basing on deeply analyzing the parallel features of SAR imaging algorithm, combining with the features of GPU high performance computing platform, this thesis designs the SAR imaging processing framework basing on the master-slaves model for heterogeneous clusters of CPU and GPU. The framework includes master control module and compute module. The master module is in charge of the control logic of master node and attributing tasks, as computing module is in charge of the control logic of computing nodes and the processing logic of GPU. With the framework, this thesis designs the implementation scheme of single node with multi-GPU and multi-node with multi-GPU in detail. After deeply analyzing the features of SAR echo data, under the requirement of realizing loading balance, it designs the choose scheme of SAR parallel granularity with multi-node and multi-GPU environment, including the parallel choose scheme among nodes and the parallel choose scheme in nodes.Second, when using GPU computing in single node, the Shared Memory and Registers of CUDA are used to optimize the kernel functions of imaging algorithm. Point to the data copy problem among CPU and GPU as data blocking, it designs copy time hide method with CUDA stream and asynchronous parallel technology. Point to communication delay problem of SAR jobs transmission among multi-node, it designs multithreading stream processing method, hiding SAR transfer time among nodes.Third, this thesis mainly focuses on frequency domain imaging algorithm which facing to the strip-map mode. It chooses the typical frequency domain imaging algorithm, RD, CS and ?Kto verify this plan, and analyzing the difference according to their results. The results prove that comparing to single GPU imaging compute, the plan with multi-GPU in single node owns higher accelerate result and better parallel effectivity, and the plan with multi-node and multi-GPU owns high scalability, and meet instantaneity meanwhile.This thesis points to practical applications, researching on the rapid imaging method under the GPU environment and it has great practical significance to research of SAR imaging algorithm and SAR imaging applications.