High Resolution And Fast Space-Borne SAR Imaging Research Based On Heterogeneous SIMD Parallelism

Synthetic Aperture Radar (SAR) has been developed really fast for recent years, higher resolution and wide swath bring more and more huge amount of SAR raw data to the ground imaging system. Meanwhile, there is an increasing demand for higher performance and efficiency of processing SAR raw data. Thus, there are more and more researches about fast or even real-time SAR imaging, kinds of high performance computing platforms and methods are applied to the SAR imaging system.Due to the driving of graphics market, there are many fast SAR imaging methods implemented with Graphics Processing Unit (GPU). Due to the massively parallel mode of GPU, it is used as the main computing processor in the classical GPU based imaging algorithms which take on most of the computing tasks during SAR imaging procedure. On contrast, the Central Processing Unit (CPU) only takes on some auxiliary works in imaging process, like SAR raw data input/output (I/O), stream controlling and so on. Hence, the computing resources and capability of CPU is underestimated or even ignored in the classical GPU based imaging algorithms.In this paper, a fast SAR imaging method based on deep CPU/GPU heterogeneously collaborative computing is proposed. By reasonable task partitioning and scheduling strategy considering collaborative computing, the whole SAR imaging algorithm can be distributed to both of CPU and GPU, then it can be implemented concurrently. The tasks that distribute to CPU can be deeply optimized by the Single Instruction Multiple Data (SIMD) based Advanced Vector Extensions (AVX) instructions, which is firstly introduced into the multi-core CPU SAR imaging algorithm. Supported by the Compute Unified Device Architecture (CUDA), the proposal in this paper has solved two problems faced by the exiting GPU imaging method, the one is that the limitation of SAR data caused by the limited device memory, and the other is the frequent memory copying operations between device and host. Beyond that, there are some other CUDA optimizing strategies applied in this proposal, like streaming technology, parallel pipeline and so on.Experimental results show that the proposed fast space-borne SAR imaging method implemented with deep CPU/GPU collaborative computing can improve the imaging efficiency about 270× times compared to the single-core CPU based SAR imaging method. The imaging efficiency based on this proposal is much higher than the data generating efficiency, so it can be seen as a real-time SAR imaging algorithm.