Scene optimization of GPU-based back-projection algorithm.

The back-projection algorithm can adapt to synthetic aperture radar imaging in any mode because of its time-domain imaging characteristics, so it has received increasing attention in scientific research and engineering practice. However, the large enormous of computation of the back-projection algorithm limits its application. Since the back-projection algorithm has a high degree of data parallelism, it is very suitable for GPU parallel processing. The data scale and hardware environment in different application scenarios will result in different bottlenecks which may prevent the back-projection algorithm from performing optimally. Thus, this paper proposes a series of strategies to aid the back-projection algorithm in achieving the best peak performance with any given problem scale on any given platform. A heuristic block strategy is proposed to optimize the peak performance of the back-projection algorithm on servers and miniaturized GPU devices, which can handle the differences in hardware platforms as well as the differences in data scales. To optimize the peak performance of the back-projection algorithm on miniaturized GPU devices, a memory management strategy is proposed by using unified memory and pinned host memory. The method proposed in this paper has achieved an acceleration ratio close to the device's peak performance on servers and miniaturized GPU devices. By using the proposed methods, any algorithm with data-independent chunking and any GPU device with a uniform memory architecture can achieve higher peak performance. [ABSTRACT FROM AUTHOR]