Implementation Of Adaptive Optics Solar Speckle Image Reconstruction Based On GPU Parallel Computing

In ground-based solar observation,when the light passes through atmosphere,it will be affected by atmospheric turbulence which will cause translation,distortion and blurring of the received image.In order to eliminate or reduce the effects of atmospheric turbulence,many techniques,including adaptive optics and post-image processing techniques,are used to obtain high-resolution solar images.The technology we used in this paper is the post-image processing which based on the speckle reconstruction algorithm.Speckle reconstruction algorithm was divided into two steps,Fourier amplitude reconstruction and Fourier phase reconstruction.Speckle interferometry is adopted to reconstruct Fourier amplitude.For phase reconstruction,speckle masking was adopted.However,due to the large data volume and complex calculation,time consumption of the two steps is pretty huge,so that the traditional serial computation is difficult to meet the real-time requirements.GPU provides a new way to improve the computational efficiency of the algorithm,because of its powerful computing power,good programmability and economy.In order to accelerate the computation process of speckle image reconstruction,solar speckle image reconstruction algorithm has been parallelized by the parallel computing architecture-CUDA on the basis of discussing the principle of the algorithm.The following works have been done in this paper:Firstly,we introduced the influence of atmospheric turbulence on solar observation,and discussed in detail the principle and process of speckle reconstruction algorithm.Secondly,aiming at the problem of long time-consuming and low efficiency of speckle reconstruction algorithm on serial computing platform,different parallel optimization schemes were proposed for different steps by GPU platform and CUDA framework.Experiment result showed that a 2304 x 1984 pixel image of TiO channel can be reconstructed within 70 s under our operating environment.Compared with the program run on CPU,the speed-up radio can up to 7.It can be seen that the reconstruction speed of the parallel algorithm has been significantly improved.Thirdly,in order to further reduce the time consuming,we proposed an acceleration method based on dual GPU and a collaborative computing method based on CPU-GPU.The acceleration method based on CPU-GPU collaborative computing can be divided into CPU-GPU collaborative computing in sub-image and CPU-GPU collaborative computing between sub-image.Experiments showed that the speed of these three acceleration methods have been improved to some extent compared with single GPU scenario.It is foreseeable that the goal of real-time reconstruction solar image can be achieved with more GPUs in the future.