Parallel Analysis And Implementation Of Phase Diversity

Affected by atmospheric turbulence, the amplitude and phase of wavefront varies randomly as beam passing through the atmosphere. In addition, there are other factors that also make the wavefront distorted, including alignment and manufacturing error of every optical element as well as environmental factor such as temperature, gravity. In order to improve the image quality, these aberrations caused by the above factors should be eliminated or reduced. Therefore, researchers proposed most of techniques, which can be mainly divided into two categories. The first one is adaptive optics which measures the wavefront aberrations in the incoming light from one object and transmits electronic signals to a deformable mirror that can change its shape rapidly to correct for the aberrations. The second one is image post-processing technology, including speckle imaging technique, deconvolution technique, blind deconvolution technique, and so on.Phase diversity method is not only used as a kind of wavefront sensing technique, but also as an image restoration technique. It has many advantages over other wavefront sensors. First, the optical system is easy to be established and it is not susceptible to systematic errors related to optical hardware. Second, it can easily be integrated into the image array to eliminate the non-common path errors of system. Finally, the technique is applied to both a point source and an extended scene. However, the computations of PD are burdensome, and thus it is mostly applied in the field of image post-processing or used with adaptive optics technique. In recent years, with the development of high performance computing technology, many large-scale computing problems are resoved. On this basis, how to speedup the computation of PD system is mainly carried out in this thesis, the detailed content of dissertation is as following:Firstly, we carried on deep discussion and analysis on the principle of PD technology to find out the performance bottlenecks, which limited the computational efficiency. In the perspective of task partitioning, the cost function and gradient function are the most time-consuming parts; in the perspective of calculated cell, FFT and reduction are the most time-consuming parts.Secondly, NVIDIA as the inventor of the GPU inventor provides a library function CUFFT for computing FFT. After several generations update of CUFFT, there are still promotion spaces and CUFFT is not suit for kernel fusing on GPU to reduce the memory access. We develop our own custom GPU FFT implementation based on the well-known Cooley-Tukey algorithm and analyze the relationship of coalesce memory access and occupancy of GPU and get the optimal configuration of thread block. The results show that the proposed method improved the computational efficiency by 1.27 times than CUFFT 6.5 for double complex data size of 512x512 pixels.Thirdly, we analyze the computing process of Optical Transfer Function(OTF), which is the most time-comsuming part of cost and gradient function. A fast OTF calculation method based on dual Fourier transform, in which the redundant computations of point spread function(PSF) and OTF are removed, is proposed in this dissertation. For an optical system with the pupil function of 512×512 pixels, the OTF computation is implemented on GPU platform. The result shows the proposed method enhances the computational efficiency about 4.88 times than the method using standard CUFFT.Fourthly, the redundancy computations for the cost and gradient function are analyzed and two kinds of implementation methods based on GPU are compared: one is the general method which is accomplished by GPU library CUFFT without precision loss(method-1) and the other one performed by our own custom FFT with a little damage of precision related to the redundant calculations(method-2). The results show that method-2 greatly speeds up the cost and gradient functions, and then the overhead of global memory accesses are reduced by kernel fusion. For the image system with the sampling factor of 3, the results reveal that method-2 achieves speedup of 1.83 compared with method-1 when the central 128×128 pixels of the point spread function are used for the image size of 256×256 pixels. And then reasonable task-assigned strategies are applied on dual GPUs, the computational efficiency of PD is incresaded by 17% using method-2.Finally, in order to study the performance of PD for static aberrations, a closed-loop adaptive optics experiment setup using LC-SLM is built based on a point source. In the experiment, the point resource is simulated by a He-Ne laser and a spatial filter, and the LC-SLM is used for creating known defocus aberrations and correcting the static aberrations. The simulation and experimental results both demonstrate that the PD can effectively correct the wavefront aberration of optical system.