Research On Optimization Of W-PROJECTION Parallel Algorithm

With the construction of large-scale synthetic aperture radio telescopes such as square kilometer array(SKA),the baseline length and observation field of view continue to increase.In order to improve imaging quality,the effects of large field of view and non-coplanar baseline effects need to be considered.W-projection has excellent computational speed and error control as an important algorithm for processing large field of view imaging and non-coplanar baseline effects.However,the algorithm has a large memory consumption and need to handle large amount of astronomical observation data.Therefore,it is necessary to improve parallel efficiency of the w-projection algorithm.In this thesis,the algorithm is optimized in terms of imaging quality and parallel efficiency of the algorithm.The parallel optimization method is used to reprogramming the algorithm and optimize the performance under CUDA.The specific research work is as follows:(1)The value of w-plane in w-projection and w-stacking algorithm determines the number of convolution kernels,and has an important influence on the imaging quality.The original w-plane experience value cannot handle the large field of view imaging well.Therefore,the w-plane optimal empirical value is obtained by experimental analyzing the value of w-plane to improve the imaging quality.(2)The w-projection algorithm is computationally complex and needs to process a large amount of observation data,so porting the algorithm to the GPU to improve the parallel efficiency of the algorithm.Aiming at the problems existing in the implementation of the current algorithm on the GPU,the memory allocation,the number of threads and the register usage in the thread block are optimized respectively,and then the programming is implemented on the CUDA platform according to the new design method,which effectively improves the parallelism of the algorithm.(3)The implementation of the w-projection algorithm under the CUDA platform needs to optimize the performance of the algorithm.By using the w-plane optimal experience value,the data block size and tile ratio of the algorithm under CUDA platform are analyzed experimentally,and the block size which is suitable for processing the observation data of SKA1-low telescope is obtained.The research results of this thesis:(1)In the process of analyzing the influence of w-plane on the processing speed and imaging quality of w-projection algorithm,the optimal experience value of w-plane is obtained,and the quality of the algorithm is improved;(2)For the existing GPU algorithm implementation problems such as: GPU device memory,the number of threads and the use of registers in the thread block have been improved and programmed.(3)Optimize the running performance of the CUDA code implementation of the algorithm.Through the experimental analysis,the optimal tile partition size for the SKA1-low telescope is obtained,which effectively improves the data parallel processing speed and quality of the SKA1-low telescope.