Research On Pulsar Data Processing Technology Of Coherent De-dispersion Based On GPU

The Pulsar are fast-rotating neutron stars with stable rotation period and extreme physical conditions such as ultra-high density,ultra-high pressure,and ultra-strong magnetic field.During the propagation process,it is affected by the Interstellar Medium(ISM)to produce chromatic dispersion,which leads to the broadening of the received pulse profile.The pulsar data needs to be processed by de-dispersion technology.Incoherent de-dispersion is carried out in the time domain.The multi-channel filter bank is used to realize the post-detection de-dispersion.The coherent de-dispersion is processed by fast Fourier transform in the frequency domain.The digital signal is processed by the interstellar medium transfer function in the frequency domain to realize the entire passband.Dispersion elimination is more thorough than incoherent de-dispersion.With the deeply research on radio pulsars,coherent de-dispersion pulsar digital terminals with wide bandwidth,super-high time resolution and frequency resolution have become a new development trend.Coherent de-dispersion algorithm requires a large number of floating-point operations.Compared with incoherent de-dispersion,the data processing cycle is longer and the realization is more difficult.It puts forward higher requirements for pulsar observation and data processing capabilities.In this context,this paper studies and implements a coherent de-dispersion baseband data processing technology based on Graphics Processing Unit(GPU),high-speed shared memory and DSPSR(Digital Signal Processing Software for Pulsar)as the data processing core.And test and analyze the performance of the data processing flow.The specific research content and results of the thesis include:First,to understand and analyze the principle of coherent de-dispersion and the realization of the pulsar data processing process,starting from the massive baseband data processing requirements,in-depth study of the data processing principle of the pulsar digital signal processing package DSPSR and the high level of the open source software PSRDADA(Processing Software for Pulsar Distributed Acquisition and Data Analysis)The performance principle of shared memory is based on the overall design of the coherent de-dispersion baseband data processing flow,which lays the foundation for the realization of the subsequent data processing flow.Second,the implementation of GPU-based coherent de-dispersion pulsar baseband data processing flow.Write the program using the principle of high-speed shared memory to achieve the highest two-way polarized baseband data input.The subsequent decoding,de-dispersion,folding and other data processing procedures are processed by DSPSR,and GPU acceleration technology is used to accelerate the data processing flow.The data processing results are based on PSRFITS Format storage.In order to verify the data processing quality of this data processing procedure,the baseband data obtained by FAST observation of PSR1939+2134 and PSR 0534+2200 were processed,and the pulse profile was successfully obtained.Compared with the incoherent dedispersion data processing result,the dispersion elimination was more thorough A finer pulse profile can be obtained,and the signal-to-noise ratio meets the requirements of subsequent research work such as data timing.Third,in view of the characteristics of long processing cycle of coherent dedispersion data,large amount of calculation,and the processing requirements of massive pulsar observation data,a data processing flow is built on the graphics processor platform to perform performance testing.Based on the experimental analysis of the acceleration effect of the whole process of coherent de-dispersion data processing,the acceleration ratio of the fast Fourier transform of the main algorithm of coherent de-dispersion was further tested.In terms of data processing efficiency,first test the data processing time of PSR 1939+2134 baseband data in GPU and CPU(Central Processing Unit)modes.The experimental results show that dual polarization with 8bits sampling and 21.4 s observation time It takes about 124 s to process the baseband data.Compared with the CPU mode,the speedup ratio can reach 2.07.Then,starting from the principle of coherent de-dispersion,the time required for fast Fourier changes of different points in the two modes is compared,and the effect of GPU parallel computing on the efficiency of the coherent de-dispersion algorithm is analyzed.Finally,combined with the GPU hardware architecture,the reasonable allocation of thread resources is discussed.