Research On GPU-Based Broadband Beamforming Algorithm

Broadband beam processing technology with high gain,strong anti-interference and high spatial resolution,etc.,is an important means of underwater acoustic signal processing.However,the broadband processing,especially the time-domain processing,has high computational complexity,and real-time beam processing systems are more difficult to implement on traditional computing platforms.Therefore,it is necessary to carry out high performance processing techniques suitable for broadband beam processing in order to provide guidance for the design of future broadband beam processing systems.On the basis of broadband undistorted beam processing,this dissertation focuses on the real-time implementation of broadband beam array processing in arrays,and studies the high performance parallel processing architecture based on GPU to improve the existing methods of wideband beamforming and improve the real-time processing of broadband beamforming ability.The main work of the dissertation is as follows:First of all,in order to obtain a distortion-free received signal,DFT beamformer,fractional delay beamformer and constant beamwidth beamformer are analyzed.To solve the problem that the suture is not smooth in the frequency domain,the solution to the overlapping segmentation is given and the improvement is analyzed.The core of time-domain implementation is the design of Fractional Delay Filter.Two design and implementation methods are given.For the more general FIR filter design,the second-order cone programming method is given,and the design errors of several methods and Performance simulation analysis and comparison.Aiming at the waveform distortion of azimuth deviation data beam output,a design method of constant beamwidth beamformer based on distributed design is given.The design error and performance of each step are simulated and analyzed.Secondly,based on CUDA heterogeneous programming model,the optimization design method of broadband beamformer based on GPU is studied.Firstly,the task of algorithm calculation is analyzed.The calculation task in frequency domain is divided into two parts: FFT calculation and frequency-domain weighting.FFT is efficiently accelerated by CUFFT.For the step of calculating the frequency-domain weighting with lower density,by letting a single thread calculate multiple Beam output increases the computational density of this step.The computational tasks in the time domain are focused on the FIR filtering process.The overlap preserving method is used to decompose the tasks to improve the parallelism.The computation density can be effectively improved by obtaining a single thread without any dependence on multiple data to obtain higher performance.Finally,different sizes of FFT calculations and FIR filtering were tested on Intel(R)Xeon CPUs and Quadro K5200 GPUs for performance and speedup.Finally,the performance of broadband beamforming system is tested,the data of sea trial is processed,the correctness of the algorithm is verified,and the demand of real-time processing is satisfied.