Efficient Implementation Of GPU Based MUSIC Algorithm

In the current situation of electronic reconnaissance,the corresponding electric field environment is very elusive.People's daily increasing demand for direction finding technology has not been demonstrated by traditional direction finding technology.Therefore,the traditional direction finding technology has been gradually replaced by the spatial spectrum estimation technology which has strong direction finding resolution ability and relatively high precision ability.Among them,the MUSIC algorithm is one of the typical algorithms in the direction finding technology of spatial spectrum estimation.Nowadays,the MUSIC algorithm is indispensable to the rapid promotion of spatial spectrum estimation algorithm.Of course,MUSIC algorithm also has some drawbacks.For example,MUSIC algorithm has many and frequent matrix transformations,and can not process data immediately.Therefore,in order to maximize the efficiency of the MUSIC algorithm,there is a corresponding study on the efficient implementation of the MUSIC algorithm.This paper analyses and studies some drawbacks in the implementation of MUSIC algorithm,analyses and optimizes the performance of MUSIC algorithm in detail,improves the algorithm from the perspective of parallel processing,and combines with a reasonable hardware platform.The MUSIC algorithm is implemented on CPU+GPU platform.Finally,the time and efficiency of MUSIC algorithm before and after optimization are compared and analyzed by the acceleration ratio and the parallel efficiency,so as to maximize the efficiency of MUSIC algorithm.The main contents of this paper are as follows:1.According to the basic principle and process analysis of MUSIC algorithm,this paper expounds some bottlenecks encountered in the operation of MUSIC algorithm,and gives detailed analysis and solutions to bottlenecks.By testing the operation time of several important components of MUSIC algorithm,the time resource allocation of these components is analyzed in detail,and the preliminary optimization scheme of MUSIC algorithm is given.2.The function of GPU platform and the excellent parallel performance of GPU platform are elaborated.The CUDA framework,storage model and programming model based on GPU platform are briefly introduced.The MUSIC algorithm is transplanted from MATLAB language environment in CPU platform to C language environment based on CUDA framework on CPU+GPU platform for operation.The acceleration ratio and parallel efficiency are used to analyze the optimization of MUSIC algorithm.On the premise of the existing allocation of time resources,the time-consuming part of MUSIC algorithm is analyzed in detail,and the basic optimization scheme of MUSIC algorithm is given.3.Detailed analysis of the MUSIC algorithm,the ultimate efficient implementation of the MUSIC algorithm.In the operation of MUSIC algorithm,the computation of matrix in SVD is large.Therefore,the parallel processing method can be used to reduce the amount of data computation and improve the acceleration ratio and parallel efficiency of MUSIC algorithm.Then,the performance of the MUSIC algorithm's kernel function is analyzed,and the appropriate memory model is adopted to reduce the unnecessary interaction between data,thereby reducing the use time and improving the acceleration ratio of the MUSIC algorithm.Secondly,the transmission performance is analyzed,and the total time is reduced by allocating lock page memory to reduce inter-data copy.Finally,we add CUDA stream operations,aggregate the decentralized operations into a CUDA stream,and implement these stream operations in parallel to improve the acceleration ratio of MUSIC algorithm.After the above optimization,the performance of MUSIC algorithm is analyzed.By comparing the acceleration ratio and parallel efficiency before and after optimization,the performance of MUSIC algorithm is analyzed.