| In the field of radar and sonar, electromagnetic interference is intense and signalquality is poor. By adjusting the weight of antenna array element, adaptive arrayantenna can form the main beam at the coming direction of target signal and form theside lobe at the coming direction of interference signal, so as to suppress the undesiredsignal. Adaptive beam forming technology, also called as adaptive beam formingalgorithms, is the core of adaptive antenna technology. In the time filed the classicaladaptive beam forming algorithms includes the least mean square (LMS) algorithm, therecursive least square (RLS) algorithm and sample matrix inverse (SMI) algorithm. AndSMI algorithm is applied in the field of radar and sonar due to its faster convergencespeed and better robust performance. However, SMI algorithm needs to be realized withsample covariance matrix inversion. And this increases the computation cost. Thus, highperformance CPU and advance digital signal processor (ADSP) are used.This paper selects the GPU as parallel computing platform. According to theproblem model abstracted from adaptive beam forming in the actual filed, we designedthe parallel SMI algorithm based on GPU, and realized these key steps including slipmatrix construction and covariance matrix inversion, etc. And we proposed severaloptimization measures, including the execution order of SMI algorithm, storageaccessing and the data storage format. These optimization can improve the executionspeed of SMI algorithm based on GPU. The experiment results demonstrate thatCPU+GPU heterogeneous SMI algorithm can win20-22x speedups in dealing with areference signal compared to the CPU version of SMI algorithm. And by making use ofthe characteristic of the next generation GPU Fermi architecture, our program ofconcurrent kernel execution can make the computing time reduced by38.5%in dealingwith two reference signals compared to the program of sequential kernel execution. |
PDF Full Text Request