GPU Many-core Parallel Computing Of Broadband Typical Adaptive Beamforming For Acoustic Signal Processing

MVDR and MUSIC algorithm are typical beamforming methods in array underwater acoustic signal processing.For the high efficiency of MVDR and MUSIC,we propose the GPU parallel acceleration algorithms of MVDR and MUSIC,combined with beamforming algorithm,GPU architecture and CUDA(Compute Unified Device Architecture)parallel programming environment.The algorithms are designed by analyzing the complexity and numerical algorithms of the two types of beamforming algorithms and combing with thread task division of GPU architecture,storage mapping strategy and CPU / GPU cooperative computing.Our main works include:1)We introduce the GPU computing architecture and programming environment in detail and summarize the main points of CUDA program performance optimization.Then,we describe the concept of array signal beamforming and the related research of adaptive beamforming algorithm MVDR and MUSIC.Finally,we conclude the research status of Beamforming Algorithm and Parallel Computing of Beamforming.2)Based on MVDR algorithm,we propose the C++ serial program of MVDR algorithm and the parallel program based on OpenMP.According to the characteristics of GPU and CUDA,the GPU parallel algorithm(CUDA MVDR)is designed for the two dimensions of frequency and element.We provide extensive experimental evaluation to demonstrate CUDA MVDR high efficiency compared to the MVDR algorithm serial program,and OpenMP MVDR parallel program,taking the 512-based and 1000 frequency points for example.The experiment demonstrates that the MVDR algorithm GPU parallel computing obtains 38.37 x speedup over the CPU single-threaded parallel and 3.08 x speedup over the single CPU node(16 cores)OpenMP multi-threaded parallel.3)Based on MUSIC algorithm,we propose the MUSIC serial program and the multicore parallel program of OpenMP.For the two dimensions of the frequency and the array element,the azimuth spectrum calculation part is parallelized in the algorithm.And thread task division of GPU architecture,storage mapping strategy and CPU / GPU cooperative computing,the GPU parallel algorithm(CUDA-MUSIC)is proposed.We provide extensive experimental evaluation to demonstrate CUDA-MUSIC high efficiency compared to the MUSIC algorithm serial program,and OpenMP MUSIC parallel program.The experiments show that CUDA-MUSIC obtains 7.25 x speedup over the CPU single-threaded,and 1.5x speedup over the single CPU node(16-core)OpenMP multi-threaded parallel.