The Implement of Common Beam Forming Using GPU

In order to study how to use GPU in real-time signal processing system, we implement common beam forming arithmetic using it. In a GTX285 GPU, computing speed is 170-450 times faster than AD TigerSharc201. This shows good prospects of GPU. Keywordscbf, GPU, CUDA 1. Preface Beam forming (BF) is the major component of Sonar Signal Processing. It fixes the target position by delaying processing multi-sensors signals. It is the basis of achieving all functions of Sonar System. Therefore, beam forming is necessary for whatever passive sonar or active sonar. Nowadays, most of digital sonar beam forming are implemented by taking the general digital signal processing base on Harvard architecture. But its development has been significant declined. It has already cannot meet Moore's Law. The industry's fastest floating-point DSP is the Tiger Sharc201 processor launched by Analog Company in 2003. It speeds up to 600MHz, floating-point computing power only 3.6 GFLOPS, has been lower than the current mainstream general-purpose processor (CPU). With the application demand for computing power growing, the sonar system can only continue to increase the number of DSP chips to meet the application requirements, which will inevitably increase the system complexity, resulting in increased costs, reduce reliability. Therefore, more and more people have been paying attention to the new high-performance processors instead of DSP. In recent years, graphics processing unit (GPU) have developed rapidly, the mainstream GPU's floating point computing power has been more than two orders of magnitude higher than the DSP, and it has cost-effective, high-compute density, low power consumption and other advantages. How to use the powerful GPU on sonar system has become an important reality issue. GPU has a large number of parallel thread processors which share multi-level high-speed memory. It is suitable for many graphic operations and other similar high parallelism of data-intensive applications, such as sonar beam forming system. For such algorithms, NVIDIA's CUDA platform look at the GPU as a data-parallel computing device and use C-like languages to develop GPU-based parallel algorithm. CUDA can also expand computing power with Open-MP, MPI, PVM and other parallel mechanism. Currently, there are GPU-based applications using CUDA in some civilian areas, such as molecular dynamics simulation , seismic wave simulation, CT 【5】 etc.. Through the GPU algorithm optimization, the system performance improved to the dozens of times. However, real-time signal processing such as sonar + Corresponding author. E-mail address: jiangjh98@yahoo.com.cn 2010 3rd International Conference on Computer and Electrical Engineering (ICCEE 2010) IPCSIT vol. 53 (2012) © (2012) IACSIT Press, Singapore DOI: 10.7763/IPCSIT.2012.V53.No.2.35 and radar algorithms implemented using GPU has not been reported. This paper implemented GPU-based common frequency domain beam forming algorithm on the CUDA platform. Section 2 describes the frequency domain common beam forming algorithm and its GPU implementation. Section 3 tests the correctness of the implementation and comparison with the main CPU and DSP performance of the play. Section 4 concludes the paper and looks ahead the prospects of GPU application.