Research And Implementation Of Parallel Resampling Algorithm Based On GPU

In modern DSP (Digital Signal Processing) system, we often convert the sampling rate of digital signals into another sampling rate in order to meet the needs of different situations, and this is what we called resampling. With the rapid development of information technology, the workload of processing, coding, transmission, and storage of digital signals is getting larger and larger. How to resample a large number of digital signals quickly has become one of the most important issues of the modern DSP. The rapid development of GPU (Compute Unified Device Architecture) brings us a new solution to this issue. In recent years, because of the development of the chip technology and the innovative development of the parallel computing architecture, GPU’s computing power has reached a hundred times of the contemporaneous CPU’s(Central Processing Unit), or even more. So GPU provides us an efficient platform for the high-performance computing, and the researches about the parallel algorithm based on GPU have become a hot issue of the high-performance computing. As the limitations of the traditional GPU’s hardware structure, programmers are difficult to use the GPU’s resources for general-purpose computing; however CUDA (Compute Unified Device Architecture) which is launched by the NVIDIA Corporation changes the state quo. CUDA can use GPU’s computing power for general-purpose computing. It’s based on the C programming language, and extends the C programming language. Therefore, programmers can use it to easily write programs which can run on GPU.This paper begins with a research about the CUDA architecture, and deeply investigates how to achieve the parallelization of the resampling algorithm with CUDA, and makes it have higher efficiency than the serial one based on CPU. The main work includes:(1) Achieve the serial algorithm based on CPU with Sinc interpolation.(2) Generalize the parallel parts of the serial algorithm and rewrite them, and then run it on GPU to achieve the parallelization of the resampling algorithm.(3) Write resampling program with Matlab, and validate the result of the serial algorithm and the parallel one, and then make a comparison between the time cost by the serial algorithm and the parallel one. The result shows that the time cost by the parallel algorithm based on GPU is much less than the serial one based on CPU. Finally, according to the experimental results, point out the bottlenecks to achieve high-performance computing algorithm on GPU, and then give some corresponding prioritization schemes.