基于多DSP的信号处理并行算法设计与实现-模式识别与智能系统专业论文.docxVIP

华 华 中 科 技 大 学 硕 士 学 位 论 文 II II Abstract In order to provide a flexible scalability and basic algorithms library for a high performance signal processing computing platform to meet the requirement of modern signal processing, this thesis introduces the parallel technology for enhancement of the computing performance and the huge demand of speed. This thesis emphasizes on the parallel algorithm and its applications in digital signal processing. It mainly covers how to design, implement and analyze parallel algorithms, corresponding parallel system architectures, parallel algorithm models and other aspects. Firstly the basic concepts of the parallel processing technology, parallel computing model and the performances of parallel algorithms are introduced. Then the multi-DSP parallel processing system platform which is suitable for the subject is described. Finally the design methods and results analysis of parallel algorithms of Fourier transform, Hadama transform and digital filtering are discussed. In parallel processing, allocation of tasks, data storage and communication problems are keys affecting the running time of parallel algorithms. In this thesis, using the parallel partition strategy in FFT parallel algorithm, two parallel algorithms which are bit exchange and matrix transpose are designed. Implementing these two algorithms in multi-DSP system (four TigerSHARC101 DSP chips) to compute 2K points complex FFT, the running time are respectively 14.39ms and 10.92ms, and the speedup rates are respectively 2.39 and 3.15, which indicates that the method of matrix transpose effectively makes up the shortage of communication start problem in the method of bit exchange and it has higher parallel efficiency. The parallel algorithm of Fast Hadama Transform in accordance with its butterfly characteristic is also presented. Implementing this algorithm in the multi-DSP system costs 3.97ms and gets speedup rate of 2.75. Using the pipeline technology, the parallel algorith