Research And Implementation Of Gpu-based Quick Audio Retrieval Algorithms

Along with the rapid development of multimedia technology and Internet technology, multimedia information is continuous expanding. How to search information fast and accurately in mass data has become a challenge of content-based audio retrieval. Currently, audio retrieval algorithms, which are based on Central Processing Unit (CPU) architecture, have been relatively mature. Simply modify the algorithm has been unable to obtain significant performance improvement. Therefore, it is necessary to use other computing devices to reduce the audio retrieval algorithm execution time. Graphic Processing Unit (GPU) can provide a powerful floating-point computing power and high-speed data transmission, but also Compute Unified Device Architecture (CUDA) provides unified and easy operator interfaces, use of GPU accelerating existing audio retrieval algorithm is of great significance.The main purpose of this study is to design fast GPU-based audio retrieval algorithm to obtain the ideal performance. Mainly focus on the following aspects:1. This dissertation analyzes the GPU Tesla architecture and CUDA, and describes how to decompose a compute task to parallel form, and map it to Tesla GPU through the two-leveled programming model of CUDA.2. A real-time audio retrieval algorithm based on segmentation is introduced. In order to determine the core calculation steps need to accelerate, this dissertation analyses an existing audio retrieval systems combined with the CUDA programming features. Two solutions that based on thread and based on block are put forward.3. Two algorithms were tested. Experimental results show that speedup is 19.9 and 22.3 respectively.