基于CPU-GPU异构计算的小波变换研究与实现

发布时间：2018-06-05 06:59

本文选题：小波变换 + 异构并行计算　；参考：《西安电子科技大学》2013年硕士论文

【摘要】：目前，随着计算机硬件技术的飞速发展，多核中央处理器(Central ProcessingUnit, CPU)和通用计算的图形处理器(Graphic Processing Unit, GPU)更新换代层出不穷，但软件应用却跟不上其硬件的发展速度，导致大量的硬件资源浪费。为了发挥其各自的性能特点和提高资源的利用率，通过并行计算对密集型计算任务加速，具有很高的应用价值和实际意义。小波变换以其在时域和频域都具有良好的局部化特性使得其在图像压缩编码领域得到广泛的研究和应用。而随着图像信息数量的剧增，，对小波变换的加速处理进而提升图像压缩的速度具有重要的实际意义。本文详细分析了小波变换的原理及过程，并阐述了CPU-GPU异构并行计算的特点。在CPU实现小波变换基础上，研究了CPU-GPU异构并行计算的程序任务的划分，性能的评估以及性能优化策略。通过对小波变换实现的具体分析和研究，采用了基于OpenMP-CUDA混合编程模型实现了异构系统的协同运算。重点研究了9/7小波滤波在GPU上的优化，通过对小波变换算法的优化，使其更适应GPU上硬件资源结构和更能发挥并行计算性能上的优势。针对小波滤波中数据划分带来低通和高通系数相关性问题，提出在代价最小的情况下实现数据的并行划分，极大提高了程序的运行效率。基于CPU-GPU异构并行计算的小波变换相比CPU串行的处理速度提升了15-36倍。
[Abstract]:At present, with the rapid development of computer hardware technology, multi-core CPU Central processing Unit (CPU) and General Computing Graphics processor (GPU) are constantly emerging, but the application of software can not keep up with the development of its hardware. Lead to a lot of waste of hardware resources. In order to give full play to their respective performance characteristics and improve the utilization of resources, intensive computing tasks are accelerated through parallel computing. Wavelet transform has been widely studied and applied in the field of image compression because of its good localization in time domain and frequency domain. With the rapid increase of the number of image information, it is of great practical significance to accelerate the processing of wavelet transform and then accelerate the speed of image compression. In this paper, the principle and process of wavelet transform are analyzed in detail, and the characteristics of CPU-GPU heterogeneous parallel computing are expounded. Based on the wavelet transform implemented by CPU, the task partition, performance evaluation and performance optimization strategy of CPU-GPU heterogeneous parallel computing are studied. By analyzing and studying the realization of wavelet transform, a hybrid programming model based on OpenMP-CUDA is used to realize the cooperative operation of heterogeneous systems. This paper focuses on the optimization of 9 / 7 wavelet filtering on GPU. By optimizing the wavelet transform algorithm, it is more suitable for the hardware resource structure on GPU and can play the advantage of parallel computing performance. In order to solve the problem of correlation between low pass and high pass coefficients brought by data partition in wavelet filtering, the parallel partition of data is proposed in the case of minimum cost, which greatly improves the running efficiency of the program. The processing speed of wavelet transform based on CPU-GPU isomerous parallel computing is 15-36 times faster than that of CPU.
【学位授予单位】：西安电子科技大学
【学位级别】：硕士
【学位授予年份】：2013
【分类号】：TP338.6;TP391.41

【参考文献】