基于多核的Loeffler算法的并行优化与实现

发布时间：2018-06-03 21:53

本文选题：DCT变换 + 多核平台　；参考：《郑州大学》2013年硕士论文

【摘要】：随着计算机技术和通信技术的快速发展,特别是大数据时代的到来,对计算机效率、计算能力的要求也越来越高。多核处理器已经日益普及,充分利用多核提供的高性能开发并行软件,逐步取代串行软件,将是未来软件行业的发展趋势。所以,对于传统的串行算法,在多核平台上实现算法的并行优化,对于提高计算机运行效率意义重大。本文首先介绍了多核平台上并行优化快速DCT变换涉及到的多核、多线程、线程同步和并行计算等一些关键技术,以及快速DCT变换涉及到的数学基础知识；在此基础上简要描述了压缩解压缩过程,并讨论快速DCT变换的整体结构。同时,通过对变换的核心部分-循环变换的分析,提出基于数据分解的分解方式对快速DCT变换之一的Loeffler变换进行并行优化,详细描述分解过程,同时进行可行性分析。根据DCT变换的本身特征,特别是二维DCT变换的特征,采用数据分解的方法进行分析。常规的DCT变换采用双重For循环来进行DCT变换,就是采用循环嵌套来逐步实现对输入系数的转换工作。快速DCT则是利用DCT变换的可分解性,先进行行变换,再进行列变换。本文对典型的快速变换算法之一Loeffler,进行进一步的数据分解,利用多核多线程的优越性实现Loeffler算法的并行优化。实验结果表明：常规DCT变换、快速DCT变换、并行优化快速DCT变换,其执行效率在逐步递增。特别是并行化后的快速Loeffler变换在特定情况下有明显的时间优越性。在处理大数据量的图片信息时,性能提升更加明显。实验证明当数据规模在1600×1200时效率可提高30%左右,且随着数据规模的增加,效率提升趋于平稳。总之,对快速Loeffler变换的并行化所带来的性能优势基本符合理论预期目标。
[Abstract]:With the rapid development of computer technology and communication technology, especially the arrival of the era of large data, the demand for computer efficiency and computing power is becoming higher and higher. Multi core processors have become increasingly popular. It is the trend of the future software industry to make full use of high performance and parallel software provided by multi-core and gradually replace serial software. For traditional serial algorithm, parallel optimization of algorithm on multi-core platform is of great significance for improving the efficiency of computer operation.
This paper first introduces some key technologies, such as multicore, multithreading, thread synchronization and parallel computing, as well as the basic knowledge involved in fast DCT transform, and then briefly describes the compression and decompression process, and discusses the whole structure of fast DCT transform. After the analysis of the core part of the transformation - cyclic transformation, the decomposition method based on data decomposition is proposed to optimize the Loeffler transform of one of the fast DCT transformations. The decomposition process is described in detail, and the feasibility analysis is also carried out.
According to the characteristics of the DCT transformation, especially the characteristics of the two-dimensional DCT transformation, the method of data decomposition is used to analyze. The conventional DCT transformation uses a double For cycle to carry out DCT transformation. It is the use of cyclic nesting to gradually realize the conversion of the input coefficients. The fast DCT is the decomposability of the DCT transformation and the advanced line transformation. In this paper, one of the typical fast transformation algorithms, Loeffler, is further decomposed, and the parallel optimization of the Loeffler algorithm is realized by using the superiority of multi core and multithreading. The experimental results show that the efficiency of the conventional DCT transform, fast DCT transform and parallel optimization for fast DCT change is gradually increasing. The improved fast Loeffler transform has obvious time superiority in a particular case. The performance enhancement is more obvious when processing large data amount of picture information. The experiment proves that when the data scale is 1600 * 1200, the efficiency can be increased by about 30%, and the efficiency lifting tends to be stable with the increase of data size. In a word, fast Loeffler transform is used. The performance advantages brought by parallelization are basically in line with the theoretical expectations.
【学位授予单位】：郑州大学
【学位级别】：硕士
【学位授予年份】：2013
【分类号】：TP338.6

【参考文献】