摘要
随着GPU计算能力及可编程性的不断增强,采用GPU作为通用加速器对应用程序进行性能加速已经成为提升程序性能的主要模式.直方图生成算法是计算机视觉的常用算法,在图像处理、模式识别、图像搜索等领域都有着广泛的应用.随着图像处理规模的扩大和实时性要求的提高,通过GPU提升直方图生成算法性能的需求也越来越强.在GPU计算平台关键优化方法和技术的基础上,完成了直方图生成算法在GPU计算平台上的实现及优化.实验结果表明,通过使用直方图备份、访存优化、数据本地化及规约优化等优化方法,直方图生成算法在AMD HD7850 GPU计算平台上的性能相对于优化前的版本达到了1.8~13.3倍的提升;相对于CPU版本,在不同数据规模下也达到了7.2~210.8倍的性能提升.
Application developers increasingly adopt GPUs as standard computing accelerators to improve application performance with their easier programmability and increasing computing power. The histogram generation algorithm is a common algorithm of computer vision, and is widely used in image processing, pattern recognition and image search. With the scale enlargement of image processing and the demand of real-time, improving the performance of histogram generation algorithm by GPU is in increasingly high demand. We introduced the realization and optimization on GPU of histogram to research the major optimization methodologies and technologies. Experimental results show that the appli- cations of access optimization of histogram backup, memory optimization, data localization and mergence optimization, and some other optimization strategies,bring about a 1.8 ~ 13.3 times speedup for the algorithm on AMD HD 7850, than versions before optimization,and brings about a 7. 2~ 210. 8 times speedup than CPU versions.
出处
《计算机科学》
CSCD
北大核心
2015年第11期32-36,共5页
Computer Science
基金
国家自然科学基金(61272136)
国家自然科学基金创新群体(61221062)资助