基于CUDA的高速并行高斯滤波算法.pdfVIP

DOI ：CNKI:42-1658/N1122.003 网络出版时间：2011-06-13 11:22 第39卷 第5期 华 中 科 技 大 学 学 报 ( 自然 科 学 版) Vol.39 No. 5 2011年 5月 J. Huazhong Univ. of Sci. Tech. ( Natural Science dition) May 2011 网络出版地址：/kcms/detail/42.1658.N1122.201105.10_003.html C DA 卢文龙 王建军 刘晓军 ( 华中科技大学 械科学与工程学院, 湖北 武汉430074) 为加快表面三维形貌分析中高斯滤波算法的执行速度, 提出了一种基于计算统一设备构架( CUDA) 的 高斯滤波算法来实现高速并行处理.分析高斯滤波算法原理和 CU DA 并行计算体系,将 CUDA 并行计算技 术引入到表面分析领域.针对高斯滤波数据间依赖性弱和CUDA 采用单指令多线程( SIMT ) 执行模型的特 点, 总结出适合于CUDA 的并行高斯滤波算法流程. 实验证明: 该方法与 CPU 串行处理方法相比, 其加速比 达到40 倍以上, 可以有效提高数据处理能力. 高斯滤波; 计算统一设备构架; 图像处理器; 表面形貌; 并行处理 TH 132. 1 A 167124512(2011) 0520010204 C DA2basedalgorithm for high2speed parallel Gaussian filtering L u Wenlong Wang J ianj un L i u X i aoj un (School of Mechanical Science and ngineering, Huazhong University of Science and Technology, Wuhan 430074, China) Abstract In order to speed up the Gaussian filtering in three2dimensional surface texture analysis, an efficient method based on compute unified device architecture ( CUDA) implemented on graphic pro2 cessing unit (GPU) was designed. CUDA parallel computing technology was introduced into the sur2 face texture analysis area by analyzing principle of Gaussian filtering and CUDA computing architec2 ture. Parallel Gaussian filtering algorithm used in CUDA was given for the characteristics of low de2 pendency of Gaussian filtering and SIMT ( single instruction multiple thread) execution model of CUDA. xperiments prove that the calculation speed based on CUDA is 40 times faster than that of traditional sequence algorithm based on CPU , can effectively improve data processing capability. Key w