基于GPU的LLE算法加速及性能优化  

Acceleration and performance optimization of LLE algorithm based on GPU

作　　者：李繁 严星[2] 张晓宇 LI Fan;YAN Xing;ZHANG Xiao-yu(Network and Experimental Teaching Center,Xinjiang University of Finance and Economics,Urumqi 830012,China;School of Information Management,Xinjiang University of Finance and Economics,Urumqi 830012,China)

机构地区：[1]新疆财经大学网络与实验教学中心,新疆乌鲁木齐830012 [2]新疆财经大学信息管理学院,新疆乌鲁木齐830012

出　　处：《计算机工程与设计》2021年第5期1314-1322,共9页Computer Engineering and Design

基　　金：国家自然科学基金项目(41830101);新疆自治区社科基金项目(17BTQ093);新疆财经大学青年博士基金项目(2015BS003)。

摘　　要：为提高非线性数据降维算法效能,分析这类算法的特点,综合考虑KNN计算和解决Sparse特征值两个问题,提出将LLE算法中的KNN搜索算法及大型稀疏矩阵解特征值这两个部分并行在GPU的运算平台上,通过这种方法来加快所有基于LLE发展而来的数据降维技术的执行时间。仿真计算结果表明,在KNN方面整体加速可达40至50倍,在解大型稀疏矩阵特征值的部分加速至10倍左右。整体来说,数据降维算法加速10倍左右,有效运用GPU提高了LLE这类算法的性能。To increase the efficiency of the nonlinear data dimension reduction algorithm,characteristics of this kind of algorithms were analyzed,and KNN calculation and SPARSE eigen solution were also taken into consideration.The KNN search algorithm and large sparse matrix solution eigenvalue of LLE algorithm were put forward to parallel on GPU platform,the execution processes of data dimension reduction techniques developed based on LLE were speeded up using this method.Experimental simu-lation shows that the overall acceleration can reach 40 to 50 times in KNN,and the acceleration can reach about 10 times in the part of solving the eigenvalues of the large sparse matrix.On the whole,the data dimensionality reduction algorithm is accele-rated by about 10 times,and the effective use of GPU improves the performance of LLE algorithm.

关 键 词：图形处理器 统一计算设备架构 LLE算法 KNN搜索 降维 

分 类 号：TP301[自动化与计算机技术—计算机系统结构]