CUDA相邻归约与其避免线程分化算法的研究  

Research on CUDA Adjacent Reduction and Thread Divergence Avoidance Algorithms

作　　者：卫易东 WEI Yidong(China Aircraft Strength Research Institute,Xi'an Shaanxi 710065,China)

机构地区：[1]中国飞机强度研究所,陕西西安710065

出　　处：《信息与电脑》2023年第18期55-57,61,共4页Information & Computer

摘　　要：在边缘计算环境下,上层应用调度图形处理器(Graphic Processing Unit,GPU)的统一计算架构(Compute Unified Device Architecture,CUDA)进行计算时,可能会遇到CUDA线程分化问题,导致运算耗时较长或线程空置化。本研究介绍了CUDA底层开发的基础原理和概念,并解释了CUDA运算的执行流程。通过分析GPU架构原理,提出了相邻归约算法和相邻归约的避免线程分化算法的实现方式和应用方法。In the edge computing environment,when the upper layer uses the Compute Unified Device Architecture(CUDA)of the Graphic Processing Unit(GPU)for computing,it may encounter the problem of CUDA thread differentiation,resulting in long computing time or thread vacancy.This study introduces the basic principles and concepts of CUDA underlying development,and explains the execution process of CUDA operations.By analyzing the principles of GPU architecture,the implementation and application methods of adjacent reduction algorithm and adjacent reduction algorithm for avoiding thread differentiation are proposed.

关 键 词：边缘计算 统一计算架构(CUDA) 优化算法 相邻归约 避免线程分化 矩阵加法 计算效率 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]