基于OpenVX的计算图优化方法综述  

An Overview of graph optimization methods based on OpenVX

作　　者：刘振 林广栋 黄光红 毛晓琦 LIU Zhen;LIN Guang-dong;HUANG Guang-hong;MAO Xiao-qi(AnHui Siliepoch Technology Coltd)

机构地区：[1]安徽芯纪元科技有限公司

出　　处：《中国集成电路》2024年第1期38-45,63,共9页China lntegrated Circuit

摘　　要：卷积神经网络在图像识别领域取得了巨大的成功,深度学习和卷积神经网络成为了研究的热点。神经网络模型的推理部署需要高性能的异构架构芯片,OpenVX使用基于计算图的执行模型实现在异构平台高性能计算。计算图优化技术可以使得硬件平台更加高效地执行计算图。本文首先简单介绍了OpenVX编程框架,之后从节点融合,节点转换、节点删除,节点拆分和节点交换五个方面重点介绍了计算图优化技术。最后指出了计算图优化技术的发展趋势。Convolutional neural networks have achieved great success in the field of image recognition,deep learning and convolutional neural networks have become research hotspots.Neural network models are deployed by high-performance heterogeneous architecture chips.To achieve high-performance computing on heterogeneous platforms,an OpenVX execution model graph-based is provided.Graph optimization technology make the hardware platform execute graph more efficiently.This paper introduces the OpenVX programming framework,then it focuses on the graph optimization technology from five aspects:node fusion,node transformation,node deletion,node splitting and node swap.Finally,the development trend of graph optimization technology is pointed out.

关 键 词：深度学习 神经网络 计算图优化 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程] TP391.41[自动化与计算机技术—控制科学与工程]