基于OpenVX的计算图优化方法综述

An Overview of graph optimization methods based on OpenVX

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.