Convolution implementation method and convolution implementation device of neural network, and terminal equipment

Abstract

The invention is applicable to the technical field of neural networks, and provides a convolutional implementation method and a convolution implementation device of a neural network, terminal equipment and a computer readable storage medium. The method comprises the steps of inputting a to-be-processed image into the neural network; obtaining convolution groups to be partitioned from all convolution layers of the neural network; obtaining all input channel data of a first to-be-blocked convolution layer in the to-be-blocked convolution group according to the to-be-processed image; partitioningall input channel data of the first convolution layer to be partitioned; according to all blocks of all input channel data of the first to-be-blocked convolution layer, obtaining an output result ofthe last to-be-blocked convolution layer in the to-be-blocked convolution group, and storing the output result of the last to-be-blocked convolution layer into a memory; and inputting the output result of the last to-be-blocked convolutional layer into a specified network of the neural network. According to the method and the device, data access to the memory in the convolution calculation processcan be reduced, so that the convolution calculation efficiency is improved.

Description

technical field [0001] The present application belongs to the technical field of neural networks, and in particular relates to a neural network convolution implementation method, convolution implementation device, terminal equipment, and computer-readable storage medium. Background technique [0002] A neural network is an operational model consisting of a large number of nodes (or neurons) connected to each other. Neural network algorithms are complex and changeable, requiring massive calculations and massive data access. The current mainstream architecture of neural network processors is a Central Processing Unit (CPU)-like architecture, facing the problem of "storage wall" , the convolutional layer in the neural network needs to repeatedly read data from the memory when performing convolution calculations, so that the computing resources cannot be fully utilized and the computing efficiency is low. Contents of the invention [0003] The present application provides a n...

AI Technical Summary

Problems solved by technology

Neural network algorithms are complex and changeable, requiring massive calculations and massive data access. The current mainstream architecture of neural network processors is a Central Processing Unit (CPU)-like architecture, facing the problem of "storage wall" , the convolutional layer in the neural network needs to repeatedly read data from the memory when performing convolution calculations, so that the computing resources cannot be fully utilized, and the computing efficiency is low.

Images