Compressed convolutional neural network-oriented parallel convolution operation method and apparatus

Abstract

The invention provides a compressed convolutional neural network-oriented parallel convolution operation method and apparatus. The method comprises the steps of determining an adopted operation mode according to input control signal convolution data shift chain length selection, accumulated offset enabling and convolution calculation enabling; and by adopting two serial shift register chains, inputting convolution data, convolution parameters and channel offset, and performing 3X3 and 1X1 convolution operation at the same time for a same input convolution data stream. According to the method, a multiplier, an accumulator, a parameter register and an offset register are added only based on original serial shift register chain-based 3X3 convolution operation; the realization method is simple; the executive efficiency is high; and the convolution operation in a compressed neural network algorithm can be effectively accelerated. According to the apparatus, a plurality of characteristic graphs can be output at the same time through simple hardware unit expansion and copying; and the apparatus has the advantages of low power consumption, high function unit utilization rate and high processing speed.

Description

technical field [0001] The present invention relates to the field of digital signal processing and dedicated hardware accelerators, in particular to a parallel convolution operation method and device for compressed convolutional neural networks. Background technique [0002] In recent years, Convolutional Neural Network (CNN) has made significant progress in deep learning. The most famous is that in 2012, Alex Krizhevsky and others proposed a classic CNN computing structure AlexNet, which has achieved great success in image classification and recognition. Such as figure 1 As shown, the input of AlexNet is a 3-channel 227×227 image data. The entire processing process includes 8 layers of operations. The first five layers are convolutional layers, and the last three layers are fully connected layers. The first layer of convolution uses The convolution kernel with a width of 3×11×11, the number of convolution kernels is 96, the second layer of convolution uses a convolution k...

AI Technical Summary

Problems solved by technology

The traditional method is to use multiple separate convolution operation units (including 3×3 convolution operation unit, 1×1 convolution operation unit) to perform convolution operations in parallel, and each convolution operation unit needs to load the same set of input feature maps data, which may cause input data bandwidth bottleneck or repeated input of data, reducing the efficiency of the system

Images