Hardware accelerator of convolutional neural network based on parallel multiplexing and parallel multiplexing method

Abstract

The invention discloses a hardware accelerator of a convolutional neural network based on parallel multiplexing and a parallel multiplexing method. The hardware accelerator comprises a parameter storage module, an REG-FIFO module, a counting control module, an input multiplexing convolution operation module, an activation module and a pooling layer module. Wherein the parameter storage module is responsible for pre-storing picture parameters and trained weight parameters; the REG-FIFO module is responsible for generating an input matrix matched with the convolution kernel and reading matrix data; the counting control module is responsible for clock cycle counting and controls input and output of the REG-FIFO module according to the clock cycle counting. The input multiplexing convolution operation module is responsible for convolution operation of a convolution layer and a full connection layer; the activation module is responsible for output activation operation of the convolution layer and the full connection layer; and the pooling layer module is responsible for pooling operation output by the activated convolutional layer. The invention aims to realize convolutional neural network calculation with high operation parallelism, high data multiplexing and low hardware complexity.

Description

technical field [0001] The invention belongs to the field of artificial intelligence hardware design, and in particular relates to a method for implementing a convolutional neural network parallel multiplexing network computing accelerator. Background technique [0002] Convolutional Neural Network (CNN) is a popular direction in the field of Deep Learning (DL) in recent years. However, with the increasing data volume of the CNN model and the continuous improvement of computational input and accuracy requirements, the accelerated processing of convolutional neural networks has become a challenge. [0003] In general, in order to speed up the CNN algorithm, the following two characteristics are often used: (1) Sparse connection: The connection of the internal neurons in the calculation process is changed to a form of non-full connection. From the perspective of CNN visualization, it is the output feature. The corresponding node of a certain response in the graph is no longer...

AI Technical Summary

Problems solved by technology

[0010] The present invention is to solve the shortcomings of the above-mentioned prior art, and proposes a hardware accelerator and a parallel multiplexing method based on a parallel multiplexing convolutional neural network. The method of multiplexing simplifies large-scale convolution calculations in neural networks, thereby realizing convolutional neural network calculations with high computing parallelism, high data reuse, and low hardware complexity, and improving hardware to achieve large-scale convolution calculations. and waste of resources

Images