Convolutional neural network hardware accelerator with novel convolution operation acceleration module

Abstract

The invention discloses a convolutional neural network hardware accelerator with a novel convolution operation acceleration module. The convolutional neural network hardware accelerator comprises an operation process management module; a parameter table module; a feature map caching module; a novel convolution operation acceleration module. The novel convolution operation acceleration module comprises: an input feature map pre-fetching module for reading input pixel points from a feature map cache; a product module which is used for multiplying the input pixel point by the convolution kernel weight; an output feature map access module which is responsible for reading an old intermediate result of the output pixel point from the feature map cache and writing a new intermediate result of theoutput pixel point back to the feature map cache; and an accumulation module which is used for completing the accumulation work of the operation result of the product module and the old intermediateresult of the output pixel point. According to the convolutional neural network hardware accelerator with the novel convolutional operation acceleration module, the convolutional operation process inthe convolutional neural network can be accelerated, and the convolutional neural network hardware accelerator has great application value in many mobile terminal devices with the characteristics of low data transmission bandwidth, low power consumption and few logic resources.

Description

technical field [0001] The invention relates to the technical fields of electronic information and deep learning, in particular to a convolution neural network hardware accelerator with a novel convolution operation acceleration module. Background technique [0002] In recent years, deep learning technology has developed rapidly, especially the convolutional neural network with a unique receptive field structure based on the working mechanism of cat brain visual cortex nerve cells has achieved great success in the field of vision applications, such as CNN in large image classification datasets On ImageNet, the recognition accuracy rate exceeds that of the human eye. However, this powerful algorithm encounters great obstacles in the process of moving towards practical applications. This is because many real-world applications where CNN can be useful are implemented on mobile devices with few computing resources, such as AI capabilities. Smartphones and smart security cameras...

AI Technical Summary

Problems solved by technology

[0004] Purpose of the invention: the purpose of the present invention is to solve the existing convolutional neural network operation method of the three ways still exist or the number of additions becomes more, or the low bandwidth of the data supply end makes the bit width of the data bus become wider to no avail, Either the registers added for caching data will consume a lot of D flip-flop resources, which is not good for the implementation of FPGA or ASIC. The three methods cannot take into account each other, and there is a problem of mutual conflict.

Images