Design And Implementation Of A Microprocessor With CNN Extension Instructions

Convolution neural network is widely used in computer vision,speech recognition and image processing.With the continuous increase of convolution kernel number and layer number of convolutional neural network algorithm,it can not meet the demand of real-time operation that general processors are used for processing.Studying the characteristics of convolution neural network algorithm and designing the microprocessor with convolutional neural network extension instruction are very important to improve the operation speed of the convolutional neural network.The use of general processors for processing has not met the need for real-time computing.Based on 60%zero neurons data of convolutional neural network data statistical results,the convolution neural network is write using C language,and using the MXnet platform test the accuracy of zero-valued neurons in network in different compression ratios.According to the RISC-V instruction set expansion rules,the operation with more execution times in the convolutional neural network algorithm is designed to four 128-bit custom instructions,which are convolution operation instruction,pooling operation instructions,data storage and data loading instructions.Then,based on the general ZION processor,using Verilog hardware description language design a microprocessor with convolutional neural network extension instruction.The main work is adding four accelerative units of custom instructions to the execution module.The microprocessors designed in this paper support both 32-bit general instructions and 128-bit custom convolution neural network acceleration instructions.The MNIST data sets and constructed convolutional neural networks are running in the software platform and the microprocessor designed in this article,respectively.The accuracy of the microprocessor is verified by comparing the two results.After that,The MNIST data sets and constructed convolutional neural networks are executed in the microprocessor designed in this paper and the general ZION processor.Results showed that execution speed is increased by 3.79 times,area is increased by 1.43 times,the speed and area ratio is 2.65 times through comparing the microprocessor that supports the convolutional neural network extension instruction and general ZION processor.