Low-power-consumption convolution operation circuit based on approximate multiplier

Abstract

The invention discloses a low-power-consumption convolution operation circuit based on an approximate multiplier. The low-power-consumption convolution operation circuit comprises a convolution operation module, an approximate convolution calculation mode generation module, an approximate multiplier module and an approximate adder module. The convolution operation module comprises an input vectormatrix, a convolution kernel matrix, an approximate convolution calculation mode and an output approximate vector matrix; the input of the approximate convolution calculation mode generation module comprises an accurate calculation mode and approximate multiplication; wherein the input of the approximate multiplier module comprises two multipliers A and B and a multiplication approximation degreeDM, and the input of the approximate adder module comprises a partial product matrix to be accumulated and an addition approximation degree DA; according to the method, a high-precision approximate convolution calculation mode is generated by designing a high-precision approximate multiplier and an approximate adder. The approximate convolution calculation mode is used for replacing the calculation mode in the original convolution operation, so that the calculation overhead can be effectively reduced on the premise of meeting the output precision requirement of the convolution operation, and the convolution operation with low power consumption is realized.

Description

technical field [0001] The invention relates to a low-power convolution operation circuit based on an approximate multiplier, belonging to the technical field of approximate calculation. Background technique [0002] Deep learning is the most representative technology in machine learning in recent years. It has made breakthrough achievements in many key areas of pattern recognition, such as image recognition, natural language processing, speech recognition, and robotics. However, a deep learning model usually contains millions or even tens of millions of parameters and a network with dozens or even dozens of layers, which brings huge computational overhead. Therefore, approximation and acceleration are crucial for deep neural networks. [0003] Convolutional layers are an integral part of most deep neural networks. The computational implementation of the convolutional layer contains a large number of convolution operations. Therefore, proposing efficient and low-power app...

AI Technical Summary

Problems solved by technology

However, a deep learning model usually contains millions or even tens of millions of parameters and a network with dozens or even dozens of layers, which brings huge computational overhead

Images