Method and device for performing convolution operation on neural network based on Winograd transform

摘要

A method and apparatus for performing convolutional computation of a neural network based on the Winograd transform is provided. An apparatus for performing convolution operation of a neural network according to the technical idea of the present disclosure, performs a Winograd transformation on an input feature map to generate a transformed input feature map including a plurality of channels having a matrix structure ; And a plurality of multiply-accumulate circuits, each of the plurality of multiply-accumulate circuits comprising a plurality of channels of the transformed input feature map between the transformed input feature map and a weight kernel transformed based on the Winograd transform. In the feature vector unit consisting of feature values at the same location, multiplication may be performed for each element, and an operation circuit for summing the results of the multiplication may be included.