A computing device based on Winograd convolution and a neural network processor including the same

Abstract

The invention provides a convolution computing unit based on Winograd convolution and a neural network processor including the same. The convolution computing unit includes a multiplier, an accumulator and a first strobe, the multiplier is used for receiving an element to be multiplied by a matrix or an element to be multiplied by a matrix dot, The first strobe is configured to receive an output value from the multiplier and an element to be accumulated, and the convolution operation unit can be switched between a plurality of operating modes by controlling the first strobe to transfer the element to be accumulated or the output value of the multiplier to the accumulator. The convolution operation unit of the invention can be applied to the neural network processor to improve the calculation efficiency and reduce the running power consumption.

Description

technical field [0001] The present invention relates to the technical field of neural networks, in particular to a computing device based on Winograd convolution and a neural network processor comprising the device Background technique [0002] In recent years, deep learning technology has developed rapidly and has been widely used in solving advanced abstract cognitive problems, such as image recognition, speech recognition, natural language understanding, weather prediction, gene expression, content recommendation and intelligent robots, and has become an academic Research hotspots in the world and industry. [0003] Deep neural network is one of the perception models with the highest level of development in the field of artificial intelligence. It simulates the neural connection structure of the human brain by establishing a model, and describes the data characteristics hierarchically through multiple transformation stages, providing images, videos, audios, etc. Large-sc...

AI Technical Summary

Problems solved by technology

Therefore, researchers currently propose a Winograd-based convolution operation method, which can complete equivalent convolution operation tasks and greatly reduce the multiplication of the convolution operation process by performing specific matrix transformations on the input feature map and weights. However, there are differences between matrix conversion and dot multiplication in Winograd convolution. When independent dedicated computing modules are used to complete the corresponding tasks, huge computing resources are required and the computing efficiency is low.

Images