Operation device for convolution neural network

Abstract

The invention relates to an operation device for a convolution neural network. The operation device comprises a plurality of row data caching units and a plurality of core computing units, wherein each row data caching unit receives externally inputted feature map data or feature map data outputted by other row data caching units via a first data selector, and each row data caching unit is provided with at least one data window; and each row data caching unit is coupled to at least one core computing unit. The core computing unit comprises a weight register and a computing unit, wherein the weight register is used for obtaining weight data; and the computing unit performs computation according to the feature map data in the data windows and weight data in the weight register. The scheme ofthe invention can realize the configurability of the number of the row data caching units and the number of the weight registers, thereby meeting requirements of users for different window sizes to be operated.

Description

technical field [0001] The invention relates to the fields of neural network algorithms and computer hardware, in particular to a computing device for convolutional neural networks. Background technique [0002] In recent years, deep convolutional neural networks have achieved widespread success in many fields, such as face recognition, intelligent video surveillance, and assisted automatic driving. [0003] In an existing computing device for a convolutional neural network (Convolutional Neural Network, CNN), the row data cache unit is used to obtain the feature map data and output to the core calculation unit (Core-Calculate, CE), and the core calculation The weight register in the unit acquires weight data and outputs it to the calculation unit in the core calculation unit, and then uses the calculation unit to perform calculations based on the feature map data of the line data buffer unit and the weight data. [0004] However, in existing computing devices for convoluti...

AI Technical Summary

Problems solved by technology

[0004] However, in existing computing devices for convolutional neural networks, a row data cache unit with a fixed number of rows is usually used, and a weight register corresponding to the row number of the row data cache unit, resulting in only a fixed window When the user needs different window sizes to implement different neural network structures (such as convolutional neural network AlexNet or GoogleNet), or to implement different applications (such as face recognition, object detection, image classification), Unable to use the same device to meet user needs

Images