Convolutional neural network-oriented memory data access and zero inserting method and device

Abstract

The embodiment of the invention provides a convolutional neural network-oriented memory data access and zero inserting method and device, and belongs to the field of convolutional neural network computing and special hardware accelerators. The method includes: automatically reading data from a data memory in a line mode or a column mode according to a data base address and a read mode of an inputcontrol signal, and automatically fill zeros on the periphery of original two-dimensional data during data output according to the size of input data ranks and a zero inserting mode instruction. Accordingly, the convolutional neural network-oriented memory data access and zero inserting device can achieve automatic reading of two-dimensional data in a memory under the control of few input controlsignals while being used for a shifting chain based convolutional neural network structure, can automatically insert zeros on the periphery of the original two-dimensional data, and sends the data toa convolutional computing module during data output, can greatly improve the data access efficiency of the convolutional neural network computing structure, and can shorten the computing time of the whole convolutional neural network.

Description

Technical field [0001] The present invention relates to the field of XXX, in particular to a method and device for memory data access and zero insertion for convolutional neural networks. Background technique [0002] In recent years, Convolutional Neural Network (CNN) has made major breakthroughs and progress in deep learning and artificial intelligence. Typical CNN models are AlexNet, ZFNet, VGGNet, GoogleNet and SqueezeNet (compressed convolutional neural network) Wait. Now, CNN has become one of the research hotspots in many scientific fields, especially in the field of image classification and pattern recognition. Because CNN avoids complicated pre-processing of images and can directly input original image data, it has been widely used. CNN is generally composed of a multi-layer convolutional layer and a down-sampling layer (pooling). The input of the convolutional layer is usually three-dimensional data composed of multiple two-dimensional feature maps. Assuming that each...

AI Technical Summary

Problems solved by technology

The main disadvantage of this method is that it is necessary to reserve a storage space larger than the original data in the internal memory, and initialize the storage space of the zero interpolation position to zero before the data is stored in the internal memory, and at the same time due to the address of the input data in the memory It is already discontinuous, which leads to the fact that the burst mode of DMA cannot be used for continuous large-block data movement in data movement, which reduces the efficiency of data movement and affects the computing performance of the entire system.

Images