Convolution calculating apparatus and method for neural network

Abstract

The invention belongs to the digital signal processing field, and specifically relates to a convolution calculating apparatus and method for a neural network. The convolution calculating apparatus andmethod for a neural network aims at solving the problem that consumption of resources is great and the utilization rate of the read in data is low during the process of convolution calculation. The convolution calculating method for a neural network includes the steps: input data matrixes are processed in rows, and every two rows of data are input serially row by row into multiply accumulator arrays for carrying out multiply and accumulate operations; the multiply accumulator arrays perform deployment according to a convolution kernel dimension (M2, N2), and can process 2*M2*N2 times of multiplication in parallel; and by means of the convolution operation law, the two groups of multiply accumulator arrays can be shifted and added, and the data operation is accelerated. The convolution calculating apparatus and method for a neural network can excavate parallelism during the calculating process and improve the calculating efficiency of the system, and at the same time can reuse the input data and directly put the calculation results into a pooling unit, thus being able to reduce data reading and writing. Besides, the convolution calculating apparatus and method for a neural networkonly need one row cushion space, thus having a small demand for resources, can realize calculation of different dimension convolution, and have the advantages of calculating flexibility, universality,high effectiveness and low power consumption property.

Description

technical field [0001] The invention belongs to the field of digital signal processing, and in particular relates to a convolution computing device and method applied to neural networks. Background technique [0002] Convolution is an important operation in mathematics and is widely used in digital signal processing. [0003] Convolutions can be computed by time-domain or frequency-domain methods. The time-domain method mainly involves multiplication and addition operations. There is no data dependence and time correlation between data at different points, and it can be accelerated through a parallel computing structure. The frequency domain method converts the convolution sequence to the frequency domain through Fourier transform, then directly multiplies the frequency domain data, and finally inversely transforms to obtain the convolution calculation result. [0004] Convolutional Neural Network (CNN for short) is an important algorithm model in deep learning. In recent ...

AI Technical Summary

Problems solved by technology

[0006] In order to solve the above-mentioned problems in the prior art, that is, in order to solve the problems of large resource consumption and low utilization rate of read data in the convolution calculation process, one aspect of the present invention provides a convolution calculation device applied to neural networks , including: data input port, convolution computing unit CCU, line buffer space, pooling unit, data output port;

Images