Two-dimensional convolution implementation method and device

Abstract

The invention provides a two-dimensional convolution implementation method and device. The method comprises the steps: one-dimension processing of two-dimensional input data, wherein two-dimensional input data is expanded into one-dimensional data according to a set dimension; convolution core decomposition, wherein the two-dimensional convolution core is decomposed into independent rows or columns according to the expansion mode of the two-dimensional input data; sub-result calculation, wherein the one-dimensional data is subjected to one-dimensional convolution operation respectively with the decomposed convolution core to correspondingly obtain a plurality of one-dimensional sub-result sequences; and sub-result merging, wherein all corresponding data in the multiple one-dimensional sub-result sequences are merged to obtain and output a final two-dimensional convolution calculation result. The invention further provides a two-dimensional convolution implementation device. According to the method and the device, all two-dimensional data can be converted into one-dimensional data, and two-dimensional convolution operation can be completed through one-dimensional convolution operation, so that the control complexity and the on-chip data routing difficulty are greatly simplified.

Description

technical field [0001] The invention relates to data and database processing, in particular to a two-dimensional convolution realization method and device for sorting data processing. Background technique [0002] Two-dimensional convolution is one of the core operations in many scientific calculations, accounting for most of the calculations in the entire application, such as the calculation between the input feature image and the two-dimensional convolution kernel in each layer of the convolutional neural network, Occupies more than 90% of the calculations in the entire network model. Therefore, improving the computational efficiency of two-dimensional convolution is the key to the optimization of many core scientific calculations. [0003] In order to accelerate the computational efficiency of two-dimensional convolution, more and more dedicated computing architectures have been proposed. This type of computing architecture is mostly designed according to the characteris...

AI Technical Summary

Problems solved by technology

[0004] The technical problem to be solved by the present invention is to provide a two-dimensional convolution implementation method and device for the problems of complex addressing and complex control of two-dimensional convolution in the prior art

Images