Convolution operation optimization method and system, terminal and storage medium

Abstract

The invention relates to a convolution operation optimization method and system, a terminal and a storage medium. Comprising the following steps: inputting image data in a data memory module into a multi-thread data caching module, and recording data characteristics of the image data in each thread; when all the threads are filled with the image data, performing spatial-temporal similarity analysis on the data features of the at least two adjacent threads, and when the data features of the at least two adjacent threads have spatial-temporal similarity, filtering out the image data of at least one thread in the at least two adjacent threads, taking the thread after filtering the image data as an idle thread to cache the image data input by the data memory module again, and performing space-time similarity analysis again when all the threads are filled with the image data again; and performing convolution calculation according to the cached image data, and outputting new image data. According to the method, the actual convolution operation amount is greatly reduced, the data reusability is improved, the overall network calculation time is shortened, and the chip performance is improved.

Description

Technical field [0001] The present application belongs to the field of depth learning, and in particular, to a convolutionary operation optimization method, system, terminal, and storage medium. Background technique [0002] In recent years, due to the popularization of large data applications and the advancement of computer hardware, deep learning technology is used to conduct feature extraction, classification, and recursive operations in data, and have a wide range of applications in computer vision, natural language processing, and intelligent system decision. . Convolutional operation is a very important depth learning feature extraction method, such as the current lead Lenet1, AlexNet, VGG-16, VGG-19 depth learning neural network, is stacked by a layer of convolution layer With the improvement of the number of network layers, the accuracy of classification will be improved. However, due to the calculation of the general computer platform and the speed of the convolution its...

AI Technical Summary

Problems solved by technology

However, the existing architecture improves the computing power by increasing the operating frequency and increasing the number of computing and storage modules. It has already faced the problems of low utilization of computing modules, high implementation costs, limited communication bandwidth, poor scalability, and large energy waste.

Images