Method and system for segmenting image data for neural networks

Abstract

The disclosure proposes a method and system for dividing image data of a neural network into blocks, and relates to the field of computer technology. In the present disclosure, the image data stored in the external storage unit is divided into blocks in the direction of the last bit of the storage order to obtain a plurality of first blocks of the image data, wherein the last bit of the storage order is the width W or Height H, so that the image data is divided into blocks in the direction of width W or height H, and the result of calculation for each block of image data input to any network layer of the neural network is the final output result at the corresponding position, There is no need to occupy additional cache resources to cache the calculation results of each block, which can save cache resources, thereby solving the contradiction between the smaller cache resources of the system on chip and the larger cache requirements of image data. In addition, the calculation results of each block do not need to be sent back to the CPU side for reorganization, which reduces the additional time cost caused by data transmission.

Description

technical field [0001] The present disclosure relates to the field of computer technology, in particular to a method and system for dividing image data of a neural network into blocks. Background technique [0002] With the continuous development and rapid update of computer technology, neural network algorithms have greater advantages compared with traditional deep learning algorithms in image processing, speech recognition, video processing and other application scenarios, but for neural network models, The amount of data it processes is very large, resulting in a very large amount of calculation and storage requirements. Therefore, it is very important to choose a suitable computing platform for network applications based on Shenjiang. [0003] Compared with the central processing unit (Central Processing Unit, CPU) and graphics processing unit (Graphics Processing Unit, GPU), the field programmable gate array (Field Programmable Gate Array, FPGA) platform has a high ene...

AI Technical Summary

Problems solved by technology

[0002] With the continuous development and rapid update of computer technology, neural network algorithms have greater advantages compared with traditional deep learning algorithms in image processing, speech recognition, video processing and other application scenarios, but for neural network models, The amount of data it processes is very large, resulting in a very large amount of calculation and storage requirements

Images