A feature map segmentation method of neural network accelerator based on systolic array

Abstract

The invention discloses a neural network accelerator feature map segmentation method based on a systolic array, comprising: obtaining an input feature map, a convolution kernel matrix, and a convolution step; and calculating the image split step according to the filter dimension and the convolution step Long; according to the split step of the image, calculate the starting position of the row and column of the current sub-image and the starting position of the row and column of the next sub-image; if the number of rows and columns of the next split sub-image is less than that of the initial split number, indicating that the next split sub-image has reached the last side. At this time, the difference between the two is calculated to obtain the number of rows and columns; if the number of rows and columns of the next split sub-image is less than the number of rows and columns of the filter, the convolution calculation is not satisfied. , discard it; input the sub-image data obtained by splitting the feature map into the systolic array for convolution operation. In this way, the data utilization rate is improved, while the convolution calculation time is reduced, and the efficiency is improved.

Description

technical field [0001] The invention relates to the field of computer technology, in particular to a method for segmenting feature maps of neural network accelerators based on systolic arrays. Background technique [0002] With the rapid development of artificial intelligence, deep convolutional neural network (CNN) has been widely used in artificial intelligence systems due to its unprecedented accuracy in image recognition, object detection, and scene understanding. Although it has ultra-high precision, the neural network gradually becomes deeper and wider, including a large number of network layers, and each network layer has convolution operations between weight data and feature map data. When the convolution operation is performed, hundreds of filters and channels in high-dimensional convolution need to be processed at the same time, involving a large amount of data relocation and multiply-accumulate operations, so the amount of calculation has exploded. [0003] The S...

AI Technical Summary

Problems solved by technology

[0004] Although the systolic array can solve a large number of convolution operations in CNN, the high-dimensional systolic array will occupy too many on-chip resources, resulting in an increase in chip area. Therefore, it is necessary to control the dimension of the systolic array within a certain range.

Images