Method and device for performing convolution operation, method and device for generating instructions

Abstract

Disclosed are a method and device for performing convolution operations, and a method and device for generating instructions. The method for performing convolution operations includes: obtaining the nth second convolution kernel; The size in height, the size after padding with the first input feature data in width and height, the size of the nth second convolution kernel in width and height, and the nth second convolution kernel in the first volume The width starting point coordinates and the height starting point coordinates in the product kernel are obtained from the filled first input feature data to obtain the second input feature data corresponding to the nth second convolution kernel; for the nth second convolution kernel performing a convolution operation on the second input feature data corresponding to the kernel and the nth second convolution kernel; and after obtaining the N second output feature data, integrating the N second output feature data. The present application can improve the efficiency of hardware performing convolutional neural network operations without increasing hardware costs.

Description

technical field [0001] The present application relates to the technical field of neural networks, and in particular to a method and device for performing convolution operations, and a method and device for generating instructions. Background technique [0002] Convolutional neural networks have been successfully applied in many fields such as image processing and autonomous driving. With the application of convolutional neural networks in many fields, in some application scenarios, it is necessary to apply convolution operations of convolution kernels of various sizes in convolutional neural networks. However, hardware dedicated to performing neural network calculations (eg, neural network accelerators) can only directly perform convolution operations with convolution kernels of a specific size due to hardware limitations. Contents of the invention [0003] In order to solve the above-mentioned technical problems, it is desired to provide a new method and device for perfo...

AI Technical Summary

Problems solved by technology

However, hardware dedicated to performing neural network calculations (e.g., neural network accelerators) can only directly perform convolution operations for convolution kernels of a certain size due to hardware limitations

Images