Parallel processing method for multi-input multi-output matrix convolution

Abstract

The invention discloses a parallel processing method for multi-input multi-output matrix convolution. The method comprises the steps that S1, an optimal calculation scheme of output feature maps is determined according to the quantity N of vector processing elements (VPEs) of a vector processor and other parameters; S2, M input feature maps are sequentially stored into an external storage DDR, andN input convolution kernels are stitched in rows according to a third dimension; S3, a first element of the first input feature map is loaded and broadcasted to a vector register, and meanwhile elements in a first row of a convolution kernel in an AM are loaded into the vector register; S4, accumulation is performed kxk times, calculation of the first input feature map is completed, and meanwhilethe second input feature map is loaded; S5, the steps are repeated till calculation of first elements of N output feature maps is completed; S6, calculation of all elements of the N output feature maps is completed according to movement step length; and S7, the steps are repeated till the process is finally completed. The method has the advantages of being easy to implement, convenient to operate, capable of improving the parallelism of the vector processor and improving processor operation efficiency and the like.

Description

technical field [0001] The invention mainly relates to the fields of artificial intelligence, machine learning, and convolutional neural network, and in particular refers to a parallel processing method of multi-input multi-output matrix convolution. Background technique [0002] With the rise of deep learning technology, the target recognition technology based on convolutional neural network has made a breakthrough, and it has been widely used in image recognition, speech recognition, natural language processing and other fields. Matrix convolution is a calculation-intensive and memory-intensive calculation, and the matrix convolution operation in the convolutional neural network model often accounts for more than 85% of the calculation of a convolutional neural network model, so how to speed up the matrix convolution operation It is an important and difficult point of current research. [0003] For calculation and memory-intensive matrix convolution operations, the curren...

AI Technical Summary

Problems solved by technology

[0005] Matrix convolution is one of the core modules commonly used in convolutional neural network models. It is not only computationally intensive and memory-intensive, but the size of the convolution kernel in matrix convolution calculations is generally small. Therefore, if reasonable calculations cannot be taken method, even if high-performance computing equipment is used, it is difficult to exert due computing advantages

Images