Acceleration method of convolutional neural network, computer readable storage medium and application of method

Abstract

The invention discloses an acceleration method of a convolutional neural network, a computer readable storage medium and an application of the method. The acceleration method of the convolutional neural network comprises the following steps: calculating the sequence of input / output data of convolution, and enabling the input data to be continuously stored in a memory according to an access sequence during convolution; sorting the convolved coefficient data to enable the coefficient data to be continuously stored in a memory according to an access sequence during convolution; dividing the matrix into more than two sub-matrixes; and distributing SIMD registers for the input data, the coefficient data and the output data of a submatrix multiplication operation to carry out convolution calculation, and enabling the number numreg of the SIMD registers required by the convolution of the sub-matrixes to meet the equation that numreg = ri + ct + ri * sc, wherein numreg is smaller than or equalto the maximum value of the number tn of the SIMD registers of a CPU processor. The input data is continuously stored in a memory, and the matrix is divided into the sub-matrixes, so that the SIMD register is supplemented and fully used for convolution of the sub-matrixes, and the convolution efficiency of the CPU is improved.

Description

technical field [0001] The present invention relates to the technical field of neural network convolution, in particular to an acceleration method for convolutional neural network, a computer-readable storage medium and an application thereof. Background technique [0002] Convolution calculations are applied in various fields of computer vision. As the complexity of deep convolutional networks continues to increase, the amount of calculations also increases, resulting in very high requirements for hardware computing capabilities in deep convolutional neural networks. However, the computing power of embedded devices is limited, and it is difficult to achieve the computing effect of general-purpose computers. [0003] With the development of embedded CPU, ARM's CPU has occupied a huge market share, among which contex-A series processors have been widely used in various intelligent devices such as smart phones, set-top boxes and smart TVs. The 128-bit SIMD (Single Instruction...

AI Technical Summary

Problems solved by technology

[0006] To this end, it is necessary to provide an acceleration method for convolutional neural networks, which is used to solve the existing technical problems of large CPU resource occupation and low computational efficiency when processing convolution calculations through CPUs.

Images