Acceleration design method of CNN network suitable for low-resource embedded chip

Abstract

The invention discloses an acceleration design method of a CNN network suitable for a low-resource embedded chip, and relates to the technical field of neural network deep learning. The method comprises the following steps: redesigning a calculation direction of convolution / pooling operation, and adopting a calculation sequence with channel direction priority; optimizing the space distribution andcalculation process of a convolution / pooling layer in the CNN network; utilizing a Neon instruction set to improve the computing performance. According to the invention, the resources such as DMA, SRAM, Neon and the like commonly existing in the existing embedded chip are fully utilized, optimization arrangement and SRAM space reutilization are carried out on calculation steps of a convolution unit and a pooling unit in a CNN network; data moving and numerical calculation are separated, DMA is used for moving data, Neon is used for achieving numerical calculation, DDR is not written in an intermediate result, and the purpose that the CNN network can be operated more quickly and efficiently in a low-resource embedded chip is achieved.

Description

technical field [0001] The invention belongs to the technical field of neural network deep learning, and in particular relates to an accelerated design method of a CNN network suitable for low-resource embedded chips. Background technique [0002] Convolutional Neural Network (CNN) is an important innovation in the field of deep learning technology. As a typical multi-layer neural network, convolutional neural network has always been at the core of research. The local connection and weight sharing methods adopted, on the one hand, reduce the number of weights to make the network easy to optimize, and on the other hand, reduce the complexity of the model, that is, reduce the risk of overfitting. Therefore, CNN has been applied in many machine vision-related tasks and has achieved great success. [0003] The excellent effect brought by CNN has stimulated the demand of many existing embedded device manufacturers to intelligently empower their original products through softwar...

AI Technical Summary

Problems solved by technology

[0005] 1. The calculation speed is not fast and the CPU is not fully utilized;

[0006] 2. Insufficient DDR bandwidth resources, but also increased a lot of additional data read and write

Images