Hard acceleration method and device for neural network model of electronic equipment

Abstract

The invention discloses a hard acceleration method and device for a neural network model of first electronic equipment. The method and the device relate to the technical field of deep learning in artificial intelligence. The method includes the following steps: obtaining to-be-recognized data and configuration parameters of the neural network model; performing hard acceleration of convolution calculation matching the neural network model according to the configuration parameters, and obtaining a convolution result of the to-be-recognized data by the neural network model; performing the hard acceleration of function calculation on the convolution result by calling at least one function module matching the neural network model from at least one preset function module based on the configuration parameters, and obtaining a recognition result of the to-be-recognized data by the neural network model. The method and the device can support the neural network model established by various open source development environments, and also support a user-defined neural network model; and when an algorithm of the neural network model is updated, only the parameters of the first electronic equipment need to be reconfigured without changing hardware design.

Description

technical field [0001] The present invention relates to the technical field of deep learning in artificial intelligence, in particular to a hard acceleration method and device for a neural network model of a first electronic device, and an auxiliary acceleration method for a neural network model of a second electronic device . Background technique [0002] In the past few decades, although the computing performance of the CPU has been improving rapidly, due to the limitations of physical laws such as power consumption, interconnection delay, and design complexity, by 2004, the computing performance of the CPU was close to the physical limit. (The main frequency is about 3.6GHZ). In this case, heterogeneous acceleration has become one of the methods to obtain higher performance computing power. The so-called heterogeneous acceleration (Hybrid Acceleration) refers to the integration of different acceleration devices based on the CPU to achieve higher performance computing ac...

AI Technical Summary

Problems solved by technology

[0004] In the process of realizing the present invention, the inventor found that when using the acceleration device to realize the neural network processing system in the prior art, the hardware and software design of the acceleration device was carried out according to the characteristics of a specific neural network model. Although this method Better computing performance can be obtained, but because it is designed for a specific neural network model, and because there are many open source development environments in the field of deep learning, such as Tensorflow, Torch, Caffe, Theano, Mxnet, Keras, etc., once the neural network model If the algorithm has been updated or due to the different versions of the development environment, it is necessary to redesign the hardware and software of the acceleration device

Because the hardware development cycle of the acceleration device is long, usually several months or more than a year, the hardware update speed of the acceleration device is far lower than the iteration speed of the algorithm of the neural network model, which is a great obstacle Wide application of accelerated equipment

Images