Multi-parallel acceleration method for CNN full connection layer operation

Abstract

The invention discloses a multi-parallel acceleration method for CNN full-connection layer operation, and the method greatly reduces the reading operation of full-connection layer parameters, effectively saves the energy consumption, and improves the energy efficiency ratio of a system through employing the sparsity characteristics of an operation result of a convolutional neural network and interpreting the numerical value of a convolution layer processing result of the convolutional neural network in advance. By utilizing the characteristics of hardware resource reuse, strong expansion capability and the like of programmable logic devices such as an FPGA and constructing a parallel pipeline multiply-accumulate architecture, the processing time is effectively saved, and the processing efficiency is improved; by comprehensively analyzing factors such as target type data, target characteristic difference and error tolerance in an application process, a processing data format is reasonably set, so that the access efficiency of data and parameters is effectively improved while the processing precision is ensured, and the purpose of multi-parallel acceleration of a full connection layer is achieved.

Description

technical field [0001] The invention belongs to the technical field of computer architecture, and in particular relates to a multi-parallel acceleration method for CNN fully connected layer operations. Background technique [0002] As an important representative of deep learning, convolutional neural network (CNN) is widely used in object classification, recognition, video analysis and natural language processing. CNN is mainly composed of convolutional layer, pooling layer, activation function and fully connected layer. Due to the large amount of parameter data, high precision requirements, and high requirements for external storage speed, the CNN fully connected layer operation has always been the bottleneck restricting the acceleration of CNN. For example, the current typical classification network VGG-16 fully connected layer has 138M parameters. [0003] At present, the main optimization methods of the fully connected layer operation method of CNN are parameter compre...

AI Technical Summary

Problems solved by technology

[0008] (1) There is no comprehensive analysis of factors such as the number of target types, the degree of difference in target features, and the degree of error tolerance, and the data format is reasonably selected, resulting in a failure to achieve a good balance between processing accuracy, parameter storage space, and parameter data throughput.

[0009] (2) Model compression on the parameters of the fully connected layer can effectively reduce the number of parameters, but there is also a risk of reducing the accuracy rate

The sparsity of the convolutional processing results of the convolutional neural network is not used, and a large number of invalid read operations are generated during the full connection operation, which wastes energy consumption and processing time

[0010] (3) Using a dedicated processor or ASIC or FPGA to realize the fully connected layer simply increases and iterates hardware resources to achieve logic acceleration, but does not fully utilize the full pipeline performance of programmable logic circuits, occupying more hardware resources and energy efficiency ratio greatly discounted;

[0011] (4) Using GPU to realize the fully connected layer has the advantages of simple programming and fast running speed, but there are problems such as high power consumption and difficult heat dissipation

Images