Universal convolution operation device

Abstract

The invention discloses a universal convolution operation device. Comprising an operation parameter receiving and analyzing module, an image data and weight parameter receiving and analyzing module, a state control module, a first data parameter control module, a second data parameter control module, a third data parameter control module, a convolution operation module, a memory module, a convolution result processing module and a processing result output module. All the modules are designed by adopting a full parallel pipeline processing architecture, and the working process control of the convolution operation module is realized through working parameters such as the number of convolution layers, the convolution operation precision, the convolution operation mode, the number of convolution operations and a feature map data processing method, so that the convolution operation requirements of the current mainstream convolution neural network are met; according to the method, the calculation framework is simplified, the calculation efficiency and the design efficiency are improved, the generalization degree is high, the flexibility is good, the energy efficiency ratio is high, multiple convolution operations of various convolution kernel sizes can be completed in parallel, and good expansibility and tailorability are achieved.

Description

technical field [0001] The present disclosure belongs to the technical field of digital image signal processing, and in particular relates to a general convolution operation device. Background technique [0002] Convolution operation is defined as a weighted summation processing method in traditional digital image processing, which is widely used in the simulation, analysis and digital signal processing of linear systems in the fields of communication, electronics, automation and so on. In recent years, deep learning technologies based on convolutional neural networks have emerged widely, and convolution operations are the core computing of convolutional neural networks. A deep learning network forward inference model of convolutional neural networks includes a large number of convolution operations. For example, the VGG-16 model contains a total of 1,634,496 2D convolution operations with a kernel size of 3 × 3. With such a huge amount of computation, the choice of computi...

AI Technical Summary

Problems solved by technology

The technical means adopted by the existing convolution operation methods are difficult to achieve a balance in terms of versatility, flexibility, and computational efficiency.

[0009] (3) Using GPU to realize convolution operation has the advantages of simple programming and fast operation speed, but there are problems such as high power consumption and difficult heat dissipation

For example, although the NVIDIA GTX-1080Ti GPU has vector data transfer operations, it has thousands of nodes in the internal calculation process to perform parallel calculations on data, but this process occupies a huge amount of hardware resources and consumes a lot of energy and power consumption. The power consumption reaches 250w

Images