3D CNN acceleration method and system based on Winograd algorithm

Abstract

The invention discloses a 3D CNN acceleration method and system based on a Winograd algorithm. The method includes the steps of reading a feature map subblock to be transformed from an input feature map, and reading a convolution kernel subblock from a weight cache, performing the 3D Winograd algorithm on a feature map subblock Bin and the convolution kernel subblock to output results and accumulating output accumulation results, and determining whether all the input feature maps are read; if all the input feature maps are read, writing the accumulation results back to an output feature map cache Out. By extending the Winograd algorithm and using the Winograd algorithm for the 3D CNN calculation, a 2D algorithm is applied to perform CNN acceleration and achieve a good effect, the computational complexity of the CNN algorithm can be effectively reduced, and the computing performance and energy efficiency ratio of a 3D CNN accelerator based on an FPGA are improved.

Description

technical field [0001] The invention relates to 3D CNN (three-dimensional convolutional neural network) acceleration technology, in particular to a Winograd algorithm-based 3D CNN acceleration method and system for an embedded platform. Background technique [0002] With the development of the field of artificial intelligence, three-dimensional convolutional neural network (Three-dimensional Convolutional Neural Network, 3D CNN) has been widely used in many complex computer vision applications, such as video classification, human motion detection and medical image analysis. Different from the traditional two-dimensional convolutional neural network (Two-dimensional Convolutional Neural Network, 2D CNN), 3D CNN can retain the time information in the 3D image during the processing process, so it can achieve comparative results in the field of 3D image recognition and classification. 2D CNN works better. [0003] With the improvement of CNN recognition accuracy, the CNN networ...

AI Technical Summary

Problems solved by technology

[0004] According to this embodiment, the current FPGA-based CNN accelerators are all 2D CNN-oriented, and there is no published literature on FPGA-based 3D CNN acceleration.

Images