3D-CNN acceleration method and device for 3D image processing, and electronic equipment

Abstract

The invention discloses a 3D-CNN acceleration method and device for 3D image processing, and electronic equipment, and the method comprises the steps: carrying out the preprocessing of a 3D image, obtaining 3D feature map data, and storing the 3D feature map data in a global memory; carrying out implicit data conversion on the 3D feature map data in the global memory to obtain first intermediate data; writing the first intermediate data into a shared memory; a reading buffer area and a writing buffer area are distributed in the shared memory, and all first intermediate data are written into a register through alternate transmission of the first intermediate data in the reading buffer area and the writing buffer area; partitioning the first intermediate data in the register; calculating the blocked first intermediate data to obtain second intermediate data; performing an inverse process of implicit data conversion on the second intermediate data, and converting the second intermediate data into new 3D feature map data again; and storing the new 3D feature map data in the global memory.

Description

technical field [0001] The present application relates to the technical field of data processing, in particular to a 3D-CNN acceleration method and device for 3D image processing, and electronic equipment. Background technique [0002] Convolutional neural network (CNN) has been widely used in computer vision in recent years, including classification, detection, segmentation and other tasks. These tasks are generally performed on images, using two-dimensional convolution (that is, the dimension of the convolution kernel is two-dimensional). For problems based on video analysis, 2D convolution cannot capture timing information very well, so 3D convolution was proposed. [0003] With the improvement of CNN recognition accuracy, the CNN model is getting bigger and bigger, and the network structure is getting more and more complex, and the requirements for computing power are constantly increasing. Since traditional CPU processors can no longer cope with the powerful parallel ...

AI Technical Summary

Problems solved by technology

The memory space of the GPU is limited, which limits the number of channels that can be used by 3D-CNN

Through the analysis of the roofline model, when the number of channels is limited, there is a large gap between the calculation memory access ratio of 3D-CNN and the calculation memory access ratio of the new generation GPU, and the calculation resources of the Tensor-core in the new generation GPU cannot be fully utilized.

Images