Convolution acceleration method, convolution calculation processing method, devices, electronic apparatus and storage medium

Abstract

The embodiments of the invention disclose a convolution acceleration method, a convolution calculation processing method, a convolution acceleration device, a convolution calculation processing device, an electronic apparatus and a storage medium. The acceleration method includes the following steps that: the to-be-processed data of a preset size in a to-be-processed task are sequentially read from the off-chip memory of an accelerator through an FIFO (first-input-first-output) port, and are stored in the input cache regions of a first cache region in the on-chip memory of the accelerator; to-be-processed input window data are sequentially read from the input cache regions in response to a condition that the to-be-processed data are stored into the input cache regions, and convolution calculation is performed on the to-be-processed input window data through convolution kernel data, so that output window data are obtained and are stored in a third cache region; and output window data inthe third cache region are sequentially stored to the off-chip memory through the FIFO port. With the methods, devices, electronic apparatus and storage medium provided by the embodiments of the invention adopted, a condition that the on-chip memory and bandwidth of the accelerator are insufficient and a condition that processor resources are limited can be avoided, and the efficiency of convolution calculation processing is improved. The methods and devices can be applied to hardware platforms such as an FPGA and an ASIC.

Description

technical field [0001] The present invention relates to computer vision technology, especially a convolution acceleration and calculation processing method, device, electronic equipment and storage medium. Background technique [0002] As an information processing technology imitating biological neural structure, neural network has developed rapidly in recent years. Neural networks have very powerful nonlinear mapping capabilities. Among them, Deep convolutional neural network (CNN) has shown remarkable performance in various computer vision tasks such as image classification, object detection, and semantic segmentation. [0003] Due to the need to process the feature map of the region in the image, the computational complexity of CNN increases exponentially. For the huge computing pressure, hardware accelerators such as Graphics Processing Unit (GPU), Field Programmable Gate Array (Field Programmable Gate Array, FPGA), and Application Specific Integrated Circuit (ASIC) are...

AI Technical Summary

Problems solved by technology

[0003] Due to the need to process the feature map of the region in the image, the calculation accuracy of CNN is improved, and its computational complexity is also doubled.

Images