Convolution operation method and device for two-dimensional data

Abstract

The invention belongs to the technical field of convolutional neural network operations, and in particular relates to a convolution operation method and device for two-dimensional data. The row deviation address between the sub-two-dimensional data block and the original two-dimensional data block can complete the access of the sub-two-dimensional data block of any length two-dimensional data block, and perform convolution operation on the sub-two-dimensional data block and the corresponding sub-convolution kernel Obtain the convolution operation results of the sub-two-dimensional data blocks, and accumulate the convolution operation results of all sub-two-dimensional data blocks to obtain the convolution operation results of the original two-dimensional data blocks, so that the convolution operation accelerator based on the shift register chain Realize the convolution operation of any length and width convolution kernel, with high operation flexibility and simple program mapping.

Description

technical field [0001] The invention belongs to the technical field of convolutional neural network operations, and in particular relates to a convolutional operation method and device for two-dimensional data. Background technique [0002] In recent years, Convolutional Neural Network (CNN) has made major breakthroughs and progress in deep learning and artificial intelligence. In the actual operation process of CNN, it is generally calculated by convolution of two-dimensional input featuremap (feature map) As a basis, and then accumulate each input feature map and the convolution result of the corresponding two-dimensional convolution kernel to obtain an output feature map. The method of realizing CNN operation by a two-dimensional convolution operation accelerator based on a shift register chain can greatly reduce the pressure of memory access and maximize data reuse. The two-dimensional convolution operation accelerator realizes the windowing operation of input data thro...

AI Technical Summary

Problems solved by technology

In the application of the convolution operation accelerator based on the shift register chain, if the length of the convolution kernel is greater than the maximum length and width supported by the convolution operation unit, the convolution kernel needs to be split, and the input data is also the original two-dimensional data. Subsets, traditional DMA cannot move sub-two-dimensional data blocks in the original two-dimensional data memory, and can only input the original two-dimensional data and perform operations on the split sub-convolution kernels, resulting in the removal of redundant input data, The calculation delay increases and the energy consumption increases. The resulting redundant calculation results increase the cumulative logic complexity of the final calculation results.

Images