Convolution operation memory access optimization method based on GPU

Abstract

A convolution operation memory access optimization method based on a GPU relates to a convolution operation memory access optimization technology. According to the invention, the defect of high memoryaccess overhead of convolution operation in the prior art can be solved. The method is characterized by comprising the steps of loading convolution kernel data into a shared memory; dividing the convolution output into sub-blocks by taking 32 columns as units to obtain a plurality of sub-blocks containing 32 columns of data and a sub-block containing less than 32 columns of data; enabling each thread to calculate an index of first data required by the thread; enabling each thread to obtain the remaining required input data from the index of the first data through a column reuse algorithm andtransmit the obtained input data to a row reuse algorithm; calculating an output result through the row reuse algorithm and storing the output result in register data sum; writing the sum into a global memory; and calculating other to-be-calculated data in the convolution output. The method is used for performing memory access optimization on convolution operation in the fields of image processing, video processing and machine learning.

Description

technical field [0001] The invention relates to a convolution operation memory access optimization technology, in particular to a GPU-based convolution operation memory access optimization method. Background technique [0002] In the fields of image processing, video processing and machine learning, convolution operation has become a core computing mode. 2D convolution is widely used in image filtering and frame difference. Depth-wise convolution is often used in mobile neural networks. Multi-channel 2D convolution is the core operation in neural networks. However, convolution operations consume a lot of computing resources and memory resources, and convolution operations occupy 90% of the execution time in image processing and machine learning. Many optimization methods for convolution operations have been proposed, among which methods based on GEMM (matrix multiplication), FFT and Winograd are the most widely used. However, these methods need to convert the input and out...

AI Technical Summary

Problems solved by technology

[0004] The purpose of the present invention is to solve the problem that the memory access cost of the convolution operation in the prior art is large, and the number of memory accesses of the convolution is too many to reduce the performance of the convolution operation

Images