Convolution operation method based on expansion access on heterogeneous many-core architecture

Abstract

The invention discloses a convolution operation method based on expansion access on a heterogeneous many-core architecture, and the method comprises the following steps: S1, inputting an input, a weight and a stride wherein the input is Hi * Wi and the weight is K * K, and calculating the shape of an output according to the shapes of the input and the weight to obtain Ho * Wo; s2, according to the shape of the output, on the dimensions of Ho and Wo, according to the logic number of each core, the convolution calculation tasks are averagely distributed to many cores; s3, determining the size of each core according to the own task size; s4, each kernel carries out convolution calculation through the obtained input (HiBLOCK * WoBLOCK) and the obtained weight; and S5, repeating S3 and S4 until the calculation is finished. According to the method, memory bandwidth resources are saved, and meanwhile, many-core computing resources can be fully utilized.

Description

technical field [0001] The invention relates to a convolution operation method based on expansion fetching on a heterogeneous many-core architecture, and belongs to the technical field of deep learning. Background technique [0002] Convolution is one of the most important concepts in deep learning. During the training and reasoning process of the convolutional neural network, the convolution operation occupies the vast majority of calculations. High-performance computing platforms usually provide dedicated solutions. For calculation-intensive functions, such as convolution in deep learning, how to provide enough data for the powerful calculation kernel in a timely manner and improve the reusability of data is a problem that needs to be solved. [0003] The convolution operation is the core operation of the artificial intelligence CNN network. The data of each K times of the convolution operation overlaps but is not repeated. If the number of data is frequent, if the chara...

AI Technical Summary

Problems solved by technology

[0004] At present, there are some optimized convolution operation methods, such as im2col, which converts the convolution operation into matrix multiplication, and uses the optimized matrix multiplication to optimize the convolution operation, but this method needs to expand the input to the original K*K times , which puts additional pressure on system memory

Heterogeneous many-core processors contain a large number of slave cores and have powerful computing power. Memory access bandwidth is the bottleneck of the system. For this kind of computing-intensive operations, such a method will not only fail to utilize the computing resources of the processor, but will also give the system Bandwidth causes huge pressure, and the optimization effect is not good. Therefore, how to efficiently use the memory access bandwidth and reduce the memory access pressure is the key to give full play to the performance of the processor.

Images