Pruning method and system based on Crossbar architecture

Abstract

The invention relates to a pruning method and system based on a Crossbar architecture. The method comprises the following steps: carrying out structural pruning on a to-be-processed weight matrix of acurrent layer of a neural network without considering an architecture factor to obtain a first weight matrix, wherein the weight matrix is a weight matrix of the neural network; determining a pruningscheme according to the size of the first weight matrix and the size of a crossbar, wherein the pruning scheme comprises block-based structured pruning, crossed array row pruning and crossed array column pruning; pruning the first weight matrix according to the pruning scheme to obtain a second weight matrix; and mapping the second weight matrix to a crossbar array in an accelerator, and accelerating the neural network through the accelerator. The resource occupation of Crossbar can be reduced, and the waste of hardware resources can be reduced.

Description

technical field [0001] The present invention relates to the field of network acceleration, in particular to a pruning method and system based on Crossbar architecture. Background technique [0002] At present, deep neural network (DNN) is widely used in image classification, target detection, semantic segmentation, speech recognition and other fields. However, due to the large weight size and computational load of mainstream neural networks, many neural network accelerators are designed to use the parallelism of the calculation of the convolutional layer and the fully connected layer in the neural network to speed up the reasoning process of the network. One of them is A typical accelerator is a vector-matrix multiplication accelerator based on the Crossbar architecture. [0003] figure 1 It is an architectural diagram of a typical neural network accelerator based on a crossbar architecture. The chip includes a plurality of small tiles (Tile), and each tile (Tile) is compo...

AI Technical Summary

Problems solved by technology

In some existing technologies, the weight of each column in each Crossbar is used as the basic pruning granularity, and the remaining columns are left-shifted and filled. This method finally obtains the pruning result. On the one hand, the algorithm only uses The sparsity in the column direction of the neural network weight matrix cannot be used, but the sparsity in the row direction cannot be used, so its compression ratio will be limited

On the other hand, after the algorithm performs left shift and complement operation on the remaining weights in the pruning results, most of the units in some Crossbars are idle, wasting hardware resources

Images