Graph convolutional network software and hardware collaborative acceleration method based on in-memory calculation

Abstract

The invention discloses a graph convolutional network software and hardware collaborative acceleration method based on in-memory calculation, and the method comprises the steps: converting a high-bit-width floating-point number of graph data into a low-bit-width fixed-point number through a quantitative fixed-point mode in the aspect of software, segmenting an original graph of the graph data into a plurality of sub-graphs, and obtaining a clustering result; node features of the multiple sub-graphs are mapped to an RRAM cross array used for feature aggregation, and then edges connected with the different sub-graphs in a clustering result are deleted in an edge deletion mode so as to deploy graph data to hardware; in the aspect of hardware, a calculation module and a control module are configured, the calculation module comprises an aggregation core array, a vector combination array and an intermediate cache, and the control module comprises an instruction queue, a graph data decoder and a neighbor cache; through collaborative design of software and hardware, acceleration of GCN (Graph Convolutional Network) calculation is realized. According to the method, the calculation parallelism degree, the throughput rate of the architecture and the resource utilization rate of hardware are improved.

Description

technical field [0001] The invention relates to the field of in-memory computing, in particular to a method and device for co-accelerating software and hardware of a graph convolution network based on in-memory computing. Background technique [0002] A graph structure is a commonly used data structure consisting of numerous nodes and edges connecting them. Structures commonly used in daily and research, such as transportation networks, social networks, and molecular structures, can be abstractly represented by graph structures. In recent years, graph neural network models (GNNs) have demonstrated strong learning capabilities on graph-based tasks, such as recommender systems, document classification, etc. The graph convolutional network model (GCN) has become the most commonly used GNN model today due to its small computational cost and excellent accuracy performance. like figure 1 As shown, each node in the graph task will have a feature vector, which contains the inform...

AI Technical Summary

Problems solved by technology

[0016] Therefore, the purpose of the present invention is to design a GCN deployment algorithm for multi-core in-memory computing hardware at the deployment algorithm level, solve the problem of how to map irregular graph data to a hardware array of a certain scale, and simultaneously improve the calculation parallelism. At the hardware level, to solve the problem of how to design in-memory computing units and data streams, a method for synergistic acceleration of software and hardware in graph convolutional networks based on in-memory computing is proposed.

Images