Method and system for improving parallel computing efficiency related to sparse matrix

Abstract

The embodiment of the invention discloses a method and system for improving parallel computing efficiency related to a sparse matrix. The method comprises the steps: acquiring a sparse matrix, whereinthe sparse matrix is composed of a plurality of non-zero values and a plurality of coordinates corresponding to the non-zero values, and the coordinates represent the positions of the non-zero valuesin the sparse matrix, the coordinates comprise row coordinates, and the row coordinates represent the row number of the non-zero value in the sparse matrix; dividing the plurality of non-zero valuesinto a plurality of first-class data regions based on the number of the plurality of non-zero values and the number of the calculation threads; traversing the plurality of first-class data regions, and dividing the non-zero values in the same row of the sparse matrix in two adjacent first-class data regions into the same data region to generate a plurality of second-class data regions; distributing the plurality of second-class data regions to a plurality of computing threads; and enabling the plurality of computing threads to execute the computing task in parallel.

Description

technical field [0001] This description relates to the field of data processing, in particular to a method and system for improving the efficiency of parallel computing involving sparse matrices. Background technique [0002] Sparse matrices are widely used in various machine learning models. For example, the core logic of the graph neural network model (GraphNeural Network, GNN) - the aggregation of information, can be realized through sparse matrix multiplication. When the size of the sparse matrix is ​​large, the calculation of sparse matrix multiplication will take more time, which will cause the overall model training time to be too long. Most of the existing sparse matrix multiplication schemes are based on single thread, which cannot take full advantage of the advantages of multi-core CPUs. If the multi-threaded parallel computing scheme is adopted, due to the irregular distribution of non-zero values ​​in the sparse matrix, the conventional evenly divided data area ...

AI Technical Summary

Problems solved by technology

When the size of the sparse matrix is ​​large, the calculation of sparse matrix multiplication will take more time, which will cause the overall model training time to be too long

Most of the existing sparse matrix multiplication schemes are based on single thread, which cannot take full advantage of the advantages of multi-core CPUs

If the multi-threaded parallel computing scheme is adopted, due to the irregular distribution of non-zero values ​​in the sparse matrix, the conventional evenly divided data area scheme will cause overlapping parts of the data processed by each thread, which needs to be synchronized through locking operations to improve operating efficiency. will be lower

Images