Self-adaptive multi-row folding storage method suitable for GPU (Graphics Processing Unit)

Abstract

The invention discloses a self-adaptive multi-row folding storage method and system suitable for a GPU (Graphics Processing Unit), and the method comprises the steps: collecting a CSR code for a sparse matrix, and obtaining a non-zero element threshold value of the sparse matrix by setting a folding granularity and a threshold value; traversing a row pointer array of the sparse matrix, and performing multi-row folding based on the non-zero element threshold value to obtain a row index and a row pointer position of each folded row; according to the row index and the row pointer position, partitioning the sparse matrix and collecting the row index and the row pointer position of the folded row corresponding to each folding block; and traversing a column index and a value array of the sparse matrix, and based on the column index, adding a non-zero element in a blocked folded row, and constructing an adaptive multi-row folding storage structure for realizing multi-row folding storage of the sparse matrix. According to the system, the self-adaptive multi-row folding storage method is realized through mutual cooperation of the modules, and the problem of load imbalance of the SpMV based on the CMRS format on an irregular matrix is solved.

Description

technical field [0001] The invention relates to the field of sparse linear algebraic data optimization, in particular to an adaptive multi-line folding storage method suitable for GPU. Background technique [0002] Sparse matrix-vector multiplication (SpMV) is the core calculation of many iterative algorithms for solving large-scale sparse linear systems, and it is widely used in many research fields, such as linear algebra, data mining, and graph analysis. The general form of SpMV is y=A·x, where A is a sparse matrix of m×n, and x and y are dense vectors of n×1 and m×1 respectively. In recent years, optimizing SpMV kernels on modern hardware architectures has attracted extensive attention from researchers. Compared with traditional CPUs, modern general-purpose graphics processing units (GPUs) tend to have higher peak floating-point calculations due to their large number of computing cores. However, the optimization of SpMV on GPU is somewhat challenging. This is mainly b...

AI Technical Summary

Problems solved by technology

[0005] Problems with CMRS: CMRS uses a fixed k-row fold for a matrix, which may cause a large difference in the number of non-zero elements in different fold rows

If the original matrix is ​​an irregular matrix, that is, the non-zero elements of different rows differ greatly, then fixed multi-row folding will exacerbate this irregularity, resulting in more serious load imbalance and reducing the overall performance

Images