SpMV parallel block calculation method and system based on dense packing

CN116821579BActive Publication Date: 2026-06-05INST OF COMPUTING TECH CHINESE ACAD OF SCI

Patent Information

Authority / Receiving Office: CN · China
Patent Type: Patents(China)
Current Assignee / Owner: INST OF COMPUTING TECH CHINESE ACAD OF SCI
Filing Date: 2023-02-01
Publication Date: 2026-06-05

AI Technical Summary

Technical Problem

Existing technologies do not fully utilize the temporal data locality of sparse matrix iterative computation, mainly focusing on a single SpMV computation and introducing additional overhead, failing to effectively optimize the performance of multiple iterative computations.

Method used

A tiling block method is proposed to convert a sparse matrix into a graph G. Multiple subgraphs are obtained by partitioning the tiling graph, and the row order is reordered and calculated according to the update number of the subgraphs, so as to achieve efficient parallel computation of sparse matrix and dense vector.

Benefits of technology

It improves the performance of sparse matrix iterative computation, reduces redundant computation, and realizes high-concurrency SpMV iterative computation, especially showing a significant acceleration effect in large-scale sparse matrices.

✦ Generated by Eureka AI based on patent content.

Smart Images

Figure CN116821579B_ABST

Patent Text Reader

Abstract

The application provides a dense-paving-based SpMV parallel block calculation method and system, which comprises the following steps: converting a sparse matrix into a graph G; performing dense-paving-graph-based division on the graph G to obtain a plurality of subgraphs Gi; each node Vi in the subgraph Gi is updated along a time dimension until all nodes in the subgraph Gi cannot meet the dependency relationship between the node Vi and an edge Ei when being updated again, and then the subgraph Gi completes maximum update of the subgraph and obtains a final subgraph; the calculation row order of the data row of the final subgraph is reordered according to the number of times of updating each node in the final subgraph; and matrix block calculation is performed on the final subgraph reordered based on a dense vector to be subjected to SpMV, so that an SpMV calculation result is obtained. The application can expand the block dimension of SpMV multiple iteration calculation, does not contain redundant calculation, does not change the core calculation loop of SpMV, and enables the SpMV iteration calculation to be executed in high concurrency.

Need to check novelty before this filing date? Find Prior Art