一种基于任务划分与动态调度的稀疏张量并行填充方法及系统
By adopting a recursive task partitioning and dynamic scheduling mechanism based on the distribution of observed samples, the problems of unreasonable task partitioning and parallel conflicts in sparse tensor filling are solved, load balancing and efficient resource utilization are achieved, and the parallel execution efficiency and accuracy of sparse tensor filling are improved. It is suitable for large-scale sparse tensor processing.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HUNAN UNIV OF SCI & TECH
- Filing Date
- 2026-05-13
- Publication Date
- 2026-07-17
AI Technical Summary
Existing sparse tensor filling methods suffer from unreasonable task partitioning, significant parallel conflicts, and low utilization of computational resources. This leads to uneven GPU thread block load, frequent random memory access, low bandwidth utilization, and local long-tail wait issues, making it difficult to fully leverage the GPU's parallel processing capabilities while ensuring filling accuracy.
By recursively partitioning tasks based on the distribution of observed samples, load-balanced sub-tensor tasks are generated. Combined with coordinate format storage, conflict detection and grouping are performed to generate conflict-free parallel task groups. During the iterative update process, dynamic scheduling is performed based on contribution, freezing low-contribution tasks and reviving tasks that meet the conditions.
It improves the parallel execution efficiency and resource utilization in the sparse tensor filling process, reduces update conflicts, and enhances filling accuracy and overall throughput. It is suitable for large-scale sparse tensor processing scenarios such as network traffic monitoring and recommendation systems.
Smart Images
Figure CN122173305B_ABST