Speaker: Professor Weifeng Liu, Super Scientific Software Laboratory, China University of Petroleum-Beijing

Inviter: Bastien Vieublé, Doctor & Xin Liu, Professor

Title: Sparse Matrix Computations on Modern High-Performance Processors

Language: Chinese

Time & Venue: 2024.09.03 15:00-16:00 Z311

Abstract: Sparse matrix computations are one of the most important common patterns in numerical linear algebra, and play a fundamental role in scientific and engineering computing. On high-performance processors with modern architectures, e.g., tensor core units and asymmetric multi-cores, sparse matrix computations often face more challenges such as diverse sparse formats, load imbalance, low efficiency in random memory accesses, poor data locality, complex task dependencies, and difficulties in utilizing acceleration units. As the scale of the matrix order being processed continues to expand, addressing these challenges becomes increasingly urgent. This presentation will start by analyzing the computational patterns of sparse basic linear algebra subprograms, discuss how to reduce the spatial cost of storing sparse data and achieve load-balanced computation, and then introduce how to utilize sparse block data structures to reduce random memory accesses and improve data locality. Furthermore, the talk will introduce synchronization-free algorithms to avoid barrier overhead caused by task dependencies and further discuss how to accelerate sparse computation using dense tensor units.