Title: Tuning Sparse Matrix Kernel Performance via Lightweight Signatures

Anirudh Jain
Ph.D. Student in Computer Science
School of Computer Science
Georgia Institute of Technology

Date: Wednesday, September 6th, 2023
Time: 4:00 – 5:30pm EST

Location: Klaus 2100

Committee
Dr. Thomas Conte (advisor, School of Computer Science, Georgia Tech)
Dr. Alexandros Daglis (School of Computer Science, Georgia Tech)
Dr. Vivek Sarkar (School of Computer Science, Georgia Tech)
Dr. Tushar Krishna (School of Electrical and Computer Engineering, Georgia Tech)
Dr. Scott Beamer (Computer Science and Engineering, UC Santa Cruz)


Abstract
Sparse matrix operations such as Sparse-Dense Matrix multiplication and Sparse-Sparse Matrix multiplication make up a significant fraction of the applications from several domains such as graph processing, machine learning, and data analytics. The performance of these kernels on today’s architectures is severely hampered by their irregular access patterns and subsequent memory bandwidth bottleneck. One way to increase their performance is by tiling the inputs to fit in the on-chip buffers and choosing the right format to represent the sparse matrix itself. In this presentation, I introduce Residue Matrices a lightweight signature that encapsulates the sparsity and structure patterns of the sparse input allowing for tile size and format selection per region of the matrix without incurring the overhead of sweeping the parameter search space. I demonstrate the use of Residues by considering SpMM as a sample kernel.

I will also present future research directions that have the potential to further benefit SpMM and can improve automatic tiling for kernels other than SpMM.