Title: The Power of Less: Harnessing Sparsity for Performance Optimization

Abstract: Sparse matrix computations are fundamental to scientific computing and data analytics applications, such as computer graphics and machine learning. Sparsity leads to irregular memory accesses that pose challenges for code optimization. While  specialized libraries can accelerate sparse computations, costly manual tuning is required for each application and architecture, reducing programmer productivity. Additionally, in machine learning, the unstructured sparsity patterns of deep learning models have rendered many of these libraries useless, prompting practitioners to use dense routine calls. Automation approaches such as compilers and runtime systems provide portability and ease of programming, but efficiently optimizing sparse codes remains a challenge due to access pattern irregularities.

In this talk, I will introduce our work on building compilers and automation frameworks to accelerate sparse numerical kernels. I will present a class of inspection strategies that automatically analyze information such as matrix sparsity patterns and the numerical methods' properties, to generate optimized sparse codes. Additionally, I will discuss algorithmic modifications that make machine learning and graphics applications more amenable to sparse code specialization and the use of sparse compilers.

Bio: Maryam Mehri Dehnavi is an Associate Professor at the Department of Computer Science at the University of Toronto, where she also serves as the Associate Chair of Research. She holds the prestigious Canada Research Chair in Parallel and Distributed Computing and is the recipient of  the Ontario Early Researcher Award. Additionally, she has served as the General Chair of PPoPP and is the Associate Editor for the Journal of Parallel and Distributed Computing. Maryam's works on building compilers and systems  that automatically transform computation patterns to patterns that are amenable  to optimization. Her software packages have been widely adopted by domain experts, particularly in computer graphics and machine learning.