Speaker: Rahul Utkoor
Topic: WACO: Learning Workload-Aware Co-optimization of the Format and Schedule of a Sparse Tensor Program
Description: The paper “WACO: Learning Workload-Aware Co-optimization of the Format and Schedule of a Sparse Tensor Program” (ASPLOS ’23) presents a novel framework that jointly optimizes the data format and execution schedule of sparse tensor kernels through machine learning. Traditional sparse compilers such as TACO can generate code for many combinations of storage formats and traversal orders but lack an intelligent policy to select the best one for a given sparsity pattern — a problem that is combinatorial and often NP-hard. WACO addresses this challenge by introducing a deep learning–based cost model that captures the coupled behavior among the sparsity pattern, tensor format, and schedule. At its core is WA-CONet, a sparse convolutional network that directly consumes raw sparse matrices to extract meaningful workload features, paired with a SuperSchedule encoding that unifies schedule and format parameters into a learnable representation. WACO employs an Approximate Nearest Neighbor Search (ANNS) strategy to efficiently navigate the vast co-optimization space without costly online exploration.
Time: 5:30 pm to 6:30 pm (IST)
Presenter: Rahul Utkoor
About the Presenter: https://www.linkedin.com/in/rahul-utkoor-385656a6/