Explore a groundbreaking presentation from PLDI 2023 introducing Mosaic, an innovative sparse tensor algebra compiler. Discover how Mosaic enables interoperability with external tensor algebra libraries and compilers, allowing users to bind tensor expressions to external functions. Learn about the compiler's ability to generate fused code and call specialized functions, offering significant performance improvements in various scenarios. Delve into Mosaic's automatic search system for mapping expressions to registered external functions and its verification process. Understand the simplicity and generality of adding new external functions to Mosaic, with current implementations averaging just 20 lines of code. Gain insights into the potential speedups of up to 206× when utilizing hand-written CPU and specialized hardware functions, and up to 3.57× with fused code generation.
Overview
Syllabus
Intro
Tensor Algebra Systems Today
Sampling Dense Matrix Multiplication
Performance
Mosaic
Taco
Components
Scheduling
Questions
Taught by
ACM SIGPLAN