Programming Languages Techniques for FPGA Compilation | In Person Only

Please note, this is an offline event with a limited number of slots. Please RSVP only if you are reasonably sure about your attendance.

Abstract :

Modern hardware compilers have issues. They can miss optimization opportunities or produce incorrectly compiled designs. On top of that, as most hardware compilers are proprietary, they are difficult or impossible to extend and improve when bugs are found. In this talk, I present my dissertation work, which focuses on applying automated reasoning techniques from the field of Programming Languages to improve hardware compilers. In the first part, I give an overview of Lakeroad, which uses program synthesis to automatically discover how to compile to complex, programmable FPGA primitives. Lakeroad finds more opportunities to use DSPs on three major FPGA architectures, when compared to proprietary and open source tools. In the second part, I present an overview of Churchroad, which addresses some of the underlying SMT scaling issues hindering Lakeroad by breaking up large synthesis queries with the help of equality saturation. I also provide some thoughts on future work, using machine learning and large language models to improve program synthesis and equality saturation.

About the Speaker:

Gus Smith recently finished his PhD at the University of Washington's PLSE lab, where he was advised by Zachary Tatlock and Luis Ceze. His dissertation proposed automatically generating compiler backends from formal models of hardware by applying automated reasoning techniques such as program synthesis and equality saturation. He is also a recipient of UW's Bonderman Fellowship, a fellowship for extended independent travel; India is the fourth stop on his 6-month fellowship trip through Asia. He appreciates any and all suggestions for where to visit while in Bangalore and beyond. His website is https://justg.us