A Lightweight Polyhedral Abstraction for MLIR
MLIR is a new compiler construction infrastructure unifying graph representations and operators for HPC and ML. It facilitates the design and implementation of code generators, translators and optimizations at different levels of abstraction and also across levels, targets and execution environments. We will present ongoing work on the design and implementation of a mathematical abstraction for polyhedral compilation in MLIR. Covering a wide range of applications, from program synthesis to static analysis, domain-specific optimization to hardware acceleration, this abstraction aims for low-degree polynomial complexity, low engineering complexity, while retaining high expressiveness and extensibility.
Albert Cohen (Google)