LOOPer

A learned automatic code optimizer for polyhedral compilers

LOOPer: A Learned Automatic Code Optimizer for Polyhedral Compilers

A learned cost-model-driven auto-scheduler for polyhedral compilers, achieving a median 1.4× speedup over Pluto on Polybench.

Improved the existing deep learning cost model accuracy by 5% through feature engineering; expanded the optimization space via a mathematical representation of code transformations.

  • Role: Research Assistant at NYU Abu Dhabi
  • Methods: Polyhedral compilation, deep learning cost models, feature engineering
  • Status: Published, PACT 2025
  • Paper: arXiv:2403.11522

References