Regularizing the Irregular
summary
Argues that scheduling transformations work on irregular pointer‑based codes by hiding traversal logic behind a small abstraction and then reordering loops; shows applying loop tiling and a tiny “magic function” that maps points to tree nodes turns recursive KD‑tree N‑body queries into a node‑then‑points schedule that enables cache reuse, vectorization and GPU mapping, yielding about 3x CPU speedup and up to 100x on GPUs while noting safety, codegen and tuning tradeoffs.
AI-written summary · about