Presented at MCHPC 2022 at SC22 in Dallas, Texas. High-performance stencil codes are typically tuned by transforming loop structure, but this work takes a different approach: the Bricks library improves performance by transforming data layouts to match the memory hierarchy instead, enabling significant speedups for structured-grid computations without manual code restructuring. This was joint work with Dr. Tuowen Zhao, Dr. Hans Johansen, and Dr. Samuel Williams at Lawrence Berkeley National Laboratory; slides here.

Papers

The ideas in this talk are fully presented in a workshop paper.