Write your own kernel optimiser
Summary
The article delves into kernel optimization using the Cachet language, contrasting hand-written optimizations with compiler-driven transformations and proving correctness. It walks through matrix multiplication scheduling for CPU and GPU, explores blocking, tiling, packing, and traffic accounting, and discusses how explicit schedules can be verified against a reference. Targeted at developers and HPC enthusiasts, it demonstrates how explicit compiler tooling can guide high-performance kernel design.