Tile Size Selection Using Cache Organization and Data Layout

Select tile sizes that fit data working sets in cache, accounting for layout and associativity

Featured image

Venue: PLDI
Slides: Google Drive

Topic: Tile size selection for loop tiling must account for cache organization (size, associativity) and data layout to maximize cache reuse and minimize conflict misses.


Summary

Loop tiling improves cache reuse for compute-intensive kernels (matrix multiply, stencils), but only when tile sizes are chosen correctly. This paper analyzes how cache organization (capacity, associativity, cache line size) and data layout (row-major vs. column-major, padding) affect optimal tile size selection. A tile size model is proposed that selects tile sizes to fit working sets in cache while avoiding conflict misses due to associativity constraints.


Background

Loop tiling

The selection problem


Key Idea


Meeting Notes

(to be filled)