Thread Coarsening: More Work per Thread Block in CUDA
A tiled matrix multiply repeats global loads across neighboring blocks. Let one block compute several output tiles, then calculate the loads it saves and the SM slots it leaves idle.
8 min readConcept
