Comment by jafioti

could be...im not opposed to looking into this to see if there's no possible trajectory from naive to strassen's without leaving logical equivalency.

all the optimizations for matmul so far have been straightforward trajectories from naive (tiling, smem caching, tensor core offload, etc.)