Comment by GregarianChild
Comment by GregarianChild 2 days ago
If the search space never leaves the programs that are equivalent to the original specification, that will probably limit the optimisations you can discover. (E.g. if you start out with standard matmul, you will not discover Strassen's algorithm.) This is not a criticism, I'm just trying to understand your algorithm.
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.)