if the target is embedded systems, usually things like naive GEMM are faster than tiled approaches for small matrices (which are probably what's used on these s

if the target is embedded systems, usually things like naive GEMM are faster than tiled approaches for small matrices (which are probably what's used on these systems).