Comment by janwas

How can tuning be independent of devising the algorithm?

Are you really suggesting writing a variant of a kernel, tuning it to the max, then discovering a new and different way to do it, and then discarding the first implementation? That seems like a lot of wasted effort.