I wonder if they're looking to achieve easy speedups. I can see lots of value enabling performance gains where normally people wouldn't bother because it's too much effort. I don't think this will take away work from those who hand-optimise their kernels and scheduling, this is to enable GPU acceleration for those, who otherwise wouldn't.