Automatically Tuned Dense Linear Algebra for Multicore+GPU