https://hgpu.org/?p=1120
Optimal loop unrolling for GPGPU programs (thesis)