https://hgpu.org/?p=2494
Accelerating SIFT on parallel architectures