GPU reimplementation

[sesse]

Hi,

I thought I had already filed a bug about this, but evidently not, so here goes. :-)

As part of Futatabi, my realtime slow motion system, I've made a GPU reimplementation of the DIS algorithm (including variational refinement). On a recent GPU, it is more than an order of magnitude faster than the reference implementation is on a fast CPU, e.g., 1080p60 (forward flow + backward flow + interpolation) is easily realtime on my RTX 2070 at quality preset 3. In terms of EPE, it's a few percent better on the Sintel sequences—I haven't tried to track exactly why, but there are a number of smaller differences (e.g. bilinear interpolation in some places, and different edge handling). I don't know if you want to do anything with it or not, but now at least you know. :-)

Code (GPLv3), documentation, video etc. at https://nageru.sesse.net/ . Thanks for the awesome algorithm!