It certainly seems like an interesting technique, and it might be worth emailing him for some sample code. It sounds like you need to render a low res shadow map, then read that back and do CPU processing to determine the importance map. (Which can be whatever, it just weights parts of the shadow image as wanting more precision.) That implies a one frame delay in the warping to me; can't imagine why that would be a problem but it's something to keep in mind. I guess you have to upload the importance maps back to GPU and then render the warp map from that. Possibly a good candidate for compute?
Edited by Promit, 10 February 2013 - 11:16 PM.