The threshold-blur stage is now shader model 2.0.

For another reason, I changed the way I pass coarse normals (for reducing memory consumption). A positive side effect was that I don't need to reconstruct them anymore in all PP stages (because they are stored on an own 888 target). This resulted in less arithmetics. Plus, I changed the gaussian weights in the blur stage so that I get now the same image quality with 5 texture lookups now instead of 7 per pixel per pass (H and V).

I am also investigating the arithmethics involved in the bilateral upsampling in respect to scalar nature of the AO terms (they are monochromatic! - the original paper was targetted on upscaling RGB colors only). If I am right, I think the upsampling will afterwards be two times faster for the cost of one additional half-sized 8888 rendertarget.

Third, I think the AO shader can be modified so that it can be used iteratively, using one PP stage per AO-sampling. This allows a dynamic configuration of the quality (which is hardcoded at the moment) and maybe also PS 2.0. Maybe. Don't bet on it

So these are the news and my thoughts, - so lets start coding!