09-08-09 - DXTC Addendum

Ryg pointed out that there are a few very important little details that I took for granted and didn't mention in my original DXTC postings , or was just not clear about :

One is that when I try all ways of hitting two given endpoints, I try both 4-color and 3-color versions. That is, given two endpoint colors C0 and C1, I quantize them to 16 bits, then try the DXT1 palette that you get from {C0,C1} and also the one from {C1,C0} (order of DXT1 determines whether it is 3-color or 4-color).

The second and related crucial thing, is that in 3-color mode, the extra color is transparent black. If the texture has no alpha at all, I assume the user will not be using it as an alpha source, so I treat the transparent black as just black. That is, I do color palette selection with alpha just ignored.

Apparently this is pretty important. I suspect this especially helps with the "4 means" method; if a bunch of the colors are near black, you want them to be classed together and then just ignored for the endpoint selection, so that they will go to the hard-coded black in 3-color mode and your interp end points will be chosen from the remaining colors.

No comments:

old rants