Yeah, I was refering to this particular error. I just saw I didn't commit the remaining code to my public git repo (... because it's a bit ugly), this ( http://tonwas.freefronthost.com/gsocwork.tar.gz ... early snapshot of my work, which I vastly cleaned up lateron) tarball provides the remaining code at snapshot/texture.c. What this code does is if the requested format is not supported, try to find some other format which hopefully won't reduce image quality. This is done by using some heuristics, i.e. rules (some of which are more important than others). Generally speaking, FourCC and 24-bit formats are avoided unless directly requested. Also we won't change format _types_, e.g. no ARGB to DXT conversion. Then, we try to not remove any color component channel; then we try to not reduce the number of bits per requested channel; then (of the few remaining formats) we pick the one with the lowest INCREASE in number of bits per channel (i.e. the best would be the same bpc); the last heuristic is the increase in number of channels, which ofc should be kept to a minimum. btw note that the heuristics given on the MSDN site of this function aren't the real ones; they have similiarity with the real ones, but my version of them is closer to actual native behavior. I've thoroughly tested this code and it works in almost 99% of all possible cases for ARGB/QWVU/Paletted/Luminance (i.e. "simple" uncompressed formats)) and in the very few cases it doesn't copy native behavior it at least returns a usuable fallback format.
Anyway, I strongly advise you to add this stuff in a later patch and replace it with a FIXME message or some comment for now.
I was thinking about something like this. I put a FIXME if CheckDeviceFormat fails until there is a real implementation.