This is an old experiment I had lying around, I never did wind up shipping it, but hopefully someone finds it useful!

It achieves very fast downsampling by staying close to the DXT encoding throughout, avoiding excess conversions between RGBA float or 32bpp, and using only integer instructions. As a side effect, the results are very good because those conversions can introduce extra lossiness.

Caveats:

• I never finished the DXT5 big endian support; I've omitted the big-endian support from this release, but if you're interested in what I had ask me!
• DXT 1 alpha handling is poor; it becomes viral, so an image with alpha will become increasingly transparent as it is scaled down.