From: Jose G. <jos...@ju...> - 2008-10-30 13:09:15
|
>> >>>> Scaling to the nearest power-of-2 is certainly asking for horrible >>>> resuls. I also don't think the hardware acceleration will buy you >>>> much, >>>> transferr overhead is quite high and not-so-current hardware is huge >>>> limitations on maximum sizes it can handle. E.g. the given example >>>> wouldn't work with most IGD chips. >>>> >>>> >>> Power-of-2 *fraction*, ie. 1/2, 1/4, etc. of the original size. >>> >> >> Sure, but at least if you use any type of non-trivial interpolation >> algorithm to compute the final non-power-of-2-fraction, you get inferior >> results if you cut down to the next larger power-of-2-fraction first. >> Consider you want to scale down from 64 pixel to 31 pixel -- you throw a >> lot of information away by scaling down to 32 pixel first, even though >> that is cheap. >> > > The semantics isn't entirely up for grabs. Either we choose the nearest > power-of-2 fraction which is greater than the desired, or nearest period. > Usually, for this application one would take the former (though at the > moment I don't recall what Carsten threw in there), and in your example > that would mean no jpg-scaling and simply software down-scale to 31. > > Errr.. I mean in your example one *would* jpg-downscale to 32, and the software down-scale to 31. The result, with 'smooth-scaling' would be quite good. Need to take Gustavo's 12 hr. advice... :) ____________________________________________________________ Click to become an artist and quit your boring job. http://thirdpartyoffers.juno.com/TGL2141/fc/Ioyw6i3mzjsJ8nFkEvwwtvjjwTGg6ELR7S8cOVhATcPDJdfD7bDVSE/ |