Re: [Algorithms] Filtering

[Fabian G. <ry...@gm...>]

On 01.10.2010 11:04, Jeff Russell wrote:
> I'm thinking maybe older gpu's did even courser filtering than this -
> 6-bit or something instead of 8. Maybe this was the recent improvement?

There's the subpixel resolution of texture coordinates going into the 
filtering unit and there's the actual filtering precision.

The former will make blocks appear once you zoom up close enough. For 
most texture this isn't a problem as you try to keep the texel:pixel 
ratio close to 1:1 anyway, but I ran into this a while ago on some 
textures that were mostly gradients. This is extra-annoying on gradient 
textures since you can't make them too large either - if they're too 
large, adjacent pixels have the same color value sometimes, which 
bilinear filtering will expand into a constant stretch in the middle of 
your gradient - a very visible problem, unfortunately.

The actual filtering has been working on at least 8-bit values for quite 
some time (at least as far back as the GeForce3 / original Radeons, not 
sure about earlier cards).

You may get more precision on newer cards, but it's probably not much 
more (half-floats or something like that, most likely). These return 
paths aren't super-wide for a reason - you don't put 128-bit (4x32) 
buses all over your chip when you can avoid it. (Samplers most likely 
support returning full 32-bit results, but it will take more than one 
cycle to transfer results then).