[Matt]

You may remember that our AquaMark 3 preview stirred up things a bit when I did an image quality comparison with the Radeon 9800 and GeForce FX 5600 Ultra with both 45.33 and 51.75 drivers. Since we didn't have a reference image, we just couldn't tell just how far off the cards are, or how spot on. We finally got a software rasterizer reference image from Massive, and have posted a follow-up article on 3DGPU.

http://www.3dgpu.com/modules/wfsecti...p?articleid=76

You'll also find the uncompressed TIFF file that Massive sent us of the reference image, so you can compare yourself your own results.

[Pete]

Sassy.

[3dilettante]

There seems to be something very different with the way the reference rasterizer handles the sea grass from all of the video cards.

Grass clumps in different places on hardware than it does in the reference rasterizer. The reference image has some very noticeable bare patches that are inexplicably filled with grass on hardware, and vice versa.

[Pete]

Yes, I was wondering if that was a precision or architecure-specific (thus unique) optimization made by Massive. Doesn't affect IQ, IMO, so it's a non-issue. The variance in ground luminosity, however, is a point worth investigating.

[3dilettante]

It looks like all the hardware screenshots have AA on, which is kind of a shame, since I really would have liked each card to be as close to the refrast as possible.

Interestingly, it looks to me that the way Aquamark handles the sea grass is by running some kind of shader operation over a fan shaped polygon. It does save a quite bit of polygon processing in comparison to working on each blade of grass, so that sounds reasonable.

What I think I see, however, is that the grass polygons are aligned very differently in the refrast than the hardware versions (all of which seem to share an overall simularity in how they handle this). In some cases, a polygon that is perpendicular to the camera's line of sight in the hardware version--and therefore being seen at its widest--is parallel and seen almost edge on in the refrast.
(The large clump of grass at the bottom of the screenshot, about 1/3 of the way from the left side is a very obvious example)

Other polygons are at different angles, and a few seem to be 180 degrees off from the refrast. Their points of rotation seem to be non-central, which might be due to some asymmetry in the polygons used.

How is this possible that there can be such a disparity?
I think I read somewhere that the refrast runs with much higher internal precisions since it is executed on a CPU. Is this true?
Aren't there ways to limit the precision so that it is at least possible for a video card to approach the refrast?

Can a matter of precision cause such a large disparity? This isn't a matter of slightly different color or brightness, but of polygon alignment. (Not that the nvidia renderings couldn't use a bit life with their palettes.)