Anti-Aliasing and Anisotropic Filtering Comparisons
In our typical fashion we have image quality comparisons for you here, in various AA and Anisotropic Filtering modes. These shots were taken on a Radeon X800 XT with new beta Catalyst drivers and a GeForce 6800 Ultra with Forceware driver version 60.72.
When we first set out to take screen captures in UT2004, we were targeting a map area that had a very challenging aliasing problem to tackle. This area of the game's "AS-Glacier" map is perfect example of modern AA challenges as game engines become more complex with detailed high resolution textures, shaders and lighting.
|X800 No AA
||X800 4X AA
||X800 6X AA
Notice in the above shots how nicely the outside lines of the girder above and on the sides of the scene, as well as the chain link going across, are all cleaned up very neatly, in both 4X and 6X AA modes, versus the No AA shot. The edges of those structures are pretty much perfectly clean at 1024X768 resolution.
A need for Centroid Sample AA:
However, take a look at the cross members inside the girder itself. See how they are still fairly aliased and jagged? This is a good example of Multi-Sample AA not cutting the mustard with a surface texture. Epic chose to texture map those cross member images in this scene, rather than draw additional polygons to create the structure. Traditional Multi-Sample AA techniques are not able to fully clean up the texture image, as part of the texture is outside its sample area and color range. Centroid sampling aims to remedy this issue, when it is fully supported in DX 9.0c and will allow for full Anti-Aliasing of edge surfaces, as well as texture area.
Article Update -5/6/2004
Leading edge Graphics Processor technologies are becoming significantly more challenging to cover from an analyst's perspective, in both their micro-architecture and the software algorithms that drive their innovative and complex rendering techniques. Unfortunately we had a misguided understanding of Centroid AA, at the time this article was published, along with its capabilities and the problems it is targeted to solve. Shortly after the launch of this article, we were contacted by David Nalasco, Technology Marketing Manager for ATi's Desktop Products, regarding our comments in our "Need for Centroid Sample AA" section above. Dave gave us a much better understanding of aliasing artifacts we were seeing in our UT2004 screenshots and what Centroid AA can and can not do, in terms image quality benefits. Here's what we were told.
"The girders in that image are not getting anti-aliased because they are rendered using an alpha-tested texture. This means each texel is tagged as being transparent or opaque. Since multisample AA only works on polygon edges (this is what makes it so fast), it does not affect areas inside a polygon. Supersampling can address this type of edge by blurring the whole scene slightly, but this has the side effects of being much slower, and also blurring out edges you actually want (such as sharp text).
There are two simple ways to get around this issue - use polygons to represent the beams instead of alpha textures, or use alpha blending instead of alpha testing. Alpha blending supports multiple levels of transparency for each pixel (usually 256 levels), so when you use bilinear or trilinear filtering, the edges get nicely softened. This is used in most other places in UT (the grass on the ground in Antalus comes to mind... get up close to it and you see the edges stay smooth), so I don't know why it wasn't used in this particular case.
Centroid sampling addresses a completely different problem. With standard multisample AA, the color of a pixel is only sampled once at the pixel center. The other samples are used to determine how much of the pixel is covered by the current polygon, and this percentage is then used to blend the current polygon color with the background color. In some cases, only a small corner of a pixel might be covered by the polygon, so the pixel center might actually lie outside of the polygon. This can cause the color value at this point to be undefined; or, if you are packing multiple textures together into a single polygon, you can end up sampling a different texture adjacent to the one you wanted. Either way, you get artifacts.Centroid sampling fixes this by ensuring that the color is always sampled from a point in the pixel that lies within the current polygon."
|GeForce 6800U No AA
||GeForce 6800U 4X AA
||GeForce 6800U 8X AA
NVIDIA's 4X Multi-Sample AA, on the GeForce 6800 Ultra, looks right on par with ATi's 4X AA mode. On the other hand, the GeForce FX 6800 Ultra's 8X Super Sample mode looks nearly perfect across the entire scene, including the cross member textures inside the girder. Unfortunately, this mode is completely unplayable on even a powerful GPU like the NV40. At 1024X768, without a single bot in the map, 8X SS AA mode produced frame rates hovering around 30 fps. Add some bot action and a bit of gun fire to the mix, and this setting becomes a slide show. Why then have we not been exposed to 4X Super-Sample mode? Surely this would have acceptable performance at half the number of samples, yet still deliver top notch IQ, right? That's a very good question we do not have answers to unfortunately, at this time.
We then fired up an area of the "DM-Corrugation" map in UT2004, enabled various Anisotropic Filtering settings for both the Radeon X800s and the GeForce 6800 Ultra. We broke out our pixel peering eyeballs for a closer look at what each card was rendering. We've also taken shots here with two different Aniso settings for the NVIDIA card, "Trilinear Optimizations On" and Trilinear Optimizations Off". With the Opt On setting the GeForce 6800 Ultra isn't doing full Trilinear Filtering but a combination of Trilinear and Bilinear depending on the scene's requirements. Have a look for yourself and be the judge on what you think looks best.
|X800 No AF
||X800 8X AF
||X800 16X AF
8X AF Opt. Off
8X AF Opt On
We intentionally didn't go in a blow up images this time, zooming in at 200%, to see pixel level detail in these shots. You simply do not play a game that way and taking a standard sized still screen shot is typically more than adequate to highlight differences in Aniso filtering quality. In our opinion, it was a virtual toss-up between the rendering quality of the ATi Radeon X800 series cards and the NVIDIA GeForce 6800 Ultra. On one hand the lighting and ceiling textures on the Radeon 8X and 16X shots are ever so slightly more detailed than in the 6800 Ultra shots. On the other hand, the far side surface of the metal floor grates looks a bit more detailed in the NVIDIA based shots, whether they were taken in Trilinear Optimization On or Off mode. As a matter of fact the only barely visible difference we could see between these two settings for NVIDIA's card, was the pipe in the far upper right section of the ceiling. In the Trilinear Opt On shot, the pipe actually looks ever so slightly more detailed. Either way, again in our humble opinion, which isn't always easy to obtain (a humble one that is), Aniso image quality between the Radeon X800s and NVIDIA's new flagship, at least as far as UT2004 goes, is a fairly close horse-race.