Company: ATI Technologies
Authour: Alex 'Morgoth Bauglir' Voicu
Editor: Charles 'Lupine' Oliver
Date: April 10th, 2008
Due to their relative age, both AA and AF (we'll use these abbreviations throughout the rest of the article) aren't necessarily en-vogue features. Though most reviews include tests that have both enabled, they're not necessarily the subject of in-depth investigations any longer, contrary to how things were in the age of the original GeForce and the VooDoo5, when AA was the hottest new arrival and AF was starting to become somewhat interesting, with the original Radeon bringing high enough levels of filtering, albeit at the cost of aggressive angle-dependency and the limitation of it working only in tandem with bilinear filtering.
In this article, both of the aforementioned techniques will be evaluated in-depth, and we'll attempt to show you how they can improve the visual quality of your everyday gaming.
For starters, let's deal with AA, as that's arguably the most noticeable of the dynamic duo, both in terms of its visual impact as well as its ability to affect performance. For those technically oriented, this Wikipedia article provides most of the nitty-gritty details of what AA is, and what it tries to accomplish. In layman's terms, AA is a technique through which the stair-step effect, known as aliasing (there's a surprise, eh?), that is very apparent in-game along polygon-edges is eliminated/reduced significantly. Aliasing is an artifact resulting from the process of sampling and reconstructing a signal as an alias of an original signal.
The first attempt at implementing AA in 3D games employed a technique known as Supersampling, which basically involved rendering the image at a higher resolution (2 times or 4 times as high, for 2X or 4X AA respectively) and then downsampling the high resolution image. This was quite costly in terms of fillrate, memory bandwidth and memory real-estate, and the cards of the time lacked the resources required for making it a feasible option.
With the introduction of the Geforce3, the technique for tackling AA that's been the preferred solution to this day came to the scene: Multisampling. It was significantly more efficient, requiring that only the depth value be supersampled, with other aspects such as color or stencil staying at one evaluated value per pixel. Needless to say this reduced the fillrate and bandwidth hits significantly as well as being less famished for VRAM, making Multisampling a solid and desirable solution to the aliasing issue. Things have been further optimized, with the introduction of color-compression and several other tweaks, and today it seems that MSAA (Multisampling AA, another abbreviation that'll be used from here on) will be with us for quite a while.
Another important aspect that has a major influence on the effective quality of AA is the sampling pattern being employed. This refers to the placement of the samples themselves around the pixel center (or centroid). The Geforce3 used a rotated grid for 2X AA reverting to an ordered one for 4X AA. Using an ordered grid is the worst case scenario in terms of effective quality, as it has a number of sensitive angles at which the quality of AA is reduced (some actually seemed to receive no AA at all). A rotated grid is better, and the best case is using a sparse sampled grid in which no two samples run along the same line.
content not found