The pixel boundary and current down-filters

You'll notice that these hardware implementations of AA post-filtering sample the signal for a single pixel to provide data to feed the down-filter. We'll come back to that later in the article, but it's considered a requirement of a fast hardware-accelerated post-filter AA scheme that involves subpixel multisampling.

The Current Common Down-Filters

The down-filter is where the AA magic happens, and it's the final arbiter of image quality in a post-filter AA scheme. How you combine the sample information in the filter kernel is key to getting good image quality from your anti-aliased image. Get the filter wrong and you can introduce more aliasing, for example.

However for any real-time graphics application, speed and area-efficiency have to be considered when choosing which filter to accelerate in silicon.

Box Filter

A box filter is the simplest to compute, since all colour samples are given equal weight to contribute to final pixel colour: 1/num_samples. So four samples get a 25% weight each, 8 samples a 12.5% weight and so on. Even for a floating-point framebuffer representation being worked on in the pixel ROP, the hardware for filtering is fixed-function and area-efficient, so the box down-filter can be computed for a relatively high number of per-pixel samples at high speed, even for floats.

The box filter is prevalent in all the hardware down-filter implementations over the last few years, with the bottleneck not the logic for the filter calculation but the bandwidth through the ROP and any decompression of samples.

Tent Filter

The tent filter is where AMD deviate from box filtering, fuelling the argument they make that they're building a better filter. The tent filter linearly weights samples according to variance, which in this case is the sample's distance to the pixel centroid. The linear contribution therefore depends on the distance of the farthest multisample being filtered, and it's what allows them to implement wide and narrow versions of the filter, depending on the area the filter samples from.

The filter kernel weights are therefore pre-computable if you know in advance exactly where your samples will be taken from, although at the expense of absolute accuracy. The name comes from the shape of the filter if you plot weight against distance, and its linear interpolation also maps nicely to hardware accelerated lerp logic already available in the hardware, outside of the ROP.

However because the tap locations are the same for any down-filter using hardware multisamples, the filter's calculation isn't any different in terms of subpixel coverage, data which can only be bettered by taking more taps. So the tent filter doesn't approximate coverage any better than a box filter for post-filter AA, an important property to consider when deciding which down-filter to choose.