3dcgi said:
I read the following interview and I'm curious what everyone thinks about the following comment from David Kirk.
http://www.voodooextreme.com/hw/interviews/davidkirk/2.html
That brings me to a point about doing it right, versus cutting corners. Do you know that one of the hardware vendors only antialiases the top and bottom of the screen, but not the middle? That?s because in one of the main benchmarks, Quake3, you don?t see much difference. But, now you know: look for it, and you?ll see it. Here?s another one: do you know that one of the vendors who makes a lot of noise about anisotropic filtering doesn't even do it right? It only works for almost horizontal or almost vertical edges. Try this. Fire up a flight simulator, and fly at a 30 or 45 degree angle, and look how jaggy things get.
He is confusing me. He mentions Anisotropic filtering and then goes and mentions edges (would he mean edges inside the texture, but then why mention a flight sim?) ? I assume he means anti-aliasing of edges. Now as I pointed out in the original FSAA whitepaper, the 3dfx one, the quality of the anti-aliasing depends on the sample points positioning, and different patterns simply by definition will act differently with different angles of the edge. A non-rotated sample pattern does a poor job with near horizontal/vertical edges, while a rotated pattern should do them quite good, but in reality the rotation just moves the problem to different angles - so quite possibly at 30 or 45 degrees you get the bad case that non-rotated patterns have at the near horizontal/vertical edges.
Only performing AA on the top and bottom of the screen sound very unlikely to me since AA is quite tricky to enable/disable for regions of the screen. Anti-aliasing completely changes your back buffer behaviour and changing this for regions of the screen would probably lose you performance rather than gain it. The only algorithm I can see this working with would be Matrox FAA since it has less impact on the buffer behaviour. In theory a tiler could do this but I am pretty confident that PowerVR is not doing this
Anyway going back to his arguments :
- Assuming he talks about AA and 30 and 45 degrees and increased jaggies : this is expected behaviour and will happen on his own NVIDIA hardware as well, you just have the find the worst case angle for the sampling pattern used. Difffernt angles will AA better or worse, its the nature of the sample pattern (unless you go for very randomly distributed sample patterns with a LOT of samples).
- Only AA top/bottom of the screen sounds quite unlikely to me given the way most hardware implements AA by upsampling the buffers - upsampling/downsampling only part of the buffer makes NO sense and would be hard to implement on most hardware.
- Anisotropic filtering and doing it "right". There is no industry standard for anisotropic filtering so to call one technique right or wrong is always going to be tricky. The results will probably be of better/worse quality which turns into a very subjective nature. Given that anisotropic techniques are often "adaptive" it makes things even more complex.
Anisotropic and AA is always going to be apples and oranges because the techniques, sampling patterns, etc, etc... are different from vendor to vendor. Calling a technique wrong or "less good" is going to be very subjective.
For what its worth I think I noticed that one hardware vendor out there fails to render all the shadows in the 3DMark Game 1 (Quality Test is easy to spot it) using WHQL drivers... must be my eyes fooling me since no-one would dare to do that... if it's not my eyes it must be an
accidental bug that results in only the 2 big noticable shadows being rendered
K~
