Geforce FX Bilinear Anisotropic Filtering Question ??

[Tahir2]

8xAF Radeon 9700 Pro

16xAF Radeon 9700 Pro

 

[KimB]

Oh, duh. I just did some more thinking. A hyperbola will map to a square when viewed on a tunnel like that. A hyperbola is a conic section that is created by slicing the cone vertically. Since the view in that program is essentially the same as looking through the base of a cone, the hyperbola will look just like its conic section, which will look just like a square once the x/y dependence is removed.

 

[Colourless]

I just have to make a comment. Can people not test using trilinear. It makes it too hard to to see the shape of the curves, which is what we want to see.

 

[Ollo]

Bilinear and trilinear gives the same result for me on a 9700. Perhaps the app forces tri?

 

[OpenGL guy]

Oh, duh. I just did some more thinking.

Better do some more.

A hyperbola will map to a square when viewed on a tunnel like that. A hyperbola is a conic section that is created by slicing the cone vertically.

Not exactly correct. As long as the plane has an angle from the cone's axis less than the angle of the cone's sides, the intersection will be a hyperbola.

Since the view in that program is essentially the same as looking through the base of a cone, the hyperbola will look just like its conic section, which will look just like a square once the x/y dependence is removed.

The projection of a hyperbola onto a plane is still a hyperbola as long as the plane of projection isn't perpendicular to the plane the hyperbola lies in, in which case the projection is a line.

You can invert the projection (i.e. map your planar hyperbola back onto a cone) and you still won't get a square.

 

[Tahir2]

I just have to make a comment. Can people not test using trilinear. It makes it too hard to to see the shape of the curves, which is what we want to see.

Some people are never satisfied... bilinear shows virtually no curving at all.

Edit: I lied, there is some curving present:

Are you happy now or should 'we' jump through hoops as well? :?

 

[Xmas]

Bilinear and trilinear gives the same result for me on a 9700. Perhaps the app forces tri?

The app certainly doesn't. But it seems like the newer drivers do force tri if you select quality AF in the driver panel. While this isn't exactly a bad thing from a gamer's standpoint, I wish that those two things (force AF and force tri) were separate settings.

 

[Randell]

[quote="Xmas] I wish that those two things (force AF and force tri) were separate settings.[/quote]

why? the fact is you can turn AF off and use bi/tri according to the app.

 

[demalion]

I've often thought that the behavior should have an "Override" switch, so that the, with that switch cleared, set values were only used when no setting was specified by the application. I'm assuming the driver/API interaction would allow this, but perhaps some APIs don't?

But in any case, what does "LINEAR_MIPMAP_LINEAR" mean? Perhaps that setting is responsible?

 

[OpenGL guy]

But in any case, what does "LINEAR_MIPMAP_LINEAR" mean? Perhaps that setting is responsible?

LINEAR_MIPMAP_LINEAR means the texture filtering is bilinear and the mip filter is linear. This is equivalent to trilinear filtering. In D3D, you would set min filter = mag filter = mip filter = LINEAR to get trilinear. Bilinear would be min = mag = LINEAR, mip = POINT.

 

[KimB]

Not exactly correct. As long as the plane has an angle from the cone's axis less than the angle of the cone's sides, the intersection will be a hyperbola.

Ah, right, but I don't think you can do that and keep the function independent of whether x or y is selected.

 

[KimB]

why? the fact is you can turn AF off and use bi/tri according to the app.

Only most apps don't have support for adjusting the AF level. It's rather pointless to run without anisotropic filteing just so that you can let the app choose its own filtering.

 

[OpenGL guy]

Not exactly correct. As long as the plane has an angle from the cone's axis less than the angle of the cone's sides, the intersection will be a hyperbola.

Ah, right, but I don't think you can do that and keep the function independent of whether x or y is selected.

You've lost me here. For a hyperbola, x and y are always related.

 

[K.I.L.E.R]

Bah! Majority of you people are insane.
Who here plays games? I would gladly play with no FSAA or AF without bitching and still enjoy the game.

 

[Randell]

why? the fact is you can turn AF off and use bi/tri according to the app.

Only most apps don't have support for adjusting the AF level. It's rather pointless to run without anisotropic filteing just so that you can let the app choose its own filtering.

As I never play without AF I dont care, I was just wondering the need for the drivers to force bi/tri. Without forced Bi AF or Tri AF games will default now to Tri wont they? Most FPS eg QIIIA, Unreal, Lithtech, Croteams engines all allow filtering selection, so you are left with how would I force Bi if you cant select. hmm nope dont care about that

 

[KimB]

Not exactly correct. As long as the plane has an angle from the cone's axis less than the angle of the cone's sides, the intersection will be a hyperbola.

Ah, right, but I don't think you can do that and keep the function independent of whether x or y is selected.

You've lost me here. For a hyperbola, x and y are always related.

I meant that you should be able to exchange x and y without any change in the output, in other words giving no preference over any derivative.

 

[KimB]

As I never play without AF I dont care, I was just wondering the need for the drivers to force bi/tri. Without forced Bi AF or Tri AF games will default now to Tri wont they? Most FPS eg QIIIA, Unreal, Lithtech, Croteams engines all allow filtering selection, so you are left with how would I force Bi if you cant select. hmm nope dont care about that

I really dislike the forcing of bilinear/trilinear with anisotropic. That is, ATI's drivers don't appear to just force bilinear/trilinear for texture that are already using one or the other, but also force them for textures that use point sampling. Why is this bad? The text in some games uses point sampling. For example, this means that some of the text in Neverwinter Nights becomes blurry when anisotropic filtering is enabled, and I have noticed that the text in Morrowind becomes blurry if anisotropic filtering is selected and the MIP LOD is lowered (which is one major reason why I continually complain about the texture aliasing of the 9700 Pro...I can't adjust the LOD to compensate properly, as it will blur text in some cases).

 

[Dio]

If you force something on the panel, you're asking the driver not to do what the application wants. In the absence of psychic PC's, it can't be certain what is the right intersection between what the app wants to do and your modifications...

 

[K.I.L.E.R]

In the absence of psychic PC's, it can't be certain what is the right intersection between what the app wants to do and your modifications...

That's why we have math.

 

[Thowllly]

If you force something on the panel, you're asking the driver not to do what the application wants. In the absence of psychic PC's, it can't be certain what is the right intersection between what the app wants to do and your modifications...

Yeah, but just because you ask the driver to change one behaviour (how to filter texels within one mipmap level) doesn't mean that the drivers should also change another behaviour (whether to interpolate between mipmaplevels or not). The point here is that the drivers shouldn't force bilinear or trilinear at all. When you force anisotropic filtering in the ATI drivers you have to choose between forcing bilinear or forcing trilinear. There should also be an 'application select' option that simply uses the filtering format the application requests (plus anisotropic, of course).

The fact is, in the first R300 drivers, selecting quality simply meant that the drivers used what the game requested. But then some stupid ass reviewers started complaining about that, and ATI changed the behaviour to force trilinear, and the users were left with no way of selecting 'application select' while forcing anisotropic.

And the drivers don’t need to be psychic to know when not to change any settings. If the application requests point sampling you can bet that there is a good reason for this (the only time any game I’ve seen use point sampling is when trying to avoid any 2d elements getting blurred), so leave it at point sampling (preferably without any AF).

I would really like NVIDA to include a force trilinear/force bilinear option in the drivers (that way I could force trilinear in Tribes2), but only as long as they also gave me the choice of 'Application select', and only as long as point sampling wasn’t affected.