9700 benchmarks

[Dio]

It's generally impossible to spot aliasing on screenshots UNLESS you are using a texture which has a regular pattern that rises above the Nyquist limit, such as a checkerboard texture.

If you sample above the Nyquist limit on a low-frequency texture, you will not see any significant artifacts. If you sample above the limit on a high-frequency but 'random' texture, you will see random pixels, which in a static shot will look the same as they would otherwise because it's still random.

In contrast, undersampling is easier to spot on screenshots; if something looks blurry in the distance it is being undersampled usually due to LOD bias tweaks. So if you have two screenshots where one looks 'aliased' and one doesn't, it probably means the one that doesn't is actually undersampled. Now, where have I seen this before....

The simple solution is to see it move. Aliasing is usually obvious even on random high-frequency textures as soon as you're in the game.

The fact that Dave hasn't seen any is, I would say, an important data point.

 

[KimB]

Can texture aliasing be exxagerated by other factors too? For example colour settings, gamma, saturation etc?

Texture aliasing mostly depends upon the contrast of the color within the texture, and how regular the texture is. So, I suppose it is possible that a higher contrast value would make texture aliasing more apparent, and adjusting anything that increases brightness also might make aliasing more apparent, since it's easier to see everything. Still, I doubt any of these options will have much effect. It mostly depends on the texture being viewed.

 

[KimB]

sorry, wrong.
It doesnt mean they will shimmer. Maybe you can see some of them, and maybe those that you can see are the onse you continue to see.

Try looking at the shot. It's absolutely obvious that the stars shown cannot be the only stars in the flag.

Either way, aliasing only shows up in MOVING images, really (texture aliasing) - at least for the most part.

Um, no. One excellent example is moire patterns. If a repeating, high-contrast texture is set to too aggressive of an LOD, it will show these, and they are quite ugly.

Now assume you take the camera view int hose pics, and slowly move TOWARDS the flag.
Your precious GF4 will have worse aliasing than ATI's because of the sudden change to seeing all the stars (according to you, only ATI's method is flawed enough to only show some fo the stars).

This is why trilinear filtering is used. Also note that I said in an ideal situation...I'm not sure whether or not a GeForce4 with the latest drivers would fail to show any stars until it got close enough to display them all (with bilinear filtering), as the LOD value apparently has been made more aggresive.

 

[darkblu]

sorry, wrong.
It doesnt mean they will shimmer. Maybe you can see some of them, and maybe those that you can see are the onse you continue to see.

Try looking at the shot. It's absolutely obvious that the stars shown cannot be the only stars in the flag.

Either way, aliasing only shows up in MOVING images, really (texture aliasing) - at least for the most part.

Um, no. One excellent example is moire patterns. If a repeating, high-contrast texture is set to too aggressive of an LOD, it will show these, and they are quite ugly.

Now assume you take the camera view int hose pics, and slowly move TOWARDS the flag.
Your precious GF4 will have worse aliasing than ATI's because of the sudden change to seeing all the stars (according to you, only ATI's method is flawed enough to only show some fo the stars).

This is why trilinear filtering is used. Also note that I said in an ideal situation...I'm not sure whether or not a GeForce4 with the latest drivers would fail to show any stars until it got close enough to display them all (with bilinear filtering), as the LOD value apparently has been made more aggresive.

..is it just me or you tend to answer only those posts you can somehow counter, while conveniently overlooking those wich on-spot show a whole in your stand?

openGL guy's and Dio's last posts address _precisely_ your "concerns" re-iterated in your last post. go read them again.

 

[KimB]

It's not as simple as that. In this case, the flag has ripples in it which can cause different mipmaps to be used for different parts of the flag.

I suppose that is possible, but I find it highly unlikely that the flag would be "just" that far away...but yes, it is a possibility (which means I am admitting that there may indeed be no aliasing in that shot...but I consider it a rather small chance).

Additionally, isn't that shot using trilinear filtering? I can't seem to see any MIP map lines...but I can't tell for certain (Most the textures are very low-contrast...).

 

[KimB]

openGL guy's and Dio's last posts address _precisely_ your "concerns" re-iterated in your last post. go read them again.

Actually, I didn't even see OpenGLguy's last post, thanks.

As for DIO's post, I thought it was very good. For example, DIO said:

It's generally impossible to spot aliasing on screenshots UNLESS you are using a texture which has a regular pattern that rises above the Nyquist limit, such as a checkerboard texture.

Notice how the only texture I was able to point out was one with a regular, high-contrast pattern? (in this case, the stars on the flag)

 

[Dio]

I suppose that is possible, but I find it highly unlikely that the flag would be "just" that far away...

The distance is not the only thing that affects mipmapping - didn't you mention you'd read the page OpenGLGuy provided on the GL specification? Have a think about what that actually means, the use of partial derivatives of u and v against x and y in screen space.

Therefore, if you tilt something at an angle non-perpendicular to the Z plane this may cause the texture to select a lower mip level. This is the whole point of anisotropic filtering, of course, to reduce the 'spreading' effect this gives due to one texture axis (that most perpendicular to the Z plane) becoming undersampled.

Or to put it very simply: a polygon may touch the near plane, be 'fully' visible, and still select the 1x1 mip level from a 256x256 texture, if it is only 2 pixels at maximum width and 512 pixels at maximum height. As the equations show, Z distance is irrelevant.

 

[Ailuros]

Just thought I'll add another "flag" to the discussion, just that you have more material to deal with.

4xS, default driver LOD setting, 30.30 (no idea why the grass turns up in wrong places; it's the demo and I refuse to reinstall that stupid game anyway).

oooops edit:

forgot the screenshot

http://users.otenet.gr/~ailuros/C4defaultLOD.jpg

 

[Doomtrooper]

Thats interesting, although the driver used in the test was NVIDIA Detonators v29.80 (WHQL Certified) and they used just 4X FSAA and not the 4 X S option.....don't know if it was any different.

 

[KimB]

The distance is not the only thing that affects mipmapping - didn't you mention you'd read the page OpenGLGuy provided on the GL specification? Have a think about what that actually means, the use of partial derivatives of u and v against x and y in screen space.

Yes...though I mostly just skimmed it...but I had known that angle can affect MIP map levels for a long time.

A few things on that.

If if it was just the angle that was causing the MIP maps to change and therefore cause some stars to be visible, you would expect those that popped in and out of visibility would be in, more or less, lines following the leading and trailing edges of each wave. This is not what is in the picture (the stars are more in "pockets"). Additionally, the "waviness" of the flag decreases as you approach the pole (something that I was quite impressed at...it's more realistic!), but the pattern of missing stars stays essentially the same.

 

[John Reynolds]

Just thought I'll add another "flag" to the discussion, just that you have more material to deal with.

4xS, default driver LOD setting, 30.30 (no idea why the grass turns up in wrong places; it's the demo and I refuse to reinstall that stupid game anyway).

oooops edit:

forgot the screenshot

http://users.otenet.gr/~ailuros/C4defaultLOD.jpg

The flag pole definitely shows the edge AA differences between the two boards.

 

[andypski]

Just thought I'll add another "flag" to the discussion, just that you have more material to deal with.

4xS, default driver LOD setting, 30.30 (no idea why the grass turns up in wrong places; it's the demo and I refuse to reinstall that stupid game anyway).

oooops edit:

forgot the screenshot

http://users.otenet.gr/~ailuros/C4defaultLOD.jpg

The flag pole definitely shows the edge AA differences between the two boards.

nVidia's 4XS mode is actually the one that gets closest to the edge quality of R9700 on near horizontal edges since it effectively uses a skewed (not rotated) grid. I believe that the edge quality of 4XS on near-vertical edges should be about the same as nVidia's ordinary 4X mode since the grid is still vertically aligned, and this shows up on the flagpole which is AA'd well by R9700 and relatively poorly by nVidia's 4X and 4XS modes.

 

[KimB]

nVidia's 4XS mode is actually the one that gets closest to the edge quality of R9700 on near horizontal edges since it effectively uses a skewed (not rotated) grid. I believe that the edge quality of 4XS on near-vertical edges should be about the same as nVidia's ordinary 4X mode since the grid is still vertically aligned, and this shows up on the flagpole which is AA'd well by R9700 and relatively poorly by nVidia's 4X and 4XS modes.

Which brings up another point...I really hope that nVidia does better than their "half-assed" 4xS mode in the NV30. I doubt that they can actually surpass the R300 in edge AA quality, but I'm hopeful that they can surpass the R300 in speed and the available number of samples per second.

And I most certainly am hoping that nVidia didn't keep the same anisotropic filtering as that available in the NV2x cards...I'm hoping that they managed to improve both performance and the maximum degree of anisotropy available.

 

[McElvis]

It's not as simple as that. In this case, the flag has ripples in it which can cause different mipmaps to be used for different parts of the flag.

I suppose that is possible, but I find it highly unlikely that the flag would be "just" that far away...but yes, it is a possibility (which means I am admitting that there may indeed be no aliasing in that shot...but I consider it a rather small chance).

Additionally, isn't that shot using trilinear filtering? I can't seem to see any MIP map lines...but I can't tell for certain (Most the textures are very low-contrast...).

Interesting thing to note, if you zoom in on the flag in the Parhelia picture, you will notice that the ends of the flag look as blured as the GF4, but the middle of the flag looks MUCH BETTER, just as good quality as the flag in the R9700 picture.

Does that lead to believe that the flag is using different mipmaps for different parts (at least on the Parhelia, anyway)?

 

[John Reynolds]

Which brings up another point...I really hope that nVidia does better than their "half-assed" 4xS mode in the NV30. I doubt that they can actually surpass the R300 in edge AA quality, but I'm hopeful that they can surpass the R300 in speed and the available number of samples per second.

And I most certainly am hoping that nVidia didn't keep the same anisotropic filtering as that available in the NV2x cards...I'm hoping that they managed to improve both performance and the maximum degree of anisotropy available.

I also hope they improve the functionality in their driver panels, because they are sorely lacking in this area!

 

[alexsok]

I also hope they improve the functionality in their driver panels, because they are sorely lacking in this area!

Definetly agree with u there!

There were a few very suspicious pics making their appearance on some websites, claiming to be from the 3xxx series of detonators, where the interface was completely revamped!

I sure hope that was a pic of the new detonator 5 drivers that will be released with the NV30 (which will probably adress all of the complaints people have right now in regards to the interface and the features provided in currect detonators - especially the aniso settings in Direct3D, which are completely missing!).

 

[Ailuros]

John,

That's what I initially said looking at the R300 shots. I can't find some older 2x2OGSS shots from the same spot, but the flag pole looked almost identical there too.

Doomtrooper,

I assume that 4xS comes also with a slight LOD offset from it's SSAA part. 4xS by the way was present many many driver revisions before 29.xx. You need a 3rd party tweaking utility to expose 4x - 9tap for both API's though.

Sidenote: does anyone actually like that game? I for one don't.

 

[alexsok]

Sidenote: does anyone actually like that game? I for one don't.

Not really...

It has some pretty cool graphics & effects (like the water effects), but it's not fun, at least for me...