Could anyone with a GFFX please...

[Xmas]

...please download the following program and run it with the following settings:

8xAF application
8xAF balanced
8xAF performance

http://demirug.bei.t-online.de/D3DAFTester.zip

and do a comparison with the program linked in my sig.

If you have the time, you could also try the 'texel analyze' function and take a screenshot of it.

If you don't have webspace for screenshots, a description of the differences would suffice. Detonator version is important.

tia

 

[Dave Baumann]

OK, I tried this on an NV31 with the det 43.45's and it shows exactly the same as as your app, in application mode. The puzzling thing is that it appears completely invariant to the 'Intellisample' mode selected in the driver control panel.

 

[Xmas]

Thanks, that's another confirmation. NVidia is obviously using another 'optimisation' method in D3D, as there is still a performance difference between the modes. But sometimes there is pretty much no visible difference. Maybe they're only doing full AF on the base texture. Demirug is currently working on a new version of the tool that also tests multitexturing.

 

[Dave Baumann]

Well, if you have 8x AF selected on the slider then I would have expected 8x to show up to show up in the app, however it doesn't. I've just tried the same with SS:SE in DX mode and that doesn't start up with the aniso mode enabled in the driver either, so I'm not convinced the drivers aren't just borked!

 

[Dave Baumann]

Nope, I'm tired. Ignore what I said about SS:SE.

 

[Xmas]

If the application sets the degree of anisotropic filtering, this will override the setting in the driver panel. That's why the slider in my program still works when you set AF in the driver panel.

(The strangest thing, however, is that my program doesn't start with the driver setting although I do not set the AF degree until the user changes that setting. Even when I comment out every line of code that is related to AF, it still doesn't start with the driver setting)

I don't have the source of the D3D app so I can't tell when it sets AF first.

 

[Bambers]

ATis texture preference slider does a similar trick to the nv30 one, lessening how much trilinear is used.

In d3d it goes down to about only 50% trilinear at high performance.

I think reviewers need to make sure that they use high quality in opengl games too though as the driver default is quality (but high quality in d3d for some reason) and the opengl trick is a little odd. Its much more extreme and deosn't vary between high performance and quality, all give only about 25% trilinear (re early 8500 reviews and the not full trilinear) and it doesn't show up on the coloured mipmaps, as soon as coloured mipmaps are turned on it reverts to full trilinear in both q3 and xmas app, can someone test this with an r300?

Normal textures, mip boundries obvious

Coloured mipmaps on tinted, full trilinear used

These were both done with texture preference set to quality in the drivers.


On the texel analyse mode in the d3d apps is the amount of green represent the number of samples being used for each pixel?

 

[Xmas]

On the texel analyse mode in the d3d apps is the amount of green represent the number of samples being used for each pixel?

Yes. The number that appears in the lower right is the maximum number of texels per pixel measured.

 

[demalion]

Bambers,

I'm puzzled by the pictures and mip map detail analysis. Also, what are you testing on? Are you asking for someone to test on R300 for confirmation, or because you are using something else?

The bottom looks the same as what I get when I set all filtering methods to linear and set the bias to 5.0 on my 8500. My control panel settings have no impact (I set both Mip Map LOD and Textures to High Performance and the test app wasn't visibly affected) with Cat 3.2 drivers. How are you getting the application not to override the control panel settings?

Also, Xmas, "normal" texture mode shows a yellow tint being blended...what does that mean? Does it indicate something in particular?

As to your (Bambers) texture setting specifications:

The difference between Quality and High Quality in the past has been compressed texture usage...I'm not sure what it means for the latest drivers, and the difference between Direct3D and OpenGL defaults may be related to how compressed textures are used for each API. Also, lowering the texture quality slider also seems to have resulted in 16 bit texture usage in the past. Since you are getting the anamolies when lowering the texture slider all the way, I'm not sure it is a "trick" (I'm only aware of that one review that lowered it for benchmarks, so I'm not sure who they'd be meaning to fool) instead of an issue with single texturing with 16-bit textures and trilinear filtering.

I'm not sure to what you are referring with your OpenGL discussion...what do you mean by "25% trilinear" and "50% trilinear", for example? That top picture looks like bilinear and textures blurred by high LOD, though I haven't replicated it on my 8500. Perhaps I'm just misunderstanding something that you said.

If you want a screenshot of the 8500 replicating the bottom shot, mention it. Except for blending grey instead of blue, it really does appear to be identical to me, and since I'm not sure what card you are using, I wasn't sure it was worth the trouble to provide it.

 

[Xmas]

Also, Xmas, "normal" texture mode shows a yellow tint being blended...what does that mean? Does it indicate something in particular?

I think that pretty much indicates incorrect rendering, for whatever reason... can you post a screenshot?

 

[Bambers]

I have an 8500 myself and I wanted to see if the same results were given by r300s.

I also wasn't using 'tricks' implying anyones trying to trick anyone, just tricks is shorter to type than optimisations or whatever anyone wants to call whatever happens when you lower the texture slider

when I was saying 50% trilinear i was meaning that the amount of screen having trilinear done was about 50% like this one here

d3d app, texture pref: high performance

in texel analyse mode (max number was 8 )

As you can see its doing trilinear over limited intervals similar to the fx in balanced/aggresive mode. Its done this since the first catalyst drivers iirc ive just never had an app to test it, (coloured mipmaps wont show it in opengl)

I used a high lod (it was around 5) on the opengl app shots to show the mip boundries up more easily. In opengl the texture pref slider doesn't seem to work so well handling this method. Any setting other than high quality does only 25% with trilinear filtering, the problem is if coloured mipmaps are turned on. If they are then the card switches to full trilinear. This is why I used the tinted mode as you can clearly see than the textures underneath are smooth, while in the picture above without coloured mipmaps the textures have obvious mip boundries in them, both shots were taken with the same driver settings and in the same running of the program, the only difference is one is set to tinted and the other is normal.

The two things that interest me is why does it go to full trilinear filtering when coloured mipmaps are turned on and does the same apply to the r300 especially in its trilinear af.

Also in ati drivers the defualt texture preference under opengl is only quality not high quality like it is in d3d, in this mode on my 8500 it does less trilinear than the high performance mode of d3d, this could affect review results in opengl.

I think that pretty much indicates incorrect rendering, for whatever reason... can you post a screenshot?

the d3d app seems to use yellow and black check patterns for all the lower detail miplevels and this shows through in the tinted mode, I assumed it was designed this way.


Max number of texels on my 8500 in 16xAF is 36.

 

[demalion]

BTW, it doesn't appear on the version Bambers seems to be using (I just ran TextureFilteringTest.exe to check). This is what it looks like with stock settings on launch after "Normal" is selected instead of "Tinted":

 

[Bambers]

texturefilteringtest.exe is opengl, the one from the top of this thread is d3d. I've been using both in my posts, probably why they're written slightly confusingly :|

 

[Dave Baumann]

Bambers:

http://www.beyond3d.com//reviews/ati/r350/index.php?p=18

Those are the OGL results on R350 at the default settings - is that what you were after?

 

[Bambers]

Not quite, I'd like to see a shot like this one, coloured mipmaps off and a high lod (4.7 in that one).

While the coloured mipmaps in the pictures in the review show normal trilinear they do on my 8500 too. Turn the coloured mipmaps off and it reverts to the 'fake' form of trilinear it was original critised for on its release. It only keeps doing full trilinear on the high quality setting in the drivers.

Texture pref:
quality
high quality

both are set to linear_mipmap_linear, the difference is pretty noticable.

 

[OpenGL guy]

ATis texture preference slider does a similar trick to the nv30 one, lessening how much trilinear is used.

In d3d it goes down to about only 50% trilinear at high performance.

Considering that this is not possible on the 8500 or 9700, I'm going to blame the test here without even examining it.

 

[Tahir2]

ATis texture preference slider does a similar trick to the nv30 one, lessening how much trilinear is used.

In d3d it goes down to about only 50% trilinear at high performance.

Considering that this is not possible on the 8500 or 9700, I'm going to blame the test here without even examining it.

It could be the fault of the conclusion rather than the test application too so examining it may still be prudent?

 

[Dave Baumann]

This is the default OGL setting with an LOD of 4.7 on an R350

 

[Bambers]

thanks for that.

Considering that this is not possible on the 8500 or 9700, I'm going to blame the test here without even examining it.

It shows up in games too however, best example I could find atm, it really doesn't show up well unless you're moving.

looking at the righthand vertical rung of the ladder.

d3d - linear_mipmap_linear - high quality
d3d - linear_mipmap_linear - high performance
d3d - linear_mipmap_nearest
opengl - linear_mipmap_linear - high quality
opengl - linear_mipmap_linear - quality (as far as mipmaps are concered though this looks the same with performance and high performance too)

These are pngs and are ~300kb each.

Btw my terminology of 50% trilinear was completely homemade , however the 8500 does seem to me to be doing a mix of some trilinear and some bilinear when the texture preference slider is not at high quality.

 

[OpenGL guy]

Btw my terminology of 50% trilinear was completely homemade , however the 8500 does seem to me to be doing a mix of some trilinear and some bilinear when the texture preference slider is not at high quality.

You're correct that there is a difference, but your conclusion is incorrect.