ATi is ch**t**g in Filtering

mikechai said:
Quasar,

Does using directx or opengl make any difference ?
I mean, does the *optimisation* happen to both directx and opengl games ?
Click to expand...
No, one week ago, we initially reported on this issue regarding CoD, which uses OpenGL.
Also, in American McGee's Alice (if you recall that) or the OpenGL-Version of Serious Sam, a similar performance-drop can be observed (if you put enough stress on the graphics chip, that is).


991060 said:
Quasar,I just checked the first 2 shots in the article with photoshop, they're nearly identical besides the random pink light.
Click to expand...
I know, that's one problem, we where dealing with. You can either turn up gamma/contrast/brightness (i don't know sh**t about Photoshop) or use this tool from ATi: http://www.ati.com/developer/TheCompressonator.exe

Also, what we did was not comparing R360 to R420 directly but instead their respective bilinear-AF to their respective "full-tri"-AF (forced via rTool).
So you have the identical pixels in black and the different ones in white (on the ComputerBase.de Shots) - but only on the SAME Chip.
 
My point being is that saying there is a different performance drop in different architectures with AA enabled, as Jimmy's example did, isn't a particularly useful example because AA can negate the fill-rate effects of AF.
Click to expand...

It was an observation, not an example, based on actual data. So AA can negate the fill-rate effects of AF, but in what situations exactly? And why not with the NV cards? Why do we see such striking differences in performance drop off when moving from 4xAA to 4xAA/8xAF between the X800 cards and the 6800 cards? And why do we see an almost identical scenario on both UT2004 and FarCry? That's what I am trying to learn more about.

Like I said earlier, some reviewers have noted that the NV AF algorithm seems slightly sharper/clearer than ATI's AF algorithm, so this could be one explanation for the differences in performance.
 
DaveBaumann said:
in which case if you did a bit comparison of R360 and R420 with mip-map colour on then there should be no differences, however this does not appear to be the case with the test I've just run.
Click to expand...

Only in cases in which both R360 and R420 apply exactly the same trilinear filtering mehtod. Even if this requirement failed to be meet, it's also possible R420 switches off trilinear when mipmap isn't colored.
 
991060 said:
Quasar，I just checked the first 2 shots in the article with photoshop, they're nearly identical besides the random pink light.
Click to expand...

You need to do more than just set the layer Blending Options to "Difference". Trust me, I thought the same thing until I looked closer. There's a whole bunch of differences, it's just very hard to see by default.

Edit:
Merge the layer down, then go to Image->Adjustments->Threshold and move the threshold to 255. But I'm pretty worthless with PS, so that may not be valid. Gave me lots of dots though :)
 
DaveBaumann said:
Quasar said:
I can give you one A4-sized page of pure-AF Fillrate-Benchmarks with Villagemark in various resolutions, AF Levels and AF methods (bi, TS-opt. and "full-tri").

The result is sometimes much more prominent:
16xAF "forced full-tri" AF (real Full-tri on R360, bri on R420) in 1280x1024 resulted in 72fps for R360 and 110fps for R420 (same fillrate and bandwidth for both).
Click to expand...

My point being is that saying there is a different performance drop in different architectures with AA enabled, as Jimmy's example did, isn't a particularly useful example because AA can negate the fill-rate effects of AF. If we then add to that there are large architectural differences to take into account and inexcess of a 9600XT's worth of fill-rate between the two, then I don't find that example particularly useful.

However, the test you are also running doesn't necessarily point to anything because there have been changes in how the texturing is achieved from R3x0 and R420 (this was confirmed by Raja Khodri at the Technology Seminar) and you don't know how the caching and Heir-Z differences will affect things between the two.

However, my earlier point was asking whether you have done similar bit comparisons with mip-map colouring enabled. The article and 3DCenter are claiming that the driver "detects" mip colouring and turns the apparent "brilinear" filtering off - in which case if you did a bit comparison of R360 and R420 with mip-map colour on then there should be no differences, however this does not appear to be the case with the test I've just run.
Click to expand...

I think Quasar did bit-correct comparison bri and Tri in one picture. If they were the same, the pixel will be colored black.

We can clearly see 9800xt has the least same pixel, while R420 and RV360 has a lot of same pixel.
________
FERRARI 156/85 SPECIFICATIONS
 
Last edited by a moderator:
Quasar said:
I know, that's one problem, we where dealing with. You can either turn up gamma/contrast/brightness (i don't know sh**t about Photoshop) or use this tool from ATi: http://www.ati.com/developer/TheCompressonator.exe

Also, what we did was not comparing R360 to R420 directly but instead their respective bilinear-AF to their respective "full-tri"-AF (forced via rTool).
So you have the identical pixels in black and the different ones in white (on the ComputerBase.de Shots) - but only on the SAME Chip.
Click to expand...
Thanks, I already see the difference by raising contrast.

It seems 9600XT was given a more aggrasive optimization, right?
 
DaveBaumann said:
My point being is that saying there is a different performance drop in different architectures with AA enabled, as Jimmy's example did, isn't a particularly useful example because AA can negate the fill-rate effects of AF. If we then add to that there are large architectural differences to take into account and inexcess of a 9600XT's worth of fill-rate between the two, then I don't find that example particularly useful.


However, the test you are also running doesn't necessarily point to anything because there have been changes in how the texturing is achieved from R3x0 and R420 (this was confirmed by Raja Khodri at the Technology Seminar) and you don't know how the caching and Heir-Z differences will affect things between the two.
Click to expand...
We also considered that an ran separate tests with different tools (Z-overdraw and some fillrate tests - as mentioned in the article - sorry for the terrible quality most online translators make of it).
The only advantage (besides more vertexpower - which is irrelevant here) is the ability of Hyper-Z HD in R420 to use the 16x16-Stage of HierZ in higher resolutions than 1280x1024 (from those i've tested).
But since the corresponding performance drops and -levels are the same at lower resolutions, this does not seem to have any impact.

Believe me, i am aware of 4xAA and it's demands on bandwidth. But the problem is, that UT2003 is far to CPU-limited to make a reasonable difference in lower resolutions - and for the HyperZ-HD Issue i wanted to try and not rely solely on measurements in that resolution.

DaveBaumann said:
However, my earlier point was asking whether you have done similar bit comparisons with mip-map colouring enabled. The article and 3DCenter are claiming that the driver "detects" mip colouring and turns the apparent "brilinear" filtering off - in which case if you did a bit comparison of R360 and R420 with mip-map colour on then there should be no differences, however this does not appear to be the case with the test I've just run.
Click to expand...
For sure there are differences between the Chips and so are single-bit differences in many (scattered!) places. But if you do a comparison of the same chip with colored mips, then you get the same small strip of identically colored pixels as in the Tri-vs-Bi-Shot of the R360 - the way it should look.
 
Comparison - R360 and R420, trilinear filtering:

compare.jpg


Bands of differences present, most notably on the wall (in the upper half of the image).

Same comparison with colour mips on

compare_col.jpg


The banding in the comparison from the first shot appears to coincide with the mip differences between the two.
 
But if you do a comparison of the same chip with colored mips, then you get the same small strip of identically colored pixels as in the Tri-vs-Bi-Shot of the R360 - the way it should look.
Click to expand...

Sorry, take me though this - I'm not understanding what you are trying to say.
 
Same thing, i already posted on Rage3D.
__
Since our sample had to go back to the manufacturer, i can only offer you the shots we've already taken:

One and Two.

As you can see, the differences are as small as in the regular R9800 shot (mainly stemming from moving particles and clouds in the background).
 
Again, please walk me through the methodogy you used to "see" that Brilinear is evidently on with R420. I'm not trying to be ibtuse, but I'm just not getting how you see this without comparing chips. I do have the boards and I can test it.
 
Dave, can you create a little one room level in UT2003 with colored textures as walls? If this is really a cheat and being used in all games we should see the brilinear filtering in action then. Or am I off base?
 
Uhh, yeah guys... Just turn up the contrast.

The default 9800pro might have slightly higher saturation/contrast values (Ati uses a colorwheel curve). Sort of like how "digital vibrance" works albeit chunkily and clunkily on Nvidia cards.

Sharper AA? I thought the whole point of using AA was to blur the edges. I would assume the softer the AA, the better it is.
 
DaveBaumann said:
Again, please walk me through the methodogy you used to "see" that Brilinear is evidently on with R420. I'm not trying to be ibtuse, but I'm just not getting how you see this without comparing chips. I do have the boards and I can test it.
Click to expand...

Ok, i'll try my best, but please be aware that english is not my native tongue. :)

(Repeat following steps for each card individually)
- Set AF to, say 16x with rTool (or via registry) and BILINEAR
- assume individual, exactly repcroducable screenshot position ;)
- take screenshot
- Set AF to same level, but disable texture-stage optimization (called "full-tri" by some).
- assume position, dump screenshot

- produce difference-images from both screenshots and take a look at the black pixel-equivalents (depending on your method of differenciating). Those are pixels, that are exactly the same color/brightness as in the bilinear image, i.e. those, that did in no way profit from trilinear being enabled.

That's about it...
Please take into account also, that single bit-differences (i.e. black dots), if grouped and massed in broad "strips" (bri-band) like in the screenies, are an indication of brilinear being applied.
 
ZenOps said:
Uhh, yeah guys... Just turn up the contrast.

The default 9800pro might have slightly higher saturation/contrast values (Ati uses a colorwheel curve). Sort of like how "digital vibrance" works albeit chunkily and clunkily on Nvidia cards.

Sharper AA? I thought the whole point of using AA was to blur the edges. I would assume the softer the AA, the better it is.
Click to expand...

;)

RainZ
 
@quasar
you did the same thing (ie bit comparison) between bilinear/trilinear with colored mips for each chip individually as well and compared the generated images between the different chips afterwards, right?
 
You must log in or register to reply here.

Similar threads

J
Anno 1800
Replies
1
Views
2K
homerdog
homerdog
D
TSMC unveils new breakthroughs for Wafer on Wafer technology
Replies
6
Views
4K
eastmen
E
Cyan
The Mario Bros. game which made that iD Software was born
Replies
12
Views
7K
Cyan
Cyan
I
  • Locked
AMD: Speculation, Rumors, and Discussion (Archive)
256 257 258
Replies
5K
Views
931K
CSI PC
C
A
this can only happen in iceland
Replies
1
Views
1K
SimW
S
Back
Top