I'd like to see ATi provide some examples of in-game scenes in which the optimizing algorithm switches off to allow full trilinear, as it does in the case of these colored mipmap tests. Such examples would prove that the algorithm was not specifically intended to islolate colored mipmaps as a special case, and would reinforce its nature as a legitimate optimization.
So, Any Questions You'd like to ask ATI about Filtering?
- Thread starter Dave Baumann
- Start date
ARghhh ... what about the most important one: is there any guaranteed way for an application, using either OGL or DirectX# or whatever, to actually get full trilinear filtering on newer ATI chips on all texture stages ? Same for anisotropy.DaveBaumann said:
The question is especially relevant when im doing something slightly out of the ordinary, ie my 1st texture is not the base ( decal ) texture.
EDIT: oh, and if not, is it a hardware limitation or driver limitation ? We know its not the API ...
PurplePigeon
Newcomer
My questions, if possible to ask:
1) Under what circumstances would their algorithm provide the most noticeable visual discrepancies?
2) What kind of testing did they do to determine whether visual quality was maintained?
1) Under what circumstances would their algorithm provide the most noticeable visual discrepancies?
2) What kind of testing did they do to determine whether visual quality was maintained?
Did ATI do this because they feared that their new line of cards couldn't hang with Nvidias 6800 without it in high quality IQ tests? Did they not care that this would tarnish their "Pure IQ" reputation?
As a x800 Pro owner, I think they're fair questions. (Great card BTW and IQ is subjectively better than my 9800 Pro)
No, I don't expect an answer to my first question, heck maybe it's so obvious that it answers itself.
As a x800 Pro owner, I think they're fair questions. (Great card BTW and IQ is subjectively better than my 9800 Pro)
No, I don't expect an answer to my first question, heck maybe it's so obvious that it answers itself.
Your statement on this issue says that you "perform considerable on-line and off-line image analysis before implementing changes." Please describe this in greater detail (i.e. show it rather than assert it). Did you develop any in-house tools for the purpose of testing this? If so, can you make them available? Did you put together any internal whitepapers to explain to management what you were proposing to do and showing image quality comparisons between the old and new --if so, please provide from those documents the sample images used to explain and show this internally.
It's probably too late, but I wanted to ask this:
Is the adaptive tri/brilinear filtering done on a per-texture basis, or does whatever is done have to be done on the whole scene?
If whole scene, what fraction of textures in a scene have to have "different" MIP maps for bri-optimization to proceed? Just one?
Thanks,
ERK
Is the adaptive tri/brilinear filtering done on a per-texture basis, or does whatever is done have to be done on the whole scene?
If whole scene, what fraction of textures in a scene have to have "different" MIP maps for bri-optimization to proceed? Just one?
Thanks,
ERK
ERK said:
How would you restrict the filtering selection to be per-scene without defering rendering of all primitives? Programmers can change the filtering method on a per-primitive basis via a state change, and there is no conceptual difference between a driver-induced change and a program-based change, especially when considering that the programmer specified filtering selection is sometimes only used as a hint, so I don't think it is very likely that the filter selection is performed on a per-scene basis.
You could do things on a per-scene basis by using data from one scene to select the filter mode for the next scene, but you can't easily predict what textures will be used in the next scene without potentially severe worst-case scenarios. I doubt this would require less work than doing things on a per-texture basis and the flaws are too large to ignore.
Maybe to late but questions I have:
I am all for smart optimizations, this may be one of them that ATI deserves to get some credit for.
- How is this filtering different from Brilinear as is what people call NV's method?
- How much fillrate can be saved from doing this method in high anisotropic filtering conditions?
- Is the filter algorithm programmable and can be varried by the drivers?
- Is there a registry setting that can modify the adaptive filtering?
- Does this new type of filtering just affect zones or zone width for the mipmaps but still always does blending between the two mipmaps (Trilinear)?
- How many different transition occure (from similar mipmap textures to colored mipmap textures) does this filtering have or is it only two steps: Full Trilinear or Optimized only? Can it have more if only two steps at the moment?
I am all for smart optimizations, this may be one of them that ATI deserves to get some credit for.
@Riff-
I sense you may be reading too much complexity into my question; but I think you answered it. If the filtering can be done per texture, then it seems that's the best way to do it.
On a related note, I had wondered if a filter-tester program (look down a tunnel with colored MIPmaps, etc.) could be used to evaluate the dynamic filtering if it contained a mixture of colored and standard textures. The idea being that the presence of the colored MIPs would force full tri- on the standard MIPs (or vice-versa). But if filtering is done on a per-texture basis, such a test would likely be useless.
Thx,
ERK
I sense you may be reading too much complexity into my question; but I think you answered it. If the filtering can be done per texture, then it seems that's the best way to do it.
On a related note, I had wondered if a filter-tester program (look down a tunnel with colored MIPmaps, etc.) could be used to evaluate the dynamic filtering if it contained a mixture of colored and standard textures. The idea being that the presence of the colored MIPs would force full tri- on the standard MIPs (or vice-versa). But if filtering is done on a per-texture basis, such a test would likely be useless.
Thx,
ERK
g__day said:
I asked something in another thread.
WE have seen in the past (radeon 9500+ and geforce fxs) differing from refrence rastner images .
We now have this that when subtracted from the 9800xt image quality, which differs from the refrence rastner , it is unacceptable.
Why is one acceptable and the other is not ?
Is there a % diffrence that we should use as a good rule of thumb ?
(To ATI and to B3D community) What is the performance penalty of chosing ATIs adaptive trilinear filtering over full "bilinear" filtering?
At least this can be measured by conventional benchmarking tools.
If it adds something to the image quality there should be a performance cost dont you think? Or is this ATI's previously speculated lossless trilinear filtering.
At least this can be measured by conventional benchmarking tools.
If it adds something to the image quality there should be a performance cost dont you think? Or is this ATI's previously speculated lossless trilinear filtering.
digitalwanderer said:Hey, I just want 'em to start returning me e-mails again!
Seriously, it's starting to get me worried. I haven't gotten the silence treatment out of ATi in years....literally!
They probably just have them go straight to the trash and thus never even see them. Perhaps they forgot they even had the DW filter on
*bump*
What happened to these questions ? Did those ever get answered ?
EDIT: i'd still very much like to see this questions answered by both major IHVs: when they implement NEW filtering algorithms like NV "brilinear" and ATI's "trylinear" or whatever, why arent those exposed in API's in the usual fashion, like it happens with other newly implemented graphics algorithms, like different compression methods etc.
What happened to these questions ? Did those ever get answered ?
EDIT: i'd still very much like to see this questions answered by both major IHVs: when they implement NEW filtering algorithms like NV "brilinear" and ATI's "trylinear" or whatever, why arent those exposed in API's in the usual fashion, like it happens with other newly implemented graphics algorithms, like different compression methods etc.
I don’t know is it is to late but could you ask ATI to provide an example of a situation where IQ is improved by this new method. It has been mentioned, by ATI, that there are cases were there is an improvement to IQ.
ATI's David Nalasco said:
Malfunction
Newcomer
nelg said:
Hehe... ya, it's been real quite since the evidence came out which showed the differences. Infact, I've been passing by here for awhile and haven't seen any excitement or any enthusiasm since I last posted.
It's almost like ATi had the wind taken from their sails. Been real dead (imho) in regards to the topics of the latest products released. Hell, I don't even know how people feel about the latest Catalyst drivers. It's been an eary dead calm floating from site to site lately.
Been more enthusiasm for nVidia's NV4x line as far as I can see, kinda sad really...
Ahh well, such as product launches go I guess... :?
