Interesting AA quote from SIGGRAPH paper

[Fred]

on Slide 13 I believe from the future outlook paper on the dev section, Nvidia mentions use of the partial DERIV ability in there future part, and that its useful for AA.

Now, isn't this an indication that they might try something a long the lines of Z3's approach.

If I remember correctly, Z3 required Zdx, and Zdy for both the occlusion check, and for there fragment compression algorithm.

Something along the lines of storing Z grad and the center referenced Z value. That way you can store small Z grad, and approximate what the actual Z value at the specific sample point is without actually storing the full Z and color values.

 

[KimB]

I think this may be more useful for custom texture filtering algorithms.

The new PS in the NV30 allow you to read the DDX and DDY values.

 

[Reverend]

This is interesting.

Here's a question : Does anyone think someone like David Kirk wouldn't know about Z3?

 

[DemoCoder]

The partial derivatives are used for doing AA in procedural texturing. They are not used for edge AA.

 

[Grue]

The partial derivatives are used for doing AA in procedural texturing. They are not used for edge AA.

In fact, it's sort of hard to imagine how you'd do decent procedural texturing without the derivative instructions. Unless you know at which frequency to cut off, there will potentially be a great deal of shimmering (think software 3D engines without mipmaps) in procedural texturing. Aliasing is a huge concern for writers of procedural shaders for off-line rendered graphics. The addition of the derivative instructions to pixel shaders is a big step toward bringing real-time graphics up to the level of film CG.

--Grue

 

[echo]

what kind of q is that?

This is interesting.

Here's a question : Does anyone think someone like David Kirk wouldn't know about Z3?

 

[RussSchultz]

A rhetorical one?

 

[LittlePenny]

I don't see why you couldn't use it with Green's theorem. It would make surface detection on a per-pixel basis nice and easy.

 

[psurge]

Well... I am sure the people at any of the 3d chip companies know about Z3 and other AA solutions which are superior to MSAA/SSAA. I am thinking however that either these types of AA are very hard/tricky to implement, require too many transistors, or are considered overkill for the current market...

 

[Fuz]

Would some one please be kind enough to post a link to a site where I might learn more about Z3?

Thanks in advance.

Fuz

 

[Ailuros]

http://www.research.compaq.com/wrl/people/jouppi/Z3/Z3paper.pdf

 

[Reverend]

Because I was told that Kirk doesn't know wtf Z3 is, which is unbelievable to me.

Someone please state the hidden hint.

 

[Ailuros]

I can't figure any hidden hint in that. I can't connect him in any way to SA either for what it's worth.

 

[MfA]

It would be slightly dissapointing if true, still that doesnt say if the person who asked a question about it could give a good description of the method to jar his memory ... or was even given the oppurtunity to do so. He might be a busy man, but I assume the chief scientists of 3D graphics companies can be bothered to at least skim the papers from the few graphics conferences each year ... maybe he just filed it in memory as "one more modified A-buffer anti-aliasing method".

 

[Kristof]

Another option would be : is Z3 patented ? If NVIDIA ends up designing something very close to Z3 but claim they did not know about its existance, or used it as a base, they might be able to duck some patent issues... I believe patent law is more severe if something was really knowingly ripped off versus when you can claim no knowledge of the patent/technique.

K~

 

[MfA]

<A HREF=http://www.cs.unc.edu/~chang/>Apparently</A> there are patents for it, but Compaq would be more than happy to license them at a reasonable fee I assume (otherwise its just dead paper).

 

[KimB]

Very interesting...

I believe I read that the R300 uses "effectively 96-bit" color instead of 128-bit, meaning 24-bit floating-point pipelines.

If this is also true of the NV30, then it seems conceivable that either architecture could make use of Z3 or something of the sort with the remaining 32 bits of data.

 

[psurge]

Rev, you mean probably no Z3 or similar in nv30... i guess we'll see soon enough.

Chalnoth - Z3 requires a fixed number of samples each having a color value, z value, z gradients, and a coverage mask (and i think also stencil and alpha values). OTOH they get really good results with just 3 samples per pixel.

 

[Dan G]

Very interesting...

I believe I read that the R300 uses "effectively 96-bit" color instead of 128-bit, meaning 24-bit floating-point pipelines.

If this is also true of the NV30, then it seems conceivable that either architecture could make use of Z3 or something of the sort with the remaining 32 bits of data.

Alternatively, it could just be mangled reference to the fact that the alpha component isn't a color attribute? 96-bit color is an accurate description of RGB at 32-bits per channel.

 

[KimB]

Alternatively, it could just be mangled reference to the fact that the alpha component isn't a color attribute? 96-bit color is an accurate description of RGB at 32-bits per channel.

Got it from OpenGL guy:
http://www.beyond3d.com/forum/viewtopic.php?p=24231#24231