R520: 32 PS pipelines? Any truth to this?

That's simple. Taking more samples from a higher mipmap instead of just one from a lower one. Ultimately, you'd want to take all samples from the base map, since there is an information loss involved with creating mipmaps.

One example would be "fast trilinear" with 16 texels from one mipmap, with 1 texel granularity instead of the 2 texel granularity you get with mipmapping.

 

Er, I don't think so, Xmas. You'd basically just be repeating the work that MIP maps already do for you (at ~2-4x the cost).

 

But if most of the problems with texture aliasing can be taken care of by artists adjusting MIP lod properly, why would you bother? This would eat up tons of performance.

 

What I'm trying to say is that standard anisotropic texture filtering for color textures just isn't going to change (much).

What will be important for filtering in the future is finding elegant ways to deal with other types of data. An example of what I'm talking about is percentage-closer filtering for shadow maps: it's similar to texture filtering, but operates on different data.

 

ok....33 pages.... ANYTHING USEFUL?

 

Hush you, speculation threads the goal is the journey and not any specific destination.

 

It's better than normal trilinear filtering, and it's not that much more expensive. In fact, it requires less bandwidth, and there are interesting optimization possibilities when using S3TC-like texture compression. It's not like "fast trilinear" hasn't been done before (though without 1-texel granularity AFAIK).

 

If you're only sampling from one MIP map, and you want to have at least the same quality as trilinear filtering, you'll need to take 16 samples. That's quite a bit more memory bandwidth that's going to be eaten up, not to mention more averaging operations.

 

I think somewhere burried in here is the "It's going to be red." bit.

 

More cache bandwidth, but less memory bandwidth. Yes, it requires more averaging operations, but with a bigger filter kernel you can accelerate anisotropic filtering as well.

 

It's going to have some new video features

 

1. If you're doing a 4x4 filter kernel, why not just go bicubic?
2. The purpose of trilinear filtering is to remove the appearance of MIP map boundaries. This won't get rid of them.
3. Though there's a lot of shared data between pixels, I doubt it'd be more efficient than trilinear filtering. After all, trilinear filtering shouldn't average more than about 1.25-1.33 texels per pixel.

 

1. Because of 2.
2. Yes it does. Because it's like generating mipmaps on the fly. But more precise, if you do it with 1-texel granularity.
3. Normal trilinear filtering uses data from two mipmaps. It uses all texels usually, without gap. Fast trilinear uses data from the higher res mipmap only. So it requires less memory bandwidth.

 

Then the filter has to be qualitatively different from bilinear near the mipmap boundaries. If you have two mipmaps, N and N+1, and N+1 gets closer to bilinear the closer you get to the mipmap boundary, doing perfect bilinear at the mipmap boundary iself, then when you swap mipmap, the 2x2 box that was used/approximated in mipmap N+1 projects into a 4x4-texel block with 2-texel granularity in mipmap N.

If you go the other way, trying to project a 1-texel granularity 4x4-texel block from mipmap N into mipmap N+1, you may end up not covering a nice texel-aligned 2x2 block, but instead partially covering texels in a 3x3 block. It could perhaps be made to work, but the mathematics get considerably more complex than what 'fast trilinear' would suggest and I am not convinced that it would actually look better.

Also: the traditional 2-texel granularity for fast trilinear requires the texture cache to have ~4 banks per pipeline; 1-texel granularity would increase that number to ~16, which can be quite expensive.

 

unlike any other ATi board so far.....

 

Maybe it will have the special feature of being purple!

So back on topic, how many pipelines will it have?
// runs...

 

I'm sure it'll have at least 1 pipeline - i just can't go wrong with that

 

Kaotik, I agree 100%

 

Maybe they do something so revolutionary new that we can't even call it pipelines anymore. That would surely mean that it hasn't got any pipelines.

 

Or, that it has pixel shading pipelines that are split into ALU pipes and TMU pipes, similar to the arrangement of Xenos - but not unifying vertex and pixel as Xenos does.

Jawed