AMD: R8xx Speculation

[rpg.314]

Waiting patiently for dx11 gpu's

 

[Jawed]

So we now have a hard launch date for W7, 22 October, so erm, fingers-crossed...

Jawed

 

[Jawed]

AMD demonstrates tessellation on the worlds first DirectX 11 graphics processor

http://www.youtube.com/watch?v=Cw6OO4pTrkI

Jawed

 

[Silent_Buddha]

Performance seems to rapidly drop off a cliff after about 15 or so patch divisions. It would have been interesting to see the triangle count as they increased the patch divisions.

Wonder if that's final silicon or if they'll have to do any respins.

CatalystMaker also seems quite excited in their product.

Regards,
SB

 

[nAo]

That's ok, it pushes several milion subpixel sized triangles per frame, setup limited for sure.

 

[rpg.314]

http://www.youtube.com/watch?v=Cw6OO4pTrkI

Jawed

Wow, so they indeed have working silicon.

 

[Jawed]

The hardware is described as "EG BROADWAY". Evergreen Broadway?

Jawed

 

[rpg.314]

That's ok, it pushes several milion subpixel sized triangles per frame, setup limited for sure.

Question, since the tesselated geometry is generated by the hw, shouldn't primitive assembly of the intra-patch triangles (atleast) be integrated with tesselator? IOW, atleast for those intra patch triangles, setup should be nearly free right?

 

[Silent_Buddha]

That's ok, it pushes several milion subpixel sized triangles per frame, setup limited for sure.

Yup that's about what I figured.

Bah, now that we know for sure the hardware exists, I want performance numbers and information darnit. Having to wait however many months for a hard launch sucks.

Regards,
SB

 

[nAo]

While you can pre-cull patches once they are tesselated the setup unit(s) will take take care of them. Now we really don't know anything about this new GPU..so I guess we have to wait and see.

 

[Jawed]

Could you do a back-facing test and a clipping test in the HS for each patch? A conservative, crude, early cull?

I've noticed on page 19:

http://developer.download.nvidia.com/presentations/2009/GDC/GDC09_D3D11Tessellation.pdf

that the Domain Shader needs to interpolate attributes. This is normally done after setup. It implies a fair amount of redundancy, i.e. coarse interpolation is perfomed (or even sub-pixel interpolation, I guess, with a high tessellation factor) once during DS and then again later after GS while pixel shading.

Jawed

 

[nAo]

Could you do a back-facing test and a clipping test in the HS for each patch? A conservative, crude, early cull?

I've noticed on page 19:

http://developer.download.nvidia.com/presentations/2009/GDC/GDC09_D3D11Tessellation.pdf

that the Domain Shader needs to interpolate attributes. This is normally done after setup. It implies a fair amount of redundancy, i.e. coarse interpolation is perfomed (or even sub-pixel interpolation, I guess, with a high tessellation factor) once during DS and then again later after GS while pixel shading.

Jawed

Per pixel attribute interpolation is done after the setup. The domain shader interpolates attributes on a per primitive basis, before primitive setup.

And yes, in theory one could pre-cull triangles, in practice it doesn't work in the absolutely general case as you are given total freedom, that's why the hw won't do it for you. Hell..you can even endup with cracks with DX11 tesselation if your shaders are not written properly.

 

[LordEC911]

Yup that's about what I figured.

Bah, now that we know for sure the hardware exists, I want performance numbers and information darnit. Having to wait however many months for a hard launch sucks.

Regards,
SB

Not too long, ~2.5-3.5months before we start getting some decent info, if the Oct launch date is correct.

 

[Blazkowicz]

Performance seems to rapidly drop off a cliff after about 15 or so patch divisions. It would have been interesting to see the triangle count as they increased the patch divisions.

Wonder if that's final silicon or if they'll have to do any respins.

CatalystMaker also seems quite excited in their product.

Regards,
SB

I have trouble making out the visual difference between 6 and 19..
but, nice tech demo.

I hope we get a nice looking demo. I don't remember anything since Ruby and 3dmark 05/06.

 

[Jawed]

And yes, in theory one could pre-cull triangles, in practice it doesn't work in the absolutely general case as you are given total freedom, that's why the hw won't do it for you. Hell..you can even endup with cracks with DX11 tesselation if your shaders are not written properly.

Hmm, now I'm thinking that rather than trying to cull patches in HS - which might not be possible, 1-in-1-out - you'd set the tessellation factors to 1.

Just realised this presentation touches on silhouette edge enhancement, which operates on the boundary patches between back-facing and front-facing patches.

I'm not clear on how adjacent patches marry their shared-edge vertices - the cracks problem. I suppose any "culled patch", one that's kept at a TS factor of 1, will have edges that can't show cracks - otherwise it wouldn't be a "conservatively culled patch". So those cracks, if they exist, can't show.

Jawed

 

[LordEC911]

I have trouble making out the visual difference between 6 and 19..
but, nice tech demo.

I hope we get a nice looking demo. I don't remember anything since Ruby and 3dmark 05/06.

White-out, Frogs and Dog chasing Ruby, that is all I remember off the top of my head.

 

[nAo]

Hmm, now I'm thinking that rather than trying to cull patches in HS - which might not be possible, 1-in-1-out - you'd set the tessellation factors to 1.

it is possible to cull patches. Not that it would have make any difference with respect to the AMD video.

I'm not clear on how adjacent patches marry their shared-edge vertices - the cracks problem. I suppose any "culled patch", one that's kept at a TS factor of 1, will have edges that can't show cracks - otherwise it wouldn't be a "conservatively culled patch". So those cracks, if they exist, can't show.

Jawed

You can have cracks even with shared edges that share the same TS factor, as you might end up evaluating them in 2 different patches working on vertices but in a different order (gotta love floating point math )

 

[Kaotik]

Not too long, ~2.5-3.5months before we start getting some decent info, if the Oct launch date is correct.

Where did this october date pop up, after several reports of first being pushed from Q4 to Q3 and then to early Q3, it jumped as far as 2010 and now october?

 

[Jawed]

it is possible to cull patches. Not that it would have make any difference with respect to the AMD video.

Is that test application part of the D3D11 SDK?

You can have cracks even with shared edges that share the same TS factor, as you might end up evaluating them in 2 different patches working on vertices but in a different order (gotta love floating point math )

Hmm, would high precision in TS defuse this problem (making fixed-function TS a good solution)? Where's that mismatch actually introduced?

I'm so glad this is someone else's problem

Jawed

 

[Jawed]

The tessellation video has gone "private".

Jawed