Much Faster Tessellation Is Coming For AMD's RadeonSI Gallium3D

[phoronix]

Phoronix: Much Faster Tessellation Is Coming For AMD's RadeonSI Gallium3D

Bas Nieuwenhuizen has announced his work on "offchip tessellation" support for the RadeonSI Gallium3D driver to provide dramatically better OpenGL tessellation performance...

http://www.phoronix.com/scan.php?page=news_item&px=RadeonSI-Offchip-Tess

[Adarion]

Remarkable! This is definitely palpable for the users outside of pure benchmarks.

[haagch]

So I tried unigine-heaven with a 4.5 kernel and apparently with OpenGL 4.2 exposed it too doesn't work on SI:
Mai 10 14:46:52 c-l kernel: [drm:si_ib_parse [radeon]] *ERROR* Invalid GFX packet3: 0x79

*reboots to drm-next-4.7-wip*
*sigh*

Edit: Still doesn't work with drm-next-4.7-wip. Here's the start of it: https://gist.github.com/ChristophHaa...20b886ebf33c95
Edit: Looked into the kernel sources. drivers/gpu/drm/radeon/cikd.h has
#define PACKET3_SET_UCONFIG_REG 0x79
but drivers/gpu/drm/radeon/sid.h
is missing this one.

[schmidtbag]

That's pretty significant. Makes me wonder what else has room for improvement (or if tessellation can be improved further). The open source drivers aren't that far behind from the closed-source Windows drivers. The open source drivers (as-is) will not feature application-specific optimizations. It would be really interesting to see if Mesa could ever come out on top.

[BNieuwenhuizen]

So I tried unigine-heaven with a 4.5 kernel and apparently with OpenGL 4.2 exposed it too doesn't work on SI:
Mai 10 14:46:52 c-l kernel: [drm:si_ib_parse [radeon]] *ERROR* Invalid GFX packet3: 0x79

I think I missed a register that moved between SI and CIK. I have a patch that might fix the errors. You can try replacing the first patch in the series with https://basnieuwenhuizen.nl/si-tess-1.patch or use the si-offchip-tess-v2 branch of https://github.com/BNieuwenhuizen/mesa .

[haagch]

I think I missed a register that moved between SI and CIK. I have a patch that might fix the errors. You can try replacing the first patch in the series with https://basnieuwenhuizen.nl/si-tess-1.patch or use the si-offchip-tess-v2 branch of https://github.com/BNieuwenhuizen/mesa .

That works.
Doesn't help too much for shadow of mordor (lowest settings) https://youtu.be/_ZsMPaqUIgY, but I think it does help a bit.
Thank you very much!

[BNieuwenhuizen]

Doesn't help too much for shadow of mordor (lowest settings) https://youtu.be/_ZsMPaqUIgY, but I think it does help a bit.

The lowest setting doesn't enable tessellation. You can check tessellation in the advanced options menu. The number from the post is based on very high, with texture quality medium and resolution of 1920x1080 (also the internal one in video options) and mesa exposing GL4.3 with commandline overrides.

[haagch]

The lowest setting doesn't enable tessellation.

Oh, didn't know that. Then tessellation isn't the bottleneck in the first place. Average increase of 5 fps compared to last time I ran it must be other improvements then.

Well, increased performance is nice, but unfortunately not much use when games that use it are hardly playable while having something like 30-40% GPU usage.

[tajjada]

mesa exposing GL4.3 with commandline overrides.

How do you do this? And why isn't Mesa currently exposing GL4.3 by default anyway? All the functionality needed for 4.3 seems to be implemented IIRC.

Could someone explain the reasoning to me? I was confused when I upgraded to latest mesa+llvm git, that glxinfo still showed 4.2 but all extensions were present.

[illwieckz]

Michael, you Heaven screenshot for this tesselation-related article has tesselation disabled…

Your screenshot:


Almost the same with tesselation:

and without tesselation:


(taken on radeonsi but without the update yet, you can notice it with the fps numbers top-right)