Tessellation Shaders Land For RadeonSI NIR Backend

[phoronix]

Phoronix: Tessellation Shaders Land For RadeonSI NIR Backend

The work led by Valve Linux driver developer Timothy Arceri on adding tessellation shader support to RadeonSI's NIR code-path has been merged to Mesa 17.4-dev Git...

http://www.phoronix.com/scan.php?page=news_item&px=RadeonSI-Tessellation-NIR-Lands

[Azpegath]

Awesome! Really looking forward to this eventually aligning their development process on one IR, hopefully allowing more development time going into optimizations.

[xxmitsu]

Has anyone tested if there are any performance differences using RadeonSI with NIR after this code landed ? Especially using different tessellation levels ?

[Azpegath]

Has anyone tested if there are any performance differences using RadeonSI with NIR after this code landed ? Especially using different tessellation levels ?

I don't think we should have too high expectations at this point. First get it working (with a decent design), then make it fast.

[xxmitsu]

I don't think we should have too high expectations at this point. First get it working (with a decent design), then make it fast.

Yes, I agree. I was just being curious if NIR OPT Passes are influencing in any way.

[pal666]

Awesome! Really looking forward to this eventually aligning their development process on one IR, hopefully allowing more development time going into optimizations.

their optimization is done by llvm

[nuetzel]

Has anyone tested if there are any performance differences using RadeonSI with NIR after this code landed ? Especially using different tessellation levels ?

Read the code (or mesa-devel mailing list)...
@Michael: ...and it wasn't even send only one week ago:

https://lists.freedesktop.org/archives/mesa-dev/2017-December/179716.html

https://lists.freedesktop.org/archives/mesa-dev/2017-December/180188.html

Yes, I did for Timothy.
UH is _faster_ on Polaris 20 (RX580). Nearly 10 fps (first scene) on my test system.

[xxmitsu]

UH is _faster_ on Polaris 20 (RX580). Nearly 10 fps (first scene) on my test system.

That's just awesome. It helps standardize a common cone path that will get more developers attention, and even performance increase! Looking forward for the progress that will be made in 2018.

[nuetzel]

Originally posted by nuetzel View Post
UH is _faster_ on Polaris 20 (RX580). Nearly 10 fps (first scene) on my test system.

That's just awesome. It helps standardize a common cone path that will get more developers attention, and even performance increase! Looking forward for the progress that will be made in 2018.

I have to take this back.
Now, both paths are mostly _on par_. - Which is good, at this stage, too.
I'm running it (NIR) all day.

[xxmitsu]

I have to take this back.
Now, both paths are mostly _on par_. - Which is good, at this stage, too.
I'm running it (NIR) all day.

I wish they would have feature parity (as far as I remember bindless was one of the missing features) and then switch on default NIR so that it would get more attention.