Mesa 22.3 Lands Open-Source Radeon Driver CPU Overhead Optimizations

[phoronix]

Phoronix: Mesa 22.3 Lands Open-Source Radeon Driver CPU Overhead Optimizations

There have been some notable merges this week for Mesa 22.3, particularly as it concerns the open-source Radeon graphics driver code...

https://www.phoronix.com/news/Mesa-2...Lower-Overhead

[V1tol]

This MR applies cleanly over 22.2 release

EDIT: Second PR mentioned below applies cleanly too.

[Venemo]

The article is not entirely accurate. Actually these two are merged now:

• https://gitlab.freedesktop.org/mesa/...requests/18808
• https://gitlab.freedesktop.org/mesa/...requests/18807

And several others are in progress as they are cleaned up and tested individually.

[Venemo]

Also, please note that the numbers we advertise these patches with are measured using a Vulkan overhead micro benchmark and it's unclear how they will translate to real world game perf. Furthermore,

• At least a few of those patches break the driver's correctness - meaning they cause bugs in some games
• Some patches cause GPU hangs

I'm working on fixing them up, it's not yet straightforward and it's unclear how the performance will change when they are fixed.

[rhysperry111]

Oh shit... the other open-source-nerd Rhys Perry is in the news again!

I guess I gotta step up my game :P

[Venemo]

Update: A +23% draw improvement from the overhead work as part of SGPR emission has also been merged.

Note this 23% is measured by only 1 user, and on my own system I measured 0.03% for that patch.

[Linuxxx]

Note this 23% is measured by only 1 user, and on my own system I measured 0.03% for that patch.

AFAIR, you are using the schedutil governor, which seems to be a trend among Valve.

Try that same measurement with the performance governor, please!

[V1tol]

AFAIR, you are using the schedutil governor, which seems to be a trend among Valve.

Try that same measurement with the performance governor, please!

Also this change may influence discrete GPUs the most. R7 6850U used in PR maybe was tested with iGPU hence no change there.

[Venemo]

AFAIR, you are using the schedutil governor, which seems to be a trend among Valve.

Try that same measurement with the performance governor, please!

You can read the measurement methodology on the merge requests linked above. I measured on a Thinkpad Z13 with platform profile set to "performance" and CPU governor also set to "performance", turbo boost disabled. I hope this is what you expected.

[Linuxxx]

You can read the measurement methodology on the merge requests linked above. I measured on a Thinkpad Z13 with platform profile set to "performance" and CPU governor also set to "performance", turbo boost disabled. I hope this is what you expected.

Sorry, didn't know You were already testing that way...

It's just that I remembered You once stating that You saw no difference between the schedutil & performance governors, whereas countless benchmarks show that there can be quite a significant difference between the two.

That is also why I see no point in the Steam Deck shipping with schedutil by default, especially since SteamOS allows setting a desired TDP by the user, in which case the performance governor would never cause excessive heat or power-draw, either way.

Hopefully one day we will get an answer to that question...

Anyway, keep up the awesome work You are doing for all us!