5-Way Mesa 12.1-dev + Linux 4.7 Git Radeon Comparison

[phoronix]

Phoronix: 5-Way Mesa 12.1-dev + Linux 4.7 Git Radeon Comparison

Following the massive Windows 10 vs. Ubuntu 16.04 Graphics Performance With Radeon Software, AMDGPU-PRO, AMDGPU+RadeonSI article, I immediately started work on my next article... In preparation for a hardware launch Linux testing later this month, I started testing my collection of AMD cards on Linux 4.7 and Mesa 12.1-dev. Here are some of those results if you are curious, including performance-per-Watt metrics...

http://www.phoronix.com/scan.php?page=news_item&px=5-Way-Mesa-12.1-Linux-4.7

[M@GOid]

Michael, the R9 290 has a big performance regression in Linux 4.7. If you do a quick test with the stock 4.4 kernel you will see this too.

[nadro]

Today I tested R9 380 in Shadow of Mordor (of course with the latest drivers from Padoka PPA) with both kernel 4.6rc5 and 4.7 and I had following results:

Kernel 4.7:
Min: 12.03
Avg: 21.16
Max: 30.57

Kernel 4.6rc5:
Min: 15.75
Avg: 32.20
Max: 64.48

Settings:
Resolution: 1920x1080
V-Sync: Off
Lighting quality: High
Mesh quality: High
Motion blur: Off
Shadow quality: High
Texture filtering: High
Texture quality: Medium
Ambient occlusion: Off
Vegetation range: Medium
Depth of Field: Off
Tesselation: On

As you can see with kernel 4.7 results was much worse.

[SirCoolCat]

This is really weird, the R9 290 specifically is suffering from serious regressions for Michael, but my R9 290 is suffering from slight regressions when I use Mesa-git instead of the stable version (Running Arch Linux). The kernel doesn't seem to make a difference for me, got the same numbers with 4.4 and 4.6. Maybe I will give 4.7rc3 a try.

[M@GOid]

The problems started with Kernel 4.7, not 4.6.

[SirCoolCat]

The problems started with Kernel 4.7, not 4.6.

And yet I just ran some benchmarks on 4.7RC3 and I get the exact same performance as on 4.6 or 4.4. The regressions I am getting are when using Mesa-git (currently at 12.1) instead of the stable 11.2, and don't look as huge as the ones Michael seems to be getting. Maybe it's an Ubuntu issue?

[M@GOid]

And yet I just ran some benchmarks on 4.7RC3 and I get the exact same performance as on 4.6 or 4.4. The regressions I am getting are when using Mesa-git (currently at 12.1) instead of the stable 11.2, and don't look as huge as the ones Michael seems to be getting. Maybe it's an Ubuntu issue?

Can be. I am using Kubuntu 16.04, and getting the new kernels from here:

kernel.ubuntu.com/~kernel-ppa/mainline/

[Dreakon]

I'm seeing the same regression on my 390X, so I would imagine it effects all Hawaii/Greneda cards. I'm glad to see I'm not the only one experiencing this. Hopefully it gets fixed soon, but in the mean time I've had to dust off my Windows 10 partition to play games again.

[RavFX]

I did got some regretion on 4.7, mainly with EVE-Online on Gallium nine because it's the only game I run beside Flightgear who seam to be CPU capped. I fixed it be reverting a kernel commit :

https://bugzilla.kernel.org/show_bug.cgi?id=119631

But I have to test a nine patch to see if it fix it, did not do that yet because of lack of time but you might want to know if the regretion come from the same commit.

[bwyan]

This is really weird, the R9 290 specifically is suffering from serious regressions for Michael, but my R9 290 is suffering from slight regressions when I use Mesa-git instead of the stable version (Running Arch Linux). The kernel doesn't seem to make a difference for me, got the same numbers with 4.4 and 4.6. Maybe I will give 4.7rc3 a try.

I am experiencing something very similar to this with my R9 280X. Much worse performance with mesa git (via padoka on Ubuntu) as compared to current mesa 11.2.2 on Debian testing.