RadeonSI Polaris: Mesa 12.0 vs. 13.0 vs. 17.0 vs. 17.1 Git

[dungeon]

There is definitely a regression with Deux Ex. I using Padoka's PPA and, while me result is a little better, it's worse than before.

And Euro Truck Simulator 2 is running like crap with the lastest Mesa. Using Ubuntu 17.04 default (Mesa 17 and kernel 4.10) the game stutter and have slowdowns like there is no tomorrow. Only using Kernel 4.11RC6 and Mesa 17.1 with the gl_thread thingy enabled, things become tolerable. Also, I got a used GTX460 from a friend and I discovered that it get slightly better fps than my RX470 in this game using the Nvidia proprietary driver. Or the game is heavily optimized for the Nvidia driver, or radeonsi is not playing well with this game.

Euro Truck Simulator 2 is CPU bound, AMDGPU-PRO runs threaded profile on it and you say gl_thread helps with mesa, means that when CPU boundware came things became total weird.

Taking into account that we have weirdness with various CPU vendors on different game titles, it make sense maybe to say which CPU you use?

Maybe shader cache new feature in mesa introduce some stutter here and there... just i guess.

[oooverclocker]

Well I'm happy that these benchmarks sometimes show such issues. When it can be ensured that these results are reproducible, perhaps with another system, it looks like it's time to take care of these regressions before moving on with too many new implementations. The inconsistent results between the versions are very unfortunate in the last months of continuous improvements and hopefully the best performing state can be secured for future releases.

[M@GOid]

Euro Truck Simulator 2 is CPU bound, AMDGPU-PRO runs threaded profile on it and you say gl_thread helps with mesa, means that when CPU boundware came things became total weird.

Taking into account that we have weirdness with various CPU vendors on different game titles, it make sense maybe to say which CPU you use?

Maybe shader cache new feature in mesa introduce some stutter here and there... just i guess.

It's a i7 3770k running at 3,9Ghz. The other system with a GTX 460 (with 1GB of VRAM), is a i5 3550k at stock speeds, and it is in a crippled mobo with only one 4GB stick, so no dual channel, using Ubuntu 16.04 using the Nvidia proprietary driver from Ubuntu official repo. I got shocked of how well it handled the game, compared with my primary system with a better CPU and much better GPU. The only advantage the RX470 have is to crank the scale setting to 400% (without loss of FPS) to get decent AA, since checking the normal AA setting only blurries the image a little without geting rid of the aliasing.

The game do not get better using the Performance CPU governor, and I have to manually disable desktop compositing (since it do not trigger KDE switch for that) to truly disable Vsync to help with the slowdowns.

[ramrod]

Is the Deus Ex regression because of Mesa or Kernel 4.10, because the results seems to match with these results. https://www.phoronix.com/scan.php?pa...RX470-DRM-Next

Maybe try with kernel 4.11 to see if the regression goes away.

[marek]

For people who get a lot of annoying frame rate hiccups, it's been mostly fixed. The fix will appear in kernel 4.12, but for now you can get it from amd-staging-4.9 from Alex's repo, or drm-next from airlied's repo.