Trying Out The New OpenGL Threaded Dispatch In Mesa 17.1

[pal666]

I wonder if it's Micheal's LLVM version.

he also have different cpu

[mitch074]

I tried Borderlands, but since it seems to top up at 60 fps on my RX480, I can't really judge; it seemed a bit smoother though. Deus Ex was marginally more fluid in the hardest part of Prague, but Tomb Raider just wouldn't start.

[pal666]

But then why does radv performance tank so bad

because it is half-baked

[pal666]

I tried Borderlands, but since it seems to top up at 60 fps on my RX480

this is optimization for cpu-bound games, your gpu is irrelevant

[smitty3268]

There was no GL threading implementation at all when you tested with Oibaf ppa. 4 hour old ppa does have, so update your ppa and test again.

I don't think that's it, because the environment variable clearly made a difference in some of the tests. The negative impacts in some of those tests are expected, and they shouldn't have appeared it it was an old GL driver that didn't recognize the mesa_glthread variable.

That's why i think it was just a couple of the tests that were stripping the environment variable out.

[mitch074]

I also have an old Athlon X4, I'll test out Borderlands on it with reduced frequencies and core count.

[smitty3268]

Phoronix usually uses Padoka ppa, but it is slow to update. In this test phoronix uses Oibaf ppa that has llvm 4.0 and Oibaf ppa did not have GL threading implementation at the time of testing. I did follow the diff file of Oibaf ppa when GL threading implementation came.

Why did the performance change in some of the tests, then?

[mitch074]

There was an update to Oibaf's PPA today, and now Tomb Raider doesn't crash. However, GL threading seems to cause a 20-25% fps drop.
Hardware: 4.2GHz Haswell 4670K, 16 Gb 2400 MHz DDR3 dual channel, Radeon RX 480 8Gb (reference design) - and a SSD.

Software: fresh install of Tomb Raider for Linux, Xubuntu 16.04, mainline kernel packages for 4.11rc2, compoositing disabled. I reduced the GPU impact by setting resolution down to 800x600, details set to Low. GL threading enabled through driconf. I take my measurements at the foot of the radio tower: extreme edge, both looking forward to the drop and looking backward to the base. Taken from Gallium_hud. Game was restarted after changing the settings.
* Looking at the drop (very low complexity): no threading, 250fps - threading, 180 fps.
* Looking at the base (higher complexity) : no threading, 117 fps - threading, 80 fps.
At top settings, the difference between the two is negligible, but Tomb Raider sure doesn't like GL threading. It might interfere with Feral's own solution.

A rapid test in DE:MD (again, 800x600, lowest details) shows pretty much the same behaviour: a 20% drop in frame rate (under the benchmark) but I haven't tried loading a game in the hardest part of Prague again - last time, it seemed smoother.

Since Bioshock is one where results are expected to be better, I tried it - but the max fps was a bit lower with GL threading too. My small run is at the last autosave at the end of gthe ame, when discovering the sea of stars part II from the moment Elisabeth discovers the key and Booker opens the door - from then it's very reproducible. I have fps peaks at around 250fps (closer to 240 with GL threading enabled) and otherwise it's constantly around 130fps (120 with GL threading). However, CPU load was far more evenly distributed across my 4 cores during the whole run: with GL threading disabled I'd get loads from 15-90% jumping from core to core, enabled I'd get a very even 40-55%.

[suberimakuri]

Thanks Mitch. Interesting real results.

[geearf]

At top settings, the difference between the two is negligible, but Tomb Raider sure doesn't like GL threading. It might interfere with Feral's own solution.

A rapid test in DE:MD (again, 800x600, lowest details) shows pretty much the same behaviour: a 20% drop in frame rate (under the benchmark) but I haven't tried loading a game in the hardest part of Prague again - last time, it seemed smoother.

As far as I understand it, there should be no FPS gains but drops when using this feature with Feral games, as they already do thread their OGL calls.
Since Aspyr doesn't do that, their games should be for trying.