Linux 4.5 AMDGPU/Radeon vs. Catalyst OpenGL Performance

[P-I H]

On my Ubuntu 16.04 with custom kernel I got these results from the Unigine Valley Benchmark with a Radeon R9 380:

Unigine Valley Benchmark 1.0
FPS:41.0
Score:1717
Min FPS:14.9
Max FPS:62.6
System
Platform:    
Linux 4.5.0-rc1-custom x86_64
CPU model:AMD FX(tm)-8120 Eight-Core Processor (4026MHz) x8
GPU model:Unknown GPU (256MB) x1
Settings
Render:OpenGL
Mode:1920x1200 fullscreen
Preset Custom
Quality High
Powered by UNIGINE Engine
Unigine Corp. © 2005-2013

[Michael]

i suspect there is something else wrong with ubuntu 15.10 specifically(some in house ninja patch) or the way PTS register the FPS, i mean i have a r9-280(HD 7950) and the few times michael post his result on ubuntu my ArchLinux system trash it even though my CPU is quite weak compared to what michael usually uses([email protected][my mobo really really don't like overclock]).

so, if this is because of none of the above, i guess is related to the compilation process bypassing an optimization in mesa(there are several sse4.1+ paths) since i compile my mesa with -march=native or is related to the way PTS start storing the FPS, if you capture them too early mesa will stay at 0(or any random low number) while the upload occurs where FGLRX tends to stall at X fps while it uploads the CS(or crash depends the game).

if this is because of the latter, i guess someone can patch PTS to show the most common lowest FPS(should get rid of the uploads drops to some extent), the average(with only values > than the second most common lowest fps) and the highest FPS or add some form of delay to benchmarks the game when all the upload is done(is kinda though) or we wait until mesa get its shader cache and the upload manager gets its own thread to fix the stall

It's nothing to do with PTS, with all these games tested the results are what's exposed by the game, no calculations on the PTS side aside from checking the standard deviation and average of what's reported by the game engine.

[andrei_me]

On my Ubuntu 16.04 with custom kernel I got these results from the Unigine Valley Benchmark with a Radeon R9 380:

Michael 's test were at higher resolution, 1440p instead of 1080p

[F1esDgSdUTYpm0iy]

i suspect there is something else wrong with ubuntu 15.10

That could very well be actually. I experienced something really specific to 15.10 -- A certain tech demo (demoscene) running fine in 15.04 using Wine and in 15.10, same Wine version, same everything that very same tech demo suddenly demonstrated very odd rendering abnormalities.

Which is in part why I switched back to 15.04 and decided to hold out till 16.04. Something just seems off about the Wily Werewolf, in my experience anyhow.

[xpris]

Michael could you in future AMD FGLRX test use the same kernel as in open source driver? I mean you can in easy way patch fglrx to support the same kernel as AMDGPU - in this case kernel 4.5rc1.
This gives us clean image and certainty as to the performance.

BTW. Thanks for test

[darkbasic]

I mean you can in easy way patch fglrx to support the same kernel as AMDGPU - in this case kernel 4.5rc1.

If so please post a link to a website which always publish patches against fglrx for the very latest testing kernels.

[Michael]

If so please post a link to a website which always publish patches against fglrx for the very latest testing kernels.

Indeed... Has there been any 4.5 patches yet?

[bridgman]

As of this morning we have an in-house card reported to be running more slowly than it should in a compute rig, so hopefully that will help us figure out what is going on.

Bad news... turned out not to be clock related, just a leftover environment variable in .bashrc forcing the compute runtime to use a copy shader for data transfers instead of SDMA. With the environment variable removed the card is running at full speed again.

[jmsq]

Any particular reason for using Unigine Valley instead of Unigine Heaven? Valley is newer, but in my experience Heaven 4.0 is the more taxing benchmark. Also it may be worth testing radeonsi/mesa with and without tessellation enabled in the benchmarks that support it. I've noticed lately that without tessellation (on Heaven) amdgpu/radeonsi can keep up with catalyst, but as soon as you enable it you lose over half of your frame rate (60fps down to 20fps.) Compare that to the proprietary drivers where tessellation is a ~10fps hit.

[Kano]

@Michael

What did I say for every benchmark last year? Right: use a faster CPU!!!

It is not needed to use more than a quad core for gaming benchmarks but for Linux you need high clockspeed, preferred at 4500 MHz or more.

You want to compare GFX cards, what you see are is that fast cards are CPU limited with Linux, that's logicial for all games with on the fly HLSL to GLSL conversation (Source engine, all eON ports). If you use Windows then the CPU limit is most likely lower.

Can you test your K/X CPUs what they could reach for top speed and then retest: R9 290, R9 Fury, GTX 960, GTX 970, GTX 980 Ti. The other cards are too similar within a 5% range most likely. Best AMD and Nvidia in one chart.