RadeonSI and Unity performance

[grigi]

Hi

So I'm using Mesa+LLVM Git on kernel 4.7.2 with radeon kernel module on a mobile Cape Verde (SI, 512su, 675Mhz, 1GB vram)(i7-3720QM, 16GB ram), and generally the experience is fantastic. Most games run the same as it used to on Catalyst, and some such as Shadow Warrior, Tomb Raider, The Raven and the Metro series run oh so noticeably better. :-)

But it seems some of my Unity games are noticeably slower than before.
I used to be able to run War For The Overworld at full HD and good settings smoothly, whereas with RadeonSI it is quite jerky. Even lowest settings is not very smooth.
The same can be said for Dreamfall Chapters, it is completely unplayable slow, until I disable shadows. Then the frame rate is practically smooth?

Does anyone have any hints for me to get performance better on Unity games? Any debug info I can provide?

[grigi]

Update:
I tried the Gallium hub to try and see what was going on with Dreamfall. This is relatively interesting. Dreamfall seems to allocate 1.4Gigs of vram, but since my card is 1Gig, it is swapping in anything from 30-200MB per frame, which I presume is the reason it is so jerky.
Reducing texture sizes, it then only requests 900MB, and the game isn't jerking so much anymore.
Turning shadows on/off doesn't seem to show anything obvious on the hud, but framerate drops about 25%.

I think the memory management of RadeonSI + radeon is not as good as the old Catalyst drivers.

In WFTO, the gpu utilisation was sitting about 50-70%, and cpu utilisation was at about 30% (so 1 core + a little). Considering that the GPU is weak compared to the average desktop card, but the CPU is much stronger in comparison, I wonder if there is some kind of fencing issue?
It also allocates 1.1Gig of vram, but never seems to swap memory in.

I know that AMDGPU is supposed to have better fencing/memory management, so I'm wondering if the performance will be better once I run on AMDGPU instead of radeon? Seems I may have to wait for 4.9 to test anything.

[frosth]

You can try this: http://www.gearsongallium.com

But don't expect boost, probably it will even not work for your config.

[grigi]

So you suggest I should try out AMDGPU SI support? I'm all for it, but due to deadlines coming up, I'm not going to risk running bleeding edge kernel until that is done. So I'll have a go in 3 weeks.

There is some interesting benchmarks on that gearsongallium page, So AMDGPU on SI currently sometimes provide a small benefit, and sometimes loses by a small margin. This indicates that if Performance is my aim, maybe AMDGPU isn't quite mature yet? Also apparently NINE rocks, so I'll have to try that out as well.

Yes, I love how fast everything is going with the OSS AMD stack :-) It is great fun to watch the massive improvements being applied all the time :-)

[frosth]

debianxfce do you have si gfx?, amd-staging-4.6, amd-staging-4.7, drm-next-4.9 and drm-next-4.9-wip are very close for SI support, and the last one is most update. If you have SI card and Intel gfx try it, but this setup with enabled igpu doesn't work for me.

[frosth]

grigi
this is my numbers:

http://openbenchmarking.org/result/1609197-LO-TEST0022750

I don't have any unity games so I don't know how it works, but don't expect too much.

debianxfce
What do you try to say?
CIK is supported over year, SI has working dpm since 2 weeks. It's not comparable. If I run my haswell as primary gfx and verde as secondary (like laptops do) it just doesn't work. Better use the livecd with SI support for try. it's easy and safe. And DAL has nothing to offer for SI by now.

[dungeon]

CIK is supported over year...

If not even 2 years in house, as first amdgpu bring up was on CIK

[grigi]

frosth Thanks. So performance difference is a wash right now. I'll seriously have a look at it in 3 weeks.

[frosth]

In 3 weeks it'll be 4.9-rc1 i think, for archlinux even CIK isn't enable yet, so i have to compile kernel myself. Probably you won't avoid this step

[grigi]

No, compiling manually is not the issue, I don't want to potentially lose a day of productivity before shipping the project. Like to keep my work-life low stress.
I have been running manually built kernels for close on 11 years now.