Benchmarking The New RadeonSI/Gallium3D Threaded Support

[grigi]

I used the GALLIUM Hud to see if I could find obvious issues. I have discounted the following:
Memory usage is under vram limit, so no memory thrashing.
tc-num-syncs is steady at 6, and setting the env var of GALLIUM_THREAD=0 also made no difference.
GPU-load is steady at 100.

So at least the threaded gallium isn't breaking it.

I also forgot to say I'm using the radeon kernel module (since the gpu is a CAPE VERDE), and on kernel 4.10.8

[grigi]

marek I tried the latest Mesa Git, and noticed a massive performance drop in Shadow of Mordor.
This is my specs:
Gentoo Linux
LLVM 3.9.1-r1 (because when I tried LLVM 4.0 I had tons of regressions at the time)
Mesa Git (0ca5bdb330d6b928c1320e5829906b195bd2c4b8)
Radeon HD M7770 1GB GDDR5
Intel i7-3720QM (2.6GHz - 3.4GHz)

My comparison was a git build just after the 17.1 split (don't have the hash, sorry)
I used to get 30fps average, now I get 17fps average.

I also dropped the texture size from medium to small in case the textures were not fitting anymore, but no change in result.

The only change was a Mesa git rebuild between the two tests.

What info do you need from me to try and find the performance issue?

[Ansla]

Michael +1 for testing with a CPU with lower IPC. Ideally a Kavery APU with its integrated GPU plus also a Fury or RX480/RX580. And the tests performed at 1080p.

[Marc.2377]

Great work, Marek.

I'm curious what the results for the Bulldozer/Piledriver architectures would look like.

[spstarr]

I don't think there would be any difference, even Arch uses 4.0, they don't do it for no reason i guess .

LLVM 5.0 has some improvements to the sisched scheduler, but if he isn't using that the default gcn scheduler also gained improvements in master unless backported to 4.0(?) I don't think Michael uses sisched aka (R600_DEBUG=sisched)

[LeJimster]

BTW, I don't want to seem ungrateful for the benchmarking Michael does. I actually look forward to it. Especially when i hear about performance patches. Its that I don't understand the testing methodology at times.

[leipero]

Fix:
ArchLinux have 3 ways of working

1.) When you use only the default repositories (Core / Extra / Community) you have Archlinux in stable mode, is like having an rolling debian because only stable software is shown to you and it takes some test to ensure all is alright before it reaches you (best setup for servers, in 10 years never had an issue with an upgrade)

2.) When you enable other repositories like staging(recommended for beta testers only)/testing/multilib-staging/etc. you enter into the rolling bleeding edge side of ArchLinux but you still have access only to latest release code too new for stable repos, so no straight SVN/Git code yet

3.) Additional repositories, here you can fine tune how and where you wanna be bleeding edge, so lcarlier repos for example enable you to get daily git/svn packages for the whole OSS graphic stack from LLVM all the way to DDX and kernels every 20 hours(like amd-staging-4.9 from ag5df where you have the extra stuff like DAL code working from amdgpu)

I personally use 1 +2 +3 on my desktops with AMD cards

Yeah but i was talking about stable repositories, testing is mainly for, well testing. There's actually repository only for internal usage, before testing, but no one should enable that. There are reasons why they keep LLVM 4.0 ins table and not newer i guess.

[jrch2k8]

I don't think there would be any difference, even Arch uses 4.0, they don't do it for no reason i guess .

Fix:
ArchLinux have 3 ways of working

1.) When you use only the default repositories (Core / Extra / Community) you have Archlinux in stable mode, is like having an rolling debian because only stable software is shown to you and it takes some test to ensure all is alright before it reaches you (best setup for servers, in 10 years never had an issue with an upgrade)

2.) When you enable other repositories like staging(recommended for beta testers only)/testing/multilib-staging/etc. you enter into the rolling bleeding edge side of ArchLinux but you still have access only to latest release code too new for stable repos, so no straight SVN/Git code yet

3.) Additional repositories, here you can fine tune how and where you wanna be bleeding edge, so lcarlier repos for example enable you to get daily git/svn packages for the whole OSS graphic stack from LLVM all the way to DDX and kernels every 20 hours(like amd-staging-4.9 from ag5df where you have the extra stuff like DAL code working from amdgpu)

I personally use 1 +2 +3 on my desktops with AMD cards

[leipero]

Michael time for you to switch LLVM 5.0 SVN/Git.... 4.0 is old ;p

I don't think there would be any difference, even Arch uses 4.0, they don't do it for no reason i guess .

[dungeon]

This response not merit his question

If not obvious, source of that article is on Intel's forum