AMD Radeon Graphics Performance With The Linux 3.19 Kernel

[Michael]

Thanks Michael, this means you will finally start using LLVM git.

I haven't been using the Oibaf PPA in weeks but have been using his better repo with LLVM 3.6 SVN, as mentioned in numerous articles.

[darkbasic]

I haven't been using the Oibaf PPA in weeks but have been using his better repo with LLVM 3.6 SVN, as mentioned in numerous articles.

I saw a couple of articles with LLVM 3.6 SVN in the past weeks but I tought they were exceptions, I guess I should look more carefully next time
Anyway as I said thanks for using live LLVM, this is *very* appreciated!

just saw the commits, they are 48 hours old, should be in mesa master soon. once they hit, my mesa compilation will kick in. at least llvm will be ready (compiling it now).

This is a separate branch which will probably NOT merge into master for a long time. I used such branches lots of time and I don't suggest you to use them for your ppa because there are *high* chances they will break something and that will mean Michael reverting to Oibaf which in turn means stable LLVM
Instead I suggest you to make another, even more experimental ppa.

[dungeon]

This is fucking interesting. Please benchmark ASAP. If I wasn't so much busy this month I would have done it myself.

I actually seems to had tested some different combination... sad to say, rendering is still broken in perf-Jan-08-2015

[groo_pcd]

I saw a couple of articles with LLVM 3.6 SVN in the past weeks but I tought they were exceptions, I guess I should look more carefully next time
Anyway as I said thanks for using live LLVM, this is *very* appreciated!



This is a separate branch which will probably NOT merge into master for a long time. I used such branches lots of time and I don't suggest you to use them for your ppa because there are *high* chances they will break something and that will mean Michael reverting to Oibaf which in turn means stable LLVM
Instead I suggest you to make another, even more experimental ppa.

i only compile from the oficial masters (intel/llvm/mesa/clc/drm/the like). if this commits arent coming to mesa master soon, then ill pass. if theres enough demand i can make a new ppa with tstellar work. Since i know him from freenode, ill ask him if its worth the trouble. tks for pointing out this commit though.

[darkbasic]

It's a waste of time right now, dungeon just said rendering is still broken

[mannerov]

It's a waste of time right now, dungeon just said rendering is still broken

I have tried it too with the mesa branch, and it works pretty good for me, and get perf gains on everything (+15% or more compared to llvm 3.5)

[dungeon]

I have tried it too with the mesa branch, and it works pretty good for me, and get perf gains on everything (+15% or more compared to llvm 3.5)

Yeah performance is there, but broken rendering too... some games are fine, but did you tried unigine valley, tesseract... those have very visible artifacts ?

[Veerappan]

hi Michael, padoka ppa maintainer here (https://launchpad.net/~paulo-miguel-...e/ubuntu/mesa/)

If you find interesting, i fixed the opencl packages in my ppa, they are kinda broken with oibafs, missing amd64 build and amdgcn support (oibafs only provides r600).

Anyway, AMD clover/opencl is working in my PPA, so feel free to fire opencl 1.1 benchmarks at will.

I would like to seize the opportunity to reiterate that has of now, my ppa is less and less a derivative of oibafs, since ive been deviating from his work and adding more cutting edge stuff to mine (LLVM 3.6, clover/opencl with amdgcn support, intel experimental modules (rendernode,create2,async-swap, dri3 is off for now), and a few well tested apps (cmplayer, transmageddon). we all stand on the shoulders of giants (oibaf)

regards

FYI: The amdgcn target machine was just added to LLVM a few days ago (and libclc/mesa as a result). Before that, all GCN chips (radeon si/ci) used the R600 llvm back-end whose name is shared with the back-end for r600 through cayman. GCN-based chips worked with OpenCL before that, just using the r600-- target.

As of right now, I think that most of the OpenCL tests that are in PTS will still fail to run, usually becuase there's a hand-ful of built-in functions in libclc that are not yet implemented that the specific benchmarks use, or in some cases because the benchmarks assume that the OpenCL implementation/device has image support (which libclc doesn't have yet).

[mannerov]

Yeah performance is there, but broken rendering too... some games are fine, but did you tried unigine valley, tesseract... those have very visible artifacts ?

No, I did test nine games. Red Alert 3, Mass effect 2, Heroes 5, Borderlands 2, Skyrim, 3Dmark05, Half life 2 lost coast, Arma 2.
Unigine heaven had compilation failure with missing instruction (but that's probably just a llvm 3.6 bug), but with opengl backend it was fine.

[dungeon]

Unigine heaven had compilation failure with missing instruction (but that's probably just a llvm 3.6 bug), but with opengl backend it was fine.

With llvm 3.6 master Heaven works fine ... rendering issues are only there for me with those llvm perf branches and it seems only with more modern opengl like unigine demos wanted, tesseract, etc... well nine probably might not be affected, but dx9 level is not that modern .