Radeon + AMDGPU Performance On Linux 4.6

[tomtomme]

Could AMD make an official naming system to help us tell all these drivers from each other in a clear manner that doesn't require forensics of the benchmarks? I can tell the difference between Catalyst & "AMDGPU" but inside of AMDGPU there appear to be multiple flavors that are not clearly distinguished from each other and that have radically different performance levels.

For example, just the other day the so-called "initial release" of AMDGPU that includes the binary blobs finally showed stronger performance for the Fury vs. a lower-end GPU like the R9-290X. Now we have a "new" release of... AMDGPU.. that doesn't include the binary blobs where the situation is right back to the way it was before. The problem is that both sets of very different drivers are being called AMDGPU in a very loose manner that leads to all sorts of inconsistencies.

the partly-binary amdgpu-hybrid-stack is called "amdgpu pro"
the non-binary just "amdgpu"

yes, it is that easy.

[tomtomme]

Michael
Oibaf kernel or whole ppa seems to be crap for tonga and fiji. Look at your older Benchmarks where e.g. fury ran 160 fps for tesseract and now 56 fps...

https://www.phoronix.com/scan.php?page=article&item=nvidia-amd-xenial&num=1

Seems like tonga and fury do not recklock with oibaf kernel, but they do with official (amdgpu backported) ubuntu 16.04 kernel 4.4 (and amdgpu-pro)

compare: https://openbenchmarking.org/result/...GA-RADEONLIN58

Really Michael, you should be the first to spot this, not me.

[bridgman]

Public OpenCL does not use HSA - what made you think it was required ?

Which userspace code were you running - the all-open stack this article is talking about, something newer/older, or the hybrid/Pro preview driver ?

[Brane215]

I decided to try it on Kabini ( Athlon 5350) on relatively fresh gentoo-spoources-4.5.0 kernel. What a load of crap. I managed to get graphical console, but X silently stopped - no trace why in logfile.

It took me an eternity to figure out that it really is driver, screwing things up. I only wanted to use it because of better OpenCL performance and support, only to find out that:

a. it needs HSA support in kernel, which is separate, independent option
b. HSA doesn't support Kabini at present

[M@GOid]

Oibaf PPA still using old LLVM 3.8. Padoka is making packages to Xenial with the latest LLVM 3.9. I suggest you use Padoka PPA in your next benchmark Michael.

[duby229]

Unless that's the intent.
It's really a symptom of one of the hardest problems in software development, one that only has local solutions that only work for a while. The problem is called "naming things".

I'd say AMD sucks at it. fglrx was gone in 2007, and yet even to this day people are still calling Catalysts kernel driver that. Crimson was a perfect chance to give the new stack a good marketing name and instead they wasted it on Catalyst.

[juno]

No, you have to enable CIK support in the kernel config and you have do this patch manually:

Then compile the kernel and install.

Thanks, but I don't think that's what I was looking for. AFAIR there have been users writing about boot problems (so before x) with amdgpu and CIK since 4.5(?)

[bug77]

Could AMD make an official naming system to help us tell all these drivers from each other in a clear manner that doesn't require forensics of the benchmarks? I can tell the difference between Catalyst & "AMDGPU" but inside of AMDGPU there appear to be multiple flavors that are not clearly distinguished from each other and that have radically different performance levels.

For example, just the other day the so-called "initial release" of AMDGPU that includes the binary blobs finally showed stronger performance for the Fury vs. a lower-end GPU like the R9-290X. Now we have a "new" release of... AMDGPU.. that doesn't include the binary blobs where the situation is right back to the way it was before. The problem is that both sets of very different drivers are being called AMDGPU in a very loose manner that leads to all sorts of inconsistencies.

Unless that's the intent.
It's really a symptom of one of the hardest problems in software development, one that only has local solutions that only work for a while. The problem is called "naming things".

[darkbasic]

Why does Fury still suck that much? Still no fix?

[juno]

Is amdgpu CIK support already fixed in 4.6?