Linux and Intel GPUs

Jonathan Jeckell
Jonathan Jeckell
Joined: 11 Nov 04
Posts: 114
Credit: 1341945207
RAC: 0
Topic 198290

Just wasted a bunch of time getting OpenCL working on a Linux machine so I could run the app "Linux running on an AMD x86_64 or Intel EM64T CPU 1.52 (BRP6-Beta-opencl-intel_gpu)"

Only to find out afterwards from the event log:
2015-10-30 02:42:27.4259 [PID=26469] [version] Checking plan class 'BRP6-Beta-opencl-intel_gpu'
2015-10-30 02:42:27.4260 [PID=26469] [version] parsed project prefs setting 'gpu_util_brp': 1.000000
2015-10-30 02:42:27.4260 [PID=26469] [version] Intel GPU device name: 'Intel(R) HD Graphics IvyBridge M GT2' doesn't match 'HD Graphics [4-9]|HD Graphics Haswell'

That would have been good to know a few hours ago. Will this be expanded to include this GPU, or is this just time in my life I will never get back? I'd really love to get two unused GPUs contributing to this project.

Bernd Machenschalk
Bernd Machenschalk
Moderator
Administrator
Joined: 15 Oct 04
Posts: 4273
Credit: 245276509
RAC: 12054

Linux and Intel GPUs

A Linux driver for Intel GPUs is completely news to me, that's why this configuration isn't really supported on Einstein@Home. Did you really get OpenCL to run on the (Intel) GPU, or does it only make use of the CPU?

On top of that BRP6 tasks are somewhat too demanding for the smaller Intel GPUs.

Anyway, I'll add "HD Graphics IvyBridge" to the plan class you named, so you can try for yourself.

Bm

BM

Bikeman (Heinz-Bernd Eggenstein)
Bikeman (Heinz-...
Moderator
Joined: 28 Aug 06
Posts: 3522
Credit: 691071364
RAC: 264742

Interesting... I found

Interesting...

I found this thing here describing a third-party solution for Linux:

http://arrayfire.com/opencl-on-intel-hd-iris-graphics-on-linux/

However this is only something for more experienced Linux users, note the many caveats in the requirements, the potential need for a kernel patch and fiddling with system configuration after every boot ... it might work, but if it doesn't, we will not be able to replicate this for now. Still we are eager to hear about your results, of course.

Cheers
HB

P.S.: Funny how the world is a village, the author of said HOW-TO is the lead author of a citizen science paper in astronomy I co-authored several years ago.

Bernd Machenschalk
Bernd Machenschalk
Moderator
Administrator
Joined: 15 Oct 04
Posts: 4273
Credit: 245276509
RAC: 12054

Yep, that's "Beignet".

Yep, that's "Beignet". Someone tried this a couple of months ago on E@H (can't find the thread now), but the results were unusable.

Also Intel claims there is a Linux OpenCL driver for Intel GPUs as part of their (commercial) Media Server Studio, but I am not aware of anyone who tried that for scientific calculation.

Best,
Bernd

BM

Jonathan Jeckell
Jonathan Jeckell
Joined: 11 Nov 04
Posts: 114
Credit: 1341945207
RAC: 0

Yes, I tried using the Intel

Yes, I tried using the Intel Media Server Studio and Beignet. I had trouble with several different approaches to using Beignet, and somewhere between one of them and the official Intel OpenCL driver for Linux I got it to run. Then suddenly the GPU showed up on the machine's listing as recognized by Einstein@home.

Thereafter, I checked the event log to ascertain why BRP6 workunits weren't downloading on the client and that's when I say the Ivy Bridge vs Haswell requirement.

The assessment tool showed that all of the OpenCL functions check out, so it would be very disappointing if the results here would be unusable.

It would also be very disappointing because I have two of these GPUs doing absolutely nothing, whereas I have 2 other similar Intel GPUs working on SETI & Einstein (on Windows) and for SETI (on the Mac) respectively.

I know these GPUs don't compare to the high-end NVIDIA and AMD cards, but it is still a processor (I also have a bunch of Raspberry Pi's crunching away, so maybe I am more inclined to make this work than most people).

Jonathan Jeckell
Jonathan Jeckell
Joined: 11 Nov 04
Posts: 114
Credit: 1341945207
RAC: 0

I tried several of those, and

I tried several of those, and Arrayfire did not work for me. I hit an error with a bad block in a PCH file about 15% into the make process. The Arrayfire site also called for version 1.0.0 of Beignet, and the one that finally worked for me used something like 1.0.2 I think.

Intel's Iris and HD OpenCL driver installed, but didn't get recognized by the Einstein@home server.

I also tried to use http://freedesktop.org/wiki/Software/Beignet (this was actually the first one I tried), and it didn't work for me either. Intel's own 01.org open source site points to this.

This is the one that I think (perhaps in conjunction with the official Intel driver) got things working. I am not at home, so I don't have access to the machine, and Google is not cooperating. I can't find the site that actually worked for me now, but I think it was a Git, and I am pretty sure it called upon a newer version of Beignet than Arrayfire.

Jonathan Jeckell
Jonathan Jeckell
Joined: 11 Nov 04
Posts: 114
Credit: 1341945207
RAC: 0

Well, thanks for giving this

Well, thanks for giving this a shot guys. It downloaded three Parks PMPS XT workunits for the Intel GPU and... this happened.
Fri 30 Oct 2015 06:31:58 PM CDT | Einstein@Home | Starting task PM0064_01061_90_0
Fri 30 Oct 2015 06:31:59 PM CDT | Einstein@Home | Computation for task PM0064_01061_90_0 finished
Fri 30 Oct 2015 06:31:59 PM CDT | Einstein@Home | Output file PM0064_01061_90_0_0 for task PM0064_01061_90_0 absent
Fri 30 Oct 2015 06:31:59 PM CDT | Einstein@Home | Output file PM0064_01061_90_0_1 for task PM0064_01061_90_0 absent

And here is the wreckage from the sterr.txt:
7.2.42

process exited with code 255 (0xff, -1)

[18:31:58][3311][INFO ] Application startup - thank you for supporting Einstein@Home!
[18:31:58][3311][INFO ] Starting data processing...
[18:31:58][3311][INFO ] Using OpenCL platform provided by: Intel
[18:31:58][3311][INFO ] Using OpenCL device "Intel(R) HD Graphics IvyBridge M GT2" by: Intel
store float %345, float addrspace(1)* %379, align 4, !tbaa !59
Illegal pointer which is not from a valid memory space.
Aborting...

Not sure if this is something that can be rectified on my end of the server end, but I'm game if you guys are.

Jonathan Jeckell
Jonathan Jeckell
Joined: 11 Nov 04
Posts: 114
Credit: 1341945207
RAC: 0

By the way, this is the

By the way, this is the source that I think got OpenCL working on Intel (or, rather the jury is still out on that given the errors with E@H). Seems to work otherwise and passed all of the process checks for OpenCL.

https://gist.github.com/spiralray/cae0bc235509e495fec1

Juha
Juha
Joined: 27 Nov 14
Posts: 49
Credit: 4962871
RAC: 25

Beignet does work but you

Beignet does work but you need the right version of it and on Haswell a fresh kernel.

Ivy Bridge and Haswell

astrocrab
astrocrab
Joined: 28 Jan 08
Posts: 208
Credit: 429202534
RAC: 0

so, can i crunch on my intel

so, can i crunch on my intel gpu under linux or not?

Jonathan Jeckell
Jonathan Jeckell
Joined: 11 Nov 04
Posts: 114
Credit: 1341945207
RAC: 0

From what I heard, you can if

From what I heard, you can if you have a Haswell processor (HD4600 or better I think).

Comment viewing options

Select your preferred way to display the comments and click "Save settings" to activate your changes.