Welcome to MilkyWay@home

CUDA/OpenCL on Mac?

Message boards : Application Code Discussion : CUDA/OpenCL on Mac?
Message board moderation

To post messages, you must log in.

AuthorMessage
Andrew Dicker

Send message
Joined: 9 Apr 13
Posts: 10
Credit: 27,318,556
RAC: 0
Message 58874 - Posted: 15 Jun 2013, 0:06:00 UTC

Hi all,
What chances are there of getting an OpenCL or CUDA milky way app for Mac?
I managed to compile a separation binary http://milkyway.cs.rpi.edu/milkyway/result.php?resultid=500657055 but got computation errors due to the OpenCL kernel not compiling (sorry, logs vanished, i forgot to save them before resetting the project when removing my app_info).

Regards, Andrew

(also: screensaver apps? i got nbody to compile for mac, but not separation)
ID: 58874 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Jake Weiss
Volunteer moderator
Project developer
Project tester
Project scientist

Send message
Joined: 25 Feb 13
Posts: 580
Credit: 94,200,158
RAC: 0
Message 59021 - Posted: 21 Jun 2013, 14:14:46 UTC

Hey Andrew,

I just put up an OpenCL app for Mac users with the release of the modified fit. It is still in a sort of testing phase so if you run into any problems with it please let me know.

Thank you,

Jake W
ID: 59021 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Andrew Dicker

Send message
Joined: 9 Apr 13
Posts: 10
Credit: 27,318,556
RAC: 0
Message 59030 - Posted: 22 Jun 2013, 0:05:29 UTC - in response to Message 59021.  
Last modified: 22 Jun 2013, 0:07:25 UTC

Cheers,
I noticed the new run, don't think i've gotten any GPU WUs yet to check how it runs though. Are the requirements the same as for the other (Win/Linux only) GPU apps? I have two machines, one with a 680MX, 2GB VRAM, and one 330M, 512MB. Should I keep an eye on them both, or only expect WU for the 680MX?

I'll report back when I get some on either.

EDIT:
I tell a lie, checking my task list from overnight results in seeing http://milkyway.cs.rpi.edu/milkyway/result.php?resultid=506354700 among others from the 680MX. A fail.
ID: 59030 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Jake Weiss
Volunteer moderator
Project developer
Project tester
Project scientist

Send message
Joined: 25 Feb 13
Posts: 580
Credit: 94,200,158
RAC: 0
Message 59031 - Posted: 22 Jun 2013, 1:39:26 UTC

Okay I've been seeing a lot of those. I think it might have something to do with the way Mac implemented one of the optional requirements for their GPU drivers. Will look into it a little more.

Thank you,

Jake W
ID: 59031 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Andrew Dicker

Send message
Joined: 9 Apr 13
Posts: 10
Credit: 27,318,556
RAC: 0
Message 59032 - Posted: 22 Jun 2013, 1:42:57 UTC - in response to Message 59031.  

I'm willing to fiddle around if you need some debugging info or anything.

Also, be aware that my machines are using pre-release OS, so it is entirely likely something is broken OS-side. For example, OpenCL 1.2 in the 680MX driver is very new.
ID: 59032 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Andrew Dicker

Send message
Joined: 9 Apr 13
Posts: 10
Credit: 27,318,556
RAC: 0
Message 59036 - Posted: 22 Jun 2013, 12:09:32 UTC - in response to Message 59031.  
Last modified: 22 Jun 2013, 12:18:17 UTC

google shows this for the error:
http://software.intel.com/en-us/forums/topic/280040
http://www.khronos.org/registry/cl/sdk/1.2/docs/man/xhtml/constant.html

So i guess we need (lines 175-176 of probabilities_kernel.cl):
__constant SC* sc __attribute__((max_constant_size(NSTREAM * sizeof(SC)))),
__constant real* sg_dx __attribute__((max_constant_size(256 * sizeof(real)))),

I'll check this out soonish
ID: 59036 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Andrew Dicker

Send message
Joined: 9 Apr 13
Posts: 10
Credit: 27,318,556
RAC: 0
Message 59048 - Posted: 23 Jun 2013, 0:45:39 UTC - in response to Message 59036.  

Made those pointer changes, kernel compiles and appears to work:

http://milkyway.cs.rpi.edu/milkyway/result.php?resultid=507356160
ID: 59048 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Andrew Dicker

Send message
Joined: 9 Apr 13
Posts: 10
Credit: 27,318,556
RAC: 0
Message 59050 - Posted: 23 Jun 2013, 1:03:19 UTC - in response to Message 59048.  

Seem to be reporting a minuscule difference in the result however:
http://milkyway.cs.rpi.edu/milkyway/workunit.php?wuid=387104484
ID: 59050 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Jake Weiss
Volunteer moderator
Project developer
Project tester
Project scientist

Send message
Joined: 25 Feb 13
Posts: 580
Credit: 94,200,158
RAC: 0
Message 59060 - Posted: 23 Jun 2013, 15:14:19 UTC

Strange. I will talk to our OpenCL guru about it and see what he has to say. Sorry I'm not in the office until the morning though so hopefully we can roll out a fix by the end of the week.

Jake W
ID: 59060 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Jake Weiss
Volunteer moderator
Project developer
Project tester
Project scientist

Send message
Joined: 25 Feb 13
Posts: 580
Credit: 94,200,158
RAC: 0
Message 59074 - Posted: 24 Jun 2013, 13:06:30 UTC

Looks like both of those links you sent me were validated workunits so I the difference must be very small. Probably just a little floating point error. Nothing to worry about because we already take that into account and don't try to optimize past 6 decimal places using double precision which should in theory give 15 or so. I will test the change here locally today.

Thank you,

Jake W
ID: 59074 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Andrew Dicker

Send message
Joined: 9 Apr 13
Posts: 10
Credit: 27,318,556
RAC: 0
Message 59084 - Posted: 24 Jun 2013, 21:28:06 UTC - in response to Message 59074.  

Yeah, the first link was "Validation inconclusive" when I posted it, but became valid later. 6dp sounds fine, the difference i noticed was much less significant.

I had a large number of tasks validate (~60) before removing my app_info and going back to standard apps. (note the computer may change ID, I seem to have to remove and re-attach the project and merge computers after fiddling with app_info)
ID: 59084 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Todd Madson

Send message
Joined: 16 Jun 12
Posts: 1
Credit: 1,223,497
RAC: 0
Message 61230 - Posted: 25 Feb 2014, 1:30:53 UTC

Maybe this is in the wrong area but I'm running the following:

iMac Corei7 2.8 ghz with 16 gigs of 1333 mhz DDR3 ram.
My video card is an ATI Radeon HD 4850 running OpenCL 1.2 (AUG 24 2013 21:03:27) (FULL_PROFILE) with 512 megs of video ram. According to ATI this is an R700 GPU.

I'm running BOINC Manager v7.2.39 and when I try to run Milkyway on my videocard I receive the following error:

Milkyway@Home: Notice from server
ATI GPU R600 (R38xx) does not support OpenCL.

So if the manufacturer says it's an R700 but the app says it's an R600 is there any way I can correct the problem to allow my admittedly older video card do any crunching?


ID: 61230 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote

Message boards : Application Code Discussion : CUDA/OpenCL on Mac?

©2024 Astroinformatics Group