Posts by Oliver Bock

log in
21) Message boards : News : New BRP application released: v1.22 (CPU, CUDA, OpenCL) (Message 111877)
Posted 452 days ago by Oliver Bock
Discussion of the new app release takes place over in this thread.

Notes:
* OpenCL GPU memory requirements reduced by 128 MB (to ~360 MB)
* Dropped support for OpenCL 1.0 GPUs (in favor of the above, OpenCL 1.0 only GPUs like the Radeon 4xxx are too slow anyway)
* More graceful OpenCL memory error handling
* Known issue: no OpenCL support for Mac OS X for the time being (we're looking into a potential Apple bug, no pun intended)
* CUDA apps are enabled again for the time being
* Input data download volume (per work unit) can be reduced by a factor of two (soon)
* Please use the latest Catalyst driver (>=12.1) and BOINC client (>=7.0.12).


Cheers,
Oliver
22) Message boards : Problems and Bug Reports : [New release] BRP app v1.22 feedback thread (Message 111876)
Posted 452 days ago by Oliver Bock
Hi,

We just released BRP4 v1.22 which adds a number of improvements.

Notes:
* OpenCL GPU memory requirements reduced by 128 MB (to ~360 MB)
* Dropped support for OpenCL 1.0 GPUs (in favor of the above, OpenCL 1.0 only GPUs like the Radeon 4xxx are too slow anyway)
* More graceful OpenCL memory error handling
* Known issue: no OpenCL support for Mac OS X for the time being (we're looking into a potential Apple bug, no pun intended)
* CUDA apps are enabled again for the time being
* Input data download volume (per work unit) can be reduced by a factor of two (soon)
* Please use the latest Catalyst driver (>=12.1) and BOINC client (>=7.0.12).

Let's try and collect your feedback to this specific release (and this one only) in this thread.


Thanks,
Oliver
23) Message boards : Problems and Bug Reports : [OpenCL] app v1.20/v1.21 feedback thread (Message 111872)
Posted 452 days ago by Oliver Bock
Looks like a BOINC issue. Thanks for reporting.

Oliver
24) Message boards : Problems and Bug Reports : [OpenCL] app v1.20/v1.21 feedback thread (Message 111863)
Posted 459 days ago by Oliver Bock
As it says, not enough memory...

Update: we'll improve the error handling and messaging for such cases. The next release will state it more explicitly and the app will ask BOINC to retry again after 5 minutes, as it does for other GPU memory issues...

Oliver
25) Message boards : Problems and Bug Reports : [OpenCL] app v1.20/v1.21 feedback thread (Message 111860)
Posted 462 days ago by Oliver Bock

I had no success in a non X setup. (I do not need X here)
Stopping lighdm so that all modules are loaded and DISPLAY (COMPUTE) variable set to ":0" atiOpenCL apps do not start on the commandline in a console.


Known limitation of the AMD drivers/runtime. If your machine is isolated you may also just allow all users to connect to your (running) X display, works just fine for me:

export DISPLAY=:0 && xhost +


Best,
Oliver
26) Message boards : Problems and Bug Reports : [OpenCL] app v1.20/v1.21 feedback thread (Message 111850)
Posted 469 days ago by Oliver Bock
It turned out on debian64, the AMD driver is installed on /usr/lib32/libamdocl32.so and not /usr/lib/libamdocl32.so.
I don't know why but I had an old version in /usr/lib.
Just copied the new lib over, and it's now working.


I experienced the same issue on my machine. I turns out that this is an inconsistency between the APP SDK and Catalyst. Both install this 32-bit compatibility library, but in different locations, hence the mix-up that leads to the segfault. I already filed a bug report with AMD.

I recommend to not install the APP SDK, you don't need it to run OpenCL GPU applications as long as you installed Catalyst >= 11.3 (Windows, 11.11 for Linux). The SDK is for developers only.


Cheers,
Oliver
27) Message boards : Problems and Bug Reports : [OpenCL] app v1.20/v1.21 feedback thread (Message 111839)
Posted 471 days ago by Oliver Bock
Oliver i installed ubuntu supplied ati catalyst it is 12.1 but its missing opencl


Are you sure you installed all relevant packages? Linux distributions might often split the original/upstream driver installation package into separate packages like the kernel module, the user space driver, the OpenCL runtime and the OpenCL ICD. Debian does this and Ubuntu could do it in a similar way since it's based on Debian.

Good to see that it now works for you!

Cheers,
Oliver
28) Message boards : Problems and Bug Reports : [OpenCL] app v1.20/v1.21 feedback thread (Message 111832)
Posted 472 days ago by Oliver Bock
First successfully completed opencl wu http://albert.phys.uwm.edu/result.php?resultid=109919


Great, so it turned out that the installation was flawed, most likely because of missing 32-bit compatibility libs.

Anyway, I hope we see at least some of those errors again on other systems as we'd really like to understand them and make BOINC more robust for these types of situations.

@HV Margus*: Please run at least one or two tasks before you fix your driver setup. This would really help us! Thank you!


Best,
Oliver
29) Message boards : Problems and Bug Reports : [OpenCL] app v1.20/v1.21 feedback thread (Message 111831)
Posted 472 days ago by Oliver Bock
Maybe we can get some feedback from those successfully returning results from 64bit linux machines with ATI/AMD gpus.


Guys, we just released v1.21 which adds some useful debugging info to the task/result error log. Please reset the albert project or wait until the next task is fetched. The log output of those tasks will help us to understand the problem.

Thanks,
Oliver
30) Message boards : Problems and Bug Reports : [OpenCL] app v1.20/v1.21 feedback thread (Message 111821)
Posted 473 days ago by Oliver Bock
For those of you on Linux who have the AMD APP SDK installed alongside the AMD Catalyst drivers, please make sure you follow the order as details in the AMD developer release notes (chapter 2).

HTH,
Oliver


Previous 10 | Next 10

Home · Your account · Message boards

This material is based upon work supported by the National Science Foundation (NSF) under Grant PHY-0555655 and by the Max Planck Gesellschaft (MPG). Any opinions, findings, and conclusions or recommendations expressed in this material are those of the investigators and do not necessarily reflect the views of the NSF or the MPG.

Copyright © 2013 Bruce Allen for the LIGO Scientific Collaboration