Spawned from https://bugs.mageia.org/show_bug.cgi?id=12129#c37 Description of problem: § Make sure we have the right packages § Document in wiki how to set it up, provude useful links to BOINC site -for both Nvidia and AMD GPUs I am available to test in my workstation where I already run BOINC: Mageia 5, Intel i7, Nvidia GK104 [GeForce GTX 760] Reproducible: Steps to Reproduce:
To add to this, Have Mageia 5 installed also. Having doing some hunting I discovered that it will only allow detection of my AMD gpu after doing xhost local:boinc as root. Then restart the service. I did try the X0.host trick but that failed too.
CC: (none) => worzel910
Thanks for the hints Peter. So maybe we only need to configure it correctly. Where did you put the commands?
Shooting in blindness: I have GeForce GTX 760 using Nvidia proprietary driver "NVIDIA GeForce 400 series and later" according to MCC hardware module. Have installed nvidia-current-cuda-opencl - CUDA and OpenCL libraries for NVIDIA proprietary driverâ and nvidia-cuda-toolkit - NVIDIA CUDA runtime librariesâ Tried [root@svarten morgan]# systemctl stop boinc-client.service [root@svarten morgan]# xhost local:boinc non-network local connections being added to access control list [root@svarten morgan]# systemctl start boinc-client.service still no suitable GPU found. [root@svarten morgan]# systemctl stop boinc-client.service [root@svarten morgan]# gpasswd -a boinc video Lägger till användaren boinc till gruppen video [root@svarten morgan]# systemctl start boinc-client.service still no suitable GPU found. Also tried installing lib64opencl1 - OpenCL ICL Loaderâ  ( which needed nvidia-cuda-toolkit to be uninstalled ) : No change. This thread http://www.gpugrid.net/forum_thread.php?id=3734 mentions libcuda.so need exist, and i see the installed package nvidia-current-cuda-opencl contain /usr/lib64/nvidia-current/libcuda.so.1 (and exist on system) - can the added ".1" confuse boinc? There is a recent thread about nvidia in ubuntu and fedora20, https://devtalk.nvidia.com/default/topic/734098/linux/-boinc-ubuntu-nvidia-no-usable-gpus-found-/post/4357750/#4357750 talking about nvidia-modprobe . The installed package x11-driver-video-nvidia-current is according to drakrpm supposed to contain /usr/lib64/nvidia-current/bin/nvidia-modprobe but it does not ...?
tried without success: # cd /usr/lib64/nvidia-current # ln -s libcuda.so.1 libcuda.so
First thing is to stop the service as root with: systemctl stop client-boinc.service Then do xhost local:boinc Then restart the service with systemctl start client-boinc.service Go to the boinc manager via tthe menu and look through the logs . I did pretty much the same as you Morgan, thought it was a missing lib or whatever. It's due to not been allowed access to the driver thus the xhost trick. Down side is it needs to be done after every reboot/power cycle.
I've been having similar issues trying to get CUDA/OpenCL working on my Geforce 770M. I wonder if it's the same underlying cause at a system configuration level, Morgan. In my case, I'm on an Optimus architecture, so that sometimes throws a wrench in things. If it's not too much of an inconvenience, could you guys try starting blender then go to File->User Preferences->System and see if you have a GPU option under the compute device selection at the bottom left?
CC: (none) => rick
@Peter: that did not work for me @Richard: No GPU (i suppose it would show in the now blue field saying "None" above the grey list field saying "CPU". OK thanks. So other apps also do not find GPU - probably easier to get to know the cause? I have no idea though.
@Morgan: Not a clue then, am running a AMD gpu so cant investigate , though I could see on my lads. @Rick Blender does not show GPU for me neither , Luxmark works fine with and without the xhost trick Boinc works with the fix above but not blender.
It does work in blender! I needed python-opencl installing So yes it works Rick
changed temporarily to nouveau and uninstalled and reinstalled nvidia driver, installed nvidia-current-cuda-opencl and nvidia-cuda-toolkit. # xhost local:boinc # gpasswd -a boinc video -> still no useable GPU in boinc installed python-opencl -> still no GPU in blender. Who is expert on GPU processing in mageia - especially Nvidia ? # nvidia-smi Sun Jan 11 17:09:11 2015 +------------------------------------------------------+ | NVIDIA-SMI 340.65 Driver Version: 340.65 | |-------------------------------+----------------------+----------------------+ | GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC | | Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. | |===============================+======================+======================| | 0 GeForce GTX 760 Off | 0000:01:00.0 N/A | N/A | | 29% 33C P8 N/A / N/A | 529MiB / 2047MiB | N/A Default | +-------------------------------+----------------------+----------------------+ +-----------------------------------------------------------------------------+ | Compute processes: GPU Memory | | GPU PID Process name Usage | |=============================================================================| | 0 Not Supported | +-----------------------------------------------------------------------------+
I do not know what to look for in logs, but here is some: # journalctl -ab | grep -i -e cuda -e opencl # journalctl -ab | grep nvidia jan 11 16:46:11 svarten kernel: nvidia: module license 'NVIDIA' taints kernel. jan 11 16:46:11 svarten kernel: [drm] Initialized nvidia-drm 0.0.0 20130102 for 0000:01:00.0 on minor 0 jan 11 16:46:21 svarten kernel: nvidia 0000:01:00.0: irq 52 for MSI/MSI-X
Reading at link below it seems that some API may get blocked when "taint" happens. http://stackoverflow.com/questions/16809378/getting-message-module-license-unspecified-taints-kernel-despite-setting-mo So maybe this is because some license file (however it works) missing that is the cause? So i set this bug against rpm "nvidia-current-kernel-3.18.1-desktop-4.mga5, x86_64, Version: 340.65-8.mga5.nonfree" ( I did not figure out what to put there exactly...) Lets not care specifically for BOINC (which i originally entered this bug for) until *any* program can use nvidia GPU. (Also tried darktable now and found OpenCL is greyed out in its settings)
Summary: Make sure BOINC can use GPU for number crunching => Fail to get CUDA or OpenCL working (because nvidia taint kernel?)Source RPM: (none) => nvidia-current-kernel-3.18.1-desktop-4.mga5, x86_64, Version: 340.65-8.mga5.nonfree
According to this forum post OpenCL was working for that user with mga5 this september, nvidia 340.58 , but not before. https://forums.mageia.org/en/viewtopic.php?f=15&t=8711&p=53189&hilit=cuda#p53189 To test, how do i go about to downgrade to that versions exactly ?
...i have some activity now in that forum thread...
It works to build locally using files from Nvidia and script from Anssi https://bugs.mageia.org/show_bug.cgi?id=15328
Does it mean there's something we can do?
I do not know how these things work, i just followed instructions. I suggest to ask Anssi.
Summary: Fail to get CUDA or OpenCL working (because nvidia taint kernel?) => Mageia nvidia do not work with CUDA or OpenCL but locally built with mageia script it works.
Anssi, would you like to comment?
CC: (none) => anssi.hannula
Whiteboard: (none) => MGA5TOO
boinc search for libcuda.so and libOpenCL.so : they are in nvidia-current-devel BUT installing the package doesn't trigger ldconfig, so the lib aren't found until, some day, another package installation triggers ldconfig and update ld.so.cache... Fix : urpmi nvidia-current-devel ldconfig
CC: (none) => yann.cantin
Still same problem, and that fix in comment 19 still works - thank you Yann ! Closing this, summing info in Bug 15328 *** This bug has been marked as a duplicate of bug 15328 ***
Status: NEW => RESOLVEDResolution: (none) => DUPLICATE