You are not logged in.
./deviceQuery Starting...
CUDA Device Query (Runtime API) version (CUDART static linking)
cudaGetDeviceCount returned 30
-> unknown error
Result = FAIL
Is it a Bug?Or I need to wait new version of cuda?
Last edited by wpjsolo (2014-03-05 09:29:26)
Offline
launch with sudo/su
http://sl1pkn07.no-ip.com/paste/view/a78f18c8
Last edited by sl1pkn07 (2014-03-04 14:34:51)
Offline
Similar issue, CUDA 5.5 samples fail with the latest nvidia update:
./deviceQueryDrv Starting...
CUDA Device Query (Driver API) statically linked version
modprobe: FATAL: Module nvidia-uvm not found.
cuInit(0) returned 999
-> CUDA_ERROR_UNKNOWN
Result = FAIL
Other samples also fail:
sudo ./particles
CUDA Particles Simulation Starting...
grid: 64 x 64 x 64 = 262144 cells
particles: 16384
modprobe: FATAL: Module nvidia-uvm not found.
CUDA error at ../../common/inc/helper_cuda.h:898 code=30(cudaErrorUnknown) "cudaGetDeviceCount(&device_count)"
These worked prior to the update. I'm guessing I need to seek out the nvidia-uvm module?
Offline
Another clue to my problem - when running the deviceQueryDrv example with sudo, I get this additional line in the output:
modprobe: FATAL: Module nvidia-uvm not found.
Without sudo, this does not occur. I've not had to use sudo with these examples before...
Offline
I've posted a comment on a related thread in the Nvidia Developer forums.
Offline
I also have this problem after the upgrade to 334.21 unfortunately. Hope there's a logical explanation and fix soon.
Offline
numasan - Take a look at the Nvidia Developer forums link I posted. Apparently the new drivers work with the CUDA 6.0 release candidate. I've applied for access to the RCs, but that probably won't go through for another day or two :-/
I'll definitely reply back with whether it worked or not after I get access. If anyone else has access to the 6.0 RC, do the rest of us a favor and try it out :-)
Offline
I update to nvidia 334.21-2, and the errors go away.
Offline
@rdahlgren
Great if it works with CUDA 6.0 but I imagine it will be quite a while before we see that in Arch.
Like wpjsolo said, some of the errors go away in 334.21-2 but with caveats. I have to run some CUDA code as root, before it becomes available for users. It is not enough to 'modprobe nvidia_uvm' before using CUDA software as a user; I _have_ to execute some CUDA code as root first. So close but not optimal.
Offline
numasan - Unfortunately the 'run some code as root' workaround isn't working for me, which is unfortunate as I was just getting started with CUDA, haha. Do you have the nvidia-uvm module at all? Probing for it fails for me and `lsmod | grep nvidia` only shows the `nvidia` module.
As for CUDA 6.0 release candidate, it requires an old version of GCC to build and the samples fail if nvcc isn't up to date. I really don't feel like going down the "multiple versions of gcc" path right now :-/
What steps exactly does your workaround involve? Does running the deviceQueryDrv sample as root work for you?
Last edited by rdahlgren (2014-03-06 00:16:24)
Offline
I had previously installed CUDA from the nvidia developers site. I installed nvidia-utils-beta and nvidia-beta-dkms from the AUR. After that, I have the behavior the other posters in this thread are describing:
1. Run any sample as a regular user, it fails on cuInit
2. Run the sample as root, it succeeds
3. Run the sample as a regular user, it succeeds
Note that this is after having loaded the nvidia-uvm module via modprobe
Offline
I use cuda-5.5.22-1 and nvidia-334.21-2 and nothing from AUR. With nvidia-334.21-1 nothing CUDA would run as either user or root, but the update from yesterday works with the workaround.
I tried manually loading nvidia-uvm but it doesn't seem to matter when running CUDA software as a user. Looks like the nvidia-uvm module gets loaded automatically when running CUDA software as root (again, not sure nvidia-uvm is the key here). Only tested with Blender and Cudaminer:
user: Start Blender/Cudaminer = no compute-device recognized.
root: Start Cudaminer, stop when it continues normally.
user: Start Blender = Nvidia compute-device recognized and works normally.
Last edited by numasan (2014-03-06 02:06:36)
Offline
I use OpenCL and not cuda but the problem is the same. After loading the nvidia_uvm module and running once as root I can use the GPU in OpenCL as a regular user.
Last edited by wwmm (2014-03-07 12:37:02)
Offline