cuda 5.5 or 5.0 fails deviceQuery on fedora19

Accelerated Computing CUDA CUDA Setup and Installation
1

after installation of cuda toolkit either in 5.5 or 5.0 , I can compile deviceQuery, but it fails to detect my nvidia GPU

./deviceQuery

cudaGetDeviceCount returned 35
→ CUDA driver version is insufficient for CUDA runtime version
Result = FAIL

However nvidia driver is loaded and operational

lspci -v | grep nvidia

Kernel driver in use: nvidia

cat /proc/driver/nvidia/version

NVRM version: NVIDIA UNIX x86_64 Kernel Module 319.32 Wed Jun 19 15:51:20 PDT 2013
GCC version: gcc version 4.8.1 20130603 (Red Hat 4.8.1-1) (GCC)

I use rpmfusion nvidia driver on fedora19
akmod-nvidia-319.32-2.fc19.x86_64

I used installer from http://developer.download.nvidia.com/compute/cuda/5_5/rel/installers/cuda_5.5.22_linux_64.run

nvcc -V

nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2013 NVIDIA Corporation
Built on Wed_Jul_17_18:36:13_PDT_2013
Cuda compilation tools, release 5.5, V5.5.0

what could be wrong ? which version of cuda toolkit should I work with ?

2

I would recommend you use the RPM drivers from the Nvidia RPM repository, or the driver included in the run file. I believe you need at least 319.37 (rather than 319.32 which is what you have).

3

I updated my akmod-nvidia to akmod-nvidia-325.15-1.fc19.1.x86_64 as available from rpmfusion non-free repo . so now I have

cat /proc/driver/nvidia/version

NVRM version: NVIDIA UNIX x86_64 Kernel Module 325.15 Wed Jul 31 18:50:56 PDT 2013
GCC version: gcc version 4.8.1 20130603 (Red Hat 4.8.1-1) (GCC)

325.15 being much higher than the “at least 319.37” you recommended , I though that would work now

unfortunatly deviceQuery still fails :-(

./deviceQuery

./deviceQuery Starting…

CUDA Device Query (Runtime API) version (CUDART static linking)

cudaGetDeviceCount returned 35
→ CUDA driver version is insufficient for CUDA runtime version
Result = FAIL

I’ve search in deviceQuery.cpp for which version that program is searching for, it is not explicitly written there, probably in a header file somewhere …

How can I know which version deviceQuery is searching for ?
thanks .

PS: cannot find a yum repo in nvidia site, is there one ? would it be better to use nvidia rpm instead of rpmfusion ones ?

4

I found a solution myself

not only there was a need to update akmod-nvidia 319.32 to 325.15
but also I notice that building the sample deviceQueryDrv failed because of -lcuda wasn’t found (ld collect error …), there do is a :
/usr/lib64/nvidia/libcuda.so.1
but no .so ! so I created the link /usr/lib64/nvidia/libcuda.so → /usr/lib64/nvidia/libcuda.so.1
then it compiled fine , I also had to manually start ldconfig so that dynamic lib are rescaned.
Then deviceQueryDrv does compile and returns correct result :

./deviceQueryDrv

Device 0: “Quadro 2000”
CUDA Driver Version: 5.5
…
Result = PASS

on my system /usr/lib64/nvidia/libcuda.so.1 comes from xorg-x11-drv-nvidia-libs-325.15-1.fc19.x86_64 (rpmfusion non-free package), it’s probably up to that package to create the link, but it didn’t, so I had to created myself .

Hence, it finally works ;-)

5

Awesome that you found a solution!

I should note (for people that stumble onto this thread with similar issues) that the driver RPMs provided in NVIDIA’s repository create the appropriate symbolic links for libcuda.so.

6

Even with the .so links correct, compilation of the CUDA 5.5 samples failed for me on Fedora 19 x86_64 because the nvcc -L path was not being set owing to a bug in the findcudalib.mk makefiles (or perhaps an “interesting feature” of Gnu make 3.82). To fix this, at line 50 (and ideally also 55 and 219)

ifeq (“$(DISTRO)”,‘’)

replace the single quotes with double ones. Otherwise the ifeq fails even when $(DISTRO) is empty: given the following, make 3.82 follows case 2…

X =
ifeq (“$(X)”,‘’)
$(warning “case 1”)
else ifeq (“$(X)”,“”)
$(warning “case 2”)
else
$(warning “case 3”)
endif