Disable CUDA in cmake

[pattacini]

@gregoire-pointeau commented on Thu Dec 01 2016

Hello, we are working on the cluster. We have the blades and the cuda machine that share the same sources.

While compiling yarp then icub-main, it works on the cuda-machine but on the blades we have a problem of cuda:

This is the cmake:

 CMAKE_BUILD_TYPE                 Release                                                               
 CMAKE_INSTALL_PREFIX             /usr/local                                                            
 CREATE_GUIS                      OFF                                                                   
 CREATE_LIB_MATH                  OFF                                                                   
 CREATE_SHARED_LIBRARY            ON                                                                    
 CREATE_YARPDATADUMPER            ON                                                                    
 CREATE_YARPMANAGER_CONSOLE       ON                                                                    
 CREATE_YARPROBOTINTERFACE        ON                                                                    
 CUDA_BUILD_CUBIN                 OFF                                                                   
 CUDA_BUILD_EMULATION             OFF                                                                   
 CUDA_SDK_ROOT_DIR                CUDA_SDK_ROOT_DIR-NOTFOUND                                            
 CUDA_TOOLKIT_ROOT_DIR            /usr/local/cuda                                                       
 CUDA_VERBOSE_BUILD               OFF                                                                   
 ENABLE_DASHBOARD_SUBMIT          OFF                                                                   
 OpenCV_DIR                       /usr/local/src/robot/opencv/build                                     
 YCM_DIR                          YCM_DIR-NOTFOUND                                                      

and this is the output:

CMake Error at /usr/share/cmake-2.8/Modules/FindPackageHandleStandardArgs.cmake:97 (MESSAGE):
   Could NOT find CUDA: Found unsuitable version "7.5", but required is exact
   version "4.2" (found /usr/local/cuda)
 Call Stack (most recent call first):
   /usr/share/cmake-2.8/Modules/FindPackageHandleStandardArgs.cmake:286 (_FPHSA_FAILURE_MESSAGE)
   /usr/share/cmake-2.8/Modules/FindCUDA.cmake:723 (find_package_handle_standard_args)
   /usr/local/src/robot/opencv/build/OpenCVConfig.cmake:48 (find_package)
   /usr/local/src/robot/opencv/build/OpenCVConfig.cmake:276 (find_host_package)
   cmake/ycm-0.2/find-modules/FindOpenCV.cmake:88 (find_package)
   cmake/YarpFindDependencies.cmake:363 (find_package)
   CMakeLists.txt:62 (include)

Is there a way to disable cuda?

Thank you

cc @AnneLaureM

[pattacini]

Hi @gregoire-pointeau

cuda and the blades should have the same OS, but of course cuda has its own stuff related to the GPU platform. Therefore, compilation should always go neatly on cuda but might bring about troubles on blades.

However, the same binaries compiled on cuda should run smoothly also if launched from blades, because they basically share the same OS. Thus, you wouldn't need to compile code on blades.

In case you really need to compile on blades, one possible workaround is to keep sharing sources, but differentiate the build process. This would entail to tailor e.g. the YARP_DIR and ICUB_DIR variables specifically for the blades.

@mbrunettini @vtikha am I wrong or missing anything?

[gregoire-pointeau]

On the blades were always set 2 directories in each project (yarp, icub-main...):

• build-cuda (to be executed on the icub-cuda)
• build-x86_64 (to be executed on the blades)

[vtikha]

@gregoire-pointeau can you confirm that the CUDA and the blades have the same OS?
If they have the same OS and mount the same filesystem, you do not need to have separate build as @pattacini correctly mentioned, but compile the required repo only from the cuda-machine. Of course cuda related modules should only run on the cuda-machine and other modules can be spread out on the various blades.

[pattacini]

Nonetheless, if they have two separate build directories, then you shouldn't experience any problem, in principle.

Are the env variables correctly set up? What are the values of YARP_DIR and ICUB_DIR, for both the systems? Perhaps, you could do which yarp on blades and check out that the path is the intended one.

Did you try out purging the cmake cache?

[gregoire-pointeau]

@AnneLaureM Could you have a look please? I'm never comfortable with everything that is architecture-related...

[drdanz]

@gregoire-pointeau did you build OpenCV manually? Apparently the warning comes from inside the OpenCVConfig.cmake file...

[vtikha]

@gregoire-pointeau let me add something to @drdanz's comment, if opencv has been built with cuda capabilities you will need to have two separate build directories for it to run on the blades.

[pattacini]

Hi @gregoire-pointeau and @AnneLaureM
Any news?

[AnneLaureM]

Hi everybody,

I am REALLY SORRY for the response time (I had my defense thesis on monday :-) ) .
A few months ago, we have compiled opencv 3.0 and by default opencv has some dependencies with cuda, so by default the cuda system librairies (in our case cuda 4.2 on Debian Wheezy...). So finally, I have just removed cuda 4.2, recompiled everything with cuda 7.0 and everything seems ok now.
Sorry for the ticket, The problem seemed rather clear with the cmake error. I did not pay enough attention to the error... :-(

I just have one last question... Which operating system do you use on your blades and the cuda machine at IIT ? Would it not be preferable to upgrade the whole cluster with Debian Jessie ? Is there a procedure to do it on the blades ?

Thanks very much for your help !!

Cheers,

Anne-Laure

[pattacini]

Congratualtions @AnneLaureM, you made the grade 🎉 🎓 😄

Actually, we've lately upgraded our system to the use of Ubuntu Xenial.

It's likely that @mbrunettini can point you at a list of useful instructions he might have for that.

[AnneLaureM]

Hi !!

Thanks @pattacini :-) Thanks to the grade, I am now able to read properly cmake errors :-D :-D.
I am surprised that the upgrade has been made with Ubuntu :-( !! @mbrunettini can you send us the list of instructions ? Do you think it is possible to upgrade the whole cluster into Debian Jessie ?

Cheers,

Anne-Laure

[pattacini]

@AnneLaureM is this still active?