[CMake] FindCUDA with ITK

[Luke Parkinson]

----- "James Bigler" <jamesbigler at gmail.com> wrote:

> On Thu, Jan 7, 2010 at 8:52 PM, Luke Parkinson < lparkinson at vpac.org >
> wrote:
>
> ----- "James Bigler" < jamesbigler at gmail.com > wrote:
> > On Wed, Jan 6, 2010 at 11:00 PM, Luke Parkinson <
> lparkinson at vpac.org > wrote:
> >
> >
> 
> Hello Everyone,
> >
> > I apologise in advance if this is the wrong way to revive an old
> topic, but I wasn't sure how to reply to old threads in the mailing
> list and I'm new to this list. I'm part of a team of developers
> looking at integrating CUDA into the Insight Toolkit (
> http://www.itk.org/ ). People subscribing to the mailing list over
> there might have heard from my colleague.
> >
> > Thanks to James Bigler and his FindCUDA.cmake module, we have been
> able to get CUDA and ITK to link into the executable, but we're having
> a similar problem to the previous post (included below just in case).
> I have created several 'build environments' using this module with my
> own modifications that include the SDK (as an aside, can anyone point
> me in the direction of code that integrates C++ and CUDA without using
> cutil.h or cutil_inline.h?). It works brilliantly with just the CUDA
> SDK, and the SDK plus our own C++ CUDA utility programs, but when we
> include the ITK in the configuration, building gives errors exactly
> like those above. I've attached my modifications to the FindCUDA.cmake
> module (with a diff), and the 'CMakeLists.txt's we're using in the
> root and source directories.
> >
> > We got the original FindCUDA.cmake from the svn repository, and
> based our code on the driver examples. When we add our own CUDA/C++
> code (files omitted as code is still under development, permission is
> required), we get the same errors as mentioned previously, along the
> lines of 'warning: comparison between signed and unsigned integer
> expressions'.
> >
> >
> 
> Just including ITK (without our own code, using the provided driver
> code examples)
> > [Attached the corresponding files as ITKOnly_FindCUDA.cmake,
> ITKOnly_root_CMakeLists.txt, ITKOnly_src_CMakeLists.txt]
> >
> > Any ideas or suggestions would be helpful. This isn't crippling
> (only warnings), but it is annoying and could lead to problems later
> on if it has to do with how the CMake module deals with C/C++ code. I
> also noticed with that same module that there are a few variables that
> don't seem to do anything. The example I found was
> CUDA_INCLUDE_DIRECTORIES (deprecated by CUDA_INCLUDE_DIRS?). I assume
> that svn version was the most up-to-date, please let me know if I am
> wrong.
> >
> > snip...
> >
> > What versions of CUDA are you using? 2.3 or a 3.0 beta?
> >
> > Also, could you send the output of 'make VERBOSE=1' with
> CUDA_NVCC_FLAGS=-v for a single object compilation that has these
> warning? I would like to determine at which point these flags are
> causing problems.
> >
> > You could also attach (if possible) the cmake build script that is
> invoked by FindCUDA for the same object?
> >
> > Also, please make sure that CUDA_HOST_COMPILATION_CPP is ON (this
> can also be verified by me with the cuda_compile....cmake file you can
> attach).
> >
> > There is also the CUDA_PROPAGATE_HOST_FLAGS that you could turn off,
> if you wish to manage all the host flag propagation manually by using
> the OPTION -Xcompiler combination.
> >
> > James
> >
> The CMake build script is identicle to the others I posted, just
> different source files and executable names. I included it anyway
> though, in case there was something I missed. I also double-checked
> the CUDA_HOST_COMPILATION flag, it's definitely set to 1. If you still
> think it would help, where can I find cuda_compile.cmake?
> 
> I've attached the complete 'make VERBOSE=1' for both the
> FindCUDA.cmake sample code, and one with some of the warnings we've
> been getting for our own code. The CMakeCache.txt is from the ITK
> build I'm using, so you can see some of the variables it had defined
> internally. If you've seen the CMakeLists.txt I sent yesterday, it
> includes an ITK file USE_ITK_FILE (it's just in the recommendded
> build, points to the attached UseITK.cmake), which is where the new
> compile flags are introduced.
> 
> Thanks for the tip about the CUDA_PROPAGATE_HOST_FLAGS and OPTION
> -Xcompiler, I'll try it out if all else fails. And thanks for the
> quick reply and you help with this,
> 
> Luke
> 
> 
> 
> Thanks. If it's possible, I need the output after adding the -v
> argument to the CUDA_NVCC_FLAGS. This will show all the compilation
> steps from nvcc and then we can see what commands nvcc is passing to
> the compiler as well as what source files its using.
> 
> Also, does it produce the warning when you don't specify -Wall? You
> may be able to suppress this flag with a -Wno-all for CUDA code.
> 
> The cuda_compile.cmake file I talked about is part of the build
> system. Each CUDA file generates its own compilation script which is
> then invoked by the build system. The one for test_bin.cu in the svn
> test system should be found here:
> 
> build/src/CMakeFiles/cuda_compile_generated_test_bin.cu.obj.cmake
> 
> This file contains all the parameters (well most of them) that are
> used to compile the CUDA file. You don't have to send it to me now. I
> can see what I need to from the verbose output.
> 
> Thanks,
> James

The output I attached was produced with the '-v' flag argument added to the CUDA_NVCC_FLAGS variable.  If the output is not what you expected, could it be that the ITK variables are overwriting the FindCUDA.cmake flags?

I tried adding -Wno-all to CUDA_NVCC_FLAGS along with -v, but it wouldn't compile, saying 'unknown flag 'v -w'' or 'v -Wno-all'.  I think there's definitely something happening with that variable, because when I added both options to the CMAKE_CXX_FLAGS variable, it worked fine.  I'll post the output with just -v in the CMAKE_CXX_FLAGS variable, just in case it helps.

By itself in the CMAKE_CXX_FLAGS variable, -w suppresses every warning but "cc1: warning: command line option "-Wno-deprecated" is valid for C++/Java/ObjC++ but not for C", and with -Wno-all I get "cc1plus: error: unrecognized command line option "-Wno-all"" with all the old warnings.  But that's somewhat helpful, since that means that passing -w in CMAKE_CXX_FLAGS makes it through to nvcc and blocks all the C++ warnings there.

Hopefully the output I got by putting -v in the CMAKE_CXX_FLAGS variable will help.

Luke
-------------- next part --------------
A non-text attachment was scrubbed...
Name: CUDA_verbose_without_make_VERBOSE.tar.gz
Type: application/x-gzip
Size: 204 bytes
Desc: not available
URL: <http://www.cmake.org/pipermail/cmake/attachments/20100111/428c21af/attachment.bin>