[CMake] [New Module] FindCUDA

[James Bigler]

On Wed, Aug 26, 2009 at 2:36 PM, Luis Ibanez<luis.ibanez at kitware.com> wrote:
> On Tue, Aug 25, 2009 at 2:11 PM, James Bigler<jamesbigler at gmail.com> wrote:
>> This is code that is up for inclusion into the CMake Modules directory.
>>
>> FindCUDA does several things.
>>
>> 1. It locates all the build tools found in the NVIDIA CUDA toolkit and SDK.
>> 2. It provides common options used to control CUDA code compilation.
>> 3. It provides macros that setup the build rules to convert CUDA files
>> into object code or PTX files.
>> 4. It maintains source level dependencies.  If you change a file that
>> a CUDA file depends on, it will cause the CUDA file to recompile.
>>
>> Attached is a zip file that contains the 4 files needed for use.  I
>> needed to use a zip file to get below the 40 KB email size limit that
>> didn't allow my email through before.
>>
>> 1. FindCUDA.cmake - main file that users call.
>> 2. run_nvcc.cmake - script used to invoke the nvcc compiler and other
>> helper commands (such as generating the dependency file).  This file
>> is configured per CUDA file, so that if you change one of the compiler
>> options it will regenerate this file causing the CUDA file to be
>> compiled anew.
>> 3. make2cmake - converts the makefile formatted dependency file that
>> nvcc produces into a form suitable for CMake to include.
>> 4. parse_cubin.cmake - parses a cubin file generated by nvcc and
>> reports statistics about the file.
>>
>> Comments are welcome.
>>
>> A version that also has some sample driver code (as well as the
>> FindCUDA scripts) can be found here:
>>
>> svn checkout https://code.sci.utah.edu/svn/findcuda/trunk FindCUDA
>>
>> Thanks,
>> James
>>
>> _______________________________________________
>> Powered by www.kitware.com
>>
>> Visit other Kitware open-source projects at http://www.kitware.com/opensource/opensource.html
>>
>> Please keep messages on-topic and check the CMake FAQ at: http://www.cmake.org/Wiki/CMake_FAQ
>>
>> Follow this link to subscribe/unsubscribe:
>> http://www.cmake.org/mailman/listinfo/cmake
>>
>
> -----------------------------------
>
>
> Hi James,
>
> This is great !
>
> Thanks for putting it together and sharing it.
>
>
> I tested it with the CUDA-ITK directory that we have
> in the NAMIC Sandbox at:
>
> http://svn.na-mic.org/NAMICSandBox/trunk/CUDA-ITK/
>
> where we used to have an older version of these files
> in the subdirectory:
>
> http://svn.na-mic.org/NAMICSandBox/trunk/CUDA-ITK/CMake/cuda/
>
>
> --
>
> Configuration with CMake went smoothly,
>
> but when building, I got the warning messages below
> (at the end of this email).
>
> It would seem that some of the ITK C++ compilation
> flags are passed to the cuda compiler, and get rejected.
>
>
> Does that ring any bell ?
>
>
>  Thanks
>
>
>     Luis
>
>
>
>
> -----------------------
> make
>
>
> ~/bin/NAMIC/NAMICSandbox/CUDA-ITK/Release
> [  7%] Building NVCC (Device) object Source/./Example1_generated_Example1.cu.o
> cc1: warning: command line option "-Wno-deprecated" is valid for
> C++/Java/ObjC++ but not for C
> cc1: warning: command line option "-ftemplate-depth-50" is valid for
> C++/ObjC++ but not for C
> cc1: warning: command line option "-Wno-deprecated" is valid for
> C++/Java/ObjC++ but not for C
> cc1: warning: command line option "-ftemplate-depth-50" is valid for
> C++/ObjC++ but not for C
> In file included from
> /home/ibanez/src/NAMIC/NAMICSandBox/CUDA-ITK/Source/Example1.cu:119:
> /tmp/tmpxft_00005d5f_00000000-1_Example1.cudafe1.stub.c: In function
> ‘void __device_stub__Z15VectorAddKernelPfS_S_i(float*, float*, float*,
> int)’:
> /tmp/tmpxft_00005d5f_00000000-1_Example1.cudafe1.stub.c:9: warning:
> dereferencing type-punned pointer will break strict-aliasing rules
> /home/ibanez/local/Cuda/cuda/bin/../include/common_functions.h: At global scope:
> /home/ibanez/local/Cuda/cuda/bin/../include/common_functions.h:70:
> warning: ‘clock_t __cuda_clock()’ defined but not used
> /home/ibanez/local/Cuda/cuda/bin/../include/common_functions.h:75:
> warning: ‘void* __cuda_memset(void*, int, size_t)’ defined but not
> used
> /home/ibanez/local/Cuda/cuda/bin/../include/common_functions.h:80:
> warning: ‘void* __cuda_memcpy(void*, const void*, size_t)’ defined but
> not used
> Scanning dependencies of target Example1
> Linking CXX executable ../bin/Example1

Yes, all C or CXX flags get passed to the host compiler by nvcc via
the -Xcompiler flag for more consistent builds.  By default
CUDA_HOST_COMPILATION_CPP is ON which should generate C++ host code
and invoke the C++ compiler instead of the C compiler.  I'm curious
why you are getting problems like this.

You can see the commands used to invoke nvcc by setting
CUDA_VERBOSE_BUILD to ON and running the build again.

James