[CMake] "modern" CUDA handling and static libraries

[Robert Maynard]

> If a .cpp file is added to library, then it won't work.

If this is occurring it is a bug. Can you fill a bug report at:
https://gitlab.kitware.com/cmake/cmake

> CUDA_PROPAGATE_HOST_FLAGS

CUDA as a first class language propagates zero C++ flags as it is a
separate language, and we expect people to set CXX_FLAGS and
CUDA_FLAGS

On Wed, Dec 20, 2017 at 3:46 PM, Crni Gorac <cgorac at gmail.com> wrote:
> Thanks for your reply.
>
> That seems to work, but only if the library consist solely of .cu
> files.  If a .cpp file is added to library, then it won't work.  Is
> this a known limitation?
>
> On a related note: what is the equivalent of
> "set(CUDA_PROPAGATE_HOST_FLAGS OFF)" now?
>
> Regards.
>
> On Tue, Dec 19, 2017 at 11:47 PM, Robert Maynard
> <robert.maynard at kitware.com> wrote:
>> Hi,
>>
>> The CUDA_RESOLVE_DEVICE_SYMBOLS property was added for this exact
>> situation. You can find it documented at
>> https://cmake.org/cmake/help/v3.10/prop_tgt/CUDA_RESOLVE_DEVICE_SYMBOLS.html
>>
>> Here is an example of it being used:
>> https://gitlab.kitware.com/cmake/cmake/commit/493671a5212c6548b2d7376c7065f5f76692a792#0eb128c2e695da6e4cedb9a408b108f117bd4fe3
>>
>> On Tue, Dec 19, 2017 at 3:14 PM, Crni Gorac <cgorac at gmail.com> wrote:
>>> Hello,
>>>
>>> My project is alike for what is described by following CMakeLists.txt
>>> for Cuda/Complex test case from CMake source distribution:
>>>
>>> #########################
>>> cmake_minimum_required(VERSION 3.8)
>>> project (CudaComplex CXX CUDA)
>>>
>>> string(APPEND CMAKE_CUDA_FLAGS " -gencode arch=compute_30,code=compute_30")
>>> set(CMAKE_CUDA_STANDARD 11)
>>> set(CMAKE_CXX_STANDARD 11)
>>> set(CMAKE_CUDA_STANDARD_REQUIRED TRUE)
>>> set(CMAKE_CXX_STANDARD_REQUIRED TRUE)
>>>
>>> add_library(CudaLib STATIC dynamic.cu file1.cu file2.cu file3.cu mixed.cu)
>>> set_target_properties(CudaLib PROPERTIES CUDA_SEPARABLE_COMPILATION OFF)
>>> set_target_properties(CudaLib PROPERTIES POSITION_INDEPENDENT_CODE OFF)
>>>
>>> add_library(CppLib STATIC dynamic.cpp mixed.cpp)
>>>
>>> add_executable(CudaComplex main.cpp)
>>> target_link_libraries(CudaComplex PUBLIC CudaLib CppLib)
>>> #########################
>>>
>>> So basically I want to build all my .cu files into one static library,
>>> and all my .cpp files (except for main.cpp) into another static
>>> library. These two libraries are then linked into main.cpp above. This
>>> works fine, and make output is:
>>>
>>> #########################
>>> Scanning dependencies of target CppLib
>>> Scanning dependencies of target CudaLib
>>> [  8%] Building CXX object CMakeFiles/CppLib.dir/mixed.cpp.o
>>> [ 25%] Building CUDA object CMakeFiles/CudaLib.dir/file1.cu.o
>>> [ 25%] Building CUDA object CMakeFiles/CudaLib.dir/file2.cu.o
>>> [ 33%] Building CXX object CMakeFiles/CppLib.dir/dynamic.cpp.o
>>> [ 41%] Building CUDA object CMakeFiles/CudaLib.dir/dynamic.cu.o
>>> [ 50%] Building CUDA object CMakeFiles/CudaLib.dir/file3.cu.o
>>> [ 58%] Building CUDA object CMakeFiles/CudaLib.dir/mixed.cu.o
>>> [ 66%] Linking CXX static library libCppLib.a
>>> [ 66%] Built target CppLib
>>> [ 75%] Linking CUDA static library libCudaLib.a
>>> [ 75%] Built target CudaLib
>>> Scanning dependencies of target CudaComplex
>>> [ 83%] Building CXX object CMakeFiles/CudaComplex.dir/main.cpp.o
>>> [ 91%] Linking CUDA device code CMakeFiles/CudaComplex.dir/cmake_device_link.o
>>> [100%] Linking CXX executable CudaComplex
>>> #########################
>>>
>>> However, in my case these static libraries are actually to be
>>> delivered to a client.  The problem is that the dynamic linking of
>>> CUDA code (this is this line "Linking CUDA device code
>>> CMakeFiles/CudaComplex.dir/cmake_device_link.o" in the make output
>>> above) would still have to get done on client side, i.e. they would
>>> have to install CUDA SDK and provide that this step is accomplished
>>> through their build system.  So my question is: is it possible to
>>> somehow generate static library with CUDA code, using this new
>>> "...LANGUAGE CUDA..." CMake functionality, but that CUDA linking is
>>> completed on my side (note that on client side no CUDA code will be
>>> added to project)?
>>>
>>> Thanks.
>>> --
>>>
>>> Powered by www.kitware.com
>>>
>>> Please keep messages on-topic and check the CMake FAQ at: http://www.cmake.org/Wiki/CMake_FAQ
>>>
>>> Kitware offers various services to support the CMake community. For more information on each offering, please visit:
>>>
>>> CMake Support: http://cmake.org/cmake/help/support.html
>>> CMake Consulting: http://cmake.org/cmake/help/consulting.html
>>> CMake Training Courses: http://cmake.org/cmake/help/training.html
>>>
>>> Visit other Kitware open-source projects at http://www.kitware.com/opensource/opensource.html
>>>
>>> Follow this link to subscribe/unsubscribe:
>>> https://cmake.org/mailman/listinfo/cmake
> --
>
> Powered by www.kitware.com
>
> Please keep messages on-topic and check the CMake FAQ at: http://www.cmake.org/Wiki/CMake_FAQ
>
> Kitware offers various services to support the CMake community. For more information on each offering, please visit:
>
> CMake Support: http://cmake.org/cmake/help/support.html
> CMake Consulting: http://cmake.org/cmake/help/consulting.html
> CMake Training Courses: http://cmake.org/cmake/help/training.html
>
> Visit other Kitware open-source projects at http://www.kitware.com/opensource/opensource.html
>
> Follow this link to subscribe/unsubscribe:
> https://cmake.org/mailman/listinfo/cmake