Search code examples
c++cmakecudalinker-errorsnvidia-jetson-nano

C++ with CUDA project in CMake gives error: nvlink fatal : Could not open input file 'CMakeFiles/MY_APP.dir/src/MY_APP.cpp.o' (target: sm_35)

Hello I was working on a C++ project, and had a cmake file that was working just fine, until I tried to add cuda into the C++ project. I am building this project on the NVIDIA Jetson Nano.

I get this error while building:

nvlink fatal : Could not open input file 'CMakeFiles/MY_APP.dir/src/MY_APP.cpp.o' (target: sm_35)

The rest of the error underneath that looks like this:

CMakeFiles/MY_APP.dir/build.make:552: recipe for target 

'CMakeFiles/MY_APP.dir/cmake_device_link.o' failed
make[2]: *** [CMakeFiles/MY_APP.dir/cmake_device_link.o] Error 1
make[2]: Leaving directory '/home/me/Code/MyApp/build'
CMakeFiles/Makefile2:127: recipe for target 'CMakeFiles/MY_APP.dir/all' failed
make[1]: *** [CMakeFiles/MY_APP.dir/all] Error 2
make[1]: Leaving directory '/home/me/Code/MY_APP/build'
Makefile:155: recipe for target 'all' failed
make: *** [all] Error 2
make: Leaving directory '/home/me/Code/MY_APP/build'

I run my cmake file using a script I called confgure.sh, which looks like this:

#!/bin/sh

cmake -S . -B build -DCUDA_TOOLKIT_ROOT_DIR=/usr/local/cuda-10.2 -DCMAKE_CUDA_COMPILER=/usr/local/cuda-10.2/bin/nvcc

I run my make file using a script I called build.sh, which looks like this:

#!/bin/sh

make -C build

My Cmake File looks like this:

cmake_minimum_required(VERSION 3.21.0)
project(MY_APP VERSION 0.0.0)

set(CMAKE_CXX_STANDARD 14)
set(CMAKE_CXX_STANDARD_REQUIRED ON)
set(CMAKE_CXX_EXTENSIONS OFF)

enable_language(CUDA)

# Pass options to NVCC
set(
    CUDA_NVCC_FLAGS
    ${CUDA_NVCC_FLAGS};
    -O3 -gencode arch=compute_35,code=sm_35
    )

set(CMAKE_CUDA_COMPILER /usr/local/cuda/bin/nvcc)
FILE(GLOB_RECURSE MY_CUDA_SRCS src/*.cu)

configure_file(src/MyAppConfig.h.in MyAppConfig.h)

#collect cpp files
FILE(GLOB_RECURSE SRC src/*.cpp)

find_package(CUDA QUIET)
if(CUDA_FOUND)
    SET(CMAKE_CUDA_COMPILER /usr/local/cuda/bin/nvcc)
    include_directories(${CUDA_INCLUDE_DIRS})
    get_filename_component(CUDA_LIBRARY_DIR ${CUDA_CUDART_LIBRARY} DIRECTORY)
    set(CMAKE_EXE_LINKER_FLAGS ${CMAKE_EXE_LINKER_FLAGS} "-L${CUDA_LIBRARY_DIR}")
    SET(ALL_CUDA_LIBS ${CUDA_LIBRARIES} ${CUDA_cusparse_LIBRARY} ${CUDA_cublas_LIBRARY})
    #${CUDA_CUDART_LIBRARY}
    #${CMAKE_CUDA_RUNTIME_LIBRARY}
    #)
    SET(LIBS ${LIBS} ${ALL_CUDA_LIBS})
    message(STATUS "CUDA_LIBRARIES: ${CUDA_INCLUDE_DIRS} ${ALL_CUDA_LIBS}")
    set(CUDA_PROPAGATE_HOST_FLAGS ON)
    set(CUDA_SEPARABLE_COMPILATION ON)
    list(APPEND CUDA_NVCC_FLAGS -gencode=arch=compute_35,code=sm_35)

    #collect CUDA files
    FILE(GLOB_RECURSE CUDA_SRC src/*.cu)

    #build static library
    #CUDA_ADD_LIBRARY(my_cuda_lib ${CUDA_SRC} STATIC)

    cuda_compile(cuda_objs ${CUDA_SRC})
    SET(SRC ${cuda_objs} ${SRC})

    SET(LIBS ${LIBS} ${my_cuda_lib})
endif()

link_libraries(${cuda_objs})

set_source_files_properties(${SRC} PROPERTIES LANGUAGE CUDA)

message("using cuda_add_executable")
cuda_add_executable(${PROJECT_NAME} ${SRC})

target_include_directories(${PROJECT_NAME} PUBLIC ${PROJECT_BINARY_DIR})

target_link_libraries(${PROJECT_NAME} ${LIBS})

#DOWNLOAD ALL THE SUBMODULES
find_package(Git QUIET)
if (GIT_FOUND AND EXISTS "${PROJECT_SOURCE_DIR}/.git")
# Update submodules as needed
    option(GIT_SUBMODULE, "Check submodules during build" ON)
    if (GIT_SUBMODULE)
        message(STATUS "Submodule update")
        execute_process(COMMAND ${GIT_EXECUTABLE}
                        submodule update --init --recursvie
                        WORKING_DIRECTORY {CMAKE_CURRENT_SOURCE_DIR}
                        RESULT_VARIABLE_GIT_SUBMOD_RESULT)
        if (NOT GIT_SUBMOD_RESULT EQUAL "0")
            message(FATAL_ERROR
                "git submodule update --init failed with ${GIT_SUMOD_RESULT},
                please check submodule")
        endif()
    endif()
endif()

#CHECK ALL THE SUBMODULES
if (NOT EXISTS
"${PROJECT_SOURCE_DIR}/external/Simple-Websocket-Server/CMakeLists.txt")
    message(FATAL_ERROR
            "The Simple-Websocket-Server submodule was not downloaded!
            GIT_SUBMODULE was turned off or failed. Please update submodule")
endif()

add_subdirectory(external/Simple-Websocket-Server)

include_directories(PUBLIC external/Simple-Websocket-Server)

find_package(PythonLibs REQUIRED)
find_package(pybind11 REQUIRED)

include_directories(${PYTHON_INCLUDE_DIRS})

target_link_libraries(${PROJECT_NAME}
    curl pthread crypto boost_system jsoncpp ${PYTHON_LIBRARIES} cudart 
    #<some-of-my-other-libraries>
)

install(TARGETS ${PROJECT_NAME} DESTINATION bin)
install(FILES "${PROJECT_BINARY_DIR}/MyAppConfig.h" DESTINATION include)

include(InstallRequiredSystemLibraries)
set(CPACK_RESOURCE_FILE_LICENSE "${CMAKE_CURRENT_SOURCE_DIR}/License.txt")
set(CPACK_PACKAGE_VERSION_MAJOR "${MY_APP_VERSION_MAJOR}")
set(CPACK_PACKAGE_VERSION_MINOR "${MY_APP_VERSION_MAJOR}")
set(CPACK_PACKAGE_VERSION_PATCH "${MY_APP_VERSION_PATCH}")
include(CPack)

set(CMAKE_CUDA_COMPILE_WHOLE_COMPILATION 
    "${CMAKE_CUDA_COMPILER} ${_CMAKE_CUDA_EXTRA_FLAGS} -c ${MY_CUDA_SRCS}")

message(CMAKE_CUDA_COMPILE_WHOLE_COMPILATION)

I am lost on how to get CUDA added to my project that already contains a bunch of C++ files, and I need to be able to call my .cu files from a .cpp file other then main.cpp, and I need to get this building in CMake, and I am doing it on the jetson nano. Any help on solving this error?

Solution

  • You are mixing up a lot of CUDA-related definitions and commands, including some from an earlier "era" of CUDA support in CMake.

    Among other things:

    1. Your CMakeLists.txt is overriding your environment setting for the CUDA compiler location.
    2. ... and actually, you shouldn't bother setting that location anyway, since you've already set the CUDA toolkit root.
    3. Don't use find_package(CUDA) with CMake versions of 3.17 or 3.18, or later. For all relevant toolkit-related paths, use find_package(CUDAToolkit)`, which does... well, less but also more.
    4. Don't use cuda_add_+suffix commands. Since CMake supports CUDA natively, you use regular add_executable, add_library etc.

    There are further issues with your CMakeLists.txt file - not all of them CUDA-related, but that should be enough to get you started. It may not in itself resolve the specific bottom-line problem you have, though.

    You may want to have a look at public repositories using CUDA and recent CMake versions to get an idea of how this is done.