OpenCVDetectCUDA.cmake 20.8 KB

Edit Raw Blame History Permalink

if((WIN32 AND NOT MSVC) OR OPENCV_CMAKE_FORCE_CUDA)
  message(STATUS "CUDA compilation is disabled (due to only Visual Studio compiler supported on your platform).")
  return()
endif()

if((NOT UNIX AND CV_CLANG) OR OPENCV_CMAKE_FORCE_CUDA)
  message(STATUS "CUDA compilation is disabled (due to Clang unsupported on your platform).")
  return()
endif()

#set(OPENCV_CMAKE_CUDA_DEBUG 1)

if(((NOT CMAKE_VERSION VERSION_LESS "3.9.0")  # requires https://gitlab.kitware.com/cmake/cmake/merge_requests/663
      OR OPENCV_CUDA_FORCE_EXTERNAL_CMAKE_MODULE)
    AND NOT OPENCV_CUDA_FORCE_BUILTIN_CMAKE_MODULE)
  ocv_update(CUDA_LINK_LIBRARIES_KEYWORD "PRIVATE")
  find_host_package(CUDA "${MIN_VER_CUDA}" QUIET)
else()
  # Use OpenCV's patched "FindCUDA" module
  set(CMAKE_MODULE_PATH "${OpenCV_SOURCE_DIR}/cmake" ${CMAKE_MODULE_PATH})

  if(ANDROID)
    set(CUDA_TARGET_OS_VARIANT "Android")
  endif()
  find_host_package(CUDA "${MIN_VER_CUDA}" QUIET)

  list(REMOVE_AT CMAKE_MODULE_PATH 0)
endif()

if(CUDA_FOUND)
  set(HAVE_CUDA 1)
  if(NOT CUDA_VERSION VERSION_LESS 11.0)
    # CUDA 11.0 removes nppicom
    ocv_list_filterout(CUDA_nppi_LIBRARY "nppicom")
    ocv_list_filterout(CUDA_npp_LIBRARY "nppicom")
  endif()

  if(WITH_CUFFT)
    set(HAVE_CUFFT 1)
  endif()

  if(WITH_CUBLAS)
    set(HAVE_CUBLAS 1)
  endif()

  if(WITH_CUDNN)
      set(CMAKE_MODULE_PATH "${OpenCV_SOURCE_DIR}/cmake" ${CMAKE_MODULE_PATH})
      find_host_package(CUDNN "${MIN_VER_CUDNN}")
      list(REMOVE_AT CMAKE_MODULE_PATH 0)

      if(CUDNN_FOUND)
        set(HAVE_CUDNN 1)
      endif()
  endif()

  if(WITH_NVCUVID)
    macro(ocv_cuda_SEARCH_NVCUVID_HEADER _filename _result)
      # place header file under CUDA_TOOLKIT_TARGET_DIR or CUDA_TOOLKIT_ROOT_DIR
      find_path(_header_result
        ${_filename}
        PATHS "${CUDA_TOOLKIT_TARGET_DIR}" "${CUDA_TOOLKIT_ROOT_DIR}"
        ENV CUDA_PATH
        ENV CUDA_INC_PATH
        PATH_SUFFIXES include
        NO_DEFAULT_PATH
        )
      if("x${_header_result}" STREQUAL "x_header_result-NOTFOUND")
        set(${_result} 0)
      else()
        set(${_result} 1)
      endif()
      unset(_header_result CACHE)
    endmacro()
    ocv_cuda_SEARCH_NVCUVID_HEADER("nvcuvid.h" HAVE_NVCUVID_HEADER)
    ocv_cuda_SEARCH_NVCUVID_HEADER("dynlink_nvcuvid.h" HAVE_DYNLINK_NVCUVID_HEADER)
    find_cuda_helper_libs(nvcuvid)
    if(WIN32)
      find_cuda_helper_libs(nvcuvenc)
    endif()
    if(CUDA_nvcuvid_LIBRARY AND (${HAVE_NVCUVID_HEADER} OR ${HAVE_DYNLINK_NVCUVID_HEADER}))
      # make sure to have both header and library before enabling
      set(HAVE_NVCUVID 1)
    endif()
    if(CUDA_nvcuvenc_LIBRARY)
      set(HAVE_NVCUVENC 1)
    endif()
  endif()

  message(STATUS "CUDA detected: " ${CUDA_VERSION})

  OCV_OPTION(CUDA_ENABLE_DEPRECATED_GENERATION "Enable deprecated generations in the list" OFF)
  set(_generations "Maxwell" "Pascal" "Volta" "Turing" "Ampere")
  if(CUDA_ENABLE_DEPRECATED_GENERATION)
    set(_generations "Fermi" "${_generations}")
    set(_generations "Kepler" "${_generations}")
  endif()
  set(_arch_fermi   "2.0")
  set(_arch_kepler  "3.0;3.5;3.7")
  set(_arch_maxwell "5.0;5.2")
  set(_arch_pascal  "6.0;6.1")
  set(_arch_volta   "7.0")
  set(_arch_turing  "7.5")
  set(_arch_ampere  "8.0;8.6")
  if(NOT CMAKE_CROSSCOMPILING)
    list(APPEND _generations "Auto")
  endif()
  set(CUDA_GENERATION "" CACHE STRING "Build CUDA device code only for specific GPU architecture. Leave empty to build for all architectures.")
  if( CMAKE_VERSION VERSION_GREATER "2.8" )
    set_property( CACHE CUDA_GENERATION PROPERTY STRINGS "" ${_generations} )
  endif()

  if(CUDA_GENERATION)
    if(NOT ";${_generations};" MATCHES ";${CUDA_GENERATION};")
      string(REPLACE ";" ", " _generations "${_generations}")
      message(FATAL_ERROR "ERROR: ${_generations} Generations are supported.")
    endif()
    unset(CUDA_ARCH_BIN CACHE)
    unset(CUDA_ARCH_PTX CACHE)
  endif()

  if(OPENCV_CUDA_DETECTION_NVCC_FLAGS MATCHES "-ccbin")
    # already specified by user
  elseif(CUDA_HOST_COMPILER AND EXISTS "${CUDA_HOST_COMPILER}")
    get_filename_component(c_compiler_realpath "${CMAKE_C_COMPILER}" REALPATH)
    # C compiler doesn't work with --run option, forcing C++ compiler instead
    if(CUDA_HOST_COMPILER STREQUAL c_compiler_realpath OR CUDA_HOST_COMPILER STREQUAL CMAKE_C_COMPILER)
      if(DEFINED CMAKE_CXX_COMPILER)
        get_filename_component(cxx_compiler_realpath "${CMAKE_CXX_COMPILER}" REALPATH)
        LIST(APPEND OPENCV_CUDA_DETECTION_NVCC_FLAGS -ccbin "${cxx_compiler_realpath}")
      else()
        message(STATUS "CUDA: CMAKE_CXX_COMPILER is not available. You may need to specify CUDA_HOST_COMPILER.")
      endif()
    else()
      LIST(APPEND OPENCV_CUDA_DETECTION_NVCC_FLAGS -ccbin "${CUDA_HOST_COMPILER}")
    endif()
  elseif(WIN32 AND CMAKE_LINKER) # Workaround for VS cl.exe not being in the env. path
    get_filename_component(host_compiler_bindir ${CMAKE_LINKER} DIRECTORY)
    LIST(APPEND OPENCV_CUDA_DETECTION_NVCC_FLAGS -ccbin "${host_compiler_bindir}")
  else()
    if(CUDA_HOST_COMPILER)
      message(STATUS "CUDA: CUDA_HOST_COMPILER='${CUDA_HOST_COMPILER}' is not valid, autodetection may not work. Specify OPENCV_CUDA_DETECTION_NVCC_FLAGS with -ccbin option for fix that")
    endif()
  endif()

  macro(ocv_filter_available_architecture result_list)
    set(__cache_key_check "${ARGN} : ${CUDA_NVCC_EXECUTABLE} ${OPENCV_CUDA_DETECTION_NVCC_FLAGS}")
    if(DEFINED OPENCV_CACHE_CUDA_SUPPORTED_CC AND OPENCV_CACHE_CUDA_SUPPORTED_CC_check STREQUAL __cache_key_check)
      set(${result_list} "${OPENCV_CACHE_CUDA_SUPPORTED_CC}")
    else()
      set(CC_LIST ${ARGN})
      foreach(target_arch ${CC_LIST})
        string(REPLACE "." "" target_arch_short "${target_arch}")
        set(NVCC_OPTION "-gencode;arch=compute_${target_arch_short},code=sm_${target_arch_short}")
        set(_cmd "${CUDA_NVCC_EXECUTABLE}" ${OPENCV_CUDA_DETECTION_NVCC_FLAGS} ${NVCC_OPTION} "${OpenCV_SOURCE_DIR}/cmake/checks/OpenCVDetectCudaArch.cu" --compile)
        execute_process(
            COMMAND ${_cmd}
            WORKING_DIRECTORY "${CMAKE_BINARY_DIR}${CMAKE_FILES_DIRECTORY}/CMakeTmp/"
            RESULT_VARIABLE _nvcc_res
            OUTPUT_VARIABLE _nvcc_out
            ERROR_VARIABLE _nvcc_err
            #ERROR_QUIET
            OUTPUT_STRIP_TRAILING_WHITESPACE
        )
        if(OPENCV_CMAKE_CUDA_DEBUG)
          message(WARNING "COMMAND: ${_cmd}")
          message(STATUS "Result: ${_nvcc_res}")
          message(STATUS "Out: ${_nvcc_out}")
          message(STATUS "Err: ${_nvcc_err}")
        endif()
        if(_nvcc_res EQUAL 0)
          LIST(APPEND ${result_list} "${target_arch}")
        endif()
      endforeach()
      string(STRIP "${${result_list}}" ${result_list})
      if(" ${${result_list}}" STREQUAL " ")
        message(WARNING "CUDA: Autodetection arch list is empty. Please enable OPENCV_CMAKE_CUDA_DEBUG=1 and check/specify OPENCV_CUDA_DETECTION_NVCC_FLAGS variable")
      endif()

      # cache detected values
      set(OPENCV_CACHE_CUDA_SUPPORTED_CC ${${result_list}} CACHE INTERNAL "")
      set(OPENCV_CACHE_CUDA_SUPPORTED_CC_check "${__cache_key_check}" CACHE INTERNAL "")
    endif()
  endmacro()

  macro(ocv_detect_native_cuda_arch status output)
    set(OPENCV_CUDA_DETECT_ARCHS_COMMAND "${CUDA_NVCC_EXECUTABLE}" ${OPENCV_CUDA_DETECTION_NVCC_FLAGS} "${OpenCV_SOURCE_DIR}/cmake/checks/OpenCVDetectCudaArch.cu" "--run")
    set(__cache_key_check "${OPENCV_CUDA_DETECT_ARCHS_COMMAND}")
    if(DEFINED OPENCV_CACHE_CUDA_ACTIVE_CC AND OPENCV_CACHE_CUDA_ACTIVE_CC_check STREQUAL __cache_key_check)
      set(${output} "${OPENCV_CACHE_CUDA_ACTIVE_CC}")
      set(${status} 0)
    else()
      execute_process(
          COMMAND ${OPENCV_CUDA_DETECT_ARCHS_COMMAND}
          WORKING_DIRECTORY "${CMAKE_BINARY_DIR}${CMAKE_FILES_DIRECTORY}/CMakeTmp/"
          RESULT_VARIABLE ${status}
          OUTPUT_VARIABLE _nvcc_out
          ERROR_VARIABLE _nvcc_err
          ERROR_QUIET
          OUTPUT_STRIP_TRAILING_WHITESPACE
      )
      if(OPENCV_CMAKE_CUDA_DEBUG)
        message(WARNING "COMMAND: ${OPENCV_CUDA_DETECT_ARCHS_COMMAND}")
        message(STATUS "Result: ${${status}}")
        message(STATUS "Out: ${_nvcc_out}")
        message(STATUS "Err: ${_nvcc_err}")
      endif()
      string(REGEX REPLACE ".*\n" "" ${output} "${_nvcc_out}") #Strip leading warning messages, if any

      if(${status} EQUAL 0)
        # cache detected values
        set(OPENCV_CACHE_CUDA_ACTIVE_CC ${${output}} CACHE INTERNAL "")
        set(OPENCV_CACHE_CUDA_ACTIVE_CC_check "${__cache_key_check}" CACHE INTERNAL "")
      endif()
    endif()
  endmacro()

  set(__cuda_arch_ptx "")
  if(CUDA_GENERATION STREQUAL "Fermi")
    set(__cuda_arch_bin ${_arch_fermi})
  elseif(CUDA_GENERATION STREQUAL "Kepler")
    set(__cuda_arch_bin ${_arch_kepler})
  elseif(CUDA_GENERATION STREQUAL "Maxwell")
    set(__cuda_arch_bin ${_arch_maxwell})
  elseif(CUDA_GENERATION STREQUAL "Pascal")
    set(__cuda_arch_bin ${_arch_pascal})
  elseif(CUDA_GENERATION STREQUAL "Volta")
    set(__cuda_arch_bin ${_arch_volta})
  elseif(CUDA_GENERATION STREQUAL "Turing")
    set(__cuda_arch_bin ${_arch_turing})
  elseif(CUDA_GENERATION STREQUAL "Ampere")
    set(__cuda_arch_bin ${_arch_ampere})
  elseif(CUDA_GENERATION STREQUAL "Auto")
    ocv_detect_native_cuda_arch(_nvcc_res _nvcc_out)
    if(NOT _nvcc_res EQUAL 0)
      message(STATUS "Automatic detection of CUDA generation failed. Going to build for all known architectures.")
    else()
      string(REGEX MATCHALL "[0-9]+\\.[0-9]" __cuda_arch_bin "${_nvcc_out}")
    endif()
  elseif(CUDA_ARCH_BIN)
    message(STATUS "CUDA: Using CUDA_ARCH_BIN=${CUDA_ARCH_BIN}")
    set(__cuda_arch_bin ${CUDA_ARCH_BIN})
  endif()

  if(NOT DEFINED __cuda_arch_bin)
    if(ARM)
      set(__cuda_arch_bin "3.2")
      set(__cuda_arch_ptx "")
    elseif(AARCH64)
      if(NOT CMAKE_CROSSCOMPILING)
        ocv_detect_native_cuda_arch(_nvcc_res _nvcc_out)
      else()
        set(_nvcc_res -1)  # emulate error, see below
      endif()
      if(NOT _nvcc_res EQUAL 0)
        message(STATUS "Automatic detection of CUDA generation failed. Going to build for all known architectures.")
        # TX1 (5.3) TX2 (6.2) Xavier (7.2) V100 (7.0)
        ocv_filter_available_architecture(__cuda_arch_bin
            5.3
            6.2
            7.2
            7.0
        )
      else()
        set(__cuda_arch_bin "${_nvcc_out}")
      endif()
      set(__cuda_arch_ptx "")
    else()
      ocv_filter_available_architecture(__cuda_arch_bin
          ${_arch_fermi}
          ${_arch_kepler}
          ${_arch_maxwell}
          ${_arch_pascal}
          ${_arch_volta}
          ${_arch_turing}
          ${_arch_ampere}
      )
    endif()
  endif()

  set(CUDA_ARCH_BIN ${__cuda_arch_bin} CACHE STRING "Specify 'real' GPU architectures to build binaries for, BIN(PTX) format is supported")
  set(CUDA_ARCH_PTX ${__cuda_arch_ptx} CACHE STRING "Specify 'virtual' PTX architectures to build PTX intermediate code for")

  string(REGEX REPLACE "\\." "" ARCH_BIN_NO_POINTS "${CUDA_ARCH_BIN}")
  string(REGEX REPLACE "\\." "" ARCH_PTX_NO_POINTS "${CUDA_ARCH_PTX}")

  # Check if user specified 1.0/2.1 compute capability: we don't support it
  macro(ocv_wipeout_deprecated_cc target_cc)
    if(" ${CUDA_ARCH_BIN} ${CUDA_ARCH_PTX}" MATCHES " ${target_cc}")
      message(SEND_ERROR "CUDA: ${target_cc} compute capability is not supported - exclude it from ARCH/PTX list and re-run CMake")
    endif()
  endmacro()
  ocv_wipeout_deprecated_cc("1.0")
  ocv_wipeout_deprecated_cc("2.1")

  # NVCC flags to be set
  set(NVCC_FLAGS_EXTRA "")

  # These vars will be passed into the templates
  set(OPENCV_CUDA_ARCH_BIN "")
  set(OPENCV_CUDA_ARCH_PTX "")
  set(OPENCV_CUDA_ARCH_FEATURES "")

  # Tell NVCC to add binaries for the specified GPUs
  string(REGEX MATCHALL "[0-9()]+" ARCH_LIST "${ARCH_BIN_NO_POINTS}")
  foreach(ARCH IN LISTS ARCH_LIST)
    if(ARCH MATCHES "([0-9]+)\\(([0-9]+)\\)")
      # User explicitly specified PTX for the concrete BIN
      set(NVCC_FLAGS_EXTRA ${NVCC_FLAGS_EXTRA} -gencode arch=compute_${CMAKE_MATCH_2},code=sm_${CMAKE_MATCH_1})
      set(OPENCV_CUDA_ARCH_BIN "${OPENCV_CUDA_ARCH_BIN} ${CMAKE_MATCH_1}")
      set(OPENCV_CUDA_ARCH_FEATURES "${OPENCV_CUDA_ARCH_FEATURES} ${CMAKE_MATCH_2}")
    else()
      # User didn't explicitly specify PTX for the concrete BIN, we assume PTX=BIN
      set(NVCC_FLAGS_EXTRA ${NVCC_FLAGS_EXTRA} -gencode arch=compute_${ARCH},code=sm_${ARCH})
      set(OPENCV_CUDA_ARCH_BIN "${OPENCV_CUDA_ARCH_BIN} ${ARCH}")
      set(OPENCV_CUDA_ARCH_FEATURES "${OPENCV_CUDA_ARCH_FEATURES} ${ARCH}")
    endif()
  endforeach()
  set(NVCC_FLAGS_EXTRA ${NVCC_FLAGS_EXTRA} -D_FORCE_INLINES)

  # Tell NVCC to add PTX intermediate code for the specified architectures
  string(REGEX MATCHALL "[0-9]+" ARCH_LIST "${ARCH_PTX_NO_POINTS}")
  foreach(ARCH IN LISTS ARCH_LIST)
    set(NVCC_FLAGS_EXTRA ${NVCC_FLAGS_EXTRA} -gencode arch=compute_${ARCH},code=compute_${ARCH})
    set(OPENCV_CUDA_ARCH_PTX "${OPENCV_CUDA_ARCH_PTX} ${ARCH}")
    set(OPENCV_CUDA_ARCH_FEATURES "${OPENCV_CUDA_ARCH_FEATURES} ${ARCH}")
  endforeach()

  # These vars will be processed in other scripts
  set(CUDA_NVCC_FLAGS ${CUDA_NVCC_FLAGS} ${NVCC_FLAGS_EXTRA})
  set(OpenCV_CUDA_CC "${NVCC_FLAGS_EXTRA}")

  if(ANDROID)
    set(CUDA_NVCC_FLAGS ${CUDA_NVCC_FLAGS} "-Xptxas;-dlcm=ca")
  endif()

  message(STATUS "CUDA NVCC target flags: ${CUDA_NVCC_FLAGS}")

  OCV_OPTION(CUDA_FAST_MATH "Enable --use_fast_math for CUDA compiler " OFF)

  if(CUDA_FAST_MATH)
    set(CUDA_NVCC_FLAGS ${CUDA_NVCC_FLAGS} --use_fast_math)
  endif()

  mark_as_advanced(CUDA_BUILD_CUBIN CUDA_BUILD_EMULATION CUDA_VERBOSE_BUILD CUDA_SDK_ROOT_DIR)

  macro(ocv_cuda_filter_options)
    foreach(var CMAKE_CXX_FLAGS CMAKE_CXX_FLAGS_RELEASE CMAKE_CXX_FLAGS_DEBUG)
      set(${var}_backup_in_cuda_compile_ "${${var}}")

      if (CV_CLANG)
        # we remove -Winconsistent-missing-override and -Qunused-arguments
        # just in case we are compiling CUDA with gcc but OpenCV with clang
        string(REPLACE "-Winconsistent-missing-override" "" ${var} "${${var}}")
        string(REPLACE "-Qunused-arguments" "" ${var} "${${var}}")
      endif()

      # we remove /EHa as it generates warnings under windows
      string(REPLACE "/EHa" "" ${var} "${${var}}")

      # we remove -ggdb3 flag as it leads to preprocessor errors when compiling CUDA files (CUDA 4.1)
      string(REPLACE "-ggdb3" "" ${var} "${${var}}")

      # we remove -Wsign-promo as it generates warnings under linux
      string(REPLACE "-Wsign-promo" "" ${var} "${${var}}")

      # we remove -Wno-sign-promo as it generates warnings under linux
      string(REPLACE "-Wno-sign-promo" "" ${var} "${${var}}")

      # we remove -Wno-delete-non-virtual-dtor because it's used for C++ compiler
      # but NVCC uses C compiler by default
      string(REPLACE "-Wno-delete-non-virtual-dtor" "" ${var} "${${var}}")

      # we remove -frtti because it's used for C++ compiler
      # but NVCC uses C compiler by default
      string(REPLACE "-frtti" "" ${var} "${${var}}")

      string(REPLACE "-fvisibility-inlines-hidden" "" ${var} "${${var}}")

      # cc1: warning: command line option '-Wsuggest-override' is valid for C++/ObjC++ but not for C
      string(REPLACE "-Wsuggest-override" "" ${var} "${${var}}")

      # issue: #11552 (from OpenCVCompilerOptions.cmake)
      string(REGEX REPLACE "-Wimplicit-fallthrough(=[0-9]+)? " "" ${var} "${${var}}")

      # removal of custom specified options
      if(OPENCV_CUDA_NVCC_FILTEROUT_OPTIONS)
        foreach(__flag ${OPENCV_CUDA_NVCC_FILTEROUT_OPTIONS})
          string(REPLACE "${__flag}" "" ${var} "${${var}}")
        endforeach()
      endif()
    endforeach()
  endmacro()

  macro(ocv_cuda_compile VAR)
    ocv_cuda_filter_options()

    if(BUILD_SHARED_LIBS)
      set(CUDA_NVCC_FLAGS ${CUDA_NVCC_FLAGS} -Xcompiler -DCVAPI_EXPORTS)
    endif()

    if(UNIX OR APPLE)
      set(CUDA_NVCC_FLAGS ${CUDA_NVCC_FLAGS} -Xcompiler -fPIC)
      if(NOT " ${CMAKE_CXX_FLAGS} ${CMAKE_CXX_FLAGS_RELEASE} ${CMAKE_CXX_FLAGS_DEBUG} ${CUDA_NVCC_FLAGS}" MATCHES "-std=")
        if(CUDA_VERSION VERSION_LESS "11.0")
          list(APPEND CUDA_NVCC_FLAGS "--std=c++11")
        else()
          list(APPEND CUDA_NVCC_FLAGS "--std=c++14")
        endif()
      endif()
    endif()
    if(APPLE)
      set(CUDA_NVCC_FLAGS ${CUDA_NVCC_FLAGS} -Xcompiler -fno-finite-math-only)
    endif()

    if(CMAKE_CROSSCOMPILING AND (ARM OR AARCH64))
      set(CUDA_NVCC_FLAGS ${CUDA_NVCC_FLAGS} -Xlinker --unresolved-symbols=ignore-in-shared-libs)
    endif()

    # disabled because of multiple warnings during building nvcc auto generated files
    if(CV_GCC AND CMAKE_CXX_COMPILER_VERSION VERSION_GREATER "4.6.0")
      ocv_warnings_disable(CMAKE_CXX_FLAGS -Wunused-but-set-variable)
    endif()

    CUDA_COMPILE(${VAR} ${ARGN})

    foreach(var CMAKE_CXX_FLAGS CMAKE_CXX_FLAGS_RELEASE CMAKE_CXX_FLAGS_DEBUG)
      set(${var} "${${var}_backup_in_cuda_compile_}")
      unset(${var}_backup_in_cuda_compile_)
    endforeach()
  endmacro()
else()
  unset(CUDA_ARCH_BIN CACHE)
  unset(CUDA_ARCH_PTX CACHE)
endif()

if(HAVE_CUDA)
  set(CUDA_LIBS_PATH "")
  foreach(p ${CUDA_LIBRARIES} ${CUDA_npp_LIBRARY})
    get_filename_component(_tmp ${p} PATH)
    list(APPEND CUDA_LIBS_PATH ${_tmp})
  endforeach()

  if(HAVE_CUBLAS)
    foreach(p ${CUDA_cublas_LIBRARY})
      get_filename_component(_tmp ${p} PATH)
      list(APPEND CUDA_LIBS_PATH ${_tmp})
    endforeach()
  endif()

  if(HAVE_CUDNN)
    foreach(p ${CUDNN_LIBRARIES})
      get_filename_component(_tmp ${p} PATH)
      list(APPEND CUDA_LIBS_PATH ${_tmp})
    endforeach()
  endif()

  if(HAVE_CUFFT)
    foreach(p ${CUDA_cufft_LIBRARY})
      get_filename_component(_tmp ${p} PATH)
      list(APPEND CUDA_LIBS_PATH ${_tmp})
    endforeach()
  endif()

  list(REMOVE_DUPLICATES CUDA_LIBS_PATH)
  link_directories(${CUDA_LIBS_PATH})

  set(CUDA_LIBRARIES_ABS ${CUDA_LIBRARIES})
  ocv_convert_to_lib_name(CUDA_LIBRARIES ${CUDA_LIBRARIES})
  set(CUDA_npp_LIBRARY_ABS ${CUDA_npp_LIBRARY})
  ocv_convert_to_lib_name(CUDA_npp_LIBRARY ${CUDA_npp_LIBRARY})
  if(HAVE_CUBLAS)
    set(CUDA_cublas_LIBRARY_ABS ${CUDA_cublas_LIBRARY})
    ocv_convert_to_lib_name(CUDA_cublas_LIBRARY ${CUDA_cublas_LIBRARY})
  endif()
  if(HAVE_CUDNN)
    set(CUDNN_LIBRARIES_ABS ${CUDNN_LIBRARIES})
    ocv_convert_to_lib_name(CUDNN_LIBRARIES ${CUDNN_LIBRARIES})
  endif()
  if(HAVE_CUFFT)
    set(CUDA_cufft_LIBRARY_ABS ${CUDA_cufft_LIBRARY})
    ocv_convert_to_lib_name(CUDA_cufft_LIBRARY ${CUDA_cufft_LIBRARY})
  endif()

  if(CMAKE_GENERATOR MATCHES "Visual Studio"
      AND NOT OPENCV_SKIP_CUDA_CMAKE_SUPPRESS_REGENERATION
  )
    message(STATUS "CUDA: MSVS generator is detected. Disabling CMake re-run checks (CMAKE_SUPPRESS_REGENERATION=ON). You need to run CMake manually if updates are required.")
    set(CMAKE_SUPPRESS_REGENERATION ON)
  endif()
endif()


# ----------------------------------------------------------------------------
# Add CUDA libraries (needed for apps/tools, samples)
# ----------------------------------------------------------------------------
if(HAVE_CUDA)
  # details: https://github.com/NVIDIA/nvidia-docker/issues/775
  if(" ${CUDA_CUDA_LIBRARY}" MATCHES "/stubs/libcuda.so" AND NOT OPENCV_SKIP_CUDA_STUB_WORKAROUND)
    set(CUDA_STUB_ENABLED_LINK_WORKAROUND 1)
    if(EXISTS "${CUDA_CUDA_LIBRARY}" AND NOT OPENCV_SKIP_CUDA_STUB_WORKAROUND_RPATH_LINK)
      set(CUDA_STUB_TARGET_PATH "${CMAKE_BINARY_DIR}${CMAKE_FILES_DIRECTORY}/")
      execute_process(COMMAND ${CMAKE_COMMAND} -E create_symlink "${CUDA_CUDA_LIBRARY}" "${CUDA_STUB_TARGET_PATH}/libcuda.so.1"
          RESULT_VARIABLE CUDA_STUB_SYMLINK_RESULT)
      if(NOT CUDA_STUB_SYMLINK_RESULT EQUAL 0)
        execute_process(COMMAND ${CMAKE_COMMAND} -E copy_if_different "${CUDA_CUDA_LIBRARY}" "${CUDA_STUB_TARGET_PATH}/libcuda.so.1"
          RESULT_VARIABLE CUDA_STUB_COPY_RESULT)
        if(NOT CUDA_STUB_COPY_RESULT EQUAL 0)
          set(CUDA_STUB_ENABLED_LINK_WORKAROUND 0)
        endif()
      endif()
      if(CUDA_STUB_ENABLED_LINK_WORKAROUND)
        set(CMAKE_EXE_LINKER_FLAGS "${CMAKE_EXE_LINKER_FLAGS} -Wl,-rpath-link,\"${CUDA_STUB_TARGET_PATH}\"")
      endif()
    else()
      set(CMAKE_EXE_LINKER_FLAGS "${CMAKE_EXE_LINKER_FLAGS} -Wl,--allow-shlib-undefined")
    endif()
    if(NOT CUDA_STUB_ENABLED_LINK_WORKAROUND)
      message(WARNING "CUDA: workaround for stubs/libcuda.so.1 is not applied")
    endif()
  endif()

  set(OPENCV_LINKER_LIBS ${OPENCV_LINKER_LIBS} ${CUDA_LIBRARIES} ${CUDA_npp_LIBRARY})
  if(HAVE_CUBLAS)
    set(OPENCV_LINKER_LIBS ${OPENCV_LINKER_LIBS} ${CUDA_cublas_LIBRARY})
  endif()
  if(HAVE_CUDNN)
    set(OPENCV_LINKER_LIBS ${OPENCV_LINKER_LIBS} ${CUDNN_LIBRARIES})
  endif()
  if(HAVE_CUFFT)
    set(OPENCV_LINKER_LIBS ${OPENCV_LINKER_LIBS} ${CUDA_cufft_LIBRARY})
  endif()
  foreach(p ${CUDA_LIBS_PATH})
    if(MSVC AND CMAKE_GENERATOR MATCHES "Ninja|JOM")
      set(OPENCV_LINKER_LIBS ${OPENCV_LINKER_LIBS} ${CMAKE_LIBRARY_PATH_FLAG}"${p}")
    else()
      set(OPENCV_LINKER_LIBS ${OPENCV_LINKER_LIBS} ${CMAKE_LIBRARY_PATH_FLAG}${p})
    endif()
  endforeach()
endif()