if(NOT USE_CUDA)
  return()
endif()

include(CheckCXXCompilerFlag)
check_cxx_compiler_flag("-std=c++11"   SUPPORT_CXX11)

################################################################################################
# A function for automatic detection of GPUs installed  (if autodetection is enabled)
# Usage:
#   mshadow_detect_installed_gpus(out_variable)
function(mshadow_detect_installed_gpus out_variable)
set(CUDA_gpu_detect_output "")
  if(NOT CUDA_gpu_detect_output)
    message(STATUS "Running GPU architecture autodetection")
    set(__cufile ${PROJECT_BINARY_DIR}/detect_cuda_archs.cu)

    file(WRITE ${__cufile} ""
      "#include <cstdio>\n"
      "#include <iostream>\n"
      "using namespace std;\n"
      "int main()\n"
      "{\n"
      "  int count = 0;\n"
      "  if (cudaSuccess != cudaGetDeviceCount(&count)) { return -1; }\n"
      "  if (count == 0) { cerr << \"No cuda devices detected\" << endl; return -1; }\n"
      "  for (int device = 0; device < count; ++device)\n"
      "  {\n"
      "    cudaDeviceProp prop;\n"
      "    if (cudaSuccess == cudaGetDeviceProperties(&prop, device))\n"
      "      std::printf(\"%d.%d \", prop.major, prop.minor);\n"
      "  }\n"
      "  return 0;\n"
      "}\n")
    if(MSVC)
      #find vcvarsall.bat and run it building msvc environment
      get_filename_component(MY_COMPILER_DIR ${CMAKE_CXX_COMPILER} DIRECTORY)
      find_file(MY_VCVARSALL_BAT vcvarsall.bat "${MY_COMPILER_DIR}/.." "${MY_COMPILER_DIR}/../..")
      execute_process(COMMAND ${MY_VCVARSALL_BAT} && ${CUDA_NVCC_EXECUTABLE} -arch sm_30 --run  ${__cufile}
                      WORKING_DIRECTORY "${PROJECT_BINARY_DIR}/CMakeFiles/"
                      RESULT_VARIABLE __nvcc_res OUTPUT_VARIABLE __nvcc_out
                      OUTPUT_STRIP_TRAILING_WHITESPACE)
    else()
      if(CUDA_LIBRARY_PATH)
        set(CUDA_LINK_LIBRARY_PATH "-L${CUDA_LIBRARY_PATH}")
      endif()
      execute_process(COMMAND ${CUDA_NVCC_EXECUTABLE} -arch sm_30 --run ${__cufile} ${CUDA_LINK_LIBRARY_PATH}
                      WORKING_DIRECTORY "${PROJECT_BINARY_DIR}/CMakeFiles/"
                      RESULT_VARIABLE __nvcc_res OUTPUT_VARIABLE __nvcc_out
                      OUTPUT_STRIP_TRAILING_WHITESPACE)
    endif()
    if(__nvcc_res EQUAL 0)
      # nvcc outputs text containing line breaks when building with MSVC.
      # The line below prevents CMake from inserting a variable with line
      # breaks in the cache
      message(STATUS "Found CUDA arch ${__nvcc_out}")
      string(REGEX MATCH "([1-9].[0-9])" __nvcc_out "${__nvcc_out}")
      string(REPLACE "2.1" "2.1(2.0)" __nvcc_out "${__nvcc_out}")
      set(CUDA_gpu_detect_output ${__nvcc_out} CACHE INTERNAL "Returned GPU architetures from mshadow_detect_gpus tool" FORCE)
    else()
      message(WARNING "Running GPU detection script with nvcc failed: ${__nvcc_out}")
    endif()
  endif()

  if(NOT CUDA_gpu_detect_output)
    message(WARNING "Automatic GPU detection failed. Building for all known architectures (${mshadow_known_gpu_archs}).")
    set(${out_variable} ${mshadow_known_gpu_archs} PARENT_SCOPE)
  else()
    set(${out_variable} ${CUDA_gpu_detect_output} PARENT_SCOPE)
  endif()
endfunction()


################################################################################################
# Function for selecting GPU arch flags for nvcc based on CUDA_ARCH_NAME
# Usage:
#   mshadow_select_nvcc_arch_flags(out_variable)
function(mshadow_select_nvcc_arch_flags out_variable)
  # List of arch names
  set(__archs_names "Fermi" "Kepler" "Maxwell" "Pascal" "Volta" "All" "Manual")
  set(__archs_name_default "All")
  if(NOT CMAKE_CROSSCOMPILING)
    list(APPEND __archs_names "Auto")
    set(__archs_name_default "Auto")
  endif()

  # set CUDA_ARCH_NAME strings (so it will be seen as dropbox in CMake-Gui)
  set(CUDA_ARCH_NAME ${__archs_name_default} CACHE STRING "Select target NVIDIA GPU achitecture.")
  set_property( CACHE CUDA_ARCH_NAME PROPERTY STRINGS "" ${__archs_names} )
  mark_as_advanced(CUDA_ARCH_NAME)

  # verify CUDA_ARCH_NAME value
  if(NOT ";${__archs_names};" MATCHES ";${CUDA_ARCH_NAME};")
    string(REPLACE ";" ", " __archs_names "${__archs_names}")
    message(FATAL_ERROR "Only ${__archs_names} architeture names are supported.")
  endif()

  if(${CUDA_ARCH_NAME} STREQUAL "Manual")
    set(CUDA_ARCH_BIN ${mshadow_known_gpu_archs} CACHE STRING "Specify 'real' GPU architectures to build binaries for, BIN(PTX) format is supported")
    set(CUDA_ARCH_PTX "50"                     CACHE STRING "Specify 'virtual' PTX architectures to build PTX intermediate code for")
    mark_as_advanced(CUDA_ARCH_BIN CUDA_ARCH_PTX)
  else()
    unset(CUDA_ARCH_BIN CACHE)
    unset(CUDA_ARCH_PTX CACHE)
  endif()

  if(${CUDA_ARCH_NAME} STREQUAL "Fermi")
    set(__cuda_arch_bin "20 21(20)")
  elseif(${CUDA_ARCH_NAME} STREQUAL "Kepler")
    set(__cuda_arch_bin "30 35")
  elseif(${CUDA_ARCH_NAME} STREQUAL "Maxwell")
    set(__cuda_arch_bin "50")
  elseif(${CUDA_ARCH_NAME} STREQUAL "Pascal")
    set(__cuda_arch_bin "60 61")
  elseif(${CUDA_ARCH_NAME} STREQUAL "Volta")
    set(__cuda_arch_bin "70")
  elseif(${CUDA_ARCH_NAME} STREQUAL "All")
    set(__cuda_arch_bin ${mshadow_known_gpu_archs})
  elseif(${CUDA_ARCH_NAME} STREQUAL "Auto")
    mshadow_detect_installed_gpus(__cuda_arch_bin)
  else()  # (${CUDA_ARCH_NAME} STREQUAL "Manual")
    set(__cuda_arch_bin ${CUDA_ARCH_BIN})
  endif()

  # remove dots and convert to lists
  string(REGEX REPLACE "\\." "" __cuda_arch_bin "${__cuda_arch_bin}")
  string(REGEX REPLACE "\\." "" __cuda_arch_ptx "${CUDA_ARCH_PTX}")
  string(REGEX MATCHALL "[0-9()]+" __cuda_arch_bin "${__cuda_arch_bin}")
  string(REGEX MATCHALL "[0-9]+"   __cuda_arch_ptx "${__cuda_arch_ptx}")
  mshadow_list_unique(__cuda_arch_bin __cuda_arch_ptx)

  set(__nvcc_flags "")
  set(__nvcc_archs_readable "")

  # Tell NVCC to add binaries for the specified GPUs
  foreach(__arch ${__cuda_arch_bin})
    if(__arch MATCHES "([0-9]+)\\(([0-9]+)\\)")
      # User explicitly specified PTX for the concrete BIN
      list(APPEND __nvcc_flags -gencode arch=compute_${CMAKE_MATCH_2},code=sm_${CMAKE_MATCH_1})
      list(APPEND __nvcc_archs_readable sm_${CMAKE_MATCH_1})
    else()
      # User didn't explicitly specify PTX for the concrete BIN, we assume PTX=BIN
      list(APPEND __nvcc_flags -gencode arch=compute_${__arch},code=sm_${__arch})
      list(APPEND __nvcc_archs_readable sm_${__arch})
    endif()
  endforeach()

  # Tell NVCC to add PTX intermediate code for the specified architectures
  foreach(__arch ${__cuda_arch_ptx})
    list(APPEND __nvcc_flags -gencode arch=compute_${__arch},code=compute_${__arch})
    list(APPEND __nvcc_archs_readable compute_${__arch})
  endforeach()

  string(REPLACE ";" " " __nvcc_archs_readable "${__nvcc_archs_readable}")
  set(${out_variable}          ${__nvcc_flags}          PARENT_SCOPE)
  set(${out_variable}_readable ${__nvcc_archs_readable} PARENT_SCOPE)
endfunction()

################################################################################################
# Short command for cuda comnpilation
# Usage:
#   mshadow_cuda_compile(<objlist_variable> <cuda_files>)
macro(mshadow_cuda_compile objlist_variable)
  foreach(var CMAKE_CXX_FLAGS CMAKE_CXX_FLAGS_RELEASE CMAKE_CXX_FLAGS_DEBUG)
    set(${var}_backup_in_cuda_compile_ "${${var}}")

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

  endforeach()
  if(UNIX OR APPLE)
    list(APPEND CUDA_NVCC_FLAGS -Xcompiler -fPIC)
  endif()

  if(APPLE)
    list(APPEND CUDA_NVCC_FLAGS -Xcompiler -Wno-unused-function)
  endif()

  set(CUDA_NVCC_FLAGS_DEBUG "${CUDA_NVCC_FLAGS_DEBUG} -G")

  if(MSVC)
    # disable noisy warnings:
    # 4819: The file contains a character that cannot be represented in the current code page (number).
    list(APPEND CUDA_NVCC_FLAGS -Xcompiler "/wd4819")
    foreach(flag_var
        CMAKE_CXX_FLAGS CMAKE_CXX_FLAGS_DEBUG CMAKE_CXX_FLAGS_RELEASE
        CMAKE_CXX_FLAGS_MINSIZEREL CMAKE_CXX_FLAGS_RELWITHDEBINFO)
      if(${flag_var} MATCHES "/MD")
        string(REGEX REPLACE "/MD" "/MT" ${flag_var} "${${flag_var}}")
      endif(${flag_var} MATCHES "/MD")
    endforeach(flag_var)
  endif()

  # If the build system is a container, make sure the nvcc intermediate files
  # go into the build output area rather than in /tmp, which may run out of space
  if(IS_CONTAINER_BUILD)
    set(CUDA_NVCC_INTERMEDIATE_DIR "${CMAKE_CURRENT_BINARY_DIR}")
    message(STATUS "Container build enabled, so nvcc intermediate files in: ${CUDA_NVCC_INTERMEDIATE_DIR}")
    list(APPEND CUDA_NVCC_FLAGS "--keep --keep-dir ${CUDA_NVCC_INTERMEDIATE_DIR}")
  endif()

  cuda_compile(cuda_objcs ${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()

  set(${objlist_variable} ${cuda_objcs})
endmacro()

################################################################################################
# Short command for cuDNN detection. Believe it soon will be a part of CUDA toolkit distribution.
# That's why not FindcuDNN.cmake file, but just the macro
# Usage:
#   detect_cuDNN()
function(detect_cuDNN)
  set(CUDNN_ROOT "" CACHE PATH "CUDNN root folder")

  find_path(CUDNN_INCLUDE cudnn.h
            PATHS ${CUDNN_ROOT} $ENV{CUDNN_ROOT} ${CUDA_TOOLKIT_INCLUDE}
            DOC "Path to cuDNN include directory." )

  get_filename_component(__libpath_hist ${CUDA_CUDART_LIBRARY} PATH)
  find_library(CUDNN_LIBRARY NAMES libcudnn.so cudnn.lib # libcudnn_static.a
                             PATHS ${CUDNN_ROOT} $ENV{CUDNN_ROOT} ${CUDNN_INCLUDE} ${__libpath_hist}
                             DOC "Path to cuDNN library.")

  if(CUDNN_INCLUDE AND CUDNN_LIBRARY)
    set(HAVE_CUDNN  TRUE PARENT_SCOPE)
    set(CUDNN_FOUND TRUE PARENT_SCOPE)

    mark_as_advanced(CUDNN_INCLUDE CUDNN_LIBRARY CUDNN_ROOT)
    message(STATUS "Found cuDNN (include: ${CUDNN_INCLUDE}, library: ${CUDNN_LIBRARY})")
  endif()
endfunction()


################################################################################################
###  Non macro section
################################################################################################

# Try to prime CUDA_TOOLKIT_ROOT_DIR by looking for libcudart.so
if(NOT CUDA_TOOLKIT_ROOT_DIR)
  find_library(CUDA_LIBRARY_PATH libcudart.so PATHS ENV LD_LIBRARY_PATH PATH_SUFFIXES lib lib64)
  if(CUDA_LIBRARY_PATH)
    get_filename_component(CUDA_LIBRARY_PATH ${CUDA_LIBRARY_PATH} DIRECTORY)
    set(CUDA_TOOLKIT_ROOT_DIR "${CUDA_LIBRARY_PATH}/..")
  endif()
endif()

find_package(CUDA 5.5 QUIET REQUIRED)
find_cuda_helper_libs(curand)  # cmake 2.8.7 compartibility which doesn't search for curand

if(NOT CUDA_FOUND)
  return()
endif()

set(HAVE_CUDA TRUE)
message(STATUS "CUDA detected: " ${CUDA_VERSION})
include_directories(SYSTEM ${CUDA_INCLUDE_DIRS})
list(APPEND mshadow_LINKER_LIBS ${CUDA_CUDART_LIBRARY}
                              ${CUDA_curand_LIBRARY} ${CUDA_CUBLAS_LIBRARIES})

# Known NVIDIA GPU achitectures mshadow can be compiled for.
# This list will be used for CUDA_ARCH_NAME = All option
if(CUDA_ARCH_ALL)
  set(mshadow_known_gpu_archs "${CUDA_ARCH_ALL}")
else()
  if(${CUDA_VERSION} EQUAL 9.0 OR ${CUDA_VERSION} GREATER 9.0)
    set(mshadow_known_gpu_archs "30 35 50 52 60 61 70")
  elseif(${CUDA_VERSION} EQUAL 8.0 OR ${CUDA_VERSION} GREATER 8.0)
    set(mshadow_known_gpu_archs "30 35 50 52 60 61")
  else()
    set(mshadow_known_gpu_archs "30 35 50 52")
  endif()
endif()

# cudnn detection
if(USE_CUDNN)
  detect_cuDNN()
  if(HAVE_CUDNN)
    add_definitions(-DUSE_CUDNN)
    include_directories(SYSTEM ${CUDNN_INCLUDE})
    list(APPEND mshadow_LINKER_LIBS ${CUDNN_LIBRARY})
  endif()
endif()

# setting nvcc arch flags
mshadow_select_nvcc_arch_flags(NVCC_FLAGS_EXTRA)
list(APPEND CUDA_NVCC_FLAGS ${NVCC_FLAGS_EXTRA})
message(STATUS "Added CUDA NVCC flags for: ${NVCC_FLAGS_EXTRA_readable}")

# Boost 1.55 workaround, see https://svn.boost.org/trac/boost/ticket/9392 or
# https://github.com/ComputationalRadiationPhysics/picongpu/blob/master/src/picongpu/CMakeLists.txt
if(Boost_VERSION EQUAL 105500)
  message(STATUS "Cuda + Boost 1.55: Applying noinline work around")
  # avoid warning for CMake >= 2.8.12
  set(CUDA_NVCC_FLAGS "${CUDA_NVCC_FLAGS} \"-DBOOST_NOINLINE=__attribute__((noinline))\" ")
endif()

# disable some nvcc diagnostic that apears in boost, glog, glags, opencv, etc.
foreach(diag cc_clobber_ignored integer_sign_change useless_using_declaration set_but_not_used)
  list(APPEND CUDA_NVCC_FLAGS -Xcudafe --diag_suppress=${diag})
endforeach()

# setting default testing device
if(NOT CUDA_TEST_DEVICE)
  set(CUDA_TEST_DEVICE -1)
endif()

mark_as_advanced(CUDA_BUILD_CUBIN CUDA_BUILD_EMULATION CUDA_VERBOSE_BUILD)
mark_as_advanced(CUDA_SDK_ROOT_DIR CUDA_SEPARABLE_COMPILATION)

# Handle clang/libc++ issue
if(APPLE)
  mshadow_detect_darwin_version(OSX_VERSION)

  # OSX 10.9 and higher uses clang/libc++ by default which is incompartible with old CUDA toolkits
  if(OSX_VERSION VERSION_GREATER 10.8)
    # enabled by default if and only if CUDA version is less than 7.0
    mshadow_option(USE_libstdcpp "Use libstdc++ instead of libc++" (CUDA_VERSION VERSION_LESS 7.0))
  endif()
endif()