cmake_minimum_required(VERSION 3.15)
project(libnd4j)
set(CMAKE_VERBOSE_MAKEFILE ON)


set (CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/cmake")
message("CMAKE MODULE PATH ${CMAKE_MODULE_PATH}")

#ensure we create lib files
set(CMAKE_WINDOWS_EXPORT_ALL_SYMBOLS OFF)

option(SD_NATIVE "Optimize for build machine (might not work on others)" OFF)
option(SD_CHECK_VECTORIZATION "checks for vectorization" OFF)
option(SD_BUILD_TESTS "Build tests" OFF)
option(SD_STATIC_LIB "Build static library" OFF)
option(SD_SHARED_LIB "Build shared library" ON)
option(SD_SANITIZE "Enable Address Sanitizer" ON)

option(FLATBUFFERS_BUILD_FLATC "Enable the build of the flatbuffers compiler" OFF)
set(FLATBUFFERS_BUILD_FLATC "OFF" CACHE STRING "Hack to disable flatc build" FORCE)

set(CMAKE_CXX_STANDARD 11)

#///////////////////////////////////////////////////////////////////////////////
# genCompilation: Generates cpp, cu files
# INPUT:
# $FILE_ITEM template-configuration that utilizes libnd4j type, macros helpers
# defined inside { include/types/types.h, include/system/type_boilerplate.h}
# OUTPUT:
# $CUSTOMOPS_GENERIC_SOURCES  generated files will be added into this List
#////////////////////////////////////////////////////////////////////////////////
#  A simple template-configuration file example:
# // hints and defines what types will be generated
# #cmakedefine LIBND4J_TYPE_GEN
# #cmakedefine FLOAT_TYPE_GEN
# // below if defines blocks are needed for correctly handling multiple types
# #if  defined(LIBND4J_TYPE_GEN)
#  BUILD_DOUBLE_TEMPLATE(template void someFunc, (arg_list,..),
#                          LIBND4J_TYPES_@FL_TYPE_INDEX@, INDEXING_TYPES);
# #endif
# #if defined(FLOAT_TYPE_GEN)
#  BUILD_SINGLE_TEMPLATE(template class SomeClass,, FLOAT_TYPES_@FL_TYPE_INDEX@);
# #endif
#////////////////////////////////////////////////////////////////////////////////

set_property(GLOBAL PROPERTY JOB_POOLS one_jobs=1 two_jobs=2)




function(genCompilation FILE_ITEM)
    get_filename_component(FILE_ITEM_WE ${FL_ITEM} NAME_WE)

    set(EXTENSION "cpp")

    if(FL_ITEM MATCHES "cu.in$")
         set(EXTENSION "cu")
    endif()

    file(READ ${FL_ITEM} CONTENT_FL)
    #check content for types

    #set all to false
    set (FLOAT_TYPE_GEN     0)
    set (INT_TYPE_GEN       0)
    set (LIBND4J_TYPE_GEN   0)
    set (PAIRWISE_TYPE_GEN  0)
    set (RANGE_STOP         -1)

    string(REGEX MATCHALL "#cmakedefine[ \t]+[^_]+_TYPE_GEN" TYPE_MATCHES ${CONTENT_FL})

    foreach(TYPEX ${TYPE_MATCHES})
        set(STOP -1)
        if(TYPEX MATCHES "INT_TYPE_GEN$")
           set (INT_TYPE_GEN  1)
           set(STOP 7)
        endif()
        if(TYPEX MATCHES "LIBND4J_TYPE_GEN$")
           set (LIBND4J_TYPE_GEN 1)
           set(STOP 9)
        endif()
        if(TYPEX MATCHES "FLOAT_TYPE_GEN$")
           set (FLOAT_TYPE_GEN 1)
           set(STOP 3)
        endif()
        if(TYPEX MATCHES "PAIRWISE_TYPE_GEN$")
           set (PAIRWISE_TYPE_GEN  1)
           set(STOP 12)
        endif()
        if(STOP GREATER RANGE_STOP)
           set(RANGE_STOP ${STOP})
        endif()

    endforeach()

    if(RANGE_STOP GREATER -1)
        foreach(FL_TYPE_INDEX RANGE 0 ${RANGE_STOP})
            # set OFF if the index is above
            if(FL_TYPE_INDEX GREATER 3)
                 set (FLOAT_TYPE_GEN     0)
            endif()
            if(FL_TYPE_INDEX GREATER 7)
                 set (INT_TYPE_GEN     0)
            endif()
            if(FL_TYPE_INDEX GREATER 9)
                 set (LIBND4J_TYPE_GEN   0)
            endif()
            set(GENERATED_SOURCE  "${CMAKE_BINARY_DIR}/compilation_units/${FILE_ITEM_WE}_${FL_TYPE_INDEX}.${EXTENSION}")
            configure_file(  "${FL_ITEM}" "${GENERATED_SOURCE}" @ONLY)
            LIST(APPEND CUSTOMOPS_GENERIC_SOURCES ${GENERATED_SOURCE} )
        endforeach()
    endif()

    set(CUSTOMOPS_GENERIC_SOURCES ${CUSTOMOPS_GENERIC_SOURCES} PARENT_SCOPE)
endfunction()


if (SD_CUDA)
    enable_language(CUDA)
    set(CMAKE_CUDA_STANDARD 11)

    set(DEFAULT_ENGINE "samediff::ENGINE_CUDA")
else()
    set(DEFAULT_ENGINE "samediff::ENGINE_CPU")
endif()

# MSVC runtime lib can be either "MultiThreaded" or "MultiThreadedDLL", /MT and /MD respectively
set(MSVC_RT_LIB "MultiThreadedDLL")

set(SD_X86_BUILD false)

if (NOT SD_IOS_BUILD AND NOT SD_ANDROID_BUILD AND NOT ${SD_ARCH} MATCHES "power*" AND NOT ${SD_ARCH} MATCHES "arm*")
    set(SD_X86_BUILD true)
endif()

# -fsanitize=address
# -fsanitize=leak
if (SD_ANDROID_BUILD)
    set_property(GLOBAL PROPERTY JOB_POOLS one_job=1 two_jobs=2)
    set(CMAKE_CXX_FLAGS_RELEASE  "${CMAKE_CXX_FLAGS_RELEASE} -O3 -fPIC -Wno-braced-scalar-init -Wno-delete-non-virtual-dtor -Wno-unused-command-line-argument -Wno-dangling-else -D_RELEASE=true")
    set(CMAKE_CXX_FLAGS_DEBUG  "${CMAKE_CXX_FLAGS_DEBUG} -O0 -g -fPIC -Wno-braced-scalar-init -Wno-delete-non-virtual-dtor -Wno-unused-command-line-argument -Wno-dangling-else")
elseif (APPLE)
    set(CMAKE_CXX_FLAGS_RELEASE  "-O3 -fPIC -Wno-braced-scalar-init -Wno-delete-non-virtual-dtor -Wno-unused-command-line-argument -Wno-dangling-else -D__APPLE_OS__=true -D_RELEASE=true")
    set(CMAKE_CXX_FLAGS_DEBUG  " -O0 -g -fPIC -Wno-braced-scalar-init -Wno-delete-non-virtual-dtor -Wno-unused-command-line-argument -Wno-dangling-else -D__APPLE_OS__=true")
elseif(WIN32)
    set(SD_X86_BUILD true)
    if (SD_CUDA)
        set(CMAKE_CXX_FLAGS_RELEASE  "-D_RELEASE=true")
        set(CMAKE_CXX_FLAGS_DEBUG  "  /FS /EHsc")
        set(CMAKE_CUDA_STANDARD 14)
    else()
        set(CMAKE_CXX_FLAGS_RELEASE  "-O3 -fPIC -D_RELEASE=true")
        set(CMAKE_CXX_FLAGS_DEBUG  " -g -O2 -fPIC")
        set(CMAKE_CUDA_STANDARD 14)
    endif()
else()
    set(CMAKE_CXX_FLAGS_RELEASE  "-O3 -fPIC -D_RELEASE=true")
    set(CMAKE_CXX_FLAGS_DEBUG  " -g -O0 -fPIC")

    if (SD_CPU AND SD_SANITIZE)
        set(CMAKE_CXX_FLAGS_DEBUG "${CMAKE_CXX_FLAGS_DEBUG} -fsanitize=address")
    endif()
endif()

if(SD_NATIVE)
    IF(${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64*")
        set(SD_X86_BUILD false)
        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -mcpu=native")
    ELSE()
        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -march=native")
    ENDIF()
endif()


if(NOT SD_CUDA)
    # we need this definition to avoid global memory use within mkldnn
    add_definitions(-DDNNL_ENABLE_CONCURRENT_EXEC=true)

    # there's a chance, we have no BLAS provided externally
    if ("${OPENBLAS_PATH}" STREQUAL "")
        #we don't want OpenBLAS on Apple
        if (NOT APPLE)
            # note: this is not a typo
            set(BLA_VENDOR "OpenBLAS")
        endif()

        # look around for system blas instead, see: https://cmake.org/cmake/help/latest/module/FindBLAS.html
        find_package(BLAS REQUIRED)
        if (BLAS_FOUND)
            message("Found external BLAS implementation: ${BLAS_LIBRARIES} ")
            add_definitions(-D__EXTERNAL_BLAS__=true)
        endif()
    else()
        # if we have externally provided OPENBLAS_PATH - let's use it
        set(HAVE_OPENBLAS 1)
        message("Setting openblas")
        include_directories(${OPENBLAS_PATH}/include/)
        link_directories(${OPENBLAS_PATH} ${OPENBLAS_PATH}/lib/)
        set(OPENBLAS_LIBRARIES openblas)
    endif()

    # building cpu_features
    if (SD_X86_BUILD)
        add_definitions(-DCPU_FEATURES=true)
        set(BUILD_PIC "ON" CACHE STRING "Hack to enforce fPIC mode" FORCE)
        configure_file(./CMakeLists.txt.cpu_features.in cpu_features-download/CMakeLists.txt)
        message("CMAKE_COMMAND: ${CMAKE_COMMAND}")
        execute_process(COMMAND ${CMAKE_COMMAND} -DBUILD_PIC=ON -G "${CMAKE_GENERATOR}" .
                RESULT_VARIABLE result
                WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}/cpu_features-download )

        if(result)
            message(FATAL_ERROR "CMake step for cpu_features failed: ${result}")
        endif()
        execute_process(COMMAND ${CMAKE_COMMAND} --build .
                RESULT_VARIABLE result
                WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}/cpu_features-download )
        if(result)
            message(FATAL_ERROR "Build step for cpu_features failed: ${result}")
        endif()

        add_subdirectory(${CMAKE_CURRENT_BINARY_DIR}/cpu_features-src
                ${CMAKE_CURRENT_BINARY_DIR}/cpu_features-build
                EXCLUDE_FROM_ALL)
        set(CPUF_SOURCE_DIR ${CMAKE_CURRENT_BINARY_DIR}/cpu_features-src)
        include_directories(${CPUF_SOURCE_DIR}/include)
        set(CPU_FEATURES cpu_features)
    endif()
endif()


#arm-compute entry
if(${HELPERS_armcompute})
 find_package(ARMCOMPUTE REQUIRED)
 execute_process(COMMAND ${CMAKE_C_COMPILER} -fuse-ld=gold -Wl,--version ERROR_QUIET OUTPUT_VARIABLE ld_version)
 if ("${ld_version}" MATCHES "GNU gold")
   set(CMAKE_EXE_LINKER_FLAGS "${CMAKE_EXE_LINKER_FLAGS} -fuse-ld=gold ")
   if (CMAKE_BUILD_TYPE STREQUAL "Debug")
        add_link_options("-Wl,--long-plt")
   endif()
 endif()

 if(ARMCOMPUTE_FOUND)
    message("Found ARMCOMPUTE: ${ARMCOMPUTE_LIBRARIES}")
    set(HAVE_ARMCOMPUTE 1)
    # Add preprocessor definition for ARM Compute NEON 
    add_definitions(-DARMCOMPUTENEON_ENABLED)
    include_directories(${ARMCOMPUTE_INCLUDE})
    message("----${ARMCOMPUTE_INCLUDE}---")
 endif()


endif()



# new mkl-dnn entry
if (${HELPERS_mkldnn})
    message("Going to pull & build mkldnn")
    set(HAVE_MKLDNN 1)
    set(DNNL_LIBRARY_TYPE "STATIC" CACHE STRING "Hack to enforce static mode" FORCE)

    configure_file(./CMakeLists.txt.mkldnn.in mkldnn-download/CMakeLists.txt)
    execute_process(COMMAND ${CMAKE_COMMAND} -G "${CMAKE_GENERATOR}" .
            RESULT_VARIABLE result
            WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}/mkldnn-download )
    if(result)
        message(FATAL_ERROR "CMake step for mkldnn failed: ${result}")
    endif()
    execute_process(COMMAND ${CMAKE_COMMAND} --build .
            RESULT_VARIABLE result
            WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}/mkldnn-download )
    if(result)
        message(FATAL_ERROR "Build step for mkldnn failed: ${result}")
    endif()

    add_subdirectory(${CMAKE_CURRENT_BINARY_DIR}/mkldnn-src
            ${CMAKE_CURRENT_BINARY_DIR}/mkldnn-build
            EXCLUDE_FROM_ALL)

    set(mkldnn_SOURCE_DIR ${CMAKE_CURRENT_BINARY_DIR}/mkldnn-build)
    set(mkldnn_EXT_DIR ${CMAKE_CURRENT_BINARY_DIR}/mkldnn-src)
    set(MKLDNN_PATH "${mkldnn_SOURCE_DIR}")
    include_directories(${mkldnn_SOURCE_DIR}/include ${mkldnn_EXT_DIR}/include ${mkldnn_SOURCE_DIR})
    set(MKLDNN dnnl)
endif()


if (${HELPERS_cudnn})
    if (NOT SD_CUDA)
        message(FATAL_ERROR "Can't build cuDNN on non-CUDA platform")
    endif()

    set(CUDNN_ROOT_DIR "" CACHE PATH "Folder contains NVIDIA cuDNN")

    SET(CUDNN_LIBNAME "cudnn")
    find_path(CUDNN_INCLUDE_DIR cudnn.h
            HINTS ${CUDNN_ROOT_DIR} ${CUDA_TOOLKIT_ROOT_DIR}
            PATH_SUFFIXES cuda/include include)

    find_library(CUDNN_LIBRARY ${CUDNN_LIBNAME}
            HINTS ${CUDNN_ROOT_DIR} ${CUDA_TOOLKIT_ROOT_DIR}
            PATH_SUFFIXES lib lib64 cuda/lib cuda/lib64 lib/x64)

    #find_library(CULIBOS_LIBRARY ${CULIBOS_LIBNAME}
    #        HINTS ${CUDNN_ROOT_DIR} ${CUDA_TOOLKIT_ROOT_DIR}
    #        PATH_SUFFIXES lib lib64 cuda/lib cuda/lib64 lib/x64)


    if (CUDNN_LIBRARY)
        set(HAVE_CUDNN true)
        set(CUDNN ${CUDNN_LIBRARY})
    else()
        message(FATAL_ERROR "Unable to find cuDNN")
    endif()
endif()

# Download and unpack flatbuffers at configure time
configure_file(CMakeLists.txt.in flatbuffers-download/CMakeLists.txt)
execute_process(COMMAND ${CMAKE_COMMAND} -G "${CMAKE_GENERATOR}" .
  RESULT_VARIABLE result
  WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}/flatbuffers-download )
if(result)
  message(FATAL_ERROR "CMake step for flatbuffers failed: ${result}")
endif()
execute_process(COMMAND ${CMAKE_COMMAND} --build .
  RESULT_VARIABLE result
  WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}/flatbuffers-download )
if(result)
  message(FATAL_ERROR "Build step for flatbuffers failed: ${result}")
endif()

# Add flatbuffers directly to our build.
add_subdirectory(${CMAKE_CURRENT_BINARY_DIR}/flatbuffers-src
                 ${CMAKE_CURRENT_BINARY_DIR}/flatbuffers-build
                 EXCLUDE_FROM_ALL)

set(HAVE_FLATBUFFERS 1)
set(FLATBUFFERS_PATH ${CMAKE_CURRENT_BINARY_DIR}/flatbuffers-src)
include_directories(${FLATBUFFERS_PATH}/include)



configure_file(include/config.h.in include/config.h)
include_directories(${CMAKE_CURRENT_BINARY_DIR}/include)


include_directories(${CMAKE_CURRENT_SOURCE_DIR}/include)
add_subdirectory(blas)
if(SD_BUILD_TESTS)
    # tests are always compiled with all ops included
    set(SD_ALL_OPS true)
    set(SD_BUILD_MINIFIER true)
    add_subdirectory(tests_cpu)
endif()


if (MSVC_DEV)
    set(SD_BUILD_MINIFIER false)
endif ()

set (CMAKE_INSTALL_PREFIX $ENV{ND4J_HOME}/nd4j-native-parent/nd4j-native/src/main/resources)

# Set package information
set(CPACK_PACKAGE_DESCRIPTION_SUMMARY "Native operations for nd4j.")
set(CPACK_PACKAGE_RELEASE 1)
set(CPACK_PACKAGE_CONTACT "agibsonccc <adam@skymind.global>")
set(CPACK_PACKAGE_VENDOR "Eclipse")
set(CPACK_SETDESTDIR "false")
set(CPACK_PACKAGING_INSTALL_PREFIX "/usr/local/lib")
set(CPACK_PACKAGE_NAME "libnd4j")
set(CPACK_PACKAGE_VERSION_MAJOR "0")
set(CPACK_PACKAGE_VERSION_MINOR "8")
set(CPACK_PACKAGE_VERSION_PATCH "0")
set(CPACK_PACKAGE_VERSION "${CPACK_PACKAGE_VERSION_MAJOR}.${CPACK_PACKAGE_VERSION_MINOR}.${CPACK_PACKAGE_VERSION_PATCH}")
set(CPACK_PACKAGE_INSTALL_DIRECTORY "libnd4j")
set(CPACK_RESOURCE_FILE_README "${CMAKE_CURRENT_SOURCE_DIR}/README.md")

# Determine distribution and release — may require redhat-lsb-core installed on CentOS / RH
execute_process(COMMAND lsb_release -si OUTPUT_VARIABLE DISTRIBUTION OUTPUT_STRIP_TRAILING_WHITESPACE)
execute_process(COMMAND lsb_release -sc OUTPUT_VARIABLE RELEASE OUTPUT_STRIP_TRAILING_WHITESPACE)
execute_process(COMMAND uname -i OUTPUT_VARIABLE ARCHITECTURE)

# Set package name and type (deb vs rpm)
if(DISTRIBUTION STREQUAL "Ubuntu")

  # Set Ubuntu-specific information (see http://www.cmake.org/Wiki/CMake:CPackPackageGenerators)
  if(ARCHITECTURE MATCHES ".*x86_64.*")
    set(CPACK_DEBIAN_PACKAGE_ARCHITECTURE "amd64")
  else()
    set(CPACK_DEBIAN_PACKAGE_ARCHITECTURE "i386")
  endif()
  set(CPACK_DEBIAN_PACKAGE_MAINTAINER "raver119")
  set(CPACK_DEBIAN_PACKAGE_SECTION "devel")
  set(CPACK_DEBIAN_PACKAGE_RECOMMENDS "cuda")
  # For Ubuntu <= 12, libatlas3gf-base, liblapack3gf
  # Build deps: libatlas3-base liblapack3 libopenblas-dev libatlas-dev liblapack-dev gcc-5 g++-5
  set(CPACK_DEBIAN_PACKAGE_DEPENDS "")
  set(CPACK_DEBIAN_PACKAGE_HOMEPAGE "https://github.com/eclipse/deeplearning4j")
  set(CPACK_GENERATOR "DEB")
  set(CPACK_PACKAGE_FILE_NAME ${CPACK_PACKAGE_NAME}_${CPACK_PACKAGE_VERSION}-${RELEASE}_${CPACK_DEBIAN_PACKAGE_ARCHITECTURE})
  set(CPACK_DEBIAN_PACKAGE_CONTROL_EXTRA "${CMAKE_CURRENT_SOURCE_DIR}/cmake/postinst;${CMAKE_CURRENT_SOURCE_DIR}/cmake/postrm;" )

elseif(DISTRIBUTION STREQUAL "CentOS")

  # Set Fedora-specific information (see http://www.cmake.org/Wiki/CMake:CPackPackageGenerators)
  execute_process(COMMAND lsb_release -sr OUTPUT_VARIABLE RELEASE OUTPUT_STRIP_TRAILING_WHITESPACE)
  if(ARCHITECTURE MATCHES ".*x86_64.*")
    set(CPACK_RPM_PACKAGE_ARCHITECTURE "x86_64")
  else()
    set(CPACK_RPM_PACKAGE_ARCHITECTURE "i686")
  endif()
  set(CPACK_PACKAGE_CONTACT "agibsonccc")
  set(CPACK_RPM_PACKAGE_GROUP "Development/Tools")
  set(CPACK_RPM_PACKAGE_LICENSE "Apache-2.0")
  set(CPACK_RPM_PACKAGE_SUGGESTS "cuda")
  # Build deps: atlas blas lapack cmake3 devtoolset-4-gcc devtoolset-4-gcc-c++
  set(CPACK_RPM_PACKAGE_REQUIRES "")
  set(CPACK_RPM_PACKAGE_URL "https://github.com/eclipse/deeplearning4j/libnd4j")
  set(CPACK_GENERATOR "RPM")
  set(CPACK_PACKAGE_FILE_NAME ${CPACK_PACKAGE_NAME}-${CPACK_PACKAGE_VERSION}.fc${RELEASE}.${CPACK_RPM_PACKAGE_ARCHITECTURE})
  set(CPACK_RPM_POST_INSTALL_SCRIPT_FILE "${CMAKE_CURRENT_SOURCE_DIR}/cmake/postinst")
  set(CPACK_RPM_POST_UNINSTALL_SCRIPT_FILE "${CMAKE_CURRENT_SOURCE_DIR}/cmake/postrm")
  set(CPACK_RPM_EXCLUDE_FROM_AUTO_FILELIST_ADDITION "/usr/local/lib")

endif()

include(CPack)