# ########################################################################
# Copyright 2016-2019 Advanced Micro Devices, Inc.
# ########################################################################

# set( Boost_DEBUG ON )
set( Boost_USE_MULTITHREADED ON )
set( Boost_DETAILED_FAILURE_MSG ON )
set( Boost_ADDITIONAL_VERSIONS 1.65.1 1.65 )
set( Boost_USE_STATIC_LIBS OFF )

if(EXISTS /etc/redhat-release)
    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fopenmp=libgomp -pthread")
else()
    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fopenmp -pthread")
endif()

find_package( Boost COMPONENTS program_options )

if( NOT Boost_FOUND )
  message( STATUS "Dynamic boost libraries not found. Attempting to find static libraries " )
  set( Boost_USE_STATIC_LIBS ON )
  find_package( Boost COMPONENTS program_options )

  if( NOT Boost_FOUND )
    message( FATAL_ERROR "boost is a required dependency and is not found;  try adding boost path to CMAKE_PREFIX_PATH" )
  endif( )
endif( )

set( THREADS_PREFER_PTHREAD_FLAG ON )
find_package( Threads REQUIRED )

# Linking lapack library requires fortran flags
enable_language( Fortran )
find_package( cblas CONFIG REQUIRED )
if( NOT cblas_FOUND )
  message( FATAL_ERROR "cblas is a required dependency and is not found;  try adding cblas path to CMAKE_PREFIX_PATH" )
endif( )

if(LINK_BLIS)
  set( BLIS_CPP ../common/blis_interface.cpp )
endif()

set( rocblas_benchmark_common
      ../common/utility.cpp
      ../common/cblas_interface.cpp
      ${BLIS_CPP}
      ../common/rocblas_parse_data.cpp
    )

add_executable( rocblas-bench client.cpp ${rocblas_benchmark_common} )
target_compile_features( rocblas-bench PRIVATE cxx_static_assert cxx_nullptr cxx_auto_type)

if( BUILD_WITH_TENSILE )
    target_compile_definitions( rocblas-bench PRIVATE BUILD_WITH_TENSILE=1 )
else()
    target_compile_definitions( rocblas-bench PRIVATE BUILD_WITH_TENSILE=0 )
endif()

# Internal header includes
target_include_directories( rocblas-bench
  PRIVATE
    $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/../include>
    $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/../../library/include>
)

set( BLIS_INCLUDE_DIR ${CMAKE_SOURCE_DIR}/build/deps/blis/include/blis )
set( BLIS_LIBRARY ${CMAKE_SOURCE_DIR}/build/deps/blis/lib/libblis.so )


if( OS_ID_rhel OR OS_ID_centos OR OS_ID_sles)
    if( OS_ID_rhel OR OS_ID_centos)
        # defer OpenMP include as search order must come after clang
        set( XXX_OPENMP_INCLUDE_DIR /opt/rh/devtoolset-7/root/usr/lib/gcc/x86_64-redhat-linux/7/include )
        set( OPENMP_LIBRARY /opt/rh/devtoolset-7/root/usr/lib/gcc/x86_64-redhat-linux/7/libgomp.so )
    else()
    #SLES
        set( OPENMP_INCLUDE_DIR /usr/lib64/gcc/x86_64-suse-linux/7/include/ )
        set( OPENMP_LIBRARY /usr/lib64/gcc/x86_64-suse-linux/7/libgomp.so )
    endif()

    if(EXISTS /opt/rocm/hcc/lib/clang/10.0.0/include/immintrin.h)
        set( CLANG_INCLUDE_DIR /opt/rocm/hcc/lib/clang/10.0.0/include )
    elseif (EXISTS /opt/rocm/hcc/lib/clang/9.0.0/include/immintrin.h)
        set( CLANG_INCLUDE_DIR /opt/rocm/hcc/lib/clang/9.0.0/include )
    else()
        error("cannot find immintrin.h")
    endif()

    # External header includes included as system files
    target_include_directories( rocblas-bench
      SYSTEM PRIVATE
        $<BUILD_INTERFACE:${CLANG_INCLUDE_DIR}>
        $<BUILD_INTERFACE:${BLIS_INCLUDE_DIR}>
        $<BUILD_INTERFACE:${HIP_INCLUDE_DIRS}>
        $<BUILD_INTERFACE:${HCC_INCLUDE_DIRS}>
        $<BUILD_INTERFACE:${Boost_INCLUDE_DIRS}>
        $<BUILD_INTERFACE:${CBLAS_INCLUDE_DIRS}>
        $<BUILD_INTERFACE:${OPENMP_INCLUDE_DIR}>
        )
    if(LINK_BLIS)
      target_link_libraries( rocblas-bench PRIVATE ${Boost_LIBRARIES} ${BLIS_LIBRARY} ${OPENMP_LIBRARY} cblas lapack roc::rocblas )
    else()
      target_link_libraries( rocblas-bench PRIVATE ${Boost_LIBRARIES} ${OPENMP_LIBRARY} cblas lapack roc::rocblas )
    endif()

else()
    # External header includes included as system files
    target_include_directories( rocblas-bench
      SYSTEM PRIVATE
        $<BUILD_INTERFACE:${HIP_INCLUDE_DIRS}>
        $<BUILD_INTERFACE:${HCC_INCLUDE_DIRS}>
        $<BUILD_INTERFACE:${Boost_INCLUDE_DIRS}>
        $<BUILD_INTERFACE:${CBLAS_INCLUDE_DIRS}>
        $<BUILD_INTERFACE:${BLIS_INCLUDE_DIR}>
    )

    if(LINK_BLIS)
      target_link_libraries( rocblas-bench PRIVATE ${Boost_LIBRARIES} ${BLIS_LIBRARY} cblas lapack roc::rocblas )
    else()
      target_link_libraries( rocblas-bench PRIVATE ${Boost_LIBRARIES} cblas lapack roc::rocblas )
    endif()

endif()

get_target_property( HIPHCC_LOCATION hip::hip_hcc IMPORTED_LOCATION_RELEASE )

if( CUDA_FOUND )
  target_include_directories( rocblas-bench
    PRIVATE
      $<BUILD_INTERFACE:${CUDA_INCLUDE_DIRS}>
      $<BUILD_INTERFACE:${hip_INCLUDE_DIRS}>
    )
  target_compile_definitions( rocblas-bench PRIVATE __HIP_PLATFORM_NVCC__ )
  target_link_libraries( rocblas-bench PRIVATE ${CUDA_LIBRARIES} )
else( )
  # auto set in hip_common.h
  #target_compile_definitions( rocblas-bench PRIVATE __HIP_PLATFORM_HCC__ )
  target_link_libraries( rocblas-bench PRIVATE ${HIPHCC_LOCATION} )
endif( )

if( CMAKE_CXX_COMPILER MATCHES ".*/hcc$" )
  # Remove following when hcc is fixed; hcc emits following spurious warning ROCm v1.6.1
  # "clang-5.0: warning: argument unused during compilation: '-isystem /opt/rocm/include'"
  target_compile_options( rocblas-bench PRIVATE -Wno-unused-command-line-argument -mf16c )
  target_include_directories( rocblas-bench PRIVATE /opt/rocm/hsa/include)

elseif( CMAKE_COMPILER_IS_GNUCXX OR CXX_VERSION_STRING MATCHES "clang")
  # GCC or hip-clang needs specific flags to turn on f16c intrinsics
  target_compile_options( rocblas-bench PRIVATE -mf16c )
endif( )

if( OS_ID_rhel OR OS_ID_centos)
    # force clang includes to take precedence over devtoolset-7 which we only want for OpenMP
    set(CMAKE_CXX_FLAGS "-isystem ${CLANG_INCLUDE_DIR} -isystem ${XXX_OPENMP_INCLUDE_DIR} ${CMAKE_CXX_FLAGS}")
endif()

# include order workaround to force /opt/rocm/include later in order to ignore installed rocblas
set(CMAKE_CXX_FLAGS "-isystem /opt/rocm/include ${CMAKE_CXX_FLAGS}")

set_target_properties( rocblas-bench PROPERTIES CXX_EXTENSIONS NO )
set_target_properties( rocblas-bench PROPERTIES RUNTIME_OUTPUT_DIRECTORY "${PROJECT_BINARY_DIR}/staging" )
add_dependencies( rocblas-bench rocblas-common )
add_subdirectory ( ./perf_script )