CMakeLists.txt

cmake_minimum_required(VERSION 3.21)
cmake_policy(VERSION 3.21)

project(Hydrogen CXX)

# Setup version information
find_program(__GIT_EXECUTABLE git)
mark_as_advanced(__GIT_EXECUTABLE)
if (__GIT_EXECUTABLE)

  execute_process(
    COMMAND ${__GIT_EXECUTABLE} rev-parse --is-inside-work-tree
    WORKING_DIRECTORY "${CMAKE_SOURCE_DIR}"
    OUTPUT_VARIABLE __BUILDING_FROM_GIT_SOURCES
    OUTPUT_STRIP_TRAILING_WHITESPACE)

  if (__BUILDING_FROM_GIT_SOURCES)
    execute_process(
      COMMAND ${__GIT_EXECUTABLE} rev-parse --show-toplevel
      WORKING_DIRECTORY "${CMAKE_SOURCE_DIR}"
      OUTPUT_VARIABLE __GIT_TOPLEVEL_DIR
      OUTPUT_STRIP_TRAILING_WHITESPACE)
    execute_process(
      COMMAND ${__GIT_EXECUTABLE} rev-parse --git-dir
      WORKING_DIRECTORY "${__GIT_TOPLEVEL_DIR}"
      OUTPUT_VARIABLE __GIT_GIT_DIR
      OUTPUT_STRIP_TRAILING_WHITESPACE)
    execute_process(
      COMMAND ${__GIT_EXECUTABLE} --git-dir "${__GIT_GIT_DIR}" describe
      --abbrev=7 --always --dirty --tags
      WORKING_DIRECTORY "${__GIT_TOPLEVEL_DIR}"
      OUTPUT_VARIABLE __GIT_DESCRIBE_VERSION
      OUTPUT_STRIP_TRAILING_WHITESPACE)
    execute_process(
      COMMAND ${__GIT_EXECUTABLE} rev-list HEAD --max-count=1 --date-order
      WORKING_DIRECTORY "${__GIT_TOPLEVEL_DIR}"
      OUTPUT_VARIABLE __GIT_LATEST_COMMIT
      OUTPUT_STRIP_TRAILING_WHITESPACE)
    execute_process(
      COMMAND ${__GIT_EXECUTABLE} rev-list hydrogen --max-count=1 --date-order
      WORKING_DIRECTORY "${__GIT_TOPLEVEL_DIR}"
      OUTPUT_VARIABLE __GIT_LATEST_HYDROGEN_COMMIT
      OUTPUT_STRIP_TRAILING_WHITESPACE)

    set(HYDROGEN_GIT_VERSION "${__GIT_DESCRIBE_VERSION}")
    set(HYDROGEN_LATEST_SHA "${__GIT_LATEST_COMMIT}")
    set(HYDROGEN_LATEST_HYDROGEN_SHA "${__GIT_LATEST_HYDROGEN_COMMIT}")
  endif (__BUILDING_FROM_GIT_SOURCES)
endif (__GIT_EXECUTABLE)

# This must be set because version tags
set(HYDROGEN_VERSION_MAJOR 1)
set(HYDROGEN_VERSION_MINOR 5)
set(HYDROGEN_VERSION_PATCH 4)
set(HYDROGEN_VERSION_MAJOR_MINOR
    "${HYDROGEN_VERSION_MAJOR}.${HYDROGEN_VERSION_MINOR}")
set(HYDROGEN_VERSION
  "${HYDROGEN_VERSION_MAJOR_MINOR}.${HYDROGEN_VERSION_PATCH}")

# Back-compatibility, until it's all sorted out
set(EL_VERSION_MAJOR ${HYDROGEN_VERSION_MAJOR})
set(EL_VERSION_MINOR ${HYDROGEN_VERSION_MINOR})
set(EL_VERSION_PATCH ${HYDROGEN_VERSION_PATCH})

string(TOUPPER "${PROJECT_NAME}" UPPER_PROJECT_NAME)

if (NOT CMAKE_BUILD_TYPE MATCHES "Debug")
  set(HYDROGEN_RELEASE_BUILD TRUE)
endif ()

# Prevent in-source builds
if (PROJECT_SOURCE_DIR STREQUAL PROJECT_BINARY_DIR)
  message(FATAL_ERROR
    "In-source build attempted; please clean the CMake cache and then "
    "switch to an out-of-source build, e.g.,\n"
    "rm CMakeCache.txt && rm -Rf CMakeFiles/\nmkdir build/ && "
    "cd build/ && cmake ..\n")
endif ()

# Set the module path
list(APPEND CMAKE_MODULE_PATH "${CMAKE_SOURCE_DIR}/cmake")
list(APPEND CMAKE_MODULE_PATH "${CMAKE_SOURCE_DIR}/cmake/modules")

# Add warning flags
include(CheckCompilerFlag)
set(__CXX_FLAGS "-Wno-unused-parameter")
set(__CXX_DEVELOPER_FLAGS "-Wall" "-Wextra" "-Wpedantic" "-pedantic")
set(__CUDA_FLAGS)
set(__CUDA_DEVELOPER_FLAGS)

# Handle some CMake basics.
#
# These are highlighted because they default to OFF, but we prefer
# that they default to ON.
if (NOT DEFINED CMAKE_EXPORT_NO_PACKAGE_REGISTRY)
  set(CMAKE_EXPORT_NO_PACKAGE_REGISTRY ON)
endif ()

if (NOT DEFINED CMAKE_POSITION_INDEPENDENT_CODE)
  set(CMAKE_POSITION_INDEPENDENT_CODE ON)
endif ()

# Feature-related options
include(CMakeDependentOption)

option(Hydrogen_ENABLE_CUDA
  "Search for CUDA support and enable related features if found."
  OFF)

option(Hydrogen_ENABLE_ROCM
  "Search for ROCm/HIP support and enable related features if found."
  OFF)

cmake_dependent_option(Hydrogen_ENABLE_ROCTRACER
  "Search for Roctracer and enable related features if found."
  OFF
  "Hydrogen_ENABLE_ROCM"
  OFF)

if (Hydrogen_ENABLE_ROCM AND Hydrogen_ENABLE_CUDA)
  message(FATAL_ERROR
    "ROCm and CUDA code paths are mutually exclusive. "
    "Please enable the one that corresponds to your hardware. "
    "If you have mixed hardware, please contact the Hydrogen developers "
    "as this would be of great interest.")
endif ()

if (Hydrogen_ENABLE_CUDA OR Hydrogen_ENABLE_ROCM)
  option(Hydrogen_ENABLE_CUB
    "Search for CUB support and enable related features if found."
    ON)

  option(Hydrogen_ENABLE_GPU_TENSOR_MATH
    "Use the GPU tensor operations when available."
    OFF)

  option(Hydrogen_ENABLE_GPU_FP16
    "Enable FP16 arithmetic in GPU code."
    ON)
endif ()

option(Hydrogen_ENABLE_QUADMATH
  "Search for quadmath library and enable related features if found." OFF)

option(Hydrogen_ENABLE_QD
  "Search for QD library and enable related features if found." OFF)

option(Hydrogen_ENABLE_MPC
  "Search for MPC(+MPFR+GMP) library and enable related features if found."
  OFF)

#
# MEMORY-RELATED OPTIONS
#

option(Hydrogen_ENABLE_HALF
  "Enable the use of \"half\" library." OFF)
# This refers to the half.sourceforge.net type "half"

option(Hydrogen_ENABLE_BFLOAT16
  "Enable the use of Intel's bfloat16 type." OFF)
if (Hydrogen_ENABLE_BFLOAT16)
  message(WARNING
    "At time of writing, we are unaware of any implementation "
    "of the \"bfloat16\". This is a placeholder option. If our "
    "information is out of date, please open an issue and point "
    "us to an implementation.")
endif ()

option(Hydrogen_USE_64BIT_INTS
  "Use 64-bit integers for internal indexing" OFF)
option(Hydrogen_USE_64BIT_BLAS_INTS
  "Use 64-bit integers for BLAS/LAPACK" OFF)
if (Hydrogen_USE_64BIT_BLAS_INTS)
  set(EL_USE_64BIT_BLAS_INTS ON)
  set(Hydrogen_USE_64BIT_INTS ON)
endif ()

if (Hydrogen_USE_64BIT_INTS)
  set(EL_USE_64BIT_INTS ON)
endif ()

# Debugging and testing options

option(Hydrogen_ENABLE_BOUNDS_CHECKING
  "Do bounds checking in various routines regardless of DEBUG build."
  OFF)
set(HYDROGEN_DO_BOUNDS_CHECKING ${Hydrogen_ENABLE_BOUNDS_CHECKING})

option(Hydrogen_ENABLE_TESTING
  "Build the test suite in the tests directory." ON)

option(Hydrogen_ENABLE_UNIT_TESTS
  "Build the Catch2-based unit tests." OFF)

option(Hydrogen_ZERO_INIT "Initialize buffers to zero by default?" OFF)
mark_as_advanced(Hydrogen_ZERO_INIT)
if (Hydrogen_ZERO_INIT)
  set(EL_ZERO_INIT ${Hydrogen_ZERO_INIT})
endif ()

option(Hydrogen_ENABLE_VALGRIND
  "Search for valgrind and enable related features" OFF)
mark_as_advanced(Hydrogen_ENABLE_VALGRIND)
if (Hydrogen_ENABLE_VALGRIND)
  set(EL_ENABLE_VALGRIND ${Hydrogen_ENABLE_VALGRIND})
endif ()

# Profiling options

option(Hydrogen_ENABLE_NVPROF
  "Search for NVProf and enable related features" OFF)
option(Hydrogen_ENABLE_VTUNE
  "Search for VTune and enable related features" OFF)
option(Hydrogen_ENABLE_SYNCHRONOUS_PROFILING
  "Perform compute-stream synchronization before beginning profile regions"
  OFF)
if (Hydrogen_ENABLE_SYNCHRONOUS_PROFILING)
  set(HYDROGEN_DEFAULT_SYNC_PROFILING TRUE)
endif ()

# Thread/OpenMP parallelism options

option(Hydrogen_ENABLE_OPENMP
  "Enable OpenMP support in Hydrogen" OFF)

if (Hydrogen_ENABLE_OPENMP)
  option(Hydrogen_ENABLE_OMP_TASKLOOP
    "Use taskloops instead of parallel for" OFF)
endif ()

# MPI-related options

# CUDA-aware MPI can provide a benefit, if it works. However, it is
# often buggy. We can avoid it by setting this to ON.
option(Hydrogen_AVOID_CUDA_AWARE_MPI "Avoid CUDA-aware MPI" OFF)

option(Hydrogen_USE_CUSTOM_ALLTOALLV
  "Avoid MPI_Alltoallv for performance reasons" OFF)
mark_as_advanced(Hydrogen_USE_CUSTOM_ALLTOALLV)
if (Hydrogen_USE_CUSTOM_ALLTOALLV)
  set(EL_USE_CUSTOM_ALLTOALLV ${Hydrogen_USE_CUSTOM_ALLTOALLV})
endif ()

# Since it is surprisingly common for MPI libraries to have bugs in
# their support for complex data, the following option forces
# Elemental to cast all possible MPI communications in terms of twice
# as many real units of data.
option(Hydrogen_AVOID_COMPLEX_MPI
  "Avoid potentially buggy complex MPI routines" OFF)
mark_as_advanced(Hydrogen_AVOID_COMPLEX_MPI)
if (Hydrogen_AVOID_COMPLEX_MPI)
  set(EL_AVOID_COMPLEX_MPI ${Hydrogen_AVOID_COMPLEX_MPI})
endif ()

# Due to a subtle flaw in the Blue Gene/P extensions for MPICH2,
# treating floating-point data as a collection of byte-sized objects
# results in a better algorithm being chosen for MPI_Allgather. This
# should not effect performance on most machines.
option(Hydrogen_USE_BYTE_ALLGATHERS
  "Avoid BG/P allgather performance bug." OFF)
mark_as_advanced(Hydrogen_USE_BYTE_ALLGATHERS)
if (Hydrogen_USE_BYTE_ALLGATHERS)
  set(EL_USE_BYTE_ALLGATHERS ${Hydrogen_USE_BYTE_ALLGATHERS})
endif ()

# Runtime feedback options

# Print a warning any time a redistribution is performed which unpacks
# a large amount of data with a non-unit stride
option(Hydrogen_CACHE_WARNINGS
  "Warns when using cache-unfriendly routines" OFF)
mark_as_advanced(Hydrogen_CACHE_WARNINGS)
if (Hydrogen_CACHE_WARNINGS)
  set(EL_CACHE_WARNINGS ${Hydrogen_CACHE_WARNINGS})
endif ()

# Print a warning when an improperly aligned redistribution is
# performed, i.e., if an unnecessary permutation communication stage
# must take place
option(Hydrogen_UNALIGNED_WARNINGS
  "Warn when performing unaligned redistributions" OFF)
mark_as_advanced(Hydrogen_UNALIGNED_WARNINGS)
if (Hydrogen_CACHE_WARNINGS)
  set(EL_CACHE_WARNINGS ${Hydrogen_UNALIGNED_WARNINGS})
endif ()

# Print a warning if an opportunity was missed to implement a
# redistribution approach specifically for vectors (instead of
# matrices)
option(Hydrogen_VECTOR_WARNINGS
  "Warn when vector redistribution chances are missed" OFF)
mark_as_advanced(Hydrogen_VECTOR_WARNINGS)
if (Hydrogen_VECTOR_WARNINGS)
  set(EL_VECTOR_WARNINGS ${Hydrogen_VECTOR_WARNINGS})
endif ()

# Check the compiler features

if (Hydrogen_ENABLE_HALF)
  find_package(HALF REQUIRED)
  set(HYDROGEN_HAVE_HALF TRUE)
endif ()

if (Hydrogen_ENABLE_CUDA)
  enable_language(CUDA)

  find_package(CUDAToolkit REQUIRED)
  set(H_CUDA_CXX_LIBS
    CUDA::cudart
    CUDA::cublas
    CUDA::cusolver
    CUDA::nvml)

  if (Hydrogen_ENABLE_GPU_TENSOR_MATH)
    set(HYDROGEN_GPU_USE_TENSOR_OP_MATH TRUE)
  endif ()

  if (Hydrogen_ENABLE_CUB)
    if (CUDAToolkit_VERSION_MAJOR LESS 11)
      find_package(CUB MODULE REQUIRED)
      set(H_EXTERNAL_CUB TRUE)
      list(APPEND H_CUDA_CXX_LIBS cuda::cub)
    endif ()
    set(HYDROGEN_HAVE_CUB TRUE)
  endif ()

  if (Hydrogen_ENABLE_NVPROF)
    if (TARGET CUDA::nvToolsExt)
      list(APPEND H_CUDA_CXX_LIBS CUDA::nvToolsExt)
    else ()
      find_package(NVTX MODULE REQUIRED)
      set(H_EXTERNAL_NVTX TRUE)
      list(APPEND H_CUDA_CXX_LIBS cuda::nvtx)
    endif ()
    set(HYDROGEN_HAVE_NVPROF TRUE)
  endif ()

  set(HYDROGEN_HAVE_CUDA TRUE)

endif (Hydrogen_ENABLE_CUDA)

if (Hydrogen_ENABLE_ROCM)
  # This *must* come first!
  find_package(hip CONFIG REQUIRED)

  enable_language(HIP)

  if (Hydrogen_ENABLE_CUB)
    find_package(hipcub REQUIRED)
    set(HYDROGEN_HAVE_CUB TRUE)
  else ()
    set(HYDROGEN_HAVE_CUB FALSE)
  endif ()

  find_package(rocblas CONFIG REQUIRED)
  find_package(rocsolver CONFIG REQUIRED)
  find_package(rocthrust CONFIG REQUIRED)

  if (Hydrogen_ENABLE_ROCTRACER)
    find_package(Roctracer MODULE COMPONENTS roctx)
    set(HYDROGEN_HAVE_ROCTRACER ${Roctracer_FOUND})
  endif ()

  include(HydrogenCleanupHIPTargets)
  h_clean_hip_targets()

  set(HYDROGEN_HAVE_ROCM TRUE)
  message(STATUS "Found ROCm/HIP toolchain. Using HIP/ROCm.")

  set(H_ROCM_CXX_LIBS
    hip::host
    roc::rocblas
    roc::rocsolver
    roc::rocthrust
    ${Roctracer_LIBRARIES})
  if (HYDROGEN_HAVE_CUB)
    list(APPEND H_ROCM_CXX_LIBS hip::hipcub)
  endif ()
  set(H_ROCM_HIP_LIBS
    hip::device)
endif (Hydrogen_ENABLE_ROCM)

if (HYDROGEN_HAVE_CUDA OR HYDROGEN_HAVE_ROCM)
  set(HYDROGEN_HAVE_GPU TRUE)

  if (Hydrogen_ENABLE_GPU_FP16)
    set(HYDROGEN_GPU_USE_FP16 TRUE)
  endif ()
endif ()

find_package(Aluminum 1.0.0 CONFIG QUIET)
if (NOT Aluminum_FOUND AND Aluminum_NOT_FOUND_MESSAGE)
  message(STATUS
    "A candidate Aluminum > v1.0.0 was found, but was not selected:")
  message(STATUS
    "  ${Aluminum_NOT_FOUND_MESSAGE}")
endif ()
# Try again, since we're technically ok with >v0.7.0
if (NOT Aluminum_FOUND)
  find_package(Aluminum 0.7.0 CONFIG QUIET)
  if (NOT Aluminum_FOUND AND Aluminum_NOT_FOUND_MESSAGE)
    message(STATUS
      "A candidate Aluminum > v0.7.0 was found, but was not selected:")
    message(STATUS
      "  ${Aluminum_NOT_FOUND_MESSAGE}")
  endif ()
endif ()

if (Aluminum_FOUND)
  set(HYDROGEN_HAVE_ALUMINUM TRUE)
  message(STATUS
    "Found Aluminum@${ALUMINUM_VERSION}: ${Aluminum_DIR}")

  if (HYDROGEN_HAVE_GPU AND AL_HAS_NCCL)
    set(HYDROGEN_HAVE_NCCL2 TRUE)
    message(STATUS "Aluminum detected with NCCL2 backend support.")
  else ()
    set(HYDROGEN_HAVE_NCCL2 FALSE)
  endif (HYDROGEN_HAVE_GPU AND AL_HAS_NCCL)

  if (HYDROGEN_HAVE_GPU AND AL_HAS_HOST_TRANSFER)
    set(HYDROGEN_HAVE_AL_HOST_XFER TRUE)
    message(STATUS "Aluminum detected with HostTransfer backend support.")
  else ()
    set(HYDROGEN_HAVE_AL_HOST_XFER FALSE)
  endif (HYDROGEN_HAVE_GPU AND AL_HAS_HOST_TRANSFER)

  if (HYDROGEN_HAVE_GPU AND AL_HAS_MPI_CUDA)
    set(HYDROGEN_HAVE_AL_MPI_CUDA TRUE)
    message(STATUS "Aluminum detected with MPI-CUDA backend support.")
  else ()
    set(HYDROGEN_HAVE_AL_MPI_CUDA FALSE)
  endif (HYDROGEN_HAVE_GPU AND AL_HAS_MPI_CUDA)

  # Check for in-place SendRecv.
  if (ALUMINUM_VERSION VERSION_GREATER_EQUAL "1.3.0")
    set(HYDROGEN_AL_SUPPORTS_INPLACE_SENDRECV TRUE)
  else ()
    set(HYDROGEN_AL_SUPPORTS_INPLACE_SENDRECV FALSE)
  endif ()

  if (HYDROGEN_AL_SUPPORTS_INPLACE_SENDRECV)
    message(STATUS "Aluminum detected with in-place SendRecv support.")
  else ()
    message(STATUS "Aluminum detected WITHOUT in-place SendRecv support.")
  endif ()

else ()

  message(FATAL_ERROR "Aluminum support required but not found. "
    "Please set Aluminum_ROOT to its installation prefix or add "
    "the installation prefix to CMAKE_PREFIX_PATH.")

endif (Aluminum_FOUND)

# Sets up EL_RESTRICT and EL_HAVE_PRETTY_FUNCTION
include(detect/CXX)

# Other TPLs

if (Hydrogen_ENABLE_VTUNE)
  find_package(VTUNE REQUIRED)
  set(HYDROGEN_HAVE_VTUNE TRUE)
else ()
  set(HYDROGEN_HAVE_VTUNE FALSE)
endif ()

if (Hydrogen_ENABLE_OPENMP)
  include(FindAndVerifyOpenMP)
  if (OpenMP_FOUND)
    set(EL_HYBRID ON)
  else ()
    message(WARNING "Requested OpenMP support but OpenMP support was either "
      "not found or not functional.")
    set(EL_HYBRID OFF)
    set(Hydrogen_ENABLE_OPENMP OFF)
  endif ()
endif (Hydrogen_ENABLE_OPENMP)

include(FindAndVerifyMPI)
include(FindAndVerifyLAPACK)
include(FindAndVerifyExtendedPrecision)

# Catch2
if (Hydrogen_ENABLE_UNIT_TESTS)
  find_package(Catch2 2.0.0 CONFIG REQUIRED)
  include(Catch)
endif ()

# Get the source files

# Macro for setting up full paths
macro(set_full_path VAR)
  unset(__tmp_names)
  foreach(filename ${ARGN})
    unset(__name)
    get_filename_component(__name "${filename}" ABSOLUTE)
    list(APPEND __tmp_names "${__name}")
  endforeach()
  set(${VAR} "${__tmp_names}")
endmacro()

set(HYDROGEN_HEADERS)
set(HYDROGEN_SOURCES)
add_subdirectory(include)
add_subdirectory(src)

# Setup the configuration files for both "Elemental" and
# "Hydrogen". There's also one for Doxygen for building "build tree"
# docs (which has the advantage that preprocessing will take
# "{,hydrogen_}config.h" into consideration).
configure_file("${PROJECT_SOURCE_DIR}/cmake/configure_files/config.h.in"
  "${PROJECT_BINARY_DIR}/include/El/config.h" ESCAPE_QUOTES)
configure_file("${PROJECT_SOURCE_DIR}/cmake/configure_files/hydrogen_config.h.in"
  "${PROJECT_BINARY_DIR}/include/El/hydrogen_config.h")
configure_file("${PROJECT_SOURCE_DIR}/doxy/Doxyfile.in"
  "${PROJECT_BINARY_DIR}/doxy/Doxyfile")

add_library(Hydrogen_CXX "${HYDROGEN_SOURCES}" "${HYDROGEN_HEADERS}")
target_include_directories(Hydrogen_CXX PUBLIC
  $<BUILD_INTERFACE:${PROJECT_SOURCE_DIR}/include>
  $<BUILD_INTERFACE:${PROJECT_SOURCE_DIR}/include/El>
  $<BUILD_INTERFACE:${PROJECT_BINARY_DIR}/include>
  $<INSTALL_INTERFACE:include>)

# For now, add all flags regardless of "developer" or not.
foreach (_FLAG IN LISTS __CXX_FLAGS __CXX_DEVELOPER_FLAGS)
  check_compiler_flag(CXX "${_FLAG}" CXX_COMPLER_HAS_FLAG_${_FLAG})
  if (CXX_COMPILER_HAS_FLAG_${_FLAG})
    target_compile_options(Hydrogen_CXX PRIVATE
      $<$<COMPILE_LANGUAGE:CXX>:${_FLAG}>)
  endif ()
endforeach ()

target_link_libraries(
  Hydrogen_CXX PUBLIC
  ${Aluminum_LIBRARIES}
  ${HALF_LIBRARIES}
  ${VTUNE_LIBRARIES}
  ${NVTX_LIBRARIES}
  $<TARGET_NAME_IF_EXISTS:OpenMP::OpenMP_CXX>
  $<TARGET_NAME_IF_EXISTS:MPI::MPI_CXX>
  $<TARGET_NAME_IF_EXISTS:LAPACK::lapack>
  $<TARGET_NAME_IF_EXISTS:EP::extended_precision>

  ${H_CUDA_CXX_LIBS}
  ${H_ROCM_CXX_LIBS}
  )

set_target_properties(Hydrogen_CXX
  PROPERTIES
  VERSION ${HYDROGEN_VERSION})

if (NOT CMAKE_CXX_STANDARD)
  set_target_properties(Hydrogen_CXX
    PROPERTIES
    CXX_STANDARD 17)
endif ()

# Add the CXX library to "Hydrogen"
set(HYDROGEN_LIBRARIES Hydrogen_CXX)

if (HYDROGEN_HAVE_CUDA)
  add_library(Hydrogen_CUDA "${HYDROGEN_GPU_SOURCES}")
  target_include_directories(Hydrogen_CUDA PUBLIC
    $<BUILD_INTERFACE:${PROJECT_SOURCE_DIR}/include>
    $<BUILD_INTERFACE:${PROJECT_BINARY_DIR}/include>
    $<INSTALL_INTERFACE:include>)

  target_link_libraries(Hydrogen_CUDA
    PUBLIC
    ${HALF_LIBRARIES})

  foreach (_FLAG IN LISTS __CUDA_FLAGS __CUDA_DEVELOPER_FLAGS)
    check_compiler_flag(CUDA "${_FLAG}" CUDA_COMPLER_HAS_FLAG_${_FLAG})
    if (CUDA_COMPILER_HAS_FLAG_${_FLAG})
      target_compile_options(Hydrogen_CUDA PRIVATE
        $<$<COMPILE_LANGUAGE:CUDA>:${_FLAG}>)
    endif ()
  endforeach ()

  set_target_properties(Hydrogen_CUDA
    PROPERTIES
    VERSION ${HYDROGEN_VERSION})

  if (NOT CMAKE_CUDA_STANDARD)
    set_target_properties(Hydrogen_CUDA
      PROPERTIES CUDA_STANDARD 17)
  endif ()

  target_link_libraries(Hydrogen_CXX PUBLIC Hydrogen_CUDA)
  list(APPEND HYDROGEN_LIBRARIES Hydrogen_CUDA)
endif ()

if (HYDROGEN_HAVE_ROCM)
  add_library(Hydrogen_ROCM STATIC "${HYDROGEN_GPU_SOURCES}")

  # HIP files use the ".hip" extension. Since we don't want to copy
  # these files, we have to explicitly set their language to be HIP.
  get_target_property(H_HIP_SOURCES Hydrogen_ROCM SOURCES)
  set_source_files_properties(${H_HIP_SOURCES}
    PROPERTIES LANGUAGE HIP)

  target_include_directories(Hydrogen_ROCM PUBLIC
    $<BUILD_INTERFACE:${PROJECT_SOURCE_DIR}/include>
    $<BUILD_INTERFACE:${PROJECT_BINARY_DIR}/include>
    $<INSTALL_INTERFACE:include>
    )

  target_link_libraries(Hydrogen_ROCM PUBLIC
    ${HALF_LIBRARIES}
    ${ROCBLAS_LIBRARIES}
    )

  set_target_properties(Hydrogen_ROCM
    PROPERTIES
    VERSION ${HYDROGEN_VERSION})

  foreach (_FLAG IN LISTS __HIP_FLAGS __HIP_DEVELOPER_FLAGS)
    check_compiler_flag(HIP "${_FLAG}" HIP_COMPLER_HAS_FLAG_${_FLAG})
    if (HIP_COMPILER_HAS_FLAG_${_FLAG})
      target_compile_options(Hydrogen_ROCM PRIVATE
        $<$<COMPILE_LANGUAGE:HIP>:${_FLAG}>)
    endif ()
  endforeach ()

  #set_target_properties(Hydrogen_ROCM PROPERTIES LINKER_LANGUAGE CXX)
  list(APPEND HYDROGEN_LIBRARIES Hydrogen_ROCM)
endif ()

# Setup the tests
if (Hydrogen_ENABLE_TESTING OR Hydrogen_ENABLE_UNIT_TESTS)
  include(CTest)
endif ()

if (Hydrogen_ENABLE_TESTING)
  add_subdirectory(tests)
endif ()

if (Hydrogen_ENABLE_UNIT_TESTS)
  add_subdirectory(unit_test)
endif ()

# Setup the library install
install(TARGETS ${HYDROGEN_LIBRARIES}
  EXPORT HydrogenTargets
  RUNTIME DESTINATION bin
  ARCHIVE DESTINATION lib
  LIBRARY DESTINATION lib
  INCLUDES DESTINATION include/El
  )

# Install target stuff
include (CMakePackageConfigHelpers)

# Write the version file for both the build and the install tree
configure_file(
  "${CMAKE_SOURCE_DIR}/cmake/configure_files/HydrogenConfigVersion.cmake.in"
  "${CMAKE_BINARY_DIR}/HydrogenConfigVersion.cmake"
  @ONLY)

# Write the configure file for the build tree
set(INCLUDE_INSTALL_DIRS "${CMAKE_SOURCE_DIR}/include"
  "${CMAKE_SOURCE_DIR}/include"
  "${CMAKE_SOURCE_DIR}/include/El"
  "${CMAKE_BINARY_DIR}/include/El")
set(LIB_INSTALL_DIR "${CMAKE_BINARY_DIR}")
set(EXTRA_CMAKE_MODULE_DIR "${CMAKE_SOURCE_DIR}/cmake/modules")
configure_package_config_file(cmake/configure_files/HydrogenConfig.cmake.in
  "${CMAKE_BINARY_DIR}/HydrogenConfig.cmake"
  INSTALL_DESTINATION "${CMAKE_BINARY_DIR}"
  PATH_VARS INCLUDE_INSTALL_DIRS LIB_INSTALL_DIR)

# Build tree export
export(EXPORT HydrogenTargets NAMESPACE H:: FILE HydrogenTargets.cmake)

# Write the configure file for the install tree
set(INCLUDE_INSTALL_DIRS include)
set(LIB_INSTALL_DIR lib)
set(CMAKE_INSTALL_DIR ${LIB_INSTALL_DIR}/cmake/hydrogen)
set(EXTRA_CMAKE_MODULE_DIR)
configure_package_config_file(
  cmake/configure_files/HydrogenConfig.cmake.in
  "${CMAKE_BINARY_DIR}/cmake/HydrogenConfig.cmake"
  INSTALL_DESTINATION ${CMAKE_INSTALL_DIR}
  PATH_VARS INCLUDE_INSTALL_DIRS LIB_INSTALL_DIR)

# Install the configuration headers
install(FILES
  "${PROJECT_BINARY_DIR}/include/El/config.h"
  "${PROJECT_BINARY_DIR}/include/El/hydrogen_config.h"
  DESTINATION "include/El")

# Install tree export
install(EXPORT HydrogenTargets
  NAMESPACE H::
  DESTINATION ${CMAKE_INSTALL_DIR})

# Install the headers
install(DIRECTORY include DESTINATION .
  FILES_MATCHING PATTERN "*.hpp" PATTERN "*.h")

# Install the cmake stuff
install(FILES
  "${PROJECT_BINARY_DIR}/cmake/HydrogenConfig.cmake"
  "${PROJECT_BINARY_DIR}/HydrogenConfigVersion.cmake"
  DESTINATION ${CMAKE_INSTALL_DIR})
install(DIRECTORY cmake/modules DESTINATION ${CMAKE_INSTALL_DIR}
  FILES_MATCHING PATTERN "*.cmake")

# Summarize this configuration
if (NOT __dont_print_summary)
  include(PrintHydrogenSummary)
  print_full_hydrogen_summary(
    VERSION_VARIABLES
    HYDROGEN_VERSION
    HYDROGEN_GIT_VERSION
    HYDROGEN_LATEST_SHA
    HYDROGEN_LATEST_HYDROGEN_SHA

    PATH_VARIABLES
    CMAKE_CXX_COMPILER
    CMAKE_CUDA_COMPILER
    CMAKE_CUDA_HOST_COMPILER
    MPI_CXX_COMPILER
    CMAKE_INSTALL_PREFIX
    CMAKE_SOURCE_DIR
    CMAKE_BINARY_DIR

    STRING_VARIABLES
    CMAKE_BUILD_TYPE
    CMAKE_CUDA_FLAGS
    CMAKE_CXX_FLAGS
    HYDROGEN_BLAS_SUFFIX
    HYDROGEN_LAPACK_SUFFIX

    BOOLEAN_VARIABLES
    BUILD_SHARED_LIBS
    Hydrogen_ENABLE_TESTING
    HYDROGEN_HAVE_QUADMATH
    HYDROGEN_HAVE_QD
    HYDROGEN_HAVE_GMP
    HYDROGEN_HAVE_MPFR
    HYDROGEN_HAVE_MPC
    HYDROGEN_HAVE_ALUMINUM
    HYDROGEN_HAVE_NCCL2
    HYDROGEN_HAVE_AL_MPI_CUDA
    HYDROGEN_HAVE_CUDA
    HYDROGEN_HAVE_ROCM
    HYDROGEN_HAVE_CUB
    HYDROGEN_HAVE_OMP_TASKLOOP
    HYDROGEN_HAVE_CUDA_AWARE_MPI
    HYDROGEN_ENSURE_HOST_MPI_BUFFERS
    HYDROGEN_HAVE_OPENBLAS
    HYDROGEN_HAVE_LAPACK
    HYDROGEN_HAVE_MKL
    HYDROGEN_HAVE_MKL_GEMMT
    EL_USE_64BIT_INTS
    EL_USE_64BIT_BLAS_INTS
    EL_ZERO_INIT
    EL_HAVE_VALGRIND
    EL_HYBRID
    EL_HAVE_OPENMP
    EL_HAVE_OMP_COLLAPSE
    EL_HAVE_OMP_SIMD
    EL_USE_CUSTOM_ALLTOALLV
    EL_AVOID_COMPLEX_MPI
    EL_USE_BYTE_ALLGATHERS
    EL_CACHE_WARNINGS
    EL_VECTOR_WARNINGS

    PRINT_EMPTY_VARIABLES
    )
  set(__dont_print_summary ON CACHE INTERNAL "" FORCE)
  # Only print the configuration on initial configuration
endif ()