TheRock/rocm-systems/projects/hip-tests/catch/unit/deviceLib/CMakeLists.txt

# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
#
# SPDX-License-Identifier: MIT

# Common Tests - Test independent of all platforms
set(TEST_SRC
    floatMath.cc
    anyAll.cc
    ballot.cc
    clz.cc
    ffs.cc
    funnelshift.cc
    brev.cc
    popc.cc
    ldg.cc
    threadfence_system.cc
    syncthreadsand.cc
    syncthreadscount.cc
    syncthreadsor.cc
    deviceAllocation.cc
    Atomic_func.cc
    DoublePrecisionIntrinsics.cc
    DoublePrecisionMathDevice.cc
    DoublePrecisionMathHost.cc
    IntegerIntrinsics.cc
    SinglePrecisionIntrinsics.cc
    SinglePrecisionMathDevice.cc
    SinglePrecisionMathHost.cc
    SimpleAtomicsTest.cc
    hipTestAtomicAdd.cc
    hipStdComplex.cc
    hipTestClock.cc
    hip_trig.cc
    hipDeviceMemcpy.cc
    hipTestIncludeMath.cc
    hipTestDotFunctions.cc
    hipTestDeviceSymbol.cc
    hipTestNew.cc
    hipThreadFence.cc
    hipTestDevice.cc
    hipTestDeviceLimit.cc
    hipTestDeviceDouble.cc
    hipTestHost.cc
    hadd.cc
)
if(HIP_PLATFORM MATCHES "nvidia")
  set_source_files_properties(hipTestHost.cc PROPERTIES COMPILE_OPTIONS "--expt-relaxed-constexpr")
endif()

# AMD only tests
set(AMD_TEST_SRC
    unsafeAtomicAddDevice.cc
    mbcnt.cc
    bitExtract.cc
    bitInsert.cc
    floatTM.cc
    hipMathFunctions.cc
    hmax_hmin.cc
    hipBfloat16.cc
    hipVectorTypesHost.cc
    hipVectorTypesDevice.cc
    hipTestHalf.cc
    hipTestHalfConstexpr.cc
    hipComplex.cc
    hipTestFMA.cc
    hipTestNativeHalf.cc
    hip_test_make_type.cc
    bfloat16.cc
    AtomicsWithRandomActiveLanesInWavefront.cc
    fp16_ops.cc
    fp8_host.cc
    fp8_e8m0.cc
    fp6_ocp.cc
    fp4_ocp.cc
    memcpy_async.cc
    bfloat16_atomic_test.cc
    fp16_atomic_test.cc
)

# The OCP host reference (fcbx packed bitfield structs in amd_hip_ocp_host.hpp)
# relies on Itanium bitfield packing and fails a static_assert under the MSVC
# bitfield ABI on Windows. Only compile the fp4/fp6/fp8 OCP tests on non-Windows
# until the header is made portable.
if(NOT WIN32)
  list(APPEND AMD_TEST_SRC ext_ocp_fp6.cc ext_ocp_fp8.cc ext_ocp_fp4.cc)
endif()

set(AMD_ARCH_SPEC_TEST_SRC
    AtomicAdd_Coherent.cc
    AtomicAdd_NonCoherent.cc
    BuiltIns_fmax.cc
    BuiltIns_fmin.cc
    unsafeAtomicAdd_RTC.cc
    unsafeAtomicAdd_Coherent.cc
    unsafeAtomicAdd_NonCoherent.cc
)

set(AMD_GFX940_SPEC_TEST_SRC
    fp8_fnuz.cc
)

set(AMD_OCP_TEST_SRC
    fp8_ocp.cc
)

if (HIP_PLATFORM MATCHES "amd")
  add_custom_command(OUTPUT ${CMAKE_CURRENT_BINARY_DIR}/kerDevAllocMultCO.code
                  COMMAND ${CMAKE_HIP_COMPILER} ${HIP_DEVICE_BUILD_FLAGS_NO_ARCH} ${OFFLOAD_ARCH_LIST} --cuda-device-only
                  -x hip ${CMAKE_CURRENT_SOURCE_DIR}/kerDevAllocMultCO.cc
                  -o ${CMAKE_CURRENT_BINARY_DIR}/kerDevAllocMultCO.code
                  -I${HIP_INCLUDE_DIR}
                  ${HIP_PATH_OPT}
                  -I${CMAKE_CURRENT_SOURCE_DIR}/../../include
                  DEPENDS ${CMAKE_CURRENT_SOURCE_DIR}/kerDevAllocMultCO.cc
                  COMMENT "Compiling kerDevAllocMultCO.code")
  add_custom_command(OUTPUT ${CMAKE_CURRENT_BINARY_DIR}/kerDevWriteMultCO.code
                  COMMAND ${CMAKE_HIP_COMPILER} ${HIP_DEVICE_BUILD_FLAGS_NO_ARCH} ${OFFLOAD_ARCH_LIST} --cuda-device-only
                  -x hip ${CMAKE_CURRENT_SOURCE_DIR}/kerDevWriteMultCO.cc
                  -o ${CMAKE_CURRENT_BINARY_DIR}/kerDevWriteMultCO.code
                  -I${HIP_INCLUDE_DIR}
                  ${HIP_PATH_OPT}
                  -I${CMAKE_CURRENT_SOURCE_DIR}/../../include
                  DEPENDS ${CMAKE_CURRENT_SOURCE_DIR}/kerDevWriteMultCO.cc
                  COMMENT "Compiling kerDevWriteMultCO.code")
  add_custom_command(OUTPUT ${CMAKE_CURRENT_BINARY_DIR}/kerDevFreeMultCO.code
                  COMMAND ${CMAKE_HIP_COMPILER} ${HIP_DEVICE_BUILD_FLAGS_NO_ARCH} ${OFFLOAD_ARCH_LIST} --cuda-device-only
                  -x hip ${CMAKE_CURRENT_SOURCE_DIR}/kerDevFreeMultCO.cc
                  -o ${CMAKE_CURRENT_BINARY_DIR}/kerDevFreeMultCO.code
                  -I${HIP_INCLUDE_DIR}
                  ${HIP_PATH_OPT}
                  -I${CMAKE_CURRENT_SOURCE_DIR}/../../include
                  DEPENDS ${CMAKE_CURRENT_SOURCE_DIR}/kerDevFreeMultCO.cc
                  COMMENT "Compiling kerDevFreeMultCO.code")
  add_custom_command(OUTPUT ${CMAKE_CURRENT_BINARY_DIR}/kerDevAllocSingleKer.code
                  COMMAND ${CMAKE_HIP_COMPILER} ${HIP_DEVICE_BUILD_FLAGS_NO_ARCH} ${OFFLOAD_ARCH_LIST} --cuda-device-only
                  -x hip ${CMAKE_CURRENT_SOURCE_DIR}/kerDevAllocSingleKer.cc
                  -o ${CMAKE_CURRENT_BINARY_DIR}/kerDevAllocSingleKer.code
                  -I${HIP_INCLUDE_DIR}
                  ${HIP_PATH_OPT}
                  -I${CMAKE_CURRENT_SOURCE_DIR}/../../include
                  DEPENDS ${CMAKE_CURRENT_SOURCE_DIR}/kerDevAllocSingleKer.cc
                  COMMENT "Compiling kerDevAllocSingleKer.code")
else()
  add_custom_command(OUTPUT ${CMAKE_CURRENT_BINARY_DIR}/kerDevAllocMultCO.code
                  COMMAND ${CMAKE_HIP_COMPILER} --fatbin ${OFFLOAD_ARCH_LIST}
                  -x cu ${CMAKE_CURRENT_SOURCE_DIR}/kerDevAllocMultCO.cc
                  -o ${CMAKE_CURRENT_BINARY_DIR}/kerDevAllocMultCO.code
                  -I${HIP_INCLUDE_DIR}
                  -I${CMAKE_CURRENT_SOURCE_DIR}/../../include
                  DEPENDS ${CMAKE_CURRENT_SOURCE_DIR}/kerDevAllocMultCO.cc
                  COMMENT "Compiling kerDevAllocMultCO.code")
  add_custom_command(OUTPUT ${CMAKE_CURRENT_BINARY_DIR}/kerDevWriteMultCO.code
                  COMMAND ${CMAKE_HIP_COMPILER} --fatbin ${OFFLOAD_ARCH_LIST}
                  -x cu ${CMAKE_CURRENT_SOURCE_DIR}/kerDevWriteMultCO.cc
                  -o ${CMAKE_CURRENT_BINARY_DIR}/kerDevWriteMultCO.code
                  -I${HIP_INCLUDE_DIR}
                  -I${CMAKE_CURRENT_SOURCE_DIR}/../../include
                  DEPENDS ${CMAKE_CURRENT_SOURCE_DIR}/kerDevWriteMultCO.cc
                  COMMENT "Compiling kerDevWriteMultCO.code")
  add_custom_command(OUTPUT ${CMAKE_CURRENT_BINARY_DIR}/kerDevFreeMultCO.code
                  COMMAND ${CMAKE_HIP_COMPILER} --fatbin ${OFFLOAD_ARCH_LIST}
                  -x cu ${CMAKE_CURRENT_SOURCE_DIR}/kerDevFreeMultCO.cc
                  -o ${CMAKE_CURRENT_BINARY_DIR}/kerDevFreeMultCO.code
                  -I${HIP_INCLUDE_DIR}
                  -I${CMAKE_CURRENT_SOURCE_DIR}/../../include
                  DEPENDS ${CMAKE_CURRENT_SOURCE_DIR}/kerDevFreeMultCO.cc
                  COMMENT "Compiling kerDevFreeMultCO.code")
  add_custom_command(OUTPUT ${CMAKE_CURRENT_BINARY_DIR}/kerDevAllocSingleKer.code
                  COMMAND ${CMAKE_HIP_COMPILER} --fatbin ${OFFLOAD_ARCH_LIST}
                  -x cu ${CMAKE_CURRENT_SOURCE_DIR}/kerDevAllocSingleKer.cc
                  -o ${CMAKE_CURRENT_BINARY_DIR}/kerDevAllocSingleKer.code
                  -I${HIP_INCLUDE_DIR}
                  -I${CMAKE_CURRENT_SOURCE_DIR}/../../include
                  DEPENDS ${CMAKE_CURRENT_SOURCE_DIR}/kerDevAllocSingleKer.cc
                  COMMENT "Compiling kerDevAllocSingleKer.code")
endif()
add_custom_target(kerDevAllocMultCO_code DEPENDS ${CMAKE_CURRENT_BINARY_DIR}/kerDevAllocMultCO.code)
add_custom_target(kerDevWriteMultCO_code DEPENDS ${CMAKE_CURRENT_BINARY_DIR}/kerDevWriteMultCO.code)
add_custom_target(kerDevFreeMultCO_code DEPENDS ${CMAKE_CURRENT_BINARY_DIR}/kerDevFreeMultCO.code)
add_custom_target(kerDevAllocSingleKer_code DEPENDS ${CMAKE_CURRENT_BINARY_DIR}/kerDevAllocSingleKer.code)
set_property(GLOBAL APPEND PROPERTY G_INSTALL_CUSTOM_TARGETS ${CMAKE_CURRENT_BINARY_DIR}/kerDevAllocSingleKer.code)
set_property(GLOBAL APPEND PROPERTY G_INSTALL_CUSTOM_TARGETS ${CMAKE_CURRENT_BINARY_DIR}/kerDevFreeMultCO.code)
set_property(GLOBAL APPEND PROPERTY G_INSTALL_CUSTOM_TARGETS ${CMAKE_CURRENT_BINARY_DIR}/kerDevWriteMultCO.code)
set_property(GLOBAL APPEND PROPERTY G_INSTALL_CUSTOM_TARGETS ${CMAKE_CURRENT_BINARY_DIR}/kerDevAllocMultCO.code)

# Accepted archs to compile this cmake file
set(ACCEPTED_OFFLOAD_ARCHS gfx90a gfx942)
set(ACCEPTED_GFX940_ARCH gfx942)
set(ACCEPTED_OCP_ARCH gfx1200 gfx1201 gfx950)
function(CheckAcceptedArchs OFFLOAD_ARCH_STR_LOCAL)
  set(ARCH_CHECK -1 PARENT_SCOPE)
  set(ARCH_GFX940 -1 PARENT_SCOPE)
  set(ARCH_OCP -1 PARENT_SCOPE)
  string(REGEX MATCHALL "--offload-arch=gfx[0-9a-z]+" OFFLOAD_ARCH_LIST ${OFFLOAD_ARCH_STR_LOCAL})
  foreach(OFFLOAD_ARCH IN LISTS OFFLOAD_ARCH_LIST)
    string(REGEX MATCHALL "--offload-arch=(gfx[0-9a-z]+)" matches ${OFFLOAD_ARCH})
    if (CMAKE_MATCH_COUNT EQUAL 1)
      if (CMAKE_MATCH_1 IN_LIST ACCEPTED_OFFLOAD_ARCHS)
        set(ARCH_CHECK 1 PARENT_SCOPE)
      endif()         # CMAKE_MATCH_1
      if (CMAKE_MATCH_1 IN_LIST ACCEPTED_GFX940_ARCH)
        set(ARCH_GFX940 1 PARENT_SCOPE)
      endif()
      if (CMAKE_MATCH_1 IN_LIST ACCEPTED_OCP_ARCH)
        set(ARCH_OCP 1 PARENT_SCOPE)
      endif()
    endif()        # CMAKE_MATCH_COUNT
  endforeach()   # OFFLOAD_ARCH_LIST
endfunction()  # CheckAcceptedArchs

if(HIP_PLATFORM MATCHES "amd")
  # Do not enable arch specific tests for ASAN.
  if (DEFINED OFFLOAD_ARCH_STR AND NOT (ENABLE_ADDRESS_SANITIZER OR THEROCK_SANITIZER STREQUAL "ASAN" OR THEROCK_SANITIZER STREQUAL "HOST_ASAN"))
    CheckAcceptedArchs(${OFFLOAD_ARCH_STR})
  elseif(DEFINED $ENV{HCC_AMDGPU_TARGET})
    CheckAcceptedArchs($ENV{HCC_AMDGPU_TARGET})
  else()
    set(ARCH_CHECK -1)
    set(ARCH_GFX940 -1)
    set(ARCH_OCP -1)
  endif()
  set(TEST_SRC ${TEST_SRC} ${AMD_TEST_SRC})
  set_source_files_properties(bfloat16.cc PROPERTIES COMPILE_FLAGS "-DHIP_ENABLE_WARP_SYNC_BUILTINS")
  if(${ARCH_CHECK} GREATER_EQUAL 0)
    set(TEST_SRC ${TEST_SRC} ${AMD_ARCH_SPEC_TEST_SRC})
    set_source_files_properties(AtomicAdd_Coherent_withunsafeflag.cc PROPERTIES COMPILE_OPTIONS "-munsafe-fp-atomics")
    set_source_files_properties(AtomicAdd_NonCoherent_withunsafeflag.cc PROPERTIES COMPILE_OPTIONS "-munsafe-fp-atomics")
    set_source_files_properties(AtomicAdd_Coherent_withnoUnsafeflag.cc PROPERTIES COMPILE_OPTIONS "-mno-unsafe-fp-atomics")
    set_source_files_properties(AtomicAdd_NonCoherent_withnoUnsafeflag.cc PROPERTIES COMPILE_OPTIONS "-mno-unsafe-fp-atomics")
    set_source_files_properties(unsafeAtomicAdd_Coherent_withunsafeflag.cc PROPERTIES COMPILE_OPTIONS "-munsafe-fp-atomics")
    set_source_files_properties(unsafeAtomicAdd_NonCoherent_withunsafeflag.cc PROPERTIES COMPILE_OPTIONS "-munsafe-fp-atomics")
    set_source_files_properties(unsafeAtomicAdd_Coherent_withnounsafeflag.cc PROPERTIES COMPILE_OPTIONS "-mno-unsafe-fp-atomics")
    set_source_files_properties(unsafeAtomicAdd_NonCoherent_withnounsafeflag.cc PROPERTIES COMPILE_OPTIONS "-mno-unsafe-fp-atomics")
    set_source_files_properties(hipMathFunctions.cc PROPERTIES COMPILE_FLAGS "-Xclang -fallow-half-arguments-and-returns")
    file(GLOB AtomicAdd_files CONFIGURE_DEPENDS *AtomicAdd_*_*.cc)
    set_property(SOURCE ${AtomicAdd_files} PROPERTY COMPILE_FLAGS --save-temps)
    file(GLOB unsafeAtomicAdd_files CONFIGURE_DEPENDS *unsafeAtomicAdd_*_*.cc)
    set_property(SOURCE ${unsafeAtomicAdd_files} PROPERTY COMPILE_FLAGS --save-temps)
  endif()
  if(${ARCH_GFX940} GREATER_EQUAL 0)
    set(TEST_SRC ${TEST_SRC} ${AMD_GFX940_SPEC_TEST_SRC})
  endif()
  if(${ARCH_OCP} GREATER_EQUAL 0)
    set(TEST_SRC ${TEST_SRC} ${AMD_OCP_TEST_SRC})
  endif()
  hip_add_exe_to_target(NAME UnitDeviceTests
                      TEST_SRC ${TEST_SRC}
                      TEST_TARGET_NAME build_tests
                      LINKER_LIBS hiprtc::hiprtc)
elseif(HIP_PLATFORM MATCHES "nvidia")
  hip_add_exe_to_target(NAME UnitDeviceTests
                      TEST_SRC ${TEST_SRC}
                      TEST_TARGET_NAME build_tests
                      COMPILE_OPTIONS --Wno-deprecated-declarations)
endif()

add_dependencies(UnitDeviceTests kerDevAllocMultCO_code kerDevWriteMultCO_code kerDevFreeMultCO_code kerDevAllocSingleKer_code)