# ----------------------------------------------------------------------------
# Copyright (c) 2025-2026 Huawei Technologies Co., Ltd.
# This program is free software, you can redistribute it and/or modify it under the terms and conditions of
# CANN Open Software License Agreement Version 2.0 (the "License").
# Please refer to the License for details. You may not use this file except in compliance with the License.
# THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
# INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
# See LICENSE in the root of the software repository for the full text of the License.
# ----------------------------------------------------------------------------
# 为target link依赖库 usage: add_modules(MODE SUB_LIBS EXTERNAL_LIBS) SUB_LIBS 为内部创建的target, EXTERNAL_LIBS为外部依赖的target
function(add_modules)
set(oneValueArgs MODE)
set(multiValueArgs TARGETS SUB_LIBS EXTERNAL_LIBS)
cmake_parse_arguments(ARGS "" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
foreach(target ${ARGS_TARGETS})
if(TARGET ${target} AND TARGET ${ARGS_SUB_LIBS})
target_link_libraries(${target} ${ARGS_MODE} ${ARGS_SUB_LIBS})
endif()
if(TARGET ${target} AND ARGS_EXTERNAL_LIBS)
target_link_libraries(${target} ${ARGS_MODE} ${ARGS_EXTERNAL_LIBS})
endif()
endforeach()
endfunction()
# 添加opbase object
function(add_opbase_modules)
if(TARGET opbase_infer_objs OR TARGET opbase_tiling_objs OR TARGET opbase_util_objs)
return()
endif()
file(GLOB_RECURSE OPS_BASE_INFER_SRC
${OPBASE_SOURCE_PATH}/src/op_common/op_host/infershape_broadcast_util.cpp
${OPBASE_SOURCE_PATH}/src/op_common/op_host/infershape_elewise_util.cpp
${OPBASE_SOURCE_PATH}/src/op_common/op_host/infershape_reduce_util.cpp
)
file(GLOB_RECURSE OPS_BASE_TILING_SRC
${OPBASE_SOURCE_PATH}/src/op_common/atvoss/elewise/*.cpp
${OPBASE_SOURCE_PATH}/src/op_common/atvoss/broadcast/*.cpp
${OPBASE_SOURCE_PATH}/src/op_common/atvoss/reduce/*.cpp
${OPBASE_SOURCE_PATH}/src/op_common/op_host/tiling_util.cpp
)
file(GLOB_RECURSE OPS_BASE_UTIL_SRC
${OPBASE_SOURCE_PATH}/src/op_common/op_host/util/*.cpp
${OPBASE_SOURCE_PATH}/src/op_common/log/*.cpp
)
if(OPS_BASE_INFER_SRC)
add_library(opbase_infer_objs OBJECT ${OPS_BASE_INFER_SRC})
target_include_directories(opbase_infer_objs PRIVATE ${OP_PROTO_INCLUDE})
target_compile_options(opbase_infer_objs
PRIVATE
$<$<NOT:$<BOOL:${ENABLE_TEST}>>:-DDISABLE_COMPILE_V1> -Dgoogle=ascend_private
-fvisibility=hidden
)
target_link_libraries(
opbase_infer_objs
PRIVATE $<BUILD_INTERFACE:$<IF:$<BOOL:${ENABLE_TEST}>,intf_llt_pub_asan_cxx17,intf_pub_cxx17>>
$<BUILD_INTERFACE:dlog_headers>
)
endif()
if(OPS_BASE_TILING_SRC)
add_library(opbase_tiling_objs OBJECT ${OPS_BASE_TILING_SRC})
target_include_directories(opbase_tiling_objs PRIVATE ${OP_TILING_INCLUDE})
target_compile_options(opbase_tiling_objs
PRIVATE
$<$<NOT:$<BOOL:${ENABLE_TEST}>>:-DDISABLE_COMPILE_V1> -Dgoogle=ascend_private
-fvisibility=hidden -fno-strict-aliasing
)
target_link_libraries(
opbase_tiling_objs
PRIVATE $<BUILD_INTERFACE:$<IF:$<BOOL:${ENABLE_TEST}>,intf_llt_pub_asan_cxx17,intf_pub_cxx17>>
$<BUILD_INTERFACE:dlog_headers>
tiling_api
)
endif()
if(OPS_BASE_UTIL_SRC)
add_library(opbase_util_objs OBJECT ${OPS_BASE_UTIL_SRC})
target_include_directories(opbase_util_objs PRIVATE ${OP_TILING_INCLUDE})
target_compile_options(opbase_util_objs
PRIVATE
$<$<NOT:$<BOOL:${ENABLE_TEST}>>:-DDISABLE_COMPILE_V1> -Dgoogle=ascend_private
-fvisibility=hidden
)
target_link_libraries(
opbase_util_objs
PRIVATE $<BUILD_INTERFACE:$<IF:$<BOOL:${ENABLE_TEST}>,intf_llt_pub_asan_cxx17,intf_pub_cxx17>>
$<BUILD_INTERFACE:dlog_headers>
)
endif()
endfunction()
# 添加infer object
function(add_infer_modules)
if(NOT TARGET ${OPHOST_NAME}_infer_obj)
if(BUILD_WITH_INSTALLED_DEPENDENCY_CANN_PKG)
npu_op_library(${OPHOST_NAME}_infer_obj GRAPH)
else()
add_library(${OPHOST_NAME}_infer_obj OBJECT)
endif()
target_include_directories(${OPHOST_NAME}_infer_obj PRIVATE ${OP_PROTO_INCLUDE})
target_compile_definitions(
${OPHOST_NAME}_infer_obj PRIVATE OPS_UTILS_LOG_SUB_MOD_NAME="OP_PROTO" OP_SUBMOD_NAME="OPS_NN"
$<$<BOOL:${ENABLE_TEST}>:ASCEND_OPSPROTO_UT> LOG_CPP
)
target_compile_options(
${OPHOST_NAME}_infer_obj PRIVATE $<$<NOT:$<BOOL:${ENABLE_TEST}>>:-DDISABLE_COMPILE_V1> -Dgoogle=ascend_private
-fvisibility=hidden
)
target_link_libraries(
${OPHOST_NAME}_infer_obj
PRIVATE $<BUILD_INTERFACE:$<IF:$<BOOL:${ENABLE_TEST}>,intf_llt_pub_asan_cxx17,intf_pub_cxx17>>
$<BUILD_INTERFACE:dlog_headers>
$<$<TARGET_EXISTS:opbase_util_objs>:$<TARGET_OBJECTS:opbase_util_objs>>
$<$<TARGET_EXISTS:opbase_infer_objs>:$<TARGET_OBJECTS:opbase_infer_objs>>
)
endif()
endfunction()
# 添加tiling object
function(add_tiling_modules)
if(NOT TARGET ${OPHOST_NAME}_tiling_obj)
if(BUILD_WITH_INSTALLED_DEPENDENCY_CANN_PKG)
npu_op_library(${OPHOST_NAME}_tiling_obj TILING)
add_dependencies(${OPHOST_NAME}_tiling_obj json)
else()
add_library(${OPHOST_NAME}_tiling_obj OBJECT)
add_dependencies(${OPHOST_NAME}_tiling_obj json)
endif()
target_include_directories(${OPHOST_NAME}_tiling_obj PRIVATE ${OP_TILING_INCLUDE})
# 始终编译 QueryBank 知识库查询实现相关源文件(file GLOB 不存在的文件会被静默跳过)
file(GLOB RUNTIME_KB_SRC ${OPS_NN_DIR}/common/src/op_host/bank_mgr/*.cpp)
if(RUNTIME_KB_SRC)
target_sources(${OPHOST_NAME}_tiling_obj PRIVATE ${RUNTIME_KB_SRC})
endif()
# 子包模式:额外编译需要 LegacyCommonMgr 的源文件
set(ENABLE_DLOPEN_LEGACY OFF)
if (BUILD_WITH_INSTALLED_DEPENDENCY_CANN_PKG AND NOT ENABLE_STATIC)
file(GLOB COMMON_SRC ${OPS_NN_DIR}/common/src/*.cpp ${OPS_NN_DIR}/common/src/op_host/*.cpp)
if(UT_TEST_ALL OR OP_HOST_UT) # ut场景下LegacyCommonMgr要打桩,通过环境变量查找legacy so
file(GLOB COMMON_SRC ${OPS_NN_DIR}/common/src/op_host/*.cpp)
endif()
list(REMOVE_ITEM COMMON_SRC ${RUNTIME_KB_SRC})
target_sources(${OPHOST_NAME}_tiling_obj PRIVATE ${COMMON_SRC})
set(ENABLE_DLOPEN_LEGACY ON)
endif()
foreach(op_category ${OP_CATEGORY_LIST})
if(ENABLE_EXPERIMENTAL)
set(op_category_dir ${OPS_NN_DIR}/experimental/${op_category})
else()
set(op_category_dir ${OPS_NN_DIR}/${op_category})
endif()
if (IS_DIRECTORY ${op_category_dir})
target_include_directories(${OPHOST_NAME}_tiling_obj PRIVATE ${op_category_dir})
endif()
endforeach()
target_compile_definitions(
${OPHOST_NAME}_tiling_obj PRIVATE OPS_UTILS_LOG_SUB_MOD_NAME="OP_TILING" OP_SUBMOD_NAME="OPS_NN"
$<$<BOOL:${ENABLE_TEST}>:ASCEND_OPTILING_UT> LOG_CPP
$<$<BOOL:${ENABLE_DLOPEN_LEGACY}>:NN_ENABLE_DLOPEN_LEGACY>
ASCEND_COMPUTE_UNIT="${ASCEND_COMPUTE_UNIT}"
)
target_compile_options(
${OPHOST_NAME}_tiling_obj PRIVATE $<$<NOT:$<BOOL:${ENABLE_TEST}>>:-DDISABLE_COMPILE_V1> -Dgoogle=ascend_private
-fvisibility=hidden -fno-strict-aliasing
)
target_link_libraries(
${OPHOST_NAME}_tiling_obj
PRIVATE $<BUILD_INTERFACE:$<IF:$<BOOL:${ENABLE_TEST}>,intf_llt_pub_asan_cxx17,intf_pub_cxx17>>
$<BUILD_INTERFACE:dlog_headers>
$<$<TARGET_EXISTS:${COMMON_NAME}_obj>:$<TARGET_OBJECTS:${COMMON_NAME}_obj>>
$<$<TARGET_EXISTS:opbase_util_objs>:$<TARGET_OBJECTS:opbase_util_objs>>
$<$<TARGET_EXISTS:opbase_tiling_objs>:$<TARGET_OBJECTS:opbase_tiling_objs>>
tiling_api
)
endif()
endfunction()
# 添加opapi object
function(add_opapi_modules)
if(NOT TARGET ${OPHOST_NAME}_opapi_obj)
if(BUILD_WITH_INSTALLED_DEPENDENCY_CANN_PKG)
npu_op_library(${OPHOST_NAME}_opapi_obj ACLNN)
else()
add_library(${OPHOST_NAME}_opapi_obj OBJECT)
endif()
set(ENABLE_DLOPEN_LEGACY OFF)
if (BUILD_WITH_INSTALLED_DEPENDENCY_CANN_PKG AND NOT ENABLE_STATIC)
if(NOT UT_TEST_ALL AND NOT OP_API_UT) # ut场景下LegacyCommonMgr要打桩,通过环境变量查找legacy so
set(LEGACY_COMMON_MANAGER_SRC ${OPS_NN_DIR}/common/src/legacy_common_manager.cpp)
target_sources(${OPHOST_NAME}_opapi_obj PRIVATE ${LEGACY_COMMON_MANAGER_SRC})
set_source_files_properties(${LEGACY_COMMON_MANAGER_SRC} PROPERTIES COMPILE_OPTIONS "-UOP_LOG_LIBOPAPI_ONLY")
endif()
set(ENABLE_DLOPEN_LEGACY ON)
endif()
if(ENABLE_TEST)
set(opapi_ut_depends_inc ${UT_PATH}/op_api/stub)
endif()
target_include_directories(${OPHOST_NAME}_opapi_obj PRIVATE
${opapi_ut_depends_inc}
${OPAPI_INCLUDE})
target_include_directories(${OPHOST_NAME}_opapi_obj PRIVATE ${OPAPI_INCLUDE})
target_compile_options(${OPHOST_NAME}_opapi_obj PRIVATE -Dgoogle=ascend_private -DACLNN_LOG_FMT_CHECK -DOP_LOG_LIBOPAPI_ONLY)
target_compile_definitions(${OPHOST_NAME}_opapi_obj PRIVATE
LOG_CPP
$<$<BOOL:${ENABLE_DLOPEN_LEGACY}>:NN_ENABLE_DLOPEN_LEGACY>
ASCEND_COMPUTE_UNIT="${ASCEND_COMPUTE_UNIT}"
)
target_link_libraries(
${OPHOST_NAME}_opapi_obj
PUBLIC $<BUILD_INTERFACE:$<IF:$<BOOL:${ENABLE_TEST}>,intf_llt_pub_asan_cxx17,intf_pub_cxx17>>
PRIVATE $<BUILD_INTERFACE:adump_headers> $<BUILD_INTERFACE:dlog_headers>
$<$<TARGET_EXISTS:opbase_util_objs>:$<TARGET_OBJECTS:opbase_util_objs>>
$<$<TARGET_EXISTS:cann_version_headers>:cann_version_headers>
)
endif()
endfunction()
# determine whether aicpu kernels skip the processing.
function(skip_aicpu_kernel op_type ascend_op_name)
set(SKIP_AICPU_FLAG FALSE PARENT_SCOPE)
if(NO_AICPU)
message(STATUS "disable aicpu kernel ${op_type}, skip it.")
set(SKIP_AICPU_FLAG TRUE PARENT_SCOPE)
return()
endif()
if(NOT "${ascend_op_name}" STREQUAL "")
set(_ascend_op_name_tmp "${ascend_op_name}")
separate_arguments(_ascend_op_name_tmp)
list(FIND _ascend_op_name_tmp "${op_type}" _index)
if(_index EQUAL -1)
message(STATUS "[${op_type}] skipped, not in ascend_op_name list: ${ascend_op_name}")
set(SKIP_AICPU_FLAG TRUE PARENT_SCOPE)
return()
else()
message(STATUS "[${op_type}] selected, in ascend_op_name list: ${ascend_op_name}")
endif()
endif()
endfunction()
# AICPU built-in package manifest: check whether <op_name> is allowed into the built-in package.
# Only gates device-side AICPU packaging (op_def/json + kernel .so); host constant folding
# and UT are not affected. ENABLE_CUSTOM bypasses the manifest.
# Sets _AICPU_BUILTIN_PKG_ALLOWED (TRUE/FALSE) in the caller's scope.
function(check_aicpu_builtin_pkg_allowed op_name)
set(_allowed FALSE)
if(AICPU_BUILTIN_PACKAGE_FORCE_ALL)
set(_allowed TRUE)
else()
foreach(_entry ${AICPU_BUILTIN_PACKAGE_MANIFEST})
if(_entry STREQUAL "${op_name}=ON")
set(_allowed TRUE)
break()
endif()
endforeach()
endif()
if(_allowed)
message(STATUS "[${op_name}] allowed in AICPU built-in package by manifest")
endif()
set(_AICPU_BUILTIN_PKG_ALLOWED ${_allowed} PARENT_SCOPE)
endfunction()
function(add_aicpu_kernel_modules)
message(STATUS "add_aicpu_kernel_modules")
if(NOT TARGET ${OPHOST_NAME}_aicpu_obj)
add_library(${OPHOST_NAME}_aicpu_obj OBJECT)
target_include_directories(${OPHOST_NAME}_aicpu_obj PRIVATE ${AICPU_INCLUDE})
target_compile_definitions(
${OPHOST_NAME}_aicpu_obj PRIVATE _FORTIFY_SOURCE=2 google=ascend_private
$<$<BOOL:${ENABLE_TEST}>:ASCEND_AICPU_UT>
)
target_compile_options(
${OPHOST_NAME}_aicpu_obj PRIVATE $<$<NOT:$<BOOL:${ENABLE_TEST}>>:-DDISABLE_COMPILE_V1> -Dgoogle=ascend_private
-fvisibility=hidden ${AICPU_DEFINITIONS}
)
target_link_libraries(
${OPHOST_NAME}_aicpu_obj
PRIVATE $<BUILD_INTERFACE:$<IF:$<BOOL:${ENABLE_TEST}>,intf_llt_pub_asan_cxx17,intf_pub_cxx17>>
$<BUILD_INTERFACE:dlog_headers>
)
endif()
endfunction()
option(PREPROCESS_ONLY "preprocess only, no cache aicpu targets" OFF)
function(add_aicpu_cust_kernel_modules target_name)
message(STATUS "add_aicpu_cust_kernel_modules for ${target_name}")
if(NOT TARGET ${target_name})
add_library(${target_name} OBJECT)
target_include_directories(${target_name} PRIVATE ${AICPU_INCLUDE})
target_compile_definitions(
${target_name} PRIVATE
_FORTIFY_SOURCE=2 _GLIBCXX_USE_CXX11_ABI=1
google=ascend_private
$<$<BOOL:${ENABLE_TEST}>:ASCEND_AICPU_UT>
)
target_compile_options(
${target_name} PRIVATE
$<$<NOT:$<BOOL:${ENABLE_TEST}>>:-DDISABLE_COMPILE_V1> -Dgoogle=ascend_private
-fvisibility=hidden ${AICPU_DEFINITIONS}
)
target_link_libraries(
${target_name}
PRIVATE $<BUILD_INTERFACE:$<IF:$<BOOL:${ENABLE_TEST}>,intf_llt_pub_asan_cxx17,intf_pub_cxx17>>
$<BUILD_INTERFACE:dlog_headers>
-Wl,--no-whole-archive
Eigen3::Eigen
)
if(NOT PREPROCESS_ONLY)
if (NOT ${target_name} IN_LIST AICPU_CUST_OBJ_TARGETS)
set(AICPU_CUST_OBJ_TARGETS ${AICPU_CUST_OBJ_TARGETS} ${target_name} CACHE INTERNAL "All aicpu cust obj targets")
endif()
endif()
endif()
endfunction()
# Compiles aicpu source as OBJECT for host side (x86).
# Collects into AICPU_HOST_OBJ_TARGETS; linking into libopconstant_folding_nn.so
# is done in symbol.cmake gen_aicpu_const_symbol().
function(add_aicpu_host_kernel_modules host_target_name)
message(STATUS "add_aicpu_host_kernel_modules for ${host_target_name}")
if(NOT TARGET ${host_target_name})
add_library(${host_target_name} OBJECT)
target_include_directories(${host_target_name} PRIVATE
${AICPU_INCLUDE}
${CANN_3RD_LIB_PATH}/eigen
)
target_compile_definitions(
${host_target_name} PRIVATE
_FORTIFY_SOURCE=2
google=ascend_private
OPS_NN_AICPU_HOST_KERNEL
)
target_compile_options(
${host_target_name} PRIVATE
-Dgoogle=ascend_private
-fvisibility=hidden ${AICPU_DEFINITIONS}
)
target_link_libraries(
${host_target_name}
PRIVATE $<BUILD_INTERFACE:$<IF:$<BOOL:${ENABLE_TEST}>,intf_llt_pub_asan_cxx14,intf_pub_cxx14>>
$<BUILD_INTERFACE:dlog_headers>
Eigen3::Eigen
)
get_property(_host_targets GLOBAL PROPERTY AICPU_HOST_OBJ_TARGETS)
if (NOT host_target_name IN_LIST _host_targets)
set_property(GLOBAL APPEND PROPERTY
AICPU_HOST_OBJ_TARGETS ${host_target_name})
message(STATUS "Aicpu register host target: ${host_target_name}")
else()
message(STATUS "Skip duplicate target: ${host_target_name}")
endif()
endif()
endfunction()
function(add_graph_plugin_modules)
if(NOT TARGET ${GRAPH_PLUGIN_NAME}_obj)
if(BUILD_WITH_INSTALLED_DEPENDENCY_CANN_PKG)
npu_op_library(${GRAPH_PLUGIN_NAME}_obj GRAPH)
else()
add_library(${GRAPH_PLUGIN_NAME}_obj OBJECT)
endif()
add_dependencies(${GRAPH_PLUGIN_NAME}_obj json)
target_include_directories(${GRAPH_PLUGIN_NAME}_obj PRIVATE
${OP_PROTO_INCLUDE}
${JSON_INCLUDE}
${PROJECT_SOURCE_DIR}/common/inc
${PROJECT_SOURCE_DIR}/common/inc/op_graph
${ASCEND_DIR}/include
${ASCEND_DIR}/include/external
${ASCEND_DIR}/include/exe_graph
${ASCEND_DIR}/include/base/context_builder
${ASCEND_DIR}/include/ge
)
target_compile_definitions(${GRAPH_PLUGIN_NAME}_obj PRIVATE OPS_UTILS_LOG_SUB_MOD_NAME="GRAPH_PLUGIN" LOG_CPP)
if(BUILD_WITH_INSTALLED_DEPENDENCY_CANN_PKG)
target_compile_options(
${GRAPH_PLUGIN_NAME}_obj PRIVATE -Dgoogle=ascend_private -fvisibility=hidden
)
else()
target_compile_options(
${GRAPH_PLUGIN_NAME}_obj PRIVATE $<$<NOT:$<BOOL:${ENABLE_TEST}>>:-DDISABLE_COMPILE_V1> -Dgoogle=ascend_private
-fvisibility=hidden
)
endif()
if(BUILD_WITH_INSTALLED_DEPENDENCY_CANN_PKG)
target_link_libraries(
${GRAPH_PLUGIN_NAME}_obj
PRIVATE $<BUILD_INTERFACE:$<IF:$<BOOL:${ENABLE_TEST}>,intf_llt_pub_asan_cxx17,intf_pub_cxx17>>
$<BUILD_INTERFACE:dlog_headers>
$<$<TARGET_EXISTS:opbase_util_objs>:$<TARGET_OBJECTS:opbase_util_objs>>
$<$<TARGET_EXISTS:opbase_infer_objs>:$<TARGET_OBJECTS:opbase_infer_objs>>
metadef
graph
register
ge_compiler
)
else()
target_link_libraries(
${GRAPH_PLUGIN_NAME}_obj
PRIVATE $<BUILD_INTERFACE:$<IF:$<BOOL:${ENABLE_TEST}>,intf_llt_pub_asan_cxx17,intf_pub_cxx17>>
$<BUILD_INTERFACE:dlog_headers>
$<$<TARGET_EXISTS:opbase_util_objs>:$<TARGET_OBJECTS:opbase_util_objs>>
$<$<TARGET_EXISTS:opbase_infer_objs>:$<TARGET_OBJECTS:opbase_infer_objs>>
)
endif()
endif()
endfunction()
macro(add_op_subdirectory)
file(GLOB CURRENT_DIRS RELATIVE ${OP_DIR} ${OP_DIR}/*)
list(FIND ASCEND_OP_NAME ${OP_NAME} INDEX)
if(NOT "${ASCEND_OP_NAME}" STREQUAL "" AND INDEX EQUAL -1)
# 非指定算子,只编译不测试
set(OP_ONLY_COMPILE on)
else()
set(OP_ONLY_COMPILE off)
endif()
if((NOT ENABLE_TEST AND NOT BENCHMARK) OR OP_ONLY_COMPILE)
list(REMOVE_ITEM CURRENT_DIRS tests)
endif()
# op_api目录已移出的算子,add_modules_sources在算子根路径CMakeLists中,且CMakeLists旧实现已删除
if(NOT EXISTS "${OP_DIR}/op_host/CMakeLists.txt")
add_subdirectory(${OP_DIR})
endif()
foreach(SUB_DIR ${CURRENT_DIRS})
if(EXISTS "${OP_DIR}/${SUB_DIR}/CMakeLists.txt")
add_subdirectory(${OP_DIR}/${SUB_DIR})
endif()
endforeach()
endmacro()
# usage: add_category_subdirectory 根据ASCEND_OP_NAME和ASCEND_COMPILE_OPS添加指定算子工程
# ASCEND_OP_NAME 指定的算子 ASCEND_COMPILE_OPS 编译需要的算子
macro(add_category_subdirectory)
foreach(op_category ${OP_CATEGORY_LIST})
if(ENABLE_EXPERIMENTAL)
set(op_category_dir ${CMAKE_CURRENT_SOURCE_DIR}/experimental/${op_category})
else()
set(op_category_dir ${CMAKE_CURRENT_SOURCE_DIR}/${op_category})
endif()
if (IS_DIRECTORY ${op_category_dir})
file(GLOB CURRENT_DIRS RELATIVE ${op_category_dir} ${op_category_dir}/*)
foreach(SUB_DIR ${CURRENT_DIRS})
set(OP_DIR ${op_category_dir}/${SUB_DIR})
set(OP_NAME "${SUB_DIR}")
if(${OP_NAME} STREQUAL "common")
set(OP_NAME "${op_category}.common")
endif()
list(FIND ASCEND_COMPILE_OPS ${OP_NAME} INDEX)
if(NOT "${ASCEND_OP_NAME}" STREQUAL "" AND INDEX EQUAL -1)
# ASCEND_OP_NAME 为空表示全部编译
continue()
endif()
if(EXISTS "${OP_DIR}/CMakeLists.txt")
file(READ "${OP_DIR}/CMakeLists.txt" OP_CMAKE_CONTENT)
if(NOT OP_CMAKE_CONTENT MATCHES "add_sources\\s*\\(")
add_op_subdirectory()
endif()
else()
if (EXISTS "${OP_DIR}/op_host/CMakeLists.txt")
add_subdirectory(${OP_DIR}/op_host)
endif()
endif()
endforeach()
endif()
endforeach()
if("${ASCEND_OP_NAME}" STREQUAL "add_example" OR "${ASCEND_OP_NAME}" STREQUAL "add_example_aicpu"
OR "${ASCEND_OP_NAME}" STREQUAL "add_example_pypto")
add_subdirectory(examples)
endif()
endmacro()
# 从两个长度一致的列表中查找相同位置的元素
function(find_value_by_key key_list value_list search_key result)
list(LENGTH key_list key_list_length)
list(LENGTH value_list value_list_length)
if(NOT ${key_list_length} EQUAL ${value_list_length})
message(FATAL_ERROR "key_list length is ${key_list_length}, value_list length is ${value_list_length}, not equal")
endif()
set(found_value "")
if(key_list_length GREATER 0)
list(FIND key_list ${search_key} index)
if(NOT ${index} EQUAL -1)
list(GET value_list ${index} found_value)
endif()
endif()
set(${result} ${found_value} PARENT_SCOPE)
endfunction()
# ------------------------------------------------------------------------------------------------------------
# require_pypto_pro(<op_name>)
# Require pypto_pro for the current operator. Call at the beginning of an operator's CMakeLists.txt, before
# registering any host or kernel build rules. If pypto_pro is unavailable in ASCEND_PYTHON_EXECUTABLE, warn
# and return from the caller's CMakeLists.txt so the whole operator is skipped.
# ------------------------------------------------------------------------------------------------------------
macro(require_pypto_pro op_name)
get_property(_pypto_pro_checked GLOBAL PROPERTY PYPTO_PRO_CHECKED)
if(NOT _pypto_pro_checked)
execute_process(
COMMAND ${ASCEND_PYTHON_EXECUTABLE} -c
"import importlib.util, sys; sys.exit(0 if importlib.util.find_spec('pypto_pro') else 1)"
RESULT_VARIABLE _pypto_pro_check_result
OUTPUT_QUIET
ERROR_QUIET
)
if(_pypto_pro_check_result EQUAL 0)
set_property(GLOBAL PROPERTY PYPTO_PRO_AVAILABLE TRUE)
else()
set_property(GLOBAL PROPERTY PYPTO_PRO_AVAILABLE FALSE)
endif()
set_property(GLOBAL PROPERTY PYPTO_PRO_CHECKED TRUE)
endif()
get_property(_pypto_pro_available GLOBAL PROPERTY PYPTO_PRO_AVAILABLE)
if(NOT _pypto_pro_available)
message(WARNING "pypto_pro is unavailable in ${ASCEND_PYTHON_EXECUTABLE}; skip operator ${op_name}.")
return()
endif()
unset(_pypto_pro_checked)
unset(_pypto_pro_check_result)
unset(_pypto_pro_available)
endmacro()
# ------------------------------------------------------------------------------------------------------------
# enable_pypto_kernel(<op_file>)
# Mark an operator whose kernel is written in PyPTO. Call ONCE from the op's CMakeLists.txt, and BEFORE
# add_all_modules_sources(...) / add_modules_sources(...) (so the tiling logic can read PYPTO_GEN_DIR).
# ------------------------------------------------------------------------------------------------------------
function(enable_pypto_kernel op_file)
set(_compute_units ${ASCEND_COMPUTE_UNIT})
list(FILTER _compute_units INCLUDE REGEX ".+")
if(NOT _compute_units)
message(FATAL_ERROR "enable_pypto_kernel: ASCEND_COMPUTE_UNIT is empty")
endif()
if(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/op_kernel/${op_file}.py")
set(_op_dir "${CMAKE_CURRENT_SOURCE_DIR}")
elseif(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/../op_kernel/${op_file}.py")
get_filename_component(_op_dir "${CMAKE_CURRENT_SOURCE_DIR}/.." REALPATH)
else()
message(FATAL_ERROR "enable_pypto_kernel: kernel python file not found: "
"${CMAKE_CURRENT_SOURCE_DIR}/op_kernel/${op_file}.py")
endif()
set(_py "${_op_dir}/op_kernel/${op_file}.py")
get_filename_component(_op_name "${_op_dir}" NAME)
list(FIND ASCEND_OP_NAME ${_op_name} _op_index)
if(NOT "${ASCEND_OP_NAME}" STREQUAL "" AND _op_index EQUAL -1)
return()
endif()
# Per-soc codegen: generated headers depend on the target arch, so emit one set per compute unit
# into binary/<soc>/<op_file>_pypto_gen. The first soc's set drives the shared host tiling include.
set(_gen_dirs)
foreach(_soc_unit ${_compute_units})
set(_gen "${CMAKE_BINARY_DIR}/binary/${_soc_unit}/${op_file}_pypto_gen")
message(STATUS "pypto codegen: ${op_file} (${_soc_unit}) -> ${_gen}")
execute_process(
COMMAND ${ASCEND_PYTHON_EXECUTABLE} ${OPS_NN_DIR}/scripts/util/pypto_codegen.py
--py-file "${_py}" --out-dir "${_gen}"
--op-file "${op_file}" --soc "${_soc_unit}"
RESULT_VARIABLE _rc
)
if(NOT _rc EQUAL 0)
message(FATAL_ERROR "pypto codegen failed for ${op_file} (${_soc_unit}) (exit ${_rc})")
endif()
list(APPEND _gen_dirs ${_gen})
endforeach()
set_property(DIRECTORY APPEND PROPERTY CMAKE_CONFIGURE_DEPENDS ${_py})
list(GET _gen_dirs 0 _primary_gen_dir)
set(PYPTO_GEN_DIR ${_primary_gen_dir} PARENT_SCOPE)
# Keyed by the current configure only: reset the cache before writing so a re-configure with a
# different --soc never accumulates stale per-soc dirs.
unset(${_op_name}_pypto_gen_dirs CACHE)
set(${_op_name}_pypto_gen_dirs "${_gen_dirs}" CACHE INTERNAL "pypto gen dirs for op ${_op_name}")
set_property(GLOBAL APPEND PROPERTY PYPTO_ENABLED_OPS ${op_file})
set_property(GLOBAL APPEND PROPERTY PYPTO_ENABLED_COMPUTE_UNITS ${_compute_units})
endfunction()
# ------------------------------------------------------------------------------------------------------------
# pypto_soc_to_arch(<soc> <out_var>)
# Resolve a soc to the pypto PYPTO_JIT_ARCH token. The authoritative mapping lives in
# scripts/util/pypto_codegen.py (single source of truth, queried once per configure via --print-arch-map)
# so the two sides cannot drift. Unknown socs fail the configure instead of silently producing a
# kernel whose arch pypto auto-detected differently.
# ------------------------------------------------------------------------------------------------------------
function(pypto_soc_to_arch soc out_var)
string(TOLOWER "${soc}" _soc)
set(${out_var} "" PARENT_SCOPE)
get_property(_pypto_arch_lines GLOBAL PROPERTY PYPTO_ARCH_MAP_LINES)
if(NOT _pypto_arch_lines)
execute_process(
COMMAND ${ASCEND_PYTHON_EXECUTABLE} ${OPS_NN_DIR}/scripts/util/pypto_codegen.py --print-arch-map
OUTPUT_VARIABLE _pypto_arch_lines
RESULT_VARIABLE _pypto_arch_map_rc
OUTPUT_STRIP_TRAILING_WHITESPACE
ERROR_QUIET
)
if(NOT _pypto_arch_map_rc EQUAL 0)
message(FATAL_ERROR "pypto: failed to query soc->arch map from pypto_codegen.py (exit ${_pypto_arch_map_rc})")
endif()
string(REPLACE "\n" ";" _pypto_arch_lines "${_pypto_arch_lines}")
set_property(GLOBAL PROPERTY PYPTO_ARCH_MAP_LINES "${_pypto_arch_lines}")
endif()
foreach(_pypto_line ${_pypto_arch_lines})
if(_pypto_line MATCHES "^${_soc}=(.+)$")
set(${out_var} "${CMAKE_MATCH_1}" PARENT_SCOPE)
return()
endif()
endforeach()
message(FATAL_ERROR "pypto: no arch mapping for soc '${soc}'; update _ARCH_MAP in "
"scripts/util/pypto_codegen.py (see build.sh SOC_TO_ARCH).")
endfunction()
# ------------------------------------------------------------------------------------------------------------
# pypto_jit_arch_env(<soc> <out_var>)
# ------------------------------------------------------------------------------------------------------------
function(pypto_jit_arch_env soc out_var)
set(${out_var} "" PARENT_SCOPE)
get_property(_pypto_ops GLOBAL PROPERTY PYPTO_ENABLED_OPS)
if(NOT _pypto_ops)
return()
endif()
# Only socs that actually compile a pypto kernel need PYPTO_JIT_ARCH; other socs in the same build
# (e.g. a regular AscendC op's compile unit) must stay untouched.
get_property(_pypto_cus GLOBAL PROPERTY PYPTO_ENABLED_COMPUTE_UNITS)
string(TOLOWER "${soc}" _soc)
list(FIND _pypto_cus ${_soc} _cu_index)
if(_cu_index EQUAL -1)
return()
endif()
pypto_soc_to_arch("${soc}" _arch)
set(${out_var} export PYPTO_JIT_ARCH=${_arch} && PARENT_SCOPE)
endfunction()
# ------------------------------------------------------------------------------------------------------------
# pypto_include_tiling_gen()
# 将 PyPTO codegen 目录(PYPTO_GEN_DIR)以 target_include_directories 的方式注入
# tiling 目标的 include 路径。目标级 include 跨目录安全(CMake targets 全局可见),
# 仅要求 CMake 3.16;tiling 源文件需显式 include 生成的 *_tilingkey.h / *_tiling.h。
# 替代旧方案(set_source_files_properties + TARGET_DIRECTORY,需 CMake >= 3.18)。
# ------------------------------------------------------------------------------------------------------------
function(pypto_include_tiling_gen)
if(NOT DEFINED PYPTO_GEN_DIR)
return()
endif()
if(NOT TARGET ${OPHOST_NAME}_tiling_obj)
return()
endif()
target_include_directories(${OPHOST_NAME}_tiling_obj PRIVATE ${PYPTO_GEN_DIR})
endfunction()
function(add_tiling_sources source_dir tiling_dir disable_in_opp)
if(NOT disable_in_opp)
set(disable_in_opp FALSE)
endif()
if(NOT BUILD_WITH_INSTALLED_DEPENDENCY_CANN_PKG AND disable_in_opp)
message(STATUS "don't need add tiling sources")
return()
endif()
if("${tiling_dir}" STREQUAL "")
file(GLOB OPTILING_SRCS ${source_dir}/*_tiling*.cpp)
else()
file(GLOB OPTILING_SRCS ${source_dir}/*_tiling*.cpp ${source_dir}/${tiling_dir}/*_tiling*.cpp)
endif()
file(GLOB_RECURSE SUB_OPTILING_SRC ${source_dir}/op_tiling/*.cpp)
if (OPTILING_SRCS OR SUB_OPTILING_SRC)
add_tiling_modules()
# 注入 PyPTO codegen 头文件的 include 路径(CMake 3.16 兼容;源文件显式 include)
pypto_include_tiling_gen()
target_sources(${OPHOST_NAME}_tiling_obj PRIVATE ${OPTILING_SRCS} ${SUB_OPTILING_SRC})
get_filename_component(_tiling_parent_inc "${source_dir}/../.." ABSOLUTE)
target_include_directories(${OPHOST_NAME}_tiling_obj PRIVATE ${_tiling_parent_inc} ${source_dir})
endif()
endfunction()
# usage: add_modules_sources(DIR OPTYPE ACLNNTYPE DEPENDENCIES NPU_ARCH TILING_DIR DISABLE_IN_OPP ALLOW_HF32) ACLNNTYPE 支持类型aclnn/aclnn_inner/aclnn_exclude OPTYPE 和 ACLNNTYPE
# DEPENDENCIES 算子依赖
# NPU_ARCH 设置支持的npuArch列表,必须与TILING_DIR一一对应,示例:dav_2201 dav_3510
# COMPUTE_UNIT 兼容参数:设置支持芯片版本号(soc列表),必须与TILING_DIR一一对应,内部逐项转换为npuArch后统一按arch匹配;与NPU_ARCH互斥
# TILING_DIR 设置所支持芯片类型对应的tiling文件目录,必须与NPU_ARCH(或COMPUTE_UNIT)一一对应,示例:arch22 arch35
# DISABLE_IN_OPP 设置是否在opp包中编译tiling文件,布尔类型:TRUE,FALSE
# ALLOW_HF32 配置算子允许HF32执行,取值:enable_hi_float_32_execution / enable_float_32_execution,
# 配置后记录OPTYPE对应的算子并按OPTYPE去重,用于生成ops_allow_hf32_nn.ini
# 需一一对应
function(add_modules_sources)
set(multiValueArgs OPTYPE ACLNNTYPE DEPENDENCIES NPU_ARCH COMPUTE_UNIT TILING_DIR DIR)
set(oneValueArgs DISABLE_IN_OPP ALLOW_HF32 HOSTCPU)
cmake_parse_arguments(MODULE "" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
if(MODULE_DIR)
set(SOURCE_DIR ${MODULE_DIR})
else()
set(SOURCE_DIR ${CMAKE_CURRENT_SOURCE_DIR})
endif()
add_opbase_modules()
# opapi 默认全部编译
file(GLOB OPAPI_SRCS ${SOURCE_DIR}/op_api/*.cpp)
if(OPAPI_SRCS)
add_opapi_modules()
target_sources(${OPHOST_NAME}_opapi_obj PRIVATE ${OPAPI_SRCS})
endif()
file(GLOB OPAPI_HEADERS ${SOURCE_DIR}/op_api/aclnn_*.h)
if(OPAPI_HEADERS)
target_sources(${OPHOST_NAME}_aclnn_exclude_headers INTERFACE ${OPAPI_HEADERS})
target_include_directories(${OPHOST_NAME}_opapi_obj PRIVATE ${SOURCE_DIR}/op_api)
endif()
# op_api目录已移出的算子,SOURCE_DIR为算子根目录路径,路径不以op_host结尾;移出前SOURCE_DIR为算子op_host目录路径
if(NOT "${SOURCE_DIR}" MATCHES "/op_host$")
set(_OP_ROOT ${SOURCE_DIR})
set(SOURCE_DIR ${SOURCE_DIR}/op_host)
else()
get_filename_component(_OP_ROOT ${SOURCE_DIR} DIRECTORY)
endif()
# 获取算子层级目录名称
get_filename_component(PARENT_DIR ${SOURCE_DIR} DIRECTORY)
get_filename_component(OP_NAME ${PARENT_DIR} NAME)
file(APPEND ${ASCEND_SUB_CONFIG_PATH} "OP_CATEGORY;${op_category};OP_NAME;${OP_NAME};${ARGN}\n")
# 记录全局的COMPILED_OPS和COMPILED_OP_DIRS,其中COMPILED_OP_DIRS只记录到算子名,例如math/abs,common不记录
if (NOT ${OP_NAME} STREQUAL "common")
set(COMPILED_OPS
${COMPILED_OPS} ${OP_NAME}
CACHE STRING "Compiled Ops" FORCE
)
set(COMPILED_OP_DIRS
${COMPILED_OP_DIRS} ${PARENT_DIR}
CACHE STRING "Compiled Ops Dirs" FORCE
)
endif()
add_tf_plugin_sources()
# op_graph: 从算子根目录下的 op_graph/ 递归收集所有 .cpp
file(GLOB_RECURSE OP_GRAPH_SRCS ${_OP_ROOT}/op_graph/*.cpp)
if(OP_GRAPH_SRCS)
add_graph_plugin_modules()
target_sources(${GRAPH_PLUGIN_NAME}_obj PRIVATE ${OP_GRAPH_SRCS})
endif()
file(GLOB OP_GRAPH_PROTO_HEADERS ${_OP_ROOT}/op_graph/*_proto*.h)
if(OP_GRAPH_PROTO_HEADERS)
target_sources(${GRAPH_PLUGIN_NAME}_proto_headers INTERFACE ${OP_GRAPH_PROTO_HEADERS})
endif()
add_onnx_plugin_sources()
file(GLOB OPINFER_SRCS ${SOURCE_DIR}/*_infershape*.cpp)
if(OPINFER_SRCS)
add_infer_modules()
target_sources(${OPHOST_NAME}_infer_obj PRIVATE ${OPINFER_SRCS})
endif()
# 添加tiling文件:NPU_ARCH 与 TILING_DIR 按位置一一配对;兼容旧 COMPUTE_UNIT 参数,
# 逐项转换为 npuArch(字符串拼接保留空元素位置,避免 list(APPEND) 丢弃空元素导致错位),
# 当前构建芯片同样转换为 npuArch 后统一按 arch 匹配
if(MODULE_NPU_ARCH)
get_npu_arch_by_soc_version("${ASCEND_COMPUTE_UNIT}" _cur_key)
set(_cur_keys "${MODULE_NPU_ARCH}")
else()
set(_cur_key "${ASCEND_COMPUTE_UNIT}")
set(_cur_keys "${MODULE_COMPUTE_UNIT}")
endif()
find_value_by_key("${_cur_keys}" "${MODULE_TILING_DIR}" "${_cur_key}" tiling_dir)
add_tiling_sources("${SOURCE_DIR}" "${tiling_dir}" "${MODULE_DISABLE_IN_OPP}")
file(GLOB AICPU_SRCS ${SOURCE_DIR}/*_aicpu.cpp)
if(AICPU_SRCS)
add_aicpu_kernel_modules()
target_sources(${OPHOST_NAME}_aicpu_obj PRIVATE ${AICPU_SRCS})
endif()
if(MODULE_HOSTCPU AND BUILD_WITH_INSTALLED_DEPENDENCY_CANN_PKG AND NOT ENABLE_CUSTOM)
file(GLOB AICPU_HOST_SRCS ${_OP_ROOT}/op_kernel_aicpu/*_aicpu.cpp)
if(AICPU_HOST_SRCS)
set(HOST_OBJ_NAME ${OP_NAME}_host_const_obj)
add_aicpu_host_kernel_modules(${HOST_OBJ_NAME})
target_sources(${HOST_OBJ_NAME} PRIVATE ${AICPU_HOST_SRCS})
endif()
endif()
if(MODULE_OPTYPE)
list(LENGTH MODULE_OPTYPE OpTypeLen)
list(LENGTH MODULE_ACLNNTYPE AclnnTypeLen)
if(NOT ${OpTypeLen} EQUAL ${AclnnTypeLen})
message(FATAL_ERROR "OPTYPE AND ACLNNTYPE Should be One-to-One")
endif()
math(EXPR index "${OpTypeLen} - 1")
foreach(i RANGE ${index})
list(GET MODULE_OPTYPE ${i} OpType)
list(GET MODULE_ACLNNTYPE ${i} AclnnType)
if(${AclnnType} STREQUAL "aclnn"
OR ${AclnnType} STREQUAL "aclnn_inner"
OR ${AclnnType} STREQUAL "aclnn_exclude"
)
file(GLOB OPDEF_SRCS ${SOURCE_DIR}/${OpType}_def*.cpp)
if(NOT ${MODULE_EXT} STREQUAL "")
file(GLOB OPDEF_EXT_SRCS ${MODULE_EXT}/${op_category}/${OP_NAME}/op_host/${OpType}*_def*.cpp)
list(APPEND OPDEF_SRCS ${OPDEF_EXT_SRCS})
endif()
if(MODULE_ALLOW_HF32)
record_allow_hf32_ops("${OpType}" "${MODULE_ALLOW_HF32}" "${OPDEF_SRCS}")
endif()
if(OPDEF_SRCS)
target_sources(${OPHOST_NAME}_opdef_${AclnnType}_obj INTERFACE ${OPDEF_SRCS})
endif()
elseif(${AclnnType} STREQUAL "no_need_aclnn")
message(STATUS "aicpu or host aicpu no need aclnn.")
else()
message(FATAL_ERROR "ACLNN TYPE UNSUPPORTED, ONLY SUPPORT aclnn/aclnn_inner/aclnn_exclude")
endif()
endforeach()
else()
file(GLOB OPDEF_SRCS ${SOURCE_DIR}/*_def*.cpp)
if(OPDEF_SRCS)
message(
FATAL_ERROR
"Should Manually specify aclnn/aclnn_inner/aclnn_exclude\n"
"usage: add_modules_sources(OPTYPE optypes ACLNNTYPE aclnntypes)\n"
"example: add_modules_sources(OPTYPE add ACLNNTYPE aclnn_exclude)"
)
endif()
endif()
endfunction()
# 记录配置了ALLOW_HF32的算子,以注册的OpType(从*_def.cpp的OP_ADD解析,大写驼峰)为键去重,用于生成ops_allow_hf32_nn.ini
# usage: record_allow_hf32_ops(op_type allow_hf32_value def_files) op_type 为add_modules_sources传入的OPTYPE(snake_case算子目录名),def_files 为对应*_def.cpp文件列表
function(record_allow_hf32_ops op_type allow_hf32_value def_files)
if(NOT allow_hf32_value MATCHES "^enable_(hi_)?float_32_execution$")
message(FATAL_ERROR "ALLOW_HF32 only supports enable_hi_float_32_execution or enable_float_32_execution, but got: ${allow_hf32_value}")
endif()
get_op_type_from_def_files("${def_files}" registered_type)
if(NOT registered_type)
message(WARNING "record_allow_hf32_ops: cannot resolve registered op type from def files of ${op_type}, skip recording")
return()
endif()
set(recorded FALSE)
foreach(entry ${ALLOW_HF32_NN_OPS})
if(entry MATCHES "^${registered_type}=")
set(recorded TRUE)
break()
endif()
endforeach()
if(NOT recorded)
list(APPEND ALLOW_HF32_NN_OPS "${registered_type}=${allow_hf32_value}")
set(ALLOW_HF32_NN_OPS "${ALLOW_HF32_NN_OPS}" CACHE INTERNAL "ops which allow hf32" FORCE)
endif()
endfunction()
# 安装待迁移的legacy impl_mode ini文件(scripts/package/legacy)到opp/built-in/op_impl/ai_core/tbe/impl_mode,与生成的ops_allow_hf32_nn.ini同目录
# 背景:allow_hf32_matmul_*_conv_*.ini从canndev仓迁移至此,canndev仓后续会删除这四个文件;
# 存量环境带有canndev安装的同名文件,安装时不覆盖、卸载时不删除(见OpsNNInfo.xml与opp_install.sh),
# 过渡期后随存量环境消化删除相应保护逻辑
function(install_allow_hf32_legacy_ini)
if(ENABLE_CUSTOM)
return()
endif()
install(DIRECTORY "${CMAKE_CURRENT_SOURCE_DIR}/scripts/package/legacy/"
DESTINATION ${OPP_PREFIX}/built-in/op_impl/ai_core/tbe/impl_mode
)
endfunction()
# 汇总ALLOW_HF32配置,生成ops_allow_hf32_nn.ini并安装到opp/built-in/op_impl/ai_core/tbe/impl_mode,在顶层CMakeLists的add_category_subdirectory之后调用
# ops_allow_hf32_nn.ini为本仓新增的默认读取文件(替代canndev迁移来的allow_hf32_matmul_*_conv_*.ini四个文件)
# ini一定会生成:未通过ALLOW_HF32配置的算子从默认legacy文件allow_hf32_matmul_f_conv_t.ini补充
function(gen_allow_hf32_ini)
if(ENABLE_CUSTOM)
return()
endif()
set(allow_hf32_ops ${ALLOW_HF32_NN_OPS})
set(legacy_ini "${CMAKE_CURRENT_SOURCE_DIR}/scripts/package/legacy/allow_hf32_matmul_f_conv_t.ini")
if(EXISTS "${legacy_ini}")
file(STRINGS "${legacy_ini}" legacy_lines)
foreach(line ${legacy_lines})
if(line MATCHES "^([A-Za-z0-9_]+)=(.+)$")
set(type "${CMAKE_MATCH_1}")
set(value "${CMAKE_MATCH_2}")
set(recorded FALSE)
foreach(entry ${allow_hf32_ops})
if(entry MATCHES "^${type}=")
set(recorded TRUE)
break()
endif()
endforeach()
if(NOT recorded)
list(APPEND allow_hf32_ops "${type}=${value}")
endif()
endif()
endforeach()
else()
message(WARNING "gen_allow_hf32_ini: legacy ini ${legacy_ini} not found, ops_allow_hf32_nn.ini only contains ALLOW_HF32 configured ops")
endif()
if(NOT allow_hf32_ops)
return()
endif()
list(REMOVE_DUPLICATES allow_hf32_ops)
list(SORT allow_hf32_ops)
string(JOIN "\n" allow_hf32_content ${allow_hf32_ops})
file(WRITE "${ASCEND_KERNEL_CONF_DST}/ops_allow_hf32_nn.ini" "${allow_hf32_content}\n")
install(FILES "${ASCEND_KERNEL_CONF_DST}/ops_allow_hf32_nn.ini"
DESTINATION ${OPP_PREFIX}/built-in/op_impl/ai_core/tbe/impl_mode
)
endfunction()
# ######################################################################################################################
# get op_type from *_def.cpp
# ######################################################################################################################
function(get_op_type_from_op_name OP_NAME OP_TYPE)
execute_process(
COMMAND
find ${OP_DIR} -name ${OP_NAME}_def.cpp -exec grep OP_ADD {} \;
OUTPUT_VARIABLE op_type
)
if(NOT op_type)
set(op_type "")
else()
# def文件头部说明的注释中也可能出现OP_ADD字样,先剥离行内注释,再提取注册行,
# 并以引号引用变量——否则多行结果会被当作list展开拼接出脏op_type(如"// -XxxYyy")
string(REGEX REPLACE "//[^\r\n]*" "" op_type "${op_type}")
string(REGEX REPLACE "[\t ]*OP_ADD\\([\t ]*" "" op_type "${op_type}")
string(REGEX REPLACE "[\t ]*\\)[^\r\n]*" "" op_type "${op_type}")
string(REGEX REPLACE "[\r\n]+" ";" op_type "${op_type}")
set(_op_type_first "")
foreach(_op_line ${op_type})
if(NOT "${_op_line}" STREQUAL "")
set(_op_type_first "${_op_line}")
break()
endif()
endforeach()
set(op_type "${_op_type_first}")
endif()
set(${OP_TYPE}
"${op_type}"
PARENT_SCOPE
)
endfunction()
# 从def文件列表解析注册的OpType(OP_ADD参数,大写驼峰),解析失败输出空串
function(get_op_type_from_def_files DEF_FILES OP_TYPE)
set(registered_type "")
foreach(def_file ${DEF_FILES})
if(EXISTS "${def_file}")
file(READ "${def_file}" def_content)
if(def_content MATCHES "OP_ADD\\([ \t\r\n]*([A-Za-z0-9_]+)")
set(registered_type "${CMAKE_MATCH_1}")
break()
endif()
endif()
endforeach()
set(${OP_TYPE} "${registered_type}" PARENT_SCOPE)
endfunction()
# usage: add_kernel_sources(OPTYPE)
# 1.调用asc_opc 工具 编译二进制kernel时, --simplified_key_mode/--impl_mode 选项中填写的值。
# 2.调用asc_opc工具 算子名.py 中auto_sync与compile_option中的配置项。
# 格式如下所示:
# [KERNEL_SRC op_name.cpp] 算子入口文件,缺省为算子名.cpp
# [COMPUTE_UNITS ascendxx] soc版本,支持配置多个unit,缺省为配置所有的soc(兼容参数,
# 内部逐项转换为npuArch后统一按arch匹配;与NPU_ARCH互斥)
# [NPU_ARCH dav_2201] npuArch版本,支持配置多个,缺省为全部支持
# [SIMPLIFIED_KEY 0/None] 缺省为0,最终的编译参数则是--simplified_key_mode=0,若设置为None,则不会携带该参数
# [AUTO_SYNC false] 同步选项,缺省为true
# [IMPL_MODE high_performance] 高性能模式,缺省为high_performance[,optional] [optional]是可选的
# [OPTIONS "option1" "option2"] 其他编译选项
function(add_kernel_sources)
if(PREPROCESS_ONLY)
return()
endif()
set(oneValueArgs KERNEL_SRC SIMPLIFIED_KEY AUTO_SYNC IMPL_MODE)
set(multiValueArgs COMPUTE_UNITS NPU_ARCH OPTIONS)
cmake_parse_arguments(MODULE "" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
# 兼容两种调用位置:在 op_kernel 目录内调用(旧)或在算子顶层调用(新)
if(CMAKE_CURRENT_SOURCE_DIR MATCHES "/op_kernel$")
set(SOURCE_DIR ${CMAKE_CURRENT_SOURCE_DIR})
else()
set(SOURCE_DIR ${CMAKE_CURRENT_SOURCE_DIR}/op_kernel)
endif()
get_filename_component(OP_DIR ${SOURCE_DIR} DIRECTORY)
get_filename_component(op_name ${OP_DIR} NAME)
# 支持芯片过滤:优先 NPU_ARCH(npuArch 列表),兼容旧 COMPUTE_UNITS(soc 列表,逐项转换为
# npuArch);当前构建芯片同样转换为 npuArch 后统一按 arch 匹配,未配置时视为全部支持。
# 注意 arch 化后同 arch 的 soc 共享过滤结果(如 ascend950 与 ascend350 同为 dav_3510)
set(_arch_list "")
if(MODULE_NPU_ARCH)
set(_arch_list "${MODULE_NPU_ARCH}")
get_npu_arch_by_soc_version("${ASCEND_COMPUTE_UNIT}" _cur_arch)
if(NOT "${_cur_arch}" IN_LIST _arch_list)
return()
endif()
elseif(MODULE_COMPUTE_UNITS)
if(NOT "${ASCEND_COMPUTE_UNIT}" IN_LIST MODULE_COMPUTE_UNITS)
return()
endif()
else()
set(MODULE_COMPUTE_UNITS "ALL")
endif()
set(cache_key "kernel_src_${op_name}_${ASCEND_COMPUTE_UNIT}")
set(op_type "")
# 检查是否已缓存
if(DEFINED ${cache_key})
set(op_type "${${cache_key}}")
endif()
if(NOT op_type)
set(binary_json ${OP_DIR}/op_host/config/${ASCEND_COMPUTE_UNIT}/${op_name}_binary.json)
if(EXISTS ${binary_json})
get_op_type_from_binary_json("${binary_json}" op_type)
if(NOT op_type)
return()
endif()
else()
get_op_type_from_op_name("${op_name}" op_type)
if(NOT op_type)
return()
endif()
endif()
endif()
set(${cache_key} "${op_type}" CACHE INTERNAL "Cached op_type for ${op_name} on ${compute_unit}")
if(NOT MODULE_KERNEL_SRC)
set(MODULE_KERNEL_SRC "${OP_NAME}")
else()
string(REGEX REPLACE "\\.cpp$" "" MODULE_KERNEL_SRC "${MODULE_KERNEL_SRC}")
endif()
set(simplified_key ${MODULE_SIMPLIFIED_KEY})
set(auto_sync ${MODULE_AUTO_SYNC})
set(impl_mode ${MODULE_IMPL_MODE})
set(options ${MODULE_OPTIONS})
if(NOT simplified_key)
set(simplified_key "0")
else()
string(TOLOWER "${simplified_key}" simplified_key_lower)
if(simplified_key_lower STREQUAL "none")
set(simplified_key "None")
endif()
endif()
if(NOT impl_mode)
set(impl_mode "high_performance,optional")
endif()
set(auto_sync_option "")
string(TOLOWER "${auto_sync}" auto_sync)
if(auto_sync STREQUAL "false")
set(auto_sync_option ${auto_sync})
endif()
string(JOIN "," unit_options ${options})
if(MODULE_COMPUTE_UNITS STREQUAL "ALL")
set(target_compute_units "ALL")
else()
set(target_compute_units "${ASCEND_COMPUTE_UNIT}")
endif()
list(APPEND kernel_src_list "${op_type} ${target_compute_units} ${MODULE_KERNEL_SRC}")
list(APPEND simplified_key_list "${op_type} ${target_compute_units} ${simplified_key}")
list(APPEND impl_mode_list "${op_type} ${target_compute_units} ${impl_mode}")
set(kernel_src_list ${kernel_src_list} CACHE INTERNAL "kernel_src")
set(simplified_key_list ${simplified_key_list} CACHE INTERNAL "simplified_key")
set(impl_mode_list ${impl_mode_list} CACHE INTERNAL "impl_mode")
if(auto_sync_option STREQUAL "false")
list(APPEND auto_sync_list "${op_type} ${target_compute_units} ${auto_sync_option}")
set(auto_sync_list ${auto_sync_list} CACHE INTERNAL "auto_sync")
endif()
if(unit_options)
list(APPEND option_list "${op_type} ${target_compute_units} ${unit_options}")
set(option_list ${option_list} CACHE INTERNAL "options")
endif()
# 标记此算子已处理完成
set(${cache_key} "${op_type}" CACHE INTERNAL "Cached op_type for ${op_name} on ${ASCEND_COMPUTE_UNIT}")
endfunction()
# usage: add_graph_plugin_sources()
function(add_graph_plugin_sources)
if(NOT "${OP_DIR}x" STREQUAL "x")
set(SOURCE_DIR ${OP_DIR}/op_graph)
else()
set(SOURCE_DIR ${CMAKE_CURRENT_SOURCE_DIR})
endif()
# 获取算子层级目录名称,判断是否编译该算子
get_filename_component(PARENT_DIR ${SOURCE_DIR} DIRECTORY)
get_filename_component(OP_NAME ${PARENT_DIR} NAME)
file(GLOB_RECURSE GRAPH_PLUGIN_SRCS ${SOURCE_DIR}/*.cpp)
if(GRAPH_PLUGIN_SRCS)
add_graph_plugin_modules()
target_sources(${GRAPH_PLUGIN_NAME}_obj PRIVATE ${GRAPH_PLUGIN_SRCS})
endif()
file(GLOB GRAPH_PLUGIN_PROTO_HEADERS ${SOURCE_DIR}/*_proto*.h)
if(GRAPH_PLUGIN_PROTO_HEADERS)
target_sources(${GRAPH_PLUGIN_NAME}_proto_headers INTERFACE ${GRAPH_PLUGIN_PROTO_HEADERS})
endif()
endfunction()
# ######################################################################################################################
# get operating system info
# ######################################################################################################################
function(get_system_info SYSTEM_INFO)
if(UNIX)
execute_process(
COMMAND
grep -i ^id= /etc/os-release
OUTPUT_VARIABLE TEMP
)
string(REGEX REPLACE "\n|id=|ID=|\"" "" SYSTEM_NAME ${TEMP})
set(${SYSTEM_INFO}
${SYSTEM_NAME}_${CMAKE_SYSTEM_PROCESSOR}
PARENT_SCOPE
)
elseif(WIN32)
message(STATUS "System is Windows. Only for pre-build.")
else()
message(FATAL_ERROR "${CMAKE_SYSTEM_NAME} not support.")
endif()
endfunction()
# ######################################################################################################################
# add compile options, e.g.: -g -O0
# ######################################################################################################################
function(add_ops_compile_options OP_TYPE)
cmake_parse_arguments(OP_COMPILE "" "OP_TYPE" "COMPUTE_UNIT;OPTIONS" ${ARGN})
execute_process(
COMMAND
${ASCEND_PYTHON_EXECUTABLE} ${CMAKE_SOURCE_DIR}/scripts/util/ascendc_gen_options.py
${ASCEND_AUTOGEN_PATH}/${CUSTOM_COMPILE_OPTIONS} ${OP_TYPE} ${OP_COMPILE_COMPUTE_UNIT} ${OP_COMPILE_OPTIONS}
RESULT_VARIABLE EXEC_RESULT
OUTPUT_VARIABLE EXEC_INFO
ERROR_VARIABLE EXEC_ERROR
)
if(${EXEC_RESULT})
message("add ops compile options info: ${EXEC_INFO}")
message("add ops compile options error: ${EXEC_ERROR}")
message(FATAL_ERROR "add ops compile options failed!")
endif()
endfunction()
###################################################################################################
# get op_type from *_binary.json
###################################################################################################
function(get_op_type_from_binary_json BINARY_JSON OP_TYPE)
execute_process(COMMAND grep -w op_type ${BINARY_JSON} OUTPUT_VARIABLE op_type)
string(REGEX REPLACE "\"op_type\"" "" op_type ${op_type})
string(REGEX MATCH "\".+\"" op_type ${op_type})
string(REGEX REPLACE "\"" "" op_type ${op_type})
set(${OP_TYPE} ${op_type} PARENT_SCOPE)
endfunction()
###################################################################################################
# convert short socVersion to long socVersion
###################################################################################################
function(map_compute_unit compute_unit compute_unit_long)
set(compute_unit_keys "ascend910b" "ascend310p" "ascend910_93" "ascend950" "ascend350" "mc62" "ascend960dt" "ascend960pr")
set(compute_unit_values "ascend910b1" "ascend310p1" "ascend910_9391" "ascend950pr_9599" "ascend350_355e" "mc62cm12aa" "ascend960dt" "ascend960pr")
list(FIND compute_unit_keys ${compute_unit} index)
if(NOT index EQUAL -1)
list(GET compute_unit_values ${index} mapped_value)
set(${compute_unit_long} ${mapped_value} PARENT_SCOPE)
else()
set(${compute_unit_long} ${compute_unit} PARENT_SCOPE)
endif()
endfunction()
###################################################################################################
# get target dir of different socVersions
###################################################################################################
function(get_target_dir compute_unit_long target_dir)
set(compute_unit_long_values "ascend910b1" "ascend310p1" "ascend910_9391" "ascend950pr_9599" "ascend350_355e" "mc62cm12aa" "ascend960dt" "ascend960pr")
set(target_dir_values "arch22" "arch22" "arch22" "arch35" "arch35" "arch35" "arch35" "arch35")
list(FIND compute_unit_long_values ${compute_unit_long} index)
if(NOT index EQUAL -1)
list(GET target_dir_values ${index} mapped_value)
set(${target_dir} ${mapped_value} PARENT_SCOPE)
else()
set(${target_dir} "" PARENT_SCOPE)
endif()
endfunction()
###################################################################################################
# get npuArch by socVersion, mapping is consistent with SOC_TO_ARCH in build.sh
# 映射列表 ASCEND_ALL_COMPUTE_UNIT/SOC_TO_ARCH_VALUES 定义在 variables.cmake,两者按位置一一对应
# usage: get_npu_arch_by_soc_version(<soc_version> <out_var>)
# e.g.: get_npu_arch_by_soc_version("ascend910b" npu_arch) -> npu_arch is dav_2201
# 未匹配的 socVersion(如 ascend910_55)返回空串
###################################################################################################
function(get_npu_arch_by_soc_version soc_version npu_arch)
string(TOLOWER "${soc_version}" soc_version_lower)
list(FIND ASCEND_ALL_COMPUTE_UNIT ${soc_version_lower} index)
if(NOT index EQUAL -1)
list(GET SOC_TO_ARCH_VALUES ${index} mapped_value)
set(${npu_arch} ${mapped_value} PARENT_SCOPE)
else()
set(${npu_arch} "" PARENT_SCOPE)
endif()
endfunction()
function(get_soc_version_by_npu_arch npu_arch soc_version)
set(index 0)
set(mapped_list)
foreach(arch ${SOC_TO_ARCH_VALUES})
if("${npu_arch}" STREQUAL "${arch}")
list(GET ASCEND_ALL_COMPUTE_UNIT ${index} mapped_value)
list(APPEND mapped_list ${mapped_value})
endif()
math(EXPR index "${index} + 1")
endforeach()
if(mapped_list)
set(${soc_version} ${mapped_list} PARENT_SCOPE)
else()
set(${soc_version} "" PARENT_SCOPE)
endif()
endfunction()
function(protobuf_generate_external comp c_var h_var)
if (NOT ARGN)
message(SEND_ERROR "Error: protobuf_generate_external() called without any proto files")
return()
endif()
set(${c_var})
set(${h_var})
set(_add_target FALSE)
set(extra_option "")
foreach(arg ${ARGN})
if ("${arg}" MATCHES "--proto_path")
set(extra_option ${arg})
endif()
endforeach()
foreach(file ${ARGN})
if ("${file}" STREQUAL "TARGET")
set(_add_target TRUE)
continue()
endif()
if ("${file}" MATCHES "--proto_path")
continue()
endif()
get_filename_component(abs_file ${file} ABSOLUTE)
get_filename_component(file_name ${file} NAME_WE)
get_filename_component(file_dir ${abs_file} PATH)
get_filename_component(parent_subdir ${file_dir} NAME)
if ("${parent_subdir}" STREQUAL "proto")
set(proto_output_path ${CMAKE_BINARY_DIR}/proto/${comp}/proto)
else()
set(proto_output_path ${CMAKE_BINARY_DIR}/proto/${comp}/proto/${parent_subdir})
endif()
list(APPEND ${c_var} "${proto_output_path}/${file_name}.pb.cc")
list(APPEND ${h_var} "${proto_output_path}/${file_name}.pb.h")
add_custom_command(
OUTPUT "${proto_output_path}/${file_name}.pb.cc" "${proto_output_path}/${file_name}.pb.h"
WORKING_DIRECTORY ${PROJECT_SOURCE_DIR}
COMMAND ${CMAKE_COMMAND} -E make_directory "${proto_output_path}"
COMMAND ${CMAKE_COMMAND} -E echo "generate proto cpp_out ${comp} by ${abs_file}"
COMMAND ${Protobuf_PROTOC_EXECUTABLE} -I${file_dir} ${extra_option} --cpp_out=${proto_output_path} ${abs_file}
DEPENDS ${abs_file} host_protoc json
COMMENT "Running C++ protocol buffer compiler on ${file}" VERBATIM)
endforeach()
if (_add_target)
add_custom_target(
${comp} DEPENDS ${${c_var}} ${${h_var}})
endif()
set_source_files_properties(${${c_var}} ${${h_var}} PROPERTIES GENERATED TRUE)
set(${c_var} ${${c_var}} PARENT_SCOPE)
set(${h_var} ${${h_var}} PARENT_SCOPE)
endfunction()
function(add_onnx_plugin_modules)
if (NOT TARGET ${ONNX_PLUGIN_NAME}_obj)
set(ge_onnx_proto_srcs
${ASCEND_DIR}/include/proto/ge_onnx.proto)
protobuf_generate_external(onnx ge_onnx_proto_cc ge_onnx_proto_h ${ge_onnx_proto_srcs})
if(BUILD_WITH_INSTALLED_DEPENDENCY_CANN_PKG)
npu_op_library(${ONNX_PLUGIN_NAME}_obj GRAPH ${ge_onnx_proto_h} )
else()
add_library(${ONNX_PLUGIN_NAME}_obj OBJECT ${ge_onnx_proto_h})
endif()
# 为特定目标设置C++14标准
set_target_properties(${ONNX_PLUGIN_NAME}_obj PROPERTIES
CXX_STANDARD 14
CXX_STANDARD_REQUIRED ON
CXX_EXTENSIONS OFF
)
target_include_directories(${ONNX_PLUGIN_NAME}_obj PRIVATE ${ONNX_PLUGIN_INCLUDE})
target_compile_definitions(${ONNX_PLUGIN_NAME}_obj PRIVATE OPS_UTILS_LOG_SUB_MOD_NAME="ONNX_PLUGIN" LOG_CPP)
if(BUILD_WITH_INSTALLED_DEPENDENCY_CANN_PKG)
target_compile_options(
${ONNX_PLUGIN_NAME}_obj PRIVATE -Dgoogle=ascend_private -fvisibility=hidden -Wno-shadow -Wno-unused-parameter
)
else()
target_compile_options(
${ONNX_PLUGIN_NAME}_obj PRIVATE $<$<NOT:$<BOOL:${ENABLE_TEST}>>:-DDISABLE_COMPILE_V1> -Dgoogle=ascend_private
-fvisibility=hidden -Wno-shadow -Wno-unused-parameter
)
endif()
target_link_libraries(
${ONNX_PLUGIN_NAME}_obj
PRIVATE $<BUILD_INTERFACE:$<IF:$<BOOL:${ENABLE_TEST}>,intf_llt_pub_asan_cxx17,intf_pub_cxx14>>
$<BUILD_INTERFACE:dlog_headers>
$<$<TARGET_EXISTS:opbase_util_objs>:$<TARGET_OBJECTS:opbase_util_objs>>
$<$<TARGET_EXISTS:opbase_infer_objs>:$<TARGET_OBJECTS:opbase_infer_objs>>
ascend_protobuf_static
)
endif()
endfunction()
# 添加 cube_utils 插件模块
function(add_cube_utils_plugin_modules)
if (NOT TARGET ${CUBE_UTILS_PLUGIN_NAME}_obj)
if(BUILD_WITH_INSTALLED_DEPENDENCY_CANN_PKG)
npu_op_library(${CUBE_UTILS_PLUGIN_NAME}_obj GRAPH)
else()
add_library(${CUBE_UTILS_PLUGIN_NAME}_obj OBJECT)
endif()
# 设置 C++14 标准
set_target_properties(${CUBE_UTILS_PLUGIN_NAME}_obj PROPERTIES
CXX_STANDARD 14
CXX_STANDARD_REQUIRED ON
CXX_EXTENSIONS OFF
)
# 设置头文件包含路径
target_include_directories(${CUBE_UTILS_PLUGIN_NAME}_obj PRIVATE
${OP_PROTO_INCLUDE}
${ASCEND_DIR}/include
${PROJECT_SOURCE_DIR}/common/inc/op_graph
${PROJECT_SOURCE_DIR}/common/inc
)
# 设置编译定义
target_compile_definitions(${CUBE_UTILS_PLUGIN_NAME}_obj PRIVATE
OPS_UTILS_LOG_SUB_MOD_NAME="CUBE_UTILS_PLUGIN"
LOG_CPP
)
# 设置编译选项
if(BUILD_WITH_INSTALLED_DEPENDENCY_CANN_PKG)
target_compile_options(
${CUBE_UTILS_PLUGIN_NAME}_obj PRIVATE
-Dgoogle=ascend_private
-fvisibility=hidden
-Wno-shadow
-Wno-unused-parameter
-Wno-deprecated-declarations
)
else()
target_compile_options(
${CUBE_UTILS_PLUGIN_NAME}_obj PRIVATE
$<$<NOT:$<BOOL:${ENABLE_TEST}>>:-DDISABLE_COMPILE_V1>
-Dgoogle=ascend_private
-fvisibility=hidden
-Wno-shadow
-Wno-unused-parameter
)
endif()
# 设置链接库
target_link_libraries(
${CUBE_UTILS_PLUGIN_NAME}_obj
PRIVATE
$<BUILD_INTERFACE:$<IF:$<BOOL:${ENABLE_TEST}>,intf_llt_pub_asan_cxx17,intf_pub_cxx14>>
$<BUILD_INTERFACE:dlog_headers>
eager_style_graph_builder_base
acl_rt
)
endif()
endfunction()
function(add_onnx_plugin_sources)
if(NOT "${OP_DIR}x" STREQUAL "x")
set(SOURCE_DIR ${OP_DIR}/framework)
else()
set(SOURCE_DIR ${CMAKE_CURRENT_SOURCE_DIR})
endif()
file(GLOB ONNX_PLUGIN_SRCS ${SOURCE_DIR}/*_onnx_plugin.cpp)
if(ONNX_PLUGIN_SRCS)
add_onnx_plugin_modules()
target_sources(${ONNX_PLUGIN_NAME}_obj PRIVATE ${ONNX_PLUGIN_SRCS})
endif()
endfunction()
# TF plugin 初始化函数(顶层调用一次,内部条件检查)
function(init_tf_plugin_modules)
# 条件不满足,直接返回,不创建 OBJECT 库(零浪费)
if(NOT BUILD_WITH_INSTALLED_DEPENDENCY_CANN_PKG OR ENABLE_TEST)
return()
endif()
# 条件满足,生成 TF protobuf
set(tf_proto_srcs
${ASCEND_DIR}/include/proto/ge_ir.proto
)
protobuf_generate_external(tf tf_proto_cc tf_proto_h ${tf_proto_srcs})
if(BUILD_WITH_INSTALLED_DEPENDENCY_CANN_PKG)
npu_op_library(${TF_PLUGIN_NAME}_obj GRAPH ${tf_proto_h} )
else()
add_library(${TF_PLUGIN_NAME}_obj OBJECT ${tf_proto_h})
endif()
# 配置编译选项
set_target_properties(${TF_PLUGIN_NAME}_obj PROPERTIES
CXX_STANDARD 17
CXX_STANDARD_REQUIRED ON
CXX_EXTENSIONS OFF
)
target_include_directories(${TF_PLUGIN_NAME}_obj PRIVATE ${TF_PLUGIN_INCLUDE})
target_compile_definitions(${TF_PLUGIN_NAME}_obj PRIVATE OPS_UTILS_LOG_SUB_MOD_NAME="TF_PLUGIN" LOG_CPP)
if(BUILD_WITH_INSTALLED_DEPENDENCY_CANN_PKG)
target_compile_options(
${TF_PLUGIN_NAME}_obj PRIVATE -Dgoogle=ascend_private -fvisibility=hidden -Wno-shadow -Wno-unused-parameter
)
else()
target_compile_options(
${TF_PLUGIN_NAME}_obj PRIVATE $<$<NOT:$<BOOL:${ENABLE_TEST}>>:-DDISABLE_COMPILE_V1> -Dgoogle=ascend_private
-fvisibility=hidden -Wno-shadow -Wno-unused-parameter
)
endif()
target_link_libraries(
${TF_PLUGIN_NAME}_obj
PRIVATE $<BUILD_INTERFACE:$<IF:$<BOOL:${ENABLE_TEST}>,intf_llt_pub_asan_cxx17,intf_pub_cxx17>>
$<BUILD_INTERFACE:dlog_headers>
$<$<TARGET_EXISTS:opbase_util_objs>:$<TARGET_OBJECTS:opbase_util_objs>>
$<$<TARGET_EXISTS:opbase_infer_objs>:$<TARGET_OBJECTS:opbase_infer_objs>>
ascend_protobuf_static
)
endfunction()
###################################################################################################
# 算子遍历时自动编译算子framework目录下_tf_plugin.cpp文件及tf_plugin目录下所有代码,无需主动调用
###################################################################################################
function(add_tf_plugin_sources)
if(NOT BUILD_WITH_INSTALLED_DEPENDENCY_CANN_PKG OR ENABLE_TEST)
return()
endif()
if(NOT "${OP_DIR}x" STREQUAL "x")
set(FRAMEWORK_DIR ${OP_DIR}/framework)
else()
set(FRAMEWORK_DIR ${CMAKE_CURRENT_SOURCE_DIR})
endif()
file(GLOB TF_PLUGIN_SRCS ${FRAMEWORK_DIR}/*_tf_plugin.cpp ${FRAMEWORK_DIR}/tf_plugin/*.cpp)
if(TF_PLUGIN_SRCS)
target_sources(${TF_PLUGIN_NAME}_obj PRIVATE ${TF_PLUGIN_SRCS})
endif()
endfunction()
# 添加 cube_utils 插件源文件
macro(add_cube_utils_plugin_sources)
# 收集当前目录(cube_utils 源码目录)下所有 .cc 文件
file(GLOB CUBE_UTILS_PLUGIN_SRCS ${CMAKE_CURRENT_SOURCE_DIR}/*.cc)
if(CUBE_UTILS_PLUGIN_SRCS)
add_cube_utils_plugin_modules()
target_sources(${CUBE_UTILS_PLUGIN_NAME}_obj PRIVATE ${CUBE_UTILS_PLUGIN_SRCS})
else()
message(STATUS "CUBE_UTILS_PLUGIN_SRCS is empty")
endif()
endmacro()
# PyTorch extension
# usage: add_sources()
# usage: add_sources("file1.cpp;file2.cpp;file3.cpp")
# 通过环境变量PE_COMPILE_ARGS获取编译参数
macro(add_sources)
# 解析参数
set(CUSTOM_SOURCES "") # 第一个参数为自定义源文件列表(可选)
# 检查是否有第一个参数
if(${ARGC} GREATER 0)
set(CUSTOM_SOURCES "${ARGV0}")
endif()
message(STATUS "CMAKE_CURRENT_SOURCE_DIR = ${CMAKE_CURRENT_SOURCE_DIR}")
# get parent dir name as OP_NAME
get_filename_component(OP_NAME ${CMAKE_CURRENT_SOURCE_DIR} NAME)
message(STATUS "OP_NAME: ${OP_NAME}")
# 从环境变量PE_COMPILE_ARGS获取编译参数
if(DEFINED ENV{PE_COMPILE_ARGS})
set(COMPILE_ARGS $ENV{PE_COMPILE_ARGS})
message(STATUS "PE_COMPILE_ARGS from environment: ${COMPILE_ARGS}")
else()
set(COMPILE_ARGS "--npu-arch=dav-2201")
message(STATUS "PE_COMPILE_ARGS environment variable not set, use default: dav-2201")
endif()
# get compile flags for current op
set(COMPILE_FLAGS "${COMPILE_ARGS} -xasc ")
message(STATUS "COMPILE FLAGS: ${COMPILE_FLAGS}")
file(GLOB_RECURSE SOURCE_FILES RELATIVE ${CMAKE_CURRENT_SOURCE_DIR} ${CMAKE_CURRENT_SOURCE_DIR}/*.cpp)
# 根据是否传入自定义源文件列表决定如何获取源文件
if(CUSTOM_SOURCES)
foreach(SRC ${CUSTOM_SOURCES})
# 确保文件存在
if(EXISTS ${SRC})
# 获取相对于当前源目录的相对路径
file(RELATIVE_PATH REL_SRC ${CMAKE_CURRENT_SOURCE_DIR} ${SRC})
list(APPEND SOURCE_FILES ${REL_SRC})
else()
message(WARNING "Source file not found: ${SRC}")
endif()
endforeach()
endif()
message(STATUS "SOURCE FILES: ${SOURCE_FILES}")
if(SOURCE_FILES STREQUAL "")
message(FATAL_ERROR "No source files found")
endif()
# set_source_files_properties
set_source_files_properties(
${SOURCE_FILES} PROPERTIES
LANGUAGE CXX
COMPILE_FLAGS "${COMPILE_FLAGS}"
)
# set target name
set(TARGET_NAME ${OP_NAME}_obj)
add_library(${TARGET_NAME} OBJECT ${SOURCE_FILES})
target_compile_options(${TARGET_NAME} PRIVATE ${COMPILE_OPTIONS})
target_include_directories(${TARGET_NAME} PRIVATE ${CMAKE_CURRENT_SOURCE_DIR} ${INCLUDE_DIRECTORIES})
# add target obj to PE_OBJECTS_LIST
set(NEW_OBJECT_EXPRESSION $<TARGET_OBJECTS:${TARGET_NAME}>)
set(TEMP_LIST ${PE_OBJECTS_LIST})
list(APPEND TEMP_LIST ${NEW_OBJECT_EXPRESSION})
set(PE_OBJECTS_LIST ${TEMP_LIST} CACHE INTERNAL "List of PyTorch extension objects" FORCE)
endmacro()
# usage: add_all_modules_sources([OPTYPE] ACLNNTYPE DEPENDENCIES COMPUTE_UNIT TILING_DIR DISABLE_IN_OPP)
# 收编算子顶层 CMakeLists 的统一入口,自适应 op_api 目录位置(算子根或 op_host 下),
# 内部复用 add_modules_sources 既有逻辑,并触发 add_all_ut_sources 收集 UT 源。
# 格式如下所示:
# [OPTYPE op_name] 算子类型名称,缺省自动取算子目录名;显式传入时以用户设置为准
# ACLNNTYPE aclnn_type aclnn类型,必填,支持 aclnn/aclnn_inner/aclnn_exclude
# [DEPENDENCIES dep_op...] 依赖的算子名称列表,缺省为空
# [COMPUTE_UNIT ascendxx...] 支持的芯片版本号,必须与TILING_DIR一一对应,缺省为配置所有的soc
# [NPU_ARCH dav_2201 dav_3510] 设置支持的npuArch列表,必须与TILING_DIR一一对应,缺省为配置所有的npuArch
# [TILING_DIR archxx...] 每种芯片类型对应的tiling文件目录,必须与COMPUTE_UNIT一一对应,缺省为空
# [DISABLE_IN_OPP TRUE/FALSE] 是否在opp包中编译tiling文件,缺省为FALSE
# [ALLOW_HF32 xxx] 配置算子允许HF32执行,取值:enable_hi_float_32_execution / enable_float_32_execution
macro(add_all_modules_sources)
set(oneValueArgs DISABLE_IN_OPP HOSTCPU ALLOW_HF32)
set(multiValueArgs OPTYPE ACLNNTYPE DEPENDENCIES COMPUTE_UNIT NPU_ARCH TILING_DIR)
cmake_parse_arguments(MODULE "" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
# 自适应 DIR:op_api 在算子根目录(模式B)则 DIR=算子根;否则 DIR=算子根/op_host(模式A)
if(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/op_api")
set(_HOST_DIR ${CMAKE_CURRENT_SOURCE_DIR})
else()
set(_HOST_DIR ${CMAKE_CURRENT_SOURCE_DIR}/op_host)
endif()
get_filename_component(_OP_NAME ${CMAKE_CURRENT_SOURCE_DIR} NAME)
# OPTYPE 未传入时,自动取算子目录名
if(NOT MODULE_OPTYPE)
set(MODULE_OPTYPE ${_OP_NAME})
endif()
add_modules_sources(DIR ${_HOST_DIR} OPTYPE ${MODULE_OPTYPE} ACLNNTYPE ${MODULE_ACLNNTYPE}
DEPENDENCIES ${MODULE_DEPENDENCIES} COMPUTE_UNIT ${MODULE_COMPUTE_UNIT} NPU_ARCH ${MODULE_NPU_ARCH}
TILING_DIR ${MODULE_TILING_DIR} DISABLE_IN_OPP ${MODULE_DISABLE_IN_OPP}
HOSTCPU ${MODULE_HOSTCPU}
ALLOW_HF32 ${MODULE_ALLOW_HF32})
add_all_ut_sources(OP_NAME ${_OP_NAME})
unset(_HOST_DIR)
unset(_OP_NAME)
endmacro()
# usage: add_all_ut_sources(OP_NAME op_name)
# 收集算子各类型 UT 源;op_api UT 同时扫描 tests/ut/op_api 与 tests/ut/op_host 两处位置以兼容历史。
macro(add_all_ut_sources)
set(oneValueArgs OP_NAME)
cmake_parse_arguments(MODULE "" "${oneValueArgs}" "" ${ARGN})
set(_UT_ROOT ${CMAKE_CURRENT_SOURCE_DIR}/tests/ut)
if(UT_TEST_ALL OR OP_API_UT)
if(EXISTS "${_UT_ROOT}/op_api")
add_modules_ut_sources(HOSTNAME ${OP_API_MODULE_NAME} MODE PRIVATE DIR ${_UT_ROOT}/op_api)
endif()
if(EXISTS "${_UT_ROOT}/op_host")
add_modules_ut_sources(HOSTNAME ${OP_API_MODULE_NAME} MODE PRIVATE DIR ${_UT_ROOT}/op_host)
endif()
# 兼容历史:部分算子 aclnn 测试放在 tests/ut/op_host/op_api 下
if(EXISTS "${_UT_ROOT}/op_host/op_api")
add_modules_ut_sources(HOSTNAME ${OP_API_MODULE_NAME} MODE PRIVATE DIR ${_UT_ROOT}/op_host/op_api)
endif()
endif()
if(UT_TEST_ALL OR OP_HOST_UT)
if(EXISTS "${_UT_ROOT}/op_host")
add_modules_ut_sources(HOSTNAME ${OP_TILING_MODULE_NAME} MODE PRIVATE DIR ${_UT_ROOT}/op_host)
add_modules_ut_sources(HOSTNAME ${OP_INFERSHAPE_MODULE_NAME} MODE PRIVATE DIR ${_UT_ROOT}/op_host)
endif()
# onnx 插件(framework)解析逻辑属 host 侧,目录结构与 tiling/infershape 逻辑类似
if(DEFINED OP_FRAMEWORK_MODULE_NAME AND EXISTS "${_UT_ROOT}/framework")
add_modules_ut_sources(HOSTNAME ${OP_FRAMEWORK_MODULE_NAME} MODE PRIVATE DIR ${_UT_ROOT}/framework)
endif()
endif()
if(UT_TEST_ALL OR OP_GRAPH_UT)
if(EXISTS "${_UT_ROOT}/op_graph")
add_modules_ut_sources(HOSTNAME ${OP_GRAPH_MODULE_NAME} MODE PRIVATE DIR ${_UT_ROOT}/op_graph)
endif()
endif()
unset(_UT_ROOT)
endmacro()