使用f2py与scikit-build-core构建时出现Undefined reference错误

问题概述 #

带有Fortran 90扩展的Python包，原基于distutils+f2py构建，因distutils弃用转用scikit-build+cmake+f2py。近期因setuptools弃用切换至无setuptools依赖的scikit-build-core后，出现链接错误：

undefined reference to '_npy_f2py_ARRAY_API'

该符号属于f2py的fortranobject库，使用numpy 1.23虽能编译，但导入模块时仍报相同错误；移除target_link_libraries命令可完成编译，但运行时会因缺失依赖报错。

解决方案 #

1. 显式获取f2py的库与包含路径 #

在CMakeLists.txt中，直接调用numpy.f2py的接口获取必要路径，避免自动检测的偏差：

find_package(Python REQUIRED COMPONENTS Interpreter Development)
find_package(Numpy REQUIRED)

# 获取f2py的包含目录和库目录
execute_process(
    COMMAND ${Python_EXECUTABLE} -c "import numpy.f2py; print(numpy.f2py.get_include())"
    OUTPUT_VARIABLE F2PY_INCLUDE_DIR
    OUTPUT_STRIP_TRAILING_WHITESPACE
)
execute_process(
    COMMAND ${Python_EXECUTABLE} -c "import numpy.f2py; print(numpy.f2py.get_lib_dir())"
    OUTPUT_VARIABLE F2PY_LIB_DIR
    OUTPUT_STRIP_TRAILING_WHITESPACE
)

# 添加包含目录
include_directories(${F2PY_INCLUDE_DIR} ${NumPy_INCLUDE_DIRS})
# 添加库搜索路径
link_directories(${F2PY_LIB_DIR})

2. 正确链接fortranobject库 #

在定义扩展目标后，显式链接fortranobject库和numpy：

add_library(your_module MODULE your_module_wrapper.c your_fortran_code.f90)
target_link_libraries(your_module PRIVATE fortranobject ${NumPy_LIBRARIES})
set_target_properties(your_module PROPERTIES PREFIX "")

3. 适配numpy版本差异 #

针对numpy 1.24+的API变化，添加版本判断调整链接策略：

if (NumPy_VERSION VERSION_GREATER_EQUAL "1.24")
    target_link_libraries(your_module PRIVATE fortranobject ${NumPy_LIBRARIES})
else()
    target_link_libraries(your_module PRIVATE ${NumPy_LIBRARIES})
endif()

4. 修正pyproject.toml配置 #

确保scikit-build-core的构建配置正确，指定numpy依赖范围：

[build-system]
requires = ["scikit-build-core>=0.7", "numpy>=1.23"]
build-backend = "scikit_build_core.build"

[project]
name = "your-package"
requires-python = ">=3.8"
dependencies = ["numpy>=1.23"]

验证步骤 #

• 删除build/、_skbuild/等缓存目录，清理旧构建产物
• 执行pip install .重新构建安装
• 运行python -c "import your_module"验证模块可正常导入

内容的提问来源于stack exchange，提问作者Adam Lanman