Cygwin环境下pip安装SciPy出现编译错误求解决方案

问题背景 #

在Cygwin环境中通过pip安装SciPy时触发编译错误，核心问题是代码中引用的M_SQRT1_2、M_PI等数学宏未被编译器识别，且无法修改SciPy源码，需通过编译参数调整解决。

错误日志 #

2023-02-11T17:48:15,190   [250/1620] Compiling C++ object scipy/special/_ufuncs_cxx.cpython-39-x86_64-cygwin.dll.p/_faddeeva.cxx.o
2023-02-11T17:48:15,191   FAILED: scipy/special/_ufuncs_cxx.cpython-39-x86_64-cygwin.dll.p/_faddeeva.cxx.o
2023-02-11T17:48:15,192   c++ -Iscipy/special/_ufuncs_cxx.cpython-39-x86_64-cygwin.dll.p -Iscipy/special -I../../scipy/special -I/tmp/pip-build-env-m1hcjhav/overlay/lib/python3.9/site-packages/numpy/core/include -I../../scipy/_lib/boost -Iscipy/_lib -I../../scipy/_lib -I../../scipy/_build_utils/src -I/usr/include/python3.9 -fvisibility=hidden -fvisibility-inlines-hidden -fdiagnostics-color=always -D_FILE_OFFSET_BITS=64 -Wall -Winvalid-pch -std=c++14 -O2 -Wno-cpp -MD -MQ scipy/special/_ufuncs_cxx.cpython-39-x86_64-cygwin.dll.p/_faddeeva.cxx.o -MF scipy/special/_ufuncs_cxx.cpython-39-x86_64-cygwin.dll.p/_faddeeva.cxx.o.d -o scipy/special/_ufuncs_cxx.cpython-39-x86_64-cygwin.dll.p/_faddeeva.cxx.o -c ../../scipy/special/_faddeeva.cxx
2023-02-11T17:48:15,193   ../../scipy/special/_faddeeva.cxx: In function ‘npy_cdouble faddeeva_ndtr(npy_cdouble)’:
2023-02-11T17:48:15,194   ../../scipy/special/_faddeeva.cxx:79:10: error: ‘M_SQRT1_2’ was not declared in this scope; did you mean ‘NPY_SQRT1_2’?
2023-02-11T17:48:15,196      79 |     z *= M_SQRT1_2;
2023-02-11T17:48:15,199         |          ^~~~~~~~~
2023-02-11T17:48:15,200         |          NPY_SQRT1_2
2023-02-11T17:48:15,201   ../../scipy/special/_faddeeva.cxx: In function ‘double faddeeva_log_ndtr(double)’:
2023-02-11T17:48:15,202   ../../scipy/special/_faddeeva.cxx:108:18: error: ‘M_SQRT1_2’ was not declared in this scope; did you mean ‘NPY_SQRT1_2’?
2023-02-11T17:48:15,203     108 |     double t = x*M_SQRT1_2;
2023-02-11T17:48:15,204         |                  ^~~~~~~~~
2023-02-11T17:48:15,205         |                  NPY_SQRT1_2
2023-02-11T17:48:15,206   ../../scipy/special/_faddeeva.cxx: In function ‘npy_cdouble faddeeva_log_ndtr_complex(npy_cdouble)’:
2023-02-11T17:48:15,207   ../../scipy/special/_faddeeva.cxx:130:53: error: ‘M_SQRT1_2’ was not declared in this scope; did you mean ‘NPY_SQRT1_2’?
2023-02-11T17:48:15,208     130 |         complex<double> w = -0.5 * Faddeeva::erfc(z*M_SQRT1_2);
2023-02-11T17:48:15,209         |                                                     ^~~~~~~~~
2023-02-11T17:48:15,210         |                                                     NPY_SQRT1_2
2023-02-11T17:48:15,211   ../../scipy/special/_faddeeva.cxx:136:11: error: ‘M_SQRT1_2’ was not declared in this scope; did you mean ‘NPY_SQRT1_2’?
2023-02-11T17:48:15,212     136 |     z *= -M_SQRT1_2;
2023-02-11T17:48:15,218         |           ^~~~~~~~~
2023-02-11T17:48:15,219         |           NPY_SQRT1_2
2023-02-11T17:48:15,220   ../../scipy/special/_faddeeva.cxx: In function ‘double faddeeva_log_ndtr_complex(npy_cdouble)’:
2023-02-11T17:48:15,221     146 |     double im = fmod(mIm_z2, 2.0*M_PI);
2023-02-11T17:48:15,223         |                                  ^~~~
2023-02-11T17:48:15,224   ../../scipy/special/_faddeeva.cxx: In function ‘double faddeeva_voigt_profile(double, double, double)’:
2023-02-11T17:48:15,225   ../../scipy/special/_faddeeva.cxx:177:24: error: ‘M_PI’ was not declared in this scope
2023-02-11T17:48:15,226     177 |         return gamma / M_PI / (x*x + gamma*gamma);
2023-02-11T17:48:15,227         |                        ^~~~

解决方案 #

方案1：通过环境变量定义缺失的宏

直接在pip安装时传递编译参数，让编译器自动识别这些数学宏：

CXXFLAGS="-DM_SQRT1_2=1/sqrt(2) -DM_PI=3.141592653589793" pip install scipy

如果需要持久化这个环境变量，可以先执行：

export CXXFLAGS="-DM_SQRT1_2=1/sqrt(2) -DM_PI=3.141592653589793"

再执行pip install scipy。

方案2：使用Cygwin包管理器安装预编译版本

跳过源码编译步骤，直接用Cygwin自带的包管理器安装现成的SciPy：

1. 如果没有apt-cyg，先安装它：

   wget https://raw.githubusercontent.com/transcode-open/apt-cyg/master/apt-cyg
   install apt-cyg /usr/local/bin
   chmod +x /usr/local/bin/apt-cyg
   

2. 安装SciPy：

   apt-cyg install python3-scipy
   

内容的提问来源于stack exchange，提问作者natan