c++ - 已编译的 boost_python 扩展无法在 python 2.7 中导入

Question

我正在尝试使用 boost 导入用 c++ 编码的 python 扩展。虽然我在使用 cmake 编译扩展时遇到了一些问题，但我设法将其链接到 boost_python27 库。然后我使用 pythons distutils 将扩展安装到 python 框架中。

但是，当我尝试导入模块时出现以下错误:

Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
ImportError: dlopen(./tools.so, 2): Library not loaded: @rpath/libboost_python.dylib
  Referenced from: /Users/DaniBook/CLionProjects/Uebung1/cmake-build-debug/tools.so
  Reason: image not found

我尝试了我在互联网上找到的所有方法，包括重新安装 boost 和使用 distutils 扩展重新编译等，以使其正常工作。任何人都可以就此提供一些帮助，并且还可以回答这个丢失的不祥图像是什么吗？

tools.h

#ifndef UEBUNG1_TOOLS_H
#define UEBUNG1_TOOLS_H

#include <string>
#include <vector>
#include <map>
#include <boost/python.hpp>
#include <boost/python/suite/indexing/vector_indexing_suite.hpp>

using namespace std;

vector<string> *product(string alphabet, int repeats);

vector<string> *product(vector<string> pools);

vector<string>* hammdist(string &pattern, int distance);

#endif //UEBUNG1_TOOLS_H

tools.cpp

#include <boost/python.hpp>
#include <boost/python/suite/indexing/vector_indexing_suite.hpp>
#include "tools.h"
#include <vector>
#include <string>
#include <map>

using namespace std;

vector<string> *product(string alphabet, int repeats) {
    //initializing vector
    auto *results = new vector<string>();
    for(auto character : alphabet) {
        string tmpstr;
        tmpstr = character;
        results->push_back(tmpstr);
    }

    //cartesian product generation
    for(int i = 1; i < repeats; i++) {
        vector<string> tmp = *results;
        results->clear();

        //iterating over temporary list adding elements from pool to each contained string
        for(auto &it : tmp) {
            for(auto &character : alphabet) {
                results->push_back(it + character);
            }
        }
    }
    return results;
}

vector<string> product(vector<string> pools) {
    //initializing vector
    auto *results = new vector<string>();
    for(auto character : pools[0]) {
        string tmpstr;
        tmpstr = character;
        results->push_back(tmpstr);
    }

    //removing the first pool container
    pools.erase(pools.begin());

    //cartesian product generation
    for(const auto &pool : pools) {
        vector<string> tmp = *results;
        results->clear();

        //iterating over temporary list adding elements from pool to each contained string
        for(auto &it : tmp) {
            for(auto character : pool) {
                results->push_back(it + character);
            }
        }
    }
    return results;
}


vector<string>* hammdist(string &pattern, int distance) {
    map<char, string> possibles = {
            {'A', "CGT"},
            {'C', "AGT"},
            {'G', "ACT"},
            {'T', "ACG"}
    };
    auto *results = new vector<string>();
    vector<string> *masks = product("01", pattern.size());
    for(auto &mask : *masks) {
        auto *permute = new vector<string>();
        auto *tmp = new vector<string>();
        if(count(mask.begin(), mask.end(), '1') == distance) {
            for(int i = 0; i < pattern.size(); i++) {
                if(mask[i] != '1') {
                    string tmpstr;
                    tmpstr = pattern[i];
                    tmp->push_back(tmpstr);
                }
                else {
                    tmp->push_back(possibles[pattern[i]]);
                }
            }
            permute = product(*tmp);
            results->insert(results->end(), permute->begin(), permute->end());
        }
        delete permute;
        delete tmp;
    }
    return results;
}

/* 
   initializing function pointers in order to tell boost that we have 
   overloaded functions to expose to python
*/
vector<string> *(*product1)(string, int) = &product;
vector<string> *(*product2)(vector<string> pools) = &product;

//include all functions that are used by the function to expose to the BOOST_PYTHON_MODULE call

using namespace boost::python;

BOOST_PYTHON_MODULE(tools) {
    //telling boost_python that we have overloaded functions which need to be called in the respective situations
    //return_value_policy<manage_new_objects> is required in order for the interface to handle the new invocation and the returned pointer

    def("product", product1, return_value_policy<manage_new_object>());
    def("product", product2, return_value_policy<manage_new_object>());

    def("hammdist", hammdist, return_value_policy<manage_new_object>());

    //vector_indexing_suite handles the wrapping of vector member functions
    //enables handling vector in a pythonic way when using in python

    class_<std::vector<string>>("string_vector")
        .def(vector_indexing_suite<std::vector<string>>());
}

CMakeLists.txt

make_minimum_required(VERSION 3.12)
project(tools)

set(CMAKE_CXX_STANDARD 17)
set(CMAKE_INCLUDE_CURRENT_DIR ON)

if(APPLE)
    set(CMAKE_SHARED_LIBRARY_SUFFIX ".so")
endif(APPLE)

find_package(PythonLibs 2.7 REQUIRED)
include_directories(${PYTHON_INCLUDE_DIRS})

set(PROJECT_SOURCE_DIR src/)
include_directories(${PROJECT_SOURCE_DIR})

set(BOOST_ROOT "/Users/DaniBook/miniconda3/pkgs/boost-1.66.0-py27_1")
set(BOOST_LIBRARYDIR "/Users/DaniBook/miniconda3/pkgs/boost-1.66.0-27_1/lib")

find_package(Boost COMPONENTS python REQUIRED)
include_directories(${Boost_INCLUDE_DIRS})

add_library(tools SHARED src/tools.cpp src/tools.h)
target_link_libraries(tools ${Boost_LIBRARIES} ${PYTHON_LIBRARIES})
set_target_properties(tools PROPERTIES PREFIX "")

setup.py

from distutils.core import setup

setup(
    name = 'tools',
    version = '0.1',
    py_modules = ['tools'])

Answer 1

好的，经过一些额外的研究，我得出了一个非常有趣的结论，当您查看错误消息时，这个结论非常明显。

问题不是缺少共享库本身，而是编译时(链接器阶段)使用的对 boost 库的引用的未定义相对路径。这可能是由于使用了驻留在 miniconda 中而不是/usr/local/lib 中的 boost 安装，后者可能位于 pythonpath 或其他东西中。但是，这可以使用 otool 和 install_name_tool 手动解决(可以通过安装 XCode 开发工具在 mac 上获取)。

因此您在命令行上执行以下操作:

otools -L ${PATH_TO_PYTHONEXTENSION}/someextension.so

像这样在终端上列出所有库及其系统路径

./build/lib.macosx-10.13-x86_64-2.7/tools.so:
    @rpath/libboost_python.dylib (compatibility version 0.0.0, current version 0.0.0)
    /usr/lib/libc++.1.dylib (compatibility version 1.0.0, current version 400.9.0)
    /usr/lib/libSystem.B.dylib (compatibility version 1.0.0, current version 1252.50.4)

可以很容易地看到我们在编译时链接的 boost 库缺少路径，我们使用 install_name_tool 解决了这个问题:

install_name_tool -change @rpath/libboost_python.dylib ${ACTUAL_PATH_TO_libbost_python.dylib} ${PATH_TO_PYTHONEXTENSION}/someextension.so

重新运行 setup.py install 现在可以解决问题并且可以成功导入扩展:

import tools
dir(tools)

给出以下输出:

['__doc__', '__file__', '__name__', '__package__', 'hammdist', 'product', 'string_vector']

外部资源链接: unsafe use of relative rpath libboost.dylib when making boost.python helloword demo?

我假设它对其他基于 Unix 的操作系统(如 Linux)同样适用。希望这可以帮助。