C++与Python交互最优方案咨询：Pybind11/Boost.Python/原生封装

Hey there! Let’s dive into the most flexible, efficient, and seamless ways to connect C++ and Python, comparing the three options you mentioned: Pybind11, Boost.Python, and raw Python C API (native wrapping).

Raw Python C API (Native Wrapping) #

This is the lowest-level approach—directly using Python’s built-in C interface.

• Flexibility: Technically unlimited, since you’re working directly with Python’s core API. But in practice, it’s extremely inflexible to use: you have to manually handle reference counting, type checking, error propagation, and convert every C++ type to Python objects yourself. There’s no abstraction to simplify binding classes, templates, or complex data structures.
• Efficiency: Runtime performance is great (no extra layers), but development efficiency is terrible. Writing even simple functions requires tons of boilerplate, and debugging memory leaks or reference count bugs is a nightmare.
• Seamlessness: Very low. The code feels like two separate languages glued together, not a smooth integration. Your example would be even more verbose if written with the raw API—you’d need to add argument parsing, error checks, and manual PyObject creation without Boost’s helpers.

Boost.Python #

Your sample code uses this, so let’s break down its pros and cons:

• Flexibility: Much better than the raw API. It wraps the Python C API to support binding C++ classes, templates, inheritance, and exceptions. It handles a lot of the tedious type conversion and reference counting under the hood.
• Efficiency: Runtime performance is solid (similar to raw API), but the tradeoff is slow compile times and large binary sizes. Boost is a massive dependency, and including Boost.Python pulls in a lot of code you might not need.
• Seamlessness: Decent, but showing its age. The syntax is clunky compared to modern alternatives, and you still have to write more boilerplate than necessary (like the PyMODINIT_FUNC setup in your example). It also doesn’t play as nicely with modern C++ features (like C++11 lambdas or smart pointers) as newer libraries.

Pybind11 #

This is the current gold standard for C++/Python integration, and for good reason:

• Flexibility: Unmatched. It’s designed for modern C++ (C11 and above) and supports almost every C feature you’d want to expose: classes, templates, lambda functions, smart pointers, enums, and even move semantics. It also lets you easily extend Python classes from C++ and vice versa.
• Efficiency: Runtime performance is identical to the raw API (since it’s just a thin wrapper around Python’s C interface). Unlike Boost.Python, it’s a header-only library—no need to compile or link against a huge dependency, which means faster compile times and smaller binaries.
• Seamlessness: Extremely high. The syntax is clean and intuitive, making the integration feel natural. For example, here’s how your strtest function would look with Pybind11:

  #include <pybind11/pybind11.h>
  #include <boost/algorithm/string.hpp>
  
  namespace py = pybind11;
  
  std::string strtest() {
      std::string s = "Boost C++ Libraries";
      boost::to_upper(s);
      return s;
  }
  
  PYBIND11_MODULE(math_demo, m) {
      m.def("strtest", &strtest, "Converts a string to uppercase");
  }
  

  Notice how you don’t have to manually create PyObjects or handle low-level setup—Pybind11 takes care of all that. It also automatically converts between C++ and Python types (like std::string to Python str) without extra code.

Final Recommendation #

If you want the best balance of flexibility, efficiency, and seamless integration, Pybind11 is the clear choice for new projects. It’s lightweight, modern, and drastically reduces the amount of boilerplate you need to write.

Boost.Python is still a valid option if you’re maintaining an existing codebase that already uses Boost, but it’s not ideal for new work. The raw Python C API should only be used if you have an extreme need to avoid all external dependencies—and even then, you’ll pay a huge price in development time.

内容的提问来源于stack exchange，提问作者user2723494