Packaging binary extensions#

Page Status:

Incomplete

Last Reviewed:

2013-12-08

One of the features of the CPython reference interpreter is that, in addition to allowing the execution of Python code, it also exposes a rich C API for use by other software. One of the most common uses of this C API is to create importable C extensions that allow things which aren’t always easy to achieve in pure Python code.

An overview of binary extensions#

Use cases#

The typical use cases for binary extensions break down into just three conventional categories:

  • accelerator modules: these modules are completely self-contained, and are created solely to run faster than the equivalent pure Python code runs in CPython. Ideally, accelerator modules will always have a pure Python equivalent to use as a fallback if the accelerated version isn’t available on a given system. The CPython standard library makes extensive use of accelerator modules. Example: When importing datetime, Python falls back to the datetime.py module if the C implementation ( _datetimemodule.c) is not available.

  • wrapper modules: these modules are created to expose existing C interfaces to Python code. They may either expose the underlying C interface directly, or else expose a more “Pythonic” API that makes use of Python language features to make the API easier to use. The CPython standard library makes extensive use of wrapper modules. Example: functools.py is a Python module wrapper for _functoolsmodule.c.

  • low-level system access: these modules are created to access lower level features of the CPython runtime, the operating system, or the underlying hardware. Through platform specific code, extension modules may achieve things that aren’t possible in pure Python code. A number of CPython standard library modules are written in C in order to access interpreter internals that aren’t exposed at the language level. Example: sys, which comes from sysmodule.c.

    One particularly notable feature of C extensions is that, when they don’t need to call back into the interpreter runtime, they can release CPython’s global interpreter lock around long-running operations (regardless of whether those operations are CPU or IO bound).

Not all extension modules will fit neatly into the above categories. The extension modules included with NumPy, for example, span all three use cases - they move inner loops to C for speed reasons, wrap external libraries written in C, FORTRAN and other languages, and use low level system interfaces for both CPython and the underlying operation system to support concurrent execution of vectorised operations and to tightly control the exact memory layout of created objects.

Disadvantages#

The main disadvantage of using binary extensions is the fact that it makes subsequent distribution of the software more difficult. One of the advantages of using Python is that it is largely cross platform, and the languages used to write extension modules (typically C or C++, but really any language that can bind to the CPython C API) typically require that custom binaries be created for different platforms.

This means that binary extensions:

  • require that end users be able to either build them from source, or else that someone publish pre-built binaries for common platforms

  • may not be compatible with different builds of the CPython reference interpreter

  • often will not work correctly with alternative interpreters such as PyPy, IronPython or Jython

  • if handcoded, make maintenance more difficult by requiring that maintainers be familiar not only with Python, but also with the language used to create the binary extension, as well as with the details of the CPython C API.

  • if a pure Python fallback implementation is provided, make maintenance more difficult by requiring that changes be implemented in two places, and introducing additional complexity in the test suite to ensure both versions are always executed.

Another disadvantage of relying on binary extensions is that alternative import mechanisms (such as the ability to import modules directly from zipfiles) often won’t work for extension modules (as the dynamic loading mechanisms on most platforms can only load libraries from disk).

Alternatives to handcoded accelerator modules#

When extension modules are just being used to make code run faster (after profiling has identified the code where the speed increase is worth additional maintenance effort), a number of other alternatives should also be considered:

  • look for existing optimised alternatives. The CPython standard library includes a number of optimised data structures and algorithms (especially in the builtins and the collections and itertools modules). The Python Package Index also offers additional alternatives. Sometimes, the appropriate choice of standard library or third party module can avoid the need to create your own accelerator module.

  • for long running applications, the JIT compiled PyPy interpreter may offer a suitable alternative to the standard CPython runtime. The main barrier to adopting PyPy is typically reliance on other binary extension modules - while PyPy does emulate the CPython C API, modules that rely on that cause problems for the PyPy JIT, and the emulation layer can often expose latent defects in extension modules that CPython currently tolerates (frequently around reference counting errors - an object having one live reference instead of two often won’t break anything, but no references instead of one is a major problem).

  • Cython is a mature static compiler that can compile most Python code to C extension modules. The initial compilation provides some speed increases (by bypassing the CPython interpreter layer), and Cython’s optional static typing features can offer additional opportunities for speed increases. Using Cython still carries the disadvantages associated with using binary extensions, but has the benefit of having a reduced barrier to entry for Python programmers (relative to other languages like C or C++).

  • Numba is a newer tool, created by members of the scientific Python community, that aims to leverage LLVM to allow selective compilation of pieces of a Python application to native machine code at runtime. It requires that LLVM be available on the system where the code is running, but can provide significant speed increases, especially for operations that are amenable to vectorisation.

Alternatives to handcoded wrapper modules#

The C ABI (Application Binary Interface) is a common standard for sharing functionality between multiple applications. One of the strengths of the CPython C API (Application Programming Interface) is allowing Python users to tap into that functionality. However, wrapping modules by hand is quite tedious, so a number of other alternative approaches should be considered.

The approaches described below don’t simplify the distribution case at all, but they can significantly reduce the maintenance burden of keeping wrapper modules up to date.

  • In addition to being useful for the creation of accelerator modules, Cython is also widely used for creating wrapper modules for C or C++ APIs. It involves wrapping the interfaces by hand, which gives a wide range of freedom in designing and optimising the wrapper code, but may not be a good choice for wrapping very large APIs quickly. See the list of third-party tools for automatic wrapping with Cython. It also supports performance-oriented Python implementations that provide a CPython-like C-API, such as PyPy and Pyston.

  • pybind11 is a pure C++11 library that provides a clean C++ interface to the CPython (and PyPy) C API. It does not require a pre-processing step; it is written entirely in templated C++. Helpers are included for Setuptools or CMake builds. It was based on Boost.Python, but doesn’t require the Boost libraries or BJam.

  • cffi is a project created by some of the PyPy developers to make it straightforward for developers that already know both Python and C to expose their C modules to Python applications. It also makes it relatively straightforward to wrap a C module based on its header files, even if you don’t know C yourself.

    One of the key advantages of cffi is that it is compatible with the PyPy JIT, allowing CFFI wrapper modules to participate fully in PyPy’s tracing JIT optimisations.

  • SWIG is a wrapper interface generator that allows a variety of programming languages, including Python, to interface with C and C++ code.

  • The standard library’s ctypes module, while useful for getting access to C level interfaces when header information isn’t available, suffers from the fact that it operates solely at the C ABI level, and thus has no automatic consistency checking between the interface actually being exported by the library and the one declared in the Python code. By contrast, the above alternatives are all able to operate at the C API level, using C header files to ensure consistency between the interface exported by the library being wrapped and the one expected by the Python wrapper module. While cffi can operate directly at the C ABI level, it suffers from the same interface inconsistency problems as ctypes when it is used that way.

Alternatives for low level system access#

For applications that need low level system access (regardless of the reason), a binary extension module often is the best way to go about it. This is particularly true for low level access to the CPython runtime itself, since some operations (like releasing the Global Interpreter Lock) are simply invalid when the interpreter is running code, even if a module like ctypes or cffi is used to obtain access to the relevant C API interfaces.

For cases where the extension module is manipulating the underlying operating system or hardware (rather than the CPython runtime), it may sometimes be better to just write an ordinary C library (or a library in another systems programming language like C++ or Rust that can export a C compatible ABI), and then use one of the wrapping techniques described above to make the interface available as an importable Python module.

Implementing binary extensions#

The CPython Extending and Embedding guide includes an introduction to writing a custom extension module in C.

FIXME: Elaborate that all this is one of the reasons why you probably don’t want to handcode your extension modules :)

Extension module lifecycle#

FIXME: This section needs to be fleshed out.

Implications of shared static state and subinterpreters#

FIXME: This section needs to be fleshed out.

Implications of the GIL#

FIXME: This section needs to be fleshed out.

Memory allocation APIs#

FIXME: This section needs to be fleshed out.

ABI Compatibility#

The CPython C API does not guarantee ABI stability between minor releases (3.2, 3.3, 3.4, etc.). This means that, typically, if you build an extension module against one version of Python, it is only guaranteed to work with the same minor version of Python and not with any other minor versions.

Python 3.2 introduced the Limited API, with is a well-defined subset of Python’s C API. The symbols needed for the Limited API form the “Stable ABI” which is guaranteed to be compatible across all Python 3.x versions. Wheels containing extensions built against the stable ABI use the abi3 ABI tag, to reflect that they’re compatible with all Python 3.x versions.

CPython’s C API stability page provides detailed information about the API / ABI stability guarantees, how to use the Limited API and the exact contents of the “Limited API”.

Building binary extensions#

FIXME: Cover the build-backends available for building extensions.

Building extensions for multiple platforms#

If you plan to distribute your extension, you should provide wheels for all the platforms you intend to support. These are usually built on continuous integration (CI) systems. There are tools to help you build highly redistributable binaries from CI; these include cibuildwheel and multibuild.

For most extensions, you will need to build wheels for all the platforms you intend to support. This means that the number of wheels you need to build is the product of:

count(Python minor versions) * count(OS) * count(architectures)

Using CPython’s Stable ABI can help significantly reduce the number of wheels you need to provide, since a single wheel on a platform can be used with all Python minor versions; eliminating one dimension of the matrix. It also removes the need to generate new wheels for each new minor version of Python.

Binary extensions for Windows#

Before it is possible to build a binary extension, it is necessary to ensure that you have a suitable compiler available. On Windows, Visual C is used to build the official CPython interpreter, and should be used to build compatible binary extensions. To set up a build environment for binary extensions, install Visual Studio Community Edition - any recent version is fine.

One caveat: if you use Visual Studio 2019 or later, your extension will depend on an “extra” file, VCRUNTIME140_1.dll, in addition to the VCRUNTIME140.dll that all previous versions back to 2015 depend on. This will add an extra requirement to using your extension on versions of CPython that do not include this extra file. To avoid this, you can add the compile-time argument /d2FH4-. Recent versions of Python may include this file.

Building for Python prior to 3.5 is discouraged, because older versions of Visual Studio are no longer available from Microsoft. If you do need to build for older versions, you can set DISTUTILS_USE_SDK=1 and MSSdk=1 to force a the currently activated version of MSVC to be found, and you should exercise care when designing your extension not to malloc/free memory across different libraries, avoid relying on changed data structures, and so on. Tools for generating extension modules usually avoid these things for you.

Binary extensions for Linux#

Linux binaries must use a sufficiently old glibc to be compatible with older distributions. The manylinux Docker images provide a build environment with a glibc old enough to support most current Linux distributions on common architectures.

Binary extensions for macOS#

Binary compatibility on macOS is determined by the target minimum deployment system, e.g. 10.9, which is often specified with the MACOSX_DEPLOYMENT_TARGET environmental variable when building binaries on macOS. When building with setuptools / distutils, the deployment target is specified with the flag --plat-name, e.g. macosx-10.9-x86_64. For common deployment targets for macOS Python distributions, see the MacPython Spinning Wheels wiki.

Publishing binary extensions#

Publishing binary extensions through PyPI uses the same upload mechanisms as publishing pure Python packages. You build a wheel file for your extension using the build-backend and upload it to PyPI using twine.

Avoid binary-only releases#

It is strongly recommended that you publish your binary extensions as well as the source code that was used to build them. This allows users to build the extension from source if they need to. Notably, this is required for certain Linux distributions that build from source within their own build systems for the distro package repositories.

Weak linking#

FIXME: This section needs to be fleshed out.

Additional resources#

Cross-platform development and distribution of extension modules is a complex topic, so this guide focuses primarily on providing pointers to various tools that automate dealing with the underlying technical challenges. The additional resources in this section are instead intended for developers looking to understand more about the underlying binary interfaces that those systems rely on at runtime.

Cross-platform wheel generation with scikit-build#

The scikit-build package helps abstract cross-platform build operations and provides additional capabilities when creating binary extension packages. Additional documentation is also available on the C runtime, compiler, and build system generator for Python binary extension modules.

Introduction to C/C++ extension modules#

For a more in depth explanation of how extension modules are used by CPython on a Debian system, see the following articles: