Skip to main content

HPy - A better C API for Python

What is HPy?

HPy provides a new API for extending Python in C. In other words, you use #include <hpy.h> instead of #include <Python.h>.

What are the advantages of HPy?

  • Zero overhead on CPython: extensions written in HPy run at the same speed as "normal" extensions.

  • Much faster on alternative implementations such as PyPy, GraalPy.

  • Universal binaries: extensions built for the HPy Universal ABI can be loaded unmodified on CPython, PyPy, GraalPython, etc.

  • A migration path for mixing legacy C-API calls with HPy API calls. Once all the code is migrated, the extension can be compiled as a universal binary that works on any CPython version, PyPy, or GraalPy.

  • Debug mode: in debug mode, you can easily identify common problems such as memory leaks, invalid lifetime of objects, invalid usage of APIs. Have you ever forgot a Py_INCREF or Py_DECREF? The HPy debug mode can be activated at runtime to detect these mistakes for you on universal binaries.

  • Nicer API: the standard Python/C API shows its age. HPy is designed to overcome some of its limitations, be more consistent, produce better quality extensions and to make it harder to introduce bugs.

  • Evolvability: As nicely summarized in PEP 620 the standard Python/C API exposes a lot of internal implementation details which makes it hard to evolve the C API. HPy doesn't have this problem because all internal implementation details are hidden.

Current status

HPy is under active development. 0.0.4 is the latest alpha release but we will soon leave the _alpha_ state and are working hard towards a stable release.

We feel that the HPy ABI is stable enough now that we can fulfill our backwards and forwards binary compatibility promises with the upcoming release, and that the API now covers enough use cases to migrate important packages (in particular, checkout our numpy port). We also provide porting guides and extensive documentation (in particular the API reference).

We are, of course, always open for design discussions and new reqiurements.

The Python/C API is huge. At the moment many popular functions are available. Is something missing for the port of your favorite extension? Please open an issue or even better a PR.

HPy-compatible extensions

The extensions that we are experimenting with include:

  • ultrajson-hpy: this was the first real-world module to be ported to HPy. It is a nice fit because it only exports functions (as opposed as custom types) and requires only a small number of API functions.

  • piconumpy: as the name suggests, this is a minimal numpy-like module which defines a custom type (similar to ndarray but with many fewer features of course).

  • numpy: one of our ambitious goals is to port numpy to HPy, and to use this experience to better understand how to design the API. This port is close to passing the test suite.

  • matplotlib: Since Matplotlib also has a dependency to NumPy, the migration to universal mode is not fully finished. HPy provides the legacy compatibility API such that we can still call legacy C API functions from HPy and successfully run the test suite.

  • kiwi-solver: A dependency of Matplotlib, it has been fully ported to universal mode.

Benchmarks and more information about the Matplotlib and kiwi-solver ports can be found in the GraalVM blog post. The benchmarks show that the HPy ports have little impact on CPython performance, while enabling close-to-cpython performance with GraalVM Python for the kiwi-solver complete HPy port.

Where we need help

  • Documentation: Our resources are very limited and we therefore concentrate on technical tasks. We already wrote a significant amount of documentation (see Documentation) but it is not complete yet. If this non-coding work is something you can do, let us know.

  • Publicity: HPy is already well received among core developers of some important projects. For instance, we are actively talking to Python, NumPy, and Cython core developers. However, we need to get more package developers to be interested in and even use HPy. We appreciate any help for advertising HPy and we are also happy to help preparing talks, demos, etc.

  • Tooling: This is also a topic we would need help. We are, for example, thinking of migration helper tools that do all the boilerplate work that can be automated when migrating a package from C API to HPy API.

  • Packaging: There are several open questions we need to discuss and answer.

    • How to package a universal extension and how to put it on PyPI?

    • How best to package HPy itself

    • How does HPy integrate with setuptools, mesonpy, or other build systems

    • Upstreaming completed forks of the packages listed below, or upstreaming parts of the forks so the code does not suffer from bit-rot

  • Website and logo design: You don't like this website and the HPy logo? Well, we have expertise in virtual machines, compilers, low-level programming, etc, but zero expertise in web development or graphic design. PRs are welcome ;)

More info

Recent blog posts