Wheel support for macOS arm64 (Apple Silicon)

[richiksc]

OpenCV 4.50 supports building for Apple Silicon (view release notes, PR opencv/opencv#18094 was merged and released in 4.50). Now that the size limit for the project on PyPI has been increased, is there any chance we get prebuilt wheels for macOS arm64? The build script supports cross compilation from x86_64 to arm64, so I don't think dedicated Apple Silicon CI hardware is needed.

[skvark]

Everything is pending this largeish CI migration task described in here: #422

This repository does not use those build scripts but it looks like the cross compilation needs just a few additional CMake flags so it shouldn't be a problem to build those wheels in the future.

[richiksc]

@skvark Thanks! Looks like Travis CI shutting off OSS repositories has significantly interfered with any work.

[skvark]

One important thing that I would like to add to my previous comment:

It will take some time to get native numpy packages for Apple silicon. In practice, this means months, maybe even longer. This package depends on numpy which means that native arm64 for Apple silicon won't be possible before a compatible numpy version has been released. Requires also Apple Silicon CI runners unless cross compilation is used.

Recommended reading: https://github.com/neurolabusc/AppleSiliconForNeuroimaging

Numpy: numpy/numpy#17807 (comment)

[richiksc]

@skvark In regards to cross compilation, I don't know whether numpy are using macOS CI runners are not, but Clang 12+ on macOS with the Xcode 12+ Command Line Tools installed supports cross compilation from Intel to arm64-apple-darwin20.2.0.

[ogrisel]

It will take some time to get native numpy packages for Apple silicon.

Note that the conda-forge community project had built a M1-native version of the numpy, scipy, pandas, scikit-learn... stack even prior to the release of the first M1 Mac Books. You can install those packages by setting up a conda env with:

https://github.com/conda-forge/miniforge

I use it on a daily basis and it works great. Note that conda-forge does not ship tensorflow though.

[skvark]

It will take some time to get native numpy packages for Apple silicon.

Note that the conda-forge community project had built a M1-native version of the numpy, scipy, pandas, scikit-learn... stack even prior to the release of the first M1 Mac Books. You can install those packages by setting up a conda env with:

https://github.com/conda-forge/miniforge

I use it on a daily basis and it works great. Note that conda-forge does not ship tensorflow though.

This is not a conda-forge project. This repository uses the standard Python package distribution tooling (pip, wheel and manylinux just to name a few). I cannot mix conda packages to these packages because conda uses their own custom package management system. Conda-forge has also a very large community and more resources at their disposal.

In short: when someone runs pip install opencv-python, pip must be able to download and install compatible numpy version from PyPI. Same applies to the build toolchain in this repository.

[richiksc]

@svkark numpy has a new release, v1.19.5, that seems to build on M1. numpy/numpy#17807 (comment)

[skvark]

This repository depends on the official binary releases which means that I need official pre-built wheel packages. There are none: https://pypi.org/project/numpy/#files

You can of course run the build for both numpy and opencv-python now manually and start using native versions.

[asmorkalov]

OpenCV team considers to migrate some of package builds to Github Action platform, first of all for Mac. M1 packages availability depends on M1 support on Github side. Related items:

• Github actions ticket about M1 support: Support for VMs on Apple M1 actions/runner-images#2187
• https://github.com/github/roadmap/projects/1

[asmorkalov]

More details: github/roadmap#95

[hannesa2]

My motivation is a working pip3 install opencv-python on Apple Silicon M1

When I understand right, it makes no sense to donate some Travis credits to make a single event macOS build for arm64 (M1), because a CI is needed with Apple Silicon M1 machines. Travis has no such machines. Am I right ?

Github actions doesn't have any such machines either. That's why a CI with Github action doesn't fulfill my needs too.
Anyway a Github CI would make sense to be ready prepared, when they would be available.
(I started to make the first steps on my fork, but I need some help)

You can of course run the build for both numpy and opencv-python now manually and start using native versions.

To build local opencv-python is currenlty the only way to use OpenCV with Python on Apple Silicon. Am I right ?

Using https://github.com/apple/tensorflow_macos install numpy, so I guess, this is no blocker anymore