BLD: cross-compilation for android using cibuildwheel by kfchou · Pull Request #29770 · numpy/numpy

kfchou · 2025-09-19T22:16:59Z

Related to: #27698.

These configs enable you to run cibuildwheel --platform android to build android-compatible .whl files.

You can use briefcase to specify these wheels when you want to build a python-based android app that uses numpy. See discussion for details.

Testing
I tested these wheels using a demo briefcase app, importing numpy and get it to print out mathematical operations.

Questions for reviewers:

Should I include the demo briefcase app for testing? Not sure how I should do that
This build requires python 3.13. I'm not sure how to specify this requirement.

Related to: numpy#27698.

- pyproject: remove unused android test package - cibw_before_build_android.sh: remove redundant code - remove android.meson.cross: cross files are dynamically generated via cibw_before_build_android.sh

andyfaff · 2025-09-20T01:34:03Z

A discussion is needed before we add wheel support for the Android platform. These discussions take place on numpy-discussion@scipy.org, as well as the regular online meetings (advertised on the mailing list).

Whilst Android is obviously quite popular we need to know that finite resources (developer time, CI time) is going spent effectively. i.e. we need to know that demand will be there.

Probably the first step would be to add a CI entry, rather than distributing a wheel.

andyfaff · 2025-09-20T01:35:40Z

Ok, so the original issue said all of that and more. We'd probably be looking for a CI entry for regular test runs, not building a wheel.

mattip · 2025-09-22T08:18:43Z

+# Override system library detection for Android - math functions are in libc
+c_args = ['-DHAVE_SIN=1', '-DHAVE_COS=1', '-DHAVE_TAN=1', '-DHAVE_SINH=1', '-DHAVE_COSH=1', '-DHAVE_TANH=1', '-DHAVE_ASIN=1', '-DHAVE_ACOS=1', '-DHAVE_ATAN=1', '-DHAVE_ATAN2=1', '-DHAVE_EXP=1', '-DHAVE_LOG=1', '-DHAVE_LOG10=1', '-DHAVE_SQRT=1', '-DHAVE_FLOOR=1', '-DHAVE_CEIL=1', '-DHAVE_FABS=1', '-DHAVE_POW=1', '-DHAVE_FMOD=1', '-DHAVE_LDEXP=1', '-DHAVE_FREXP=1', '-DHAVE_MODF=1', '-DHAVE_COPYSIGN=1', '-DHAVE_FINITE=1', '-DHAVE_ISINF=1', '-DHAVE_ISNAN=1', '-DHAVE_ISFINITE=1', '-DHAVE_EXPM1=1', '-DHAVE_LOG1P=1', '-DHAVE_ASINH=1', '-DHAVE_ACOSH=1', '-DHAVE_ATANH=1', '-DHAVE_RINT=1', '-DHAVE_TRUNC=1', '-DHAVE_EXP2=1', '-DHAVE_LOG2=1', '-DHAVE_HYPOT=1', '-DHAVE_CBRT=1', '-DHAVE_NEXTAFTER=1', '-DHAVE_CSIN=1', '-DHAVE_CCOS=1', '-DHAVE_CTAN=1', '-DHAVE_CSINH=1', '-DHAVE_CCOSH=1', '-DHAVE_CTANH=1', '-DHAVE_CASIN=1', '-DHAVE_CACOS=1', '-DHAVE_CATAN=1', '-DHAVE_CASINH=1', '-DHAVE_CACOSH=1', '-DHAVE_CATANH=1', '-DHAVE_CEXP=1', '-DHAVE_CLOG=1', '-DHAVE_CPOW=1', '-DHAVE_CSQRT=1', '-DHAVE_CABS=1', '-DHAVE_CARG=1', '-DHAVE_CIMAG=1', '-DHAVE_CREAL=1', '-DHAVE_CONJ=1', '-DHAVE_STRTOLL=1', '-DHAVE_STRTOULL=1', '-DHAVE_STRTOLD=1']
+cpp_args = ['-DHAVE_SIN=1', '-DHAVE_COS=1', '-DHAVE_TAN=1', '-DHAVE_SINH=1', '-DHAVE_COSH=1', '-DHAVE_TANH=1', '-DHAVE_ASIN=1', '-DHAVE_ACOS=1', '-DHAVE_ATAN=1', '-DHAVE_ATAN2=1', '-DHAVE_EXP=1', '-DHAVE_LOG=1', '-DHAVE_LOG10=1', '-DHAVE_SQRT=1', '-DHAVE_FLOOR=1', '-DHAVE_CEIL=1', '-DHAVE_FABS=1', '-DHAVE_POW=1', '-DHAVE_FMOD=1', '-DHAVE_LDEXP=1', '-DHAVE_FREXP=1', '-DHAVE_MODF=1', '-DHAVE_COPYSIGN=1', '-DHAVE_FINITE=1', '-DHAVE_ISINF=1', '-DHAVE_ISNAN=1', '-DHAVE_ISFINITE=1', '-DHAVE_EXPM1=1', '-DHAVE_LOG1P=1', '-DHAVE_ASINH=1', '-DHAVE_ACOSH=1', '-DHAVE_ATANH=1', '-DHAVE_RINT=1', '-DHAVE_TRUNC=1', '-DHAVE_EXP2=1', '-DHAVE_LOG2=1', '-DHAVE_HYPOT=1', '-DHAVE_CBRT=1', '-DHAVE_NEXTAFTER=1', '-DHAVE_CSIN=1', '-DHAVE_CCOS=1', '-DHAVE_CTAN=1', '-DHAVE_CSINH=1', '-DHAVE_CCOSH=1', '-DHAVE_CTANH=1', '-DHAVE_CASIN=1', '-DHAVE_CACOS=1', '-DHAVE_CATAN=1', '-DHAVE_CASINH=1', '-DHAVE_CACOSH=1', '-DHAVE_CATANH=1', '-DHAVE_CEXP=1', '-DHAVE_CLOG=1', '-DHAVE_CPOW=1', '-DHAVE_CSQRT=1', '-DHAVE_CABS=1', '-DHAVE_CARG=1', '-DHAVE_CIMAG=1', '-DHAVE_CREAL=1', '-DHAVE_CONJ=1', '-DHAVE_STRTOLL=1', '-DHAVE_STRTOULL=1', '-DHAVE_STRTOLD=1']
+EOF


This should all end up in one of the config.h files, not as command line flags

Not even that, if these checks actually fail, they can either be fixed in numpy/_core/meson.build or a simple skip for the checks can be added there. That goes for has_function_sin et al. higher up too.

mattip · 2025-09-22T08:34:35Z

Should I include the demo briefcase app for testing? Not sure how I should do that

We do have aarch64 testing base on github ubuntu-22.04-arm runners. Is there a way to leverage those to test android? I think testing would be a better first step than just wheel building.

rgommers

Thanks for working on this @kfchou! I left some more detailed comments; I'd say the overall goal is to make the custom code we need to carry for this as small and numpy-specific as possible, and having one CI job that tests that building an Android wheel actually works as intended.

Should I include the demo briefcase app for testing? Not sure how I should do that

No, that is undesirable. The right approach is running the test suite (or a relevant subset of it, depending on how show things run) under an emulator. On gh-27698 it was pointed out that this is possible.

This build requires python 3.13. I'm not sure how to specify this requirement.

The general config files (pyproject.toml, cibw_before*, repair_*) should not make such an assumption, since they should work for any Python version >=3.13. In the CI job to test this cross build, 3.13 can be picked.

https://github.com/beeware/toga has examples of CI jobs that may be helpful here.

rgommers · 2025-09-27T04:23:55Z

+set -e
+set -x
+
+# This script is designed to be called by cibuildwheecho "--- END WORKAROUND ---"


Accidental edit at the end of this sentence

rgommers · 2025-09-27T04:24:34Z

+echo "CIBW_ARCHS: ${CIBW_ARCHS:-'(not set)'}"
+echo "CIBW_ARCHS_ANDROID: ${CIBW_ARCHS_ANDROID:-'(not set)'}"
+
+# Check for architecture in standard cibuildwheel environment variablesbuilding for Android.


typo in "variablesbuilding"

rgommers · 2025-09-27T04:25:34Z

+# WORKAROUND FOR CIBUILDWHEEL ISSUE:
+# cibuildwheel documentation states that CIBW_ARCHS should be available for Android builds,
+# but cibuildwheel 3.1.4 does NOT set these environment variables for Android.
+# This appears to be a bug or incomplete implementation in cibuildwheel.


It'd be good to fix that upstream in cibuildwheel and add a link here to that issue or PR.

I think this is a misunderstanding: cibuildwheel reads the CIBW variables as one possible means of configuration, it doesn't set them itself.

rgommers · 2025-09-27T04:27:10Z

+# As a workaround, we detect the architecture from various sources:
+# 1. Check if CIBW_ARCHS is already set (future-proofing)
+# 2. Look for architecture hints in environment variables
+# 3. Check process command line for architecture indicators


You already set this in pyproject.toml, so three different ways of handling it here seems like overkill?

rgommers · 2025-09-27T04:28:14Z

+
+echo "--- Building for architecture: $TARGET_ARCH ---"
+
+# Set the target API level from cibuildwheel's environment variable  


Also already handled in pyproject.toml

rgommers · 2025-09-27T04:37:42Z

+has_function_strtold = true
+
+# Skip BLAS/LAPACK for Android to avoid complex math library dependencies
+allow-noblas = true


This goes in [tool.cibuildwheel.android] in pyproject.toml normally. There's already an example for 32-bit Windows there.

rgommers · 2025-09-27T04:39:08Z

+# This ensures symbols like PyExc_ValueError and math functions are available
+# Note: On Android, math functions are part of libc, not a separate libm
+c_link_args = ['-lpython3.13']
+cpp_link_args = ['-lpython3.13']


If this isn't handled correctly in Meson, that's where this should be fixed. That also avoids hardcoding the interpreter version here.

rgommers · 2025-09-27T04:40:34Z

+# Override system library detection for Android - math functions are in libc
+c_args = ['-DHAVE_SIN=1', '-DHAVE_COS=1', '-DHAVE_TAN=1', '-DHAVE_SINH=1', '-DHAVE_COSH=1', '-DHAVE_TANH=1', '-DHAVE_ASIN=1', '-DHAVE_ACOS=1', '-DHAVE_ATAN=1', '-DHAVE_ATAN2=1', '-DHAVE_EXP=1', '-DHAVE_LOG=1', '-DHAVE_LOG10=1', '-DHAVE_SQRT=1', '-DHAVE_FLOOR=1', '-DHAVE_CEIL=1', '-DHAVE_FABS=1', '-DHAVE_POW=1', '-DHAVE_FMOD=1', '-DHAVE_LDEXP=1', '-DHAVE_FREXP=1', '-DHAVE_MODF=1', '-DHAVE_COPYSIGN=1', '-DHAVE_FINITE=1', '-DHAVE_ISINF=1', '-DHAVE_ISNAN=1', '-DHAVE_ISFINITE=1', '-DHAVE_EXPM1=1', '-DHAVE_LOG1P=1', '-DHAVE_ASINH=1', '-DHAVE_ACOSH=1', '-DHAVE_ATANH=1', '-DHAVE_RINT=1', '-DHAVE_TRUNC=1', '-DHAVE_EXP2=1', '-DHAVE_LOG2=1', '-DHAVE_HYPOT=1', '-DHAVE_CBRT=1', '-DHAVE_NEXTAFTER=1', '-DHAVE_CSIN=1', '-DHAVE_CCOS=1', '-DHAVE_CTAN=1', '-DHAVE_CSINH=1', '-DHAVE_CCOSH=1', '-DHAVE_CTANH=1', '-DHAVE_CASIN=1', '-DHAVE_CACOS=1', '-DHAVE_CATAN=1', '-DHAVE_CASINH=1', '-DHAVE_CACOSH=1', '-DHAVE_CATANH=1', '-DHAVE_CEXP=1', '-DHAVE_CLOG=1', '-DHAVE_CPOW=1', '-DHAVE_CSQRT=1', '-DHAVE_CABS=1', '-DHAVE_CARG=1', '-DHAVE_CIMAG=1', '-DHAVE_CREAL=1', '-DHAVE_CONJ=1', '-DHAVE_STRTOLL=1', '-DHAVE_STRTOULL=1', '-DHAVE_STRTOLD=1']
+cpp_args = ['-DHAVE_SIN=1', '-DHAVE_COS=1', '-DHAVE_TAN=1', '-DHAVE_SINH=1', '-DHAVE_COSH=1', '-DHAVE_TANH=1', '-DHAVE_ASIN=1', '-DHAVE_ACOS=1', '-DHAVE_ATAN=1', '-DHAVE_ATAN2=1', '-DHAVE_EXP=1', '-DHAVE_LOG=1', '-DHAVE_LOG10=1', '-DHAVE_SQRT=1', '-DHAVE_FLOOR=1', '-DHAVE_CEIL=1', '-DHAVE_FABS=1', '-DHAVE_POW=1', '-DHAVE_FMOD=1', '-DHAVE_LDEXP=1', '-DHAVE_FREXP=1', '-DHAVE_MODF=1', '-DHAVE_COPYSIGN=1', '-DHAVE_FINITE=1', '-DHAVE_ISINF=1', '-DHAVE_ISNAN=1', '-DHAVE_ISFINITE=1', '-DHAVE_EXPM1=1', '-DHAVE_LOG1P=1', '-DHAVE_ASINH=1', '-DHAVE_ACOSH=1', '-DHAVE_ATANH=1', '-DHAVE_RINT=1', '-DHAVE_TRUNC=1', '-DHAVE_EXP2=1', '-DHAVE_LOG2=1', '-DHAVE_HYPOT=1', '-DHAVE_CBRT=1', '-DHAVE_NEXTAFTER=1', '-DHAVE_CSIN=1', '-DHAVE_CCOS=1', '-DHAVE_CTAN=1', '-DHAVE_CSINH=1', '-DHAVE_CCOSH=1', '-DHAVE_CTANH=1', '-DHAVE_CASIN=1', '-DHAVE_CACOS=1', '-DHAVE_CATAN=1', '-DHAVE_CASINH=1', '-DHAVE_CACOSH=1', '-DHAVE_CATANH=1', '-DHAVE_CEXP=1', '-DHAVE_CLOG=1', '-DHAVE_CPOW=1', '-DHAVE_CSQRT=1', '-DHAVE_CABS=1', '-DHAVE_CARG=1', '-DHAVE_CIMAG=1', '-DHAVE_CREAL=1', '-DHAVE_CONJ=1', '-DHAVE_STRTOLL=1', '-DHAVE_STRTOULL=1', '-DHAVE_STRTOLD=1']
+EOF


Not even that, if these checks actually fail, they can either be fixed in numpy/_core/meson.build or a simple skip for the checks can be added there. That goes for has_function_sin et al. higher up too.

rgommers · 2025-09-27T04:41:19Z

+import sysconfig
+import os
+
+# Check BLDLIBRARY configuration


Left-over debug code from here onwards?

rgommers · 2025-09-27T04:45:06Z

@@ -0,0 +1,142 @@
+#!/bin/bash


This looks like it should be either a separate tool or part of Briefcase. It's not specific to numpy, just a generic wheel repair utility.

kfchou · 2025-10-07T21:50:55Z

Thanks for your feedback everyone. @rgommers, I'm slowly working through your comments.

Admittedly, most of the this work is done by directing copilot/sonnet since I'm not experience with Meson or C.

mhsmith · 2025-10-12T20:07:38Z

We do have aarch64 testing base on github ubuntu-22.04-arm runners. Is there a way to leverage those to test android?

Unfortunately I found that none of GitHub's ARM runners can run the Android emulator (1, 2). They can only run it on Linux x86_64.

I think testing would be a better first step than just wheel building.

I would strongly recommend doing this through cibuildwheel, as it automaticaly handles all the details of setting up Android build and test environments, starting an emulator, etc. However, on Android it does require the test command to be a python -c or python -m command, so the shell script currently in the cibuildwheel section of pyproject.toml wouldn't work here.

mhsmith · 2025-12-09T19:06:53Z

+
+# Try different approaches to fix BLDLIBRARY issue
+# The issue is meson-python is inserting literal '$(BLDLIBRARY)' in link commands
+# For Android, we don't want to link Python extensions to libpython at all


We absolutely do want to link them to libpython, but if it's using the Python .pc file to achieve this, then this may be affected by python/cpython#138800.

mhsmith · 2025-12-17T21:58:25Z

This PR has now been overtaken by #30412, so I suggest closing it.

NumPy on Android requires complex.h for creal/cimag functions. See: numpy/numpy#29770

kfchou added 2 commits September 11, 2025 13:55

BLD: cross-compilation for android using cibuildwheel

c2cb6b5

Related to: numpy#27698.

BLD: clean up cibw_android files

3287e4e

- pyproject: remove unused android test package - cibw_before_build_android.sh: remove redundant code - remove android.meson.cross: cross files are dynamically generated via cibw_before_build_android.sh

github-actions Bot added the 36 - Build Build related PR label Sep 19, 2025

Merge branch 'main' into cibw-android

842eddb

mattip reviewed Sep 22, 2025

View reviewed changes

rgommers reviewed Sep 27, 2025

View reviewed changes

rgommers mentioned this pull request Sep 27, 2025

Add Android Support #27698

Open

mhsmith mentioned this pull request Nov 3, 2025

Feature Request: Add support for building Android wheels mesonbuild/meson-python#810

Closed

mhsmith reviewed Dec 9, 2025

View reviewed changes

mhsmith mentioned this pull request Dec 10, 2025

BLD: Add Android support #30412

Draft

torlando-tech pushed a commit to torlando-tech/android-python-wheels that referenced this pull request Jan 3, 2026

fix: add -include complex.h for NumPy Android build

25ff102

NumPy on Android requires complex.h for creal/cimag functions. See: numpy/numpy#29770


		echo "--- Building for architecture: $TARGET_ARCH ---"

		# Set the target API level from cibuildwheel's environment variable

Uh oh!

Uh oh!

Conversation

kfchou commented Sep 19, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

andyfaff commented Sep 20, 2025

Uh oh!

andyfaff commented Sep 20, 2025

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

mattip commented Sep 22, 2025

Uh oh!

rgommers left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

kfchou commented Oct 7, 2025

Uh oh!

mhsmith commented Oct 12, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

mhsmith Dec 9, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

mhsmith commented Dec 17, 2025

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

kfchou commented Sep 19, 2025 •

edited

Loading

mhsmith commented Oct 12, 2025 •

edited

Loading

mhsmith Dec 9, 2025 •

edited

Loading