conda-mobile

A collection of conda recipes for cross compiling libraries, python, and python extensions for iOS and Android.

The idea is to be able to easily create a python distribution that works on Android and iOS by simply using conda install android-<package> or conda install ios-<package> within an environment created for a specific app.

These recipes have been tested and confirmed working using enaml-native.

Visit anaconda.org/codelv/repo to see which packages are available.

Targets

Currently all recipes are built for 2.7.16 and 3.7.4 with optional support for openssl, sqlite, and ctypes for the following targets:

Android (arm (armeabi-v7a), arm64 (arm64-v8a), x86_64, x86 (i686))

At one time iPhone builds worked but they are no longer being maintained.

iPhone (armv7, aarch64)
iPhone Simulator (x86_64, i386)

Note: All libs and extensions are built as unstripped shared modules. Android strips libraries automatically.

Packages are installed into the env using a prefix for the target platform as follows

android/<arch> - Android recipes for a given arch (arm (v7a), arm64, x86_64, x86)
iphoneos - iOS recipes for the IPhone (armv7 and arm64)
iphonesimulator - iOS recipes for the IPhone simulator (x86_64 and i386)

Each target then contains it's own include, lib, and python folders as follows

# In <env>/<target> (ex miniconda3/envs/myapp/iphoneos)

<target>/include/ # All headers
<target>/libs/ # All shared libraries files
<target>/python/ # Python stdlib and site-packages

Recipes can easily reference these locations using appropriate CFLAGS and LDFLAGS.

Building packages

Only developers of packages should need to build recipes. End users should be able to simply install and use prebuilt versions.

See the conda docs to get started.

There are three general recipe types that are generally relevant for conda-mobile:

Pure python packages, no extensions
Python package with, with extensions
Non python dependent libs

Pure python packages are prefixed with pip-<package> to distinguish them between regular conda packages that may exist with the same name. You can create these package (almost) automatically using the enaml-native-cli's enaml-native make-pip-recipe <package> command.

Libraries with compiled extensions / cython components are split into separate packages with the prefix ios-<package> and android-package. Otherwise all packages would need to be built from mac osx.

To add a new recipe or to build existing recipes:

Install miniconda3
Install conda-build via conda install conda-build (recommended outside of an env)
Clone this repo or create your own recipe(s)
Setup the android standalone toolchain by running android-toolchain/build.sh
Add the requirements to your recipes as needed
If building an extension or c/c++ library install build-essential, autopoint, and texinfo with apt
Run conda-build <recipe-name> (python extensions must be built separately for 2.7 and 3.7 by specifying it as an argument ex conda-build --py 3.7 <recipe-name>)
Then either add a PR or create your own repos with recipes.

Note: If using linux with an encrypted home directory you may have to build in a different root to avoid "path to long" errors. Add the --croot=/opt/conda or some other path to fix this.

Recipe requirements

Now that both python 2.7 and 3.7 are supported all recipes that build python extensions MUST include py27 or py37 in the build string so conda knows which version to install. conda will do this automatically if you simply include python in the run requirements of your recipe (even though the ios / android version is actually used on the app). Alternatively you can manually set string: py{{CONDA_PY}} under the build section.

In addition to specifying a build string, any requirements that have python extensions must filter by the build string since conda does not do this automatically at the moment.

Note: This only applies to python extensions. Pure c/c++ libraries do not need to do this!

For example:

requirements:
  build:
    - python
    - cython
    - android-python * py{{CONDA_PY}}*
    - android-libc++
  run:
    - python # Adding python here make conda append the py27 or py37 build string
    - android-python  * py{{CONDA_PY}}* # Required so it installs the correct pkg for the py version
    - android-libc++ # Pure c/c++ libraries do not need a "build" filter

Failure to include python as a run requirement may cause conda to install the incorrect packages (ie a 2.7 package may be selected for a 3.7 python build!).

In order to make your package "installable" from other operating systems (ie windows) the recipe must use noarch:generic in the build parameters as follows.

build:
    binary_relocation: False
    noarch: generic

The binary_relocation: False prevents conda from doing it's relocation logic on the libraries created which can cause issues.

Installing packages

This requires conda. The preferred method is to install miniconda3 from conda.io/miniconda.html.

Create an env conda create -n myapp
Activate it source activate myapp
Add this channel conda config --add channels codelv
Then install your apps dependencies using the the android or ios prefix. For example conda install ios-python ios-msgpack

Pure python packages can be installed with pip using the --target to tell it where to install. pip install tornado --target=$CONDA_PREFIX/iphoneos/python/site-packages or (preferred) create a conda package for it using enaml-native make-pip-recipe <package>.

Once the packages are installed, the build system (gradle, xcode) can then easily grab these libraries from the env for the given target and use them as is.

Note: An import hook is required to load python extensions from "non" standard locations. enaml-native supports this.

Contributing

This is released under the GPL v3. See the LICENSE included adjacent to this file.

Pull requests are always welcome.

youpsla / conda-mobile