necessary build tools for the archiconda distribution
This project has been superseded by miniforge, a project by conda-forge.
See:
This repository holds many configuration scripts for the Archiconda3 distribution.
Focused on porting conda-forge’s work to 64 bit Arm processors.
The goal of this repository is develop the groundwork needed to compile conda-forge on aarch64.
Once that is layed out, we will be working toward backporting much of this work to conda-forge
conda-build and anaconda-client usingFor users to install Archiconda3, they need to be able to install the conda package.
conda-build, and anaconda-client do not need to be in that package, they can be
obtained from conda-forge directly.
What this means, is that every package that conda depends on, to run, needs to be
an arch specific package. The packages used to build those package do not need to
be arch specific packages.
For now, these are the necessary packages:
We are basically compute bound at this point. shippable gives us 1 CI to use at a time (for one organization).
1 build, even for a nearly empty pure python package, takes about 3 minutes.
See for example the package setuptools:
https://app.shippable.com/github/Archiconda/setuptools-feedstock/dashboard
The 3 minutes (180 seconds) are broken up as follows:
We can maybe cut 50 seconds of this by uploading our own container.
But ultimately, we only have 1 CI, with 1 parallel job at a time, so we cannot run too many
feedstocks at once.
It does have 96 cores, so maybe we can find a different way to parallelize things? I really feel
like that might be over complicating things.
I haven’t had many problems compiling standard software.
For example, python compiled on the first shot, and the issues were primarely due to the fact
that some software hardcodes binutil dependencies.
Therefore, we can potentially not have shippable automatically get triggered.
Users would have to wait until a regular linux 64 bit build passes, before triggering the shippable
build.
python
conda
conda constructor
selenium
chromedriver-binary
python-chromedriver-binary
He will build the following critial packages:
The compilers and libstdc
These include the following:
binutils_impl_linux-aarch64-2.29.1-hc862510_0.tar.bz2
binutils_linux-aarch64-2.29.1-h1dbaa89_0.tar.bz2
crosstool-ng-1.23.0.451_g5888cf1-5.tar.bz2
gcc_impl_linux-aarch64-7.3.0-h68995b2_0.tar.bz2
gcc_linux-aarch64-7.3.0-h98564e2_0.tar.bz2
gdb_linux-aarch64-7.12.1-h6bc79d0_0.tar.bz2
gfortran_impl_linux-aarch64-7.3.0-h5c90dd9_0.tar.bz2
gfortran_linux-aarch64-7.3.0-h98564e2_0.tar.bz2
gxx_impl_linux-aarch64-7.3.0-h5c90dd9_0.tar.bz2
gxx_linux-aarch64-7.3.0-h98564e2_0.tar.bz2
libgcc-ng-7.3.0-h5c90dd9_0.tar.bz2
libgfortran-ng-7.3.0-h6bc79d0_0.tar.bz2
libstdcxx-ng-7.3.0-h5c90dd9_0.tar.bz2
Technically you don’t need make (listed below), but i’ve had so much trouble building it,
I’m just going to take jjhelmus’
make-4.2.1-h7b6447c_1.tar.bz2
These packages are available from his c4aarch64 anaconda channel.
He also created a tag on anaconda.org, I assume through cross compilation,
where you can bootstrap a conda for aarch64.
https://anaconda.org/jjhelmus/repo?label=aarch64_bootstrap
Create a docker image that uses Centos7, and has Archiconda installed.
There is a bug in the version of conda in the bootstrap, that doesn’t allow you to have multiple channels.
The maximum number of channels is 2. This is why this next step is important.
Use that anaconda channel to create an anaconda installer. See the installer directory
in this repositiory.
Start building pacakges.
You can try and build make, but after you do that, try and build m4 with your newly cut version of make.
I couldn’t get it to work. I think jjhelmus did something special. Or I did something wrong 😕
The build order is documented https://github.com/Archiconda/build-tools/issues/4
Most of the heavy lifting is done by fork_conda_forge.py, a python script that does:
fork_conda_forge.py to believe that I’m not stealing your credentials.conda smithy rerender
Archiconda to rerender things correctly.aarch64 branch.Because the repository has been enabled, pushing to that branch will trigger a build on shippable.
The Docker image should:
# Set the locale
ENV LANG en_US.UTF-8
ENV LANGUAGE en_US:en
ENV LC_ALL en_US.UTF-8
Install a few basic pacakges. bzip2 and curl or wget are necessary to help bootstrap the process.
Install Archiconda3.
Add the archiconda3 to the path.
If you are creating the docker image while running on your personal laptop, it is useful to
use qemu-static to run an aarch64 inside x86.
Put the line
# needed to build this on x86
COPY qemu-aarch64-static /usr/bin/
and you should be gravy.
I’ve found that it can be helpful to build and test using qemu and docker.
It might be as simple as iterating quickly to find what dependencies were missing from
the original packages.
Often, binutils are hardcoded, so you have to pass compilation flags to tell the build
system what ar command to use.
It might be useful to mount a local directory to use with the docker build system.
For example, the following command mounts the registration directory which
contains a bunch of feedstocks and runs the archiconda/centos7 image.
The first command it runs is bash allows you to interact with the system.
docker run -v /home/mark2/git/aarch64/build-tools/registration:/feedstocks -i -t archiconda/centos7 bash
Be warned, CPU emulation is SLOW and will make your computer crawl.
I started compiling cmake at the same time it did on shippable, and didn’t even get through the
bootstrap when it has finished compiling.