Gitlab

The gitlab #gitforge includes tones of features. Among these, a facility called the container registry, which stores per project container images. Guix pack allows the creation of custom #reproductible environments as images. In particular, it is possible to create a docker image out of our manifest and channels files with

guix time-machine -C channels.scm -- pack --compression=xz --save-provenance -f docker -m manifest.scm

Check the documentation for options.
Remember that there are obviously alternative methods to produce docker images. The point on using guix resides on its reproducibility capabilities: you’ll be able to create a new, identical docker image, out of the manifest and channels files at any point in time. Even more: you’ll have the capacity to retrieve your manifest file out of the binary image in case your manifest file gets lost.
Then, this image must be loaded into the local docker store with

docker load < IMAGE

and renamed to something meaningful

docker tag IMAGE:latest gitlab-registry.whatever.fr/domain/group/NAME:TAG

go remote

Finally, pushed to the remote container registry of your project with

docker push gitlab-registry.whatever.fr/domain/group/NAME:TAG

At this point, you have an environment where you’ll run your tests using gitlab's ci features. You’ll set up your gitlab’s runners and manifest files to use this container to execute your jobs.
As an alternative, you could use a ssh executor running on your own fast and powerful hardware resources (dedicated machine, shared cluster, etc.). In this case, you’d rather produce an apptainer container image with:

guix time-machine -C channels.scm -- pack -f squashfs ...

scp this container file to your computing resources and call it from the #gitlab runner.

Github

The github is probably the most popular #gitforge out there. It follows a similar to #gitlab in its conception (pull requests and merge requests, you catch the idea ?). It also includes a container registry, and the set of features if offers may be exchanged with ease with any other #gitforge following the same paradigm. No need to go into more details.
There is a couple of interesting tips about using #github, though. It happens more usually than not that users encounter frequently problems of #reproducibility when using container images hosted on ghcr.io, the hosting service for user images. These images are usually employed for running #ci testing pipelines, and they usually break as upstream changes happen: updates, image definition changes, image packages upgrades, etc. If you read my dependencies hell post, this should ring a bell.
What can be done about in what concerns #modernhw ? Well, we have #guix. Let’s try a differente approach: building an image locally, and pushing it to #github registry. Let’s see how.

in practice

An example repository shows tha way to proceed. Its contents allow to create a docker container image to be hosted remotely. It includes all that’s necessary to perform remote #ci testing of a #modernhw #vhdl design.

docker pull ghcr.io/csantosb/hdl
docker images # check $ID
docker run -ti $ID bash

It includes a couple of #plaintext files to produce a #deterministic container. First, the channels.scm file with the list of guix chanels to use to pull packages from. Then, a manifest.scm, with the list of packages to be install within the container.
The image container may be build with

image=$(guix time-machine --channels=channels.scm -- \
             pack -f docker \
             -S /bin=bin \
             --save-provenance \
             -m manifest.scm)

At this point, it is to be load to the docker store with

docker load < $image
# docker images

Now it is time to tag the image

docker tag IMID ghcr.io/USER/REPO:RELEASE

and login to ghcr.io

docker login -u USER -p PASSWORD ghcr.io

Finally, the image is to be push remotely

docker push ghcr.io/USER/HDL:RELEASE

test

You’ll may test this image using the neorv32 project, for example, with:

docker pull ghcr.io/csantosb/hdl
docker run -ti ID bash
git clone --depth=1 https://github.com/stnolting/neorv32
cd neorv32
git clone --depth=1 https://github.com/stnolting/neorv32-vunit test
cd test
rm -rf neorv32
ln -sf ../../neorv32 neorv32
python3 sim/run.py --ci-mode -v

git

We said #git, then, as mandatory when tracking changes (in documentation, project development, taking notes, etc.). Meaningful changes imply new commits (and good commit messages, for what it takes), but this comes along with a risk of introducing issues. Some kind of mechanism is necessary to automatize the execution of a checkout list to be run per new commit. The list is project aware, for sure, but may also be different following the git branch, and even the kind of commit (merges are to be considered differently to regular commits in topic branches, for example). We need to consider what an issue exactly is, and then you’ll need to adopt a different perspective on kinds of checkout lists.

verification

First (ideally), one starts with clear specifications about the goals of current development effort (in practice this never happens in research, and if you ever have it, they’ll evolve with time). These specifications (you’ll figure out where to find them somehow) will define the tests to run. For example, if you need to implement in firmware a deep neural network, you’ll probably have access to a test data set to verify the outcomes are correct. You may tune, improve or even completely change the architecture of your network, at the very end, you’ll have to verify your design with help of the test data set. Additionally, you may define more sophisticated tests: consumption, area, resources, etc. These all fall into the category of verification testing.

unit tests

Secondly, you’ll be running unit tests during your whole design cycle (and they’ll evolve along with it), and target tests (the one we mentioned just before). Does this addition perform correctly ? What if we stress a module with random inputs ? Are we going through all code in a given design unit ? Do we cover all values of some input/output signal in this important module ? These are all unit testing checkouts, and they’ll help us to detect issues in an early stage of design.

codesign

Codesign falls somewhere in between the two previous: as a testing methodology, it includes concepts of verification and unit testing (and can be combined with them). It is way more ambitious and complex, but also more powerful. No matter your testing strategy, the point here is that you’ll be running these tests (fully or partially) automatically at the several different stages of your development cycle. If they fail, you’ll have to be warned.

guix

Guix, as a package manager, provides all necessary software to deploy our tests (and can be extended with additional tooling). It also includes all that's necessary to create a running environment where we will execute our tests. Most importantly, #guix does so in a #deterministic and #reproductible way: we will be able to reproduce our tests in the future under exactly the same conditions. Shell containers, profiles and the time machine mechanism allow the degree of #reproducibility we need here. All it takes is a couple of text files.

Most usually, we will focus on two strategies to seek for issues: local, and remote. Local strategies are greatly based on git hooks, and will be topic of another post. Let’s see now in practice what can be done with help of remote tools, based on #ci, understood as a methodology consisting on automatically executing a set of tests procedures on a digital design.
#ciseries