README.md

How to run locally

You need to have docker installed on your local machine. Please follow this guide to install Docker CE version.

After installing docker, please follow the commands below

# install the aha wrapper script
pip install -e .

# pull docker image from docker hub (this takes a while!)
docker pull stanfordaha/garnet:latest

# run the container in the background and delete it when it exits
# (this will print out the name of the container to attach to)
aha docker

# you can also do the following instead, but using the wrapper shim is
# suggested as it will do things like mount /cad automatically when 
# detected
# docker run -d -it --rm stanfordaha/garnet bash

# attach to the container name printed out above
docker attach <container-name>

# load required simulator (ncsim)
module load incisive

# run a small regression suite
aha regress pr

Managing Grouped Dependency Updates

It is inevitable that upstream dependencies change their APIs and break compatibility. Before we get into how to manage mass upgrades of dependencies, we'll outline a few steps that, when possible, avoid this problem.

• Upstream module adds new API features and marks the old ones as deprecated.

• Downstream module updates to use the new API.

• Upstream module makes breaking API change and removes compatibility with old API.

By doing this, CI builds will always remain successful and you avoid the problem of having to update multiple dependencies simultaneously, assuming the downstream modules are brought up-to-date in a timely manner.

Of course, this isn't always how things end up happening so here's what to do when needed:

Manually updating a branch

To manually handle the mass upgrades, we need to do a few things. First we'll set up a branch for the upgrades. If this is an upgrade expected to take a long time (e.g. weeks), you might want to set up Dependabot as explained in the later sections and name the branch starting with dev/.

By setting the branch name as dev/ we allow the GitHub Actions CI to automate Dependabot merges into the branch when at least the same number of checks pass (you might want this behavior because it's likely that most of the checks will be failing, and as dependencies update more checks should pass. Without this, you would have to manually merge in Dependabot pull requests to the branch which gets tedious very fast).

Once we have the branch set up, the plan is to have it continually fail CI checks as updates are merged in, and once the CI checks successfully complete it can be merged back into master or whichever branch was the original target.

Configuring a branch

Branches represent different sets of dependencies of the full flow. While the master branch is intended to always be the latest working versions of everything, branches have a few different uses. A branch that doesn't have automatic updates can be useful for freezing a set of dependencies entirely, like when you need to get consistent and repeatable results for a paper. Another use is to track different branches of modules with many dependencies - CoreIR does this in the coreir-dev branch to see if/how changes on the dev branch break downstream dependencies. Lastly branches can be used for coordinating grouped updates, such as when APIs break compatibility in heavily depended modules.

Each branch is by default set to build a Docker image and publish it to DockerHub, so one benefit of creating a branch is that others needing to use the same setup can easily replicate it by cloning the image.

Changing the tracked branches of submodules

To set up a new branch to track a different branch of a dependency (e.g. coreir tracking dev instead of master), create a new branch of this repo and modify the .gitmodules with the modified branch for the relevant submodules.

[submodule "coreir"]
    path = coreir
    url = https://github.com/rdaly525/coreir
    branch = dev

Enabling Dependabot Updates

To enable dependabot to automatically update the submodules in a branch, the .dependabot/config.yml on master needs to be updated. A new entry needs to be added to update_configs, and should probably just be identical to the others with a different target_branch. For example, here is the update_configs entry for master:

  - package_manager: "submodules"
    target_branch: "master"
    directory: "/"
    update_schedule: "daily"
    automerged_updates:
      - match:
          dependency_type: "all"
          update_type: "all"

Once added to the config.yml, dependabot should be set up to make pull requests for your new branch. More information about setting up the config file can be fould here. You might want to set default_reviewers or default_assignees to yourself for your branch.

Configuring a fork

Much like configuring a branch, it's possible to configure a fork. You might want to use a fork rather than a branch for things like papers so it doesn't clutter the main repo with additional noise. If you want the fork to be private rather than public, here is a guide explaining how to do so. Forks will not publish images to DockerHub unless you set them up to.

Enabling GitHub Actions CI

TODO: Document

Enabling Dependabot Updates

TODO: Document

Docker crash course

Create a new image:

• docker build .

List images:

• docker images

Start and enter an image:

• docker run -it <image-id>

List containers:

• docker container ls

Delete stopped containers:

• docker container prune

Clean up all unused docker things:

• docker system prune -a