packer-rocm/rocDecode: packages

[joshlay2amd]

Inspiration: ROCm/rocDecode@8591037

Extending the packages handled by the ROCm play while reducing those in user-data. Surprisingly, this helps iteration speed. cloud-init processes in series while Ansible can process lists of packages at once.

Other process improvements:

• The 'requirement' installation task is now delayed after repository setup. Allows the ROCm/amdgpu repositories to provide any of them, instead of just the Linux distribution.
• Package downgrades are permitted for requests with the 'extra' and 'amdgpu' packages

The very-specific path has been tested. Need to ensure defaults still behave as expected.

[joshlay2amd]

Commits 6f68a23 and beyond include a refactor from the rocm play to an amdgpu_install role.

This attempts to do the installation for selected components with the scripted method.

Choosing a specific build seemed ineffective, landing on latest despite a specific build provided. I believe simply mishandled changing the build declarations from two keys in a dictionary (amdgpu_install_repos) to independent strings.

More to follow.

[joshlay2amd]

A combination of issues - asking for versioned packages that should be handled by amdgpu-install (d5199ff)... and not passing the build numbers through correctly [as hinted in my last post - caad06b ]

Now able to build/deploy an image with very-specific release/builds and the internal package. Need to test the GA path next.

[joshlay2amd]

Built an image with mostly-defaults, example:

ansible-playbook packer-rocm/playbooks/build.yml \
        -t build \
        -e hidden=false \
        -e rocm_builder_cpus=8 \
        -e rocm_builder_disk="40G"

Inspecting the archive, I found an oversight - the release/build defaults slipped. I intended 6.2.2 (for consistency with some testing)... but had put into motion changes for 6.2.3. Now consistent! Added MAINTENANCE.md as a warning 😄

Now can do both typical paths with amdgpu-install: given tag, given branch/internal packages. I believe this can be merged technically speaking. May benefit from well-rested review - typos and such are very likely.