Re: Running autopkgtests with AMD GPUs with rocm-test-launcher
I've given this some more thought.
First and foremost, I believe I was thinking with too much flexibility
in mind. I saw potential for a pretty capable test runner, but now
believe we should keep it simple, and need a test runner just good
enough for the ROCm packages.
On 2025-06-12 17:38, Christian Kastner wrote:
> The script must also either collect some GTest options and pass them on,
> or pass all other arguments to the tests entirely (so $@ minus the
> --filter option). For example, the previously mentioned rocrand [1] and
> hipsolver [2] both use this to trigger an XML export of the test results
> as autopkgtest artifacts.
I'd still opt for consuming --filter and passing all other arguments to
the tests.
But as GTests can also be configured through the environment, so another
option could be to require that, and only accept filter.
> In the same vein, the options should probably also be supported as
> envvars, which is useful to CI environments which test Debian (or
> Ubuntu) in containers. It allows for single-container re-use.
This still makes sense.
> Finally, this script should go into pkg-rocm-tools, and other packages
> should build-depend on it and copy it to the relevant destination.
> Otherwise we'll end up with dozens of often-diverging copies of some
> runner, as we saw for upstream-binaries.
This does not. We should not be copying files; it's a needless
maintenance chore and tedious to keep in sync.
It would be simpler to keep a master copy somewhere and either use that
to run the tests, or symlink to it if necessary.
runner: /usr/share/pkg-rocm-tools/bin/run-tests
invocation: /usr/share/pkg-rocm-tools/bin/run-tests <package>
or
symlink /usr/libexec/rocm/<package>/run-tests to that, and have
<package> inferred from the pwd.
Or both.
Best,
Christian
Reply to: