Hi Fred,
I have a bunch of package, that I would like to instrument in order to use the rocm autopkgtest ci. Do you know what is the procedure to follow in order to add an autopkgtest specificaly for this rocm-ci ? And inddeed how to have them run on the rocm-ci.
That is wonderful to hear! I would be thrilled to support your
efforts.
In adding autopkgtests that depend on hardware accelerators, the
standard that has been followed for the ROCm libraries has been to
check for the availability of GPU hardware and skip the tests when
appropriate hardware is not available. For AMD GPU compute-based
libraries, this is done by checking if /dev/kfd is accessible and
returning error code 77 when it is not. An example of that can be
found in the rocrand test script [1]. The tests are marked as
'skippable' for this reason [2]. They're also marked with
'needs-sudo' because we need sudo to check for errors in dmesg.
Christian runs the scheduler for the Debian ROCm CI [3]. I believe that the way it works is that he maintains a list of ROCm packages and the scheduler will pick up tests for those packages and all their reverse dependencies. If you wish to test GPU packages that don't depend on ROCm, you may need to ask him to add them manually.
Here an non exhautive list of packages I would like to instrument. clpeak silx pyfai, pyvkfft nabu,
I would suggest clblast. It already has a set of autopkgtests based on pocl. It would be nice to check it with rocm-opencl-icd as well.
Sincerely,
Cory Bloor