Andy Porter
Andy Porter
> @arporter Should we rename the "nemo" tutorials to Code transformation before this release as well? Ideally, yes.
Thanks @LonelyCat124 and @hiker. I like "generic" but am happy to go with the consensus.
OK.
We now have a hard deadline of 13th of December for this release. We should probably aim for at least a week earlier to allow for any unforeseen problems.
The only thing outstanding now is #2701 (plus my small PR to update the OMP docs). Do you want to back-port your fix for this @sergisiso or shall we leave...
Release has been published on github and manually uploaded to pypi. RTD has updated. I've opened a PR on spack (https://github.com/spack/spack/pull/47964/files). I'll email the LFRic list now!
I think that making arrays thread-private on a GPU which has a *lot* of threads often causes problems. The solution that has worked in WW3 and which Simon is using...
@sergisiso @LonelyCat124 We do now have the 'privatise_arrays' option but this can be defeated if the DA can't be certain that a given array is not used outside the target...
The code in question is here: https://forge.nemo-ocean.eu/extdevs/nemo/-/blob/Fork-15-refactoring-si3-routines-for-gpu/src/ICE/icethd_zdf_bl99.F90?ref_type=heads#L409 and one of the troublesome arrays is `zdiagbis`.
I think the rules are: 1. A kernel must only write to the current/'owned' cell. 2. Any fields with reads from neighbouring locations must be marked in the metadata as...