NEMOv5 GPU ready transformations

[addy419]

Hi @arporter @sergisiso,

I am back working on the GPU benchmarking for NEMO and have encountered issues with the master branch of PSyclone.
The ORCA2_ICE_PISCES config produces a signal 11 with nvhpc@24.11 on bdylib.f90
BENCH config produces a signal 11 on trabbc.f90. Both on the v5.0-RC branch of NEMO.

Can you please check if the integration tests are successful as expected? Are there any changes to the branch or the compiler in the integration tests?

[sergisiso]

The integration tests are successful for v5.0-RC BENCH, but we are using nvfortran 25.1.

[sergisiso]

Oh and examples/nemo/scripts/KGOs/namelist_cfg_bench_small as namelist_cfg, I don't know if there is any difference to the one in the repo

[addy419]

Maybe it is the version then, I am still on 24.11. Will update things and try again.

Also, I am running BENCH with ORCA1-like, ORCA025-like, ORCA12-like and the new ORCA36-like namelists for these benchmarks.

[addy419]

@sergisiso I tried nvhpc 25.1 and 24.11 with both managed and unified memory (and same flags as the CI), but BENCH is unable to compile trabbc.f90 without encountering fort2: terminated by signal 11

[addy419]

From what I see, the integration tests aren't running https://github.com/stfc/PSyclone/actions/workflows/nemo_v5_tests.yml?query=is%3Asuccess, but maybe the status is private? Can you please recheck the status

[arporter]

They are private. You can see the plots here: https://psyclone.readthedocs.io/en/latest/developer_guide/integration-test.html

[addy419]

Thanks! I am unsure what the issue might be anymore. It will be very awkward if it differs on each platform. I'll try falling back to sm_70 (V100) to see if something changes.

[arporter]

The (new!) test platform has H200s and an AMD host processor.

[addy419]

I was successfully able to compile and run ORCA2_ICE_PISCES and BENCH (orca1_like) by adding trabbc.f90 to OFFLOAD_ISSUES.