Skip ACC data xfers around halo exchanges for single-GPU runs

[abishekg7]

Issue:

For GPU atmosphere model runs on a single GPU, the OpenACC host <-> device data transfers around halo exchanges are triggered unnecessarily, leading to performance degradation. This issue was first noticed during the 2026 OpenHackathon, and follows the introduction of PR #1355 into the develop branch. The config_gpu_aware_mpi namelist option is meant as a fallback option for performing halo exchanges with MPI distributions that do not support GPU-aware communications. When performing multi-GPU runs with a supported MPI distribution, the config_gpu_aware_mpi must be set to true for better performance. A side effect of this introduction is that config_gpu_aware_mpi defaults to false for the serial GPU case, triggering an extraneous set of halo exchanges, and degrading performance.

Solution

This PR introduces another logical flag, do_halo_exchange, to the if clauses of the OpenACC data transfer statements so that they only trigger when running on multiple-GPUs.

The Nsys profiles of single-GPU runs (Global, real, 30km case) show the performance improvement after the introduction of the do_halo_exchange flag. The host <-> device data transfers that occur during the course of one dynamics step has been minimized.

Before:

After:

This work was completed in part at the NCAR/NLR/NOAA Open Hackathon, part of the Open Hackathons program. The authors would like to acknowledge OpenACC-Standard.org for their support.