Description

In preparation for the transition to the next-generation CCPP framework code generatore capgen, we should try to convert the blocked GFS DDTs used in FV3atm and any code related to it to contiguous arrays. For the purpose of threading, we can use new CCPP framework capabilities in ccpp_prebuild that send chunks of the contiguous arrays to the physics during the time integration (run) phase for OpenMP parallel processing.

Most NWP models have moved to contiguous arrays + chunking as described above, because performance comparisons (for example at UCAR/CISL) showed better performance. In addition, the code is less complicated.

Solution

1. Implement the above solution. Ideally, the regression tests all pass with b4b identical results. This means making fewer changes than potentially possible to achieve b4b.
2. Compare performance of the new code against the existing code

This will open the possibilities to further optimize the code in a follow-up PR.

Alternatives

Implement blocked data structures in the next generation CCPP framework code generator.

Related to

NCAR/ccpp-framework#314