Kernel arguments about rte-rrtmgp

[m214089]

@RobertPincus The NAG compiler is exactly following the Fortran standard. You can change meanwhile some other compilers to behave similar (comply with the standard).

However, we do need a solution for the given problem: What about assumed shape/size arrays as arguments? If one defines those as pointer and uses nullify and associated, you could mimic the optional/'present' - the array bounds can be set as local variables at the beginning of the executable statements section via the size intrinsic or ubound.

Another option would be a BIND(C) container type containing the respective arrays. Just some unspecific thinking.

[peterukk]

Apparently optional arguments are C++ compatible under Fortran 2018, but I guess you need Fortran 2008 compatibility?

[RobertPincus]

@m214089 Agreed that the NAG behavior is the Fortran standard.

@peterukk We could consider to update the standard although 2018 might not yet be supported by many compilers.

The tension, it seems, is between Fortran performance, C compatibility, and flexibility. Fortran performance is best with explicit arrays (assumed-size?); C bindings (important for interoperability of kernels and front-ends) mean no optional arguments and assumed-size arrays (not assumed size, @m214089). I tried having different kernels e.g. for broadband and spectrally-detailed fluxes but a) this leads to lots of repeated code, and b) there are multiple permutations (e.g. Jacobians x broadband/spectral).

Keep the ideas coming.

[RobertPincus]

In commit 3b5d157 on the main branch we resolved this issue by allocating space for all output argument, even if they are not used. This lets the code conform to the Fortran standard for assumed-size array arguments, as noted by @m214089. It means that the memory footprint can't be reduced - but at least memory traffic can be made smaller.