{P} CPU optimization (part 1)

[FlorianDeconinck]

As of W09.24, the CPU performance is 2.5 slower than the original Fortran.

Dynamical core should be the first target of optimization:

• Flip layout to K-axis as first stride
• K loop needs to be aggregated between all stencils and move up so we have one K loop and many I/J
• Cache coherency work
• Better localization of variable
• Vectorization

Other identified issues are:

• OpenMP parallelism for model that run threaded exists but should be "automatic" depending on user configuration

---

DoD: Reach 20/30% of Fortran. Flag any further improvements for a second round of optimizations, if any

• K-loop aggresive fusion (dace orchestrated) NDSL#6
• K-first friendly memory layout (Fortran layout) NDSL#7
• OpenMP automatic configuration NDSL#8