OpenMP                    revised 7.2016, 1.2020, 10.2023

OpenMP is a shared-memory parallel-programming API.  As the grid dimensions get larger,
the need for parallel computing becomes more important.  With OpenMP, wgrib2 will use
multiple threads, typically one thread per core.  For the typical PC, wgrib2 will run 
on the number of cores that are on the CPU chip which is typically 2, 3, 4, 6 or 8.
(Yes, I've owned machines with 2, 3, 4, 6 and 8 cores.) The multi-core speedup is only 
significant when the grids have several million grid points. (Update: 2019 brings
12 and 16 core PCs. Update 2021: Alder Lakes brings more cores.)

You can control the number of cores used by the environment variable OMP_NUM_THREADS
 and the -ncpu option.  The latter overrides the former.

Status: requires OpenMP v3.1 or greater
wgrib2 v3.1.3: allows code to be written for SIMD and GPUs
               only SIMD is present in v3.1.3

Hints:
     complex-packing reading is parallelized
     complex-packing writing is partly parallelized.
     simple-packing is parallelized
     jpeg2000, png, AEC are not parallelized because they depends on external libraries
     geolocation is parallelized except when using Proj4.
     Running multiple copies of wgrib2 can be done along with OpenMP.
     Nodes with 24+ cores should have environment variable OMP_NUM_THREADS set to
      a number less than the number of cores. Little perforance gain for
      a large number of cores.
     NUMA should be considered.
     wgrib2ms/wgrib2mv is faster than using wgrib2 with OpenMP.  
        really fast: wgrib2ms/wgrib2mv with AEC compression

Future:
     wgrib2 allows writting SIMD and GPU code through OpenMP pragmas.
     Wgrib2 is frequently I/O bound, and the run time is usually quite fast.
       So speeding up wgrib2 will have little impact.  Consider writing
       SIMD and gpu code for wgrib2 an exercise on non-trivial code.