8/2021, 5/2023       Optimize for a specific CPU

The wgrib2's makefile assumes a generic CPU.  For 64-bit AMD or Intel 
cpus on a 64-bit OS, that means wgrib2 will be compiled for the 64-bit 
instruction set first introduced by AMD in 1999.  That means the latest 
instructions like AVX-2 and AVX-512 will not be used.  The advantage of 
using the older instruction set is that the executable doesn't need to
be recompiled for different flavors of the processor assuming compatible
flavors of the OS.  The disadvangtage is that the code may run slightly
slower.

Historically I didn't recommend optimizing wgrib2 to the new instructions
(typically SIMD) because the speed increases were limited by disk speeds 
and SIMD vectorization was limited in the code.  (Wgrib2 required OpenMP
3.1 which doesn't support SIMD.)  Wgrib2 v3.1.3 supports the SIMD and
GPU (not used) features of OpenMP if the compiler supports them. So
my stance towards cpu optimization is changing.

1) OpenMP SIMD is useful
    a) double loops
         for (...) {      /* parallelize by threading */
           for (...) {    /* parallelize by SIMD */
            stuff
           }
         }
    b) overhead for threading parallelization is large
       overhead for SIMD parallelization is small
       therefore SIMD is better for short loops

2) File systems can be fast.
   I have NMVE storage at home (PCIe-3 speeds, fast).
   I have lustre storage on my HPC (hard disks)
   I have network storage on my workstation (1GB/s network speed limited, slow)
   So SIMD can have real speed ups on some systems.

3) Optimizing to the CPU is useful for SIMD because
    a) original vector length was 128 bits. AVX-2 is 256 bits and AVX-512
        is 512 bits. Bigger is faster.  
    b) AVX-2 and AVX-512 eliminate the requirement for memory alignment.
        This, in my opinion, is more important.  

4) Optimize for AVX-512?
   AVX-512 is the fastest but has limited support.  For Intel, the
   support is limited to servers and 11th generation chips (12th and
   current 13th generation does not support AVX-512). For AMD, the 
   support is limited to Zen4 chips (5/2023). Intel defined many 
   AVX-512 subsets, so you have to specify the subsets of AVX-512
   that are required. 
5) Optimize for AVX-2?
   AVX-2 is available for every modern x86 CPU for a long time.  The
   only excepsions are the older low-end Atom/Celeron/Pentium (10th,
   11th gen?) from Intel. For AMD, you have to piledriver chips.
   These non-AVX-2 cpus exist, I use a Gemini-Lake notebook for travel.


If you want to try optimizing wgrib2, add the following before compiling


export CPPFLAGS='-march=native -mtune=native'
export FFLAGS='-march=native -mtune=native'
export CC=gcc
export FC=gfortran