CCPP acceptance: fv3_stochy / fv3_ccpp stochy bit-for-bit identical

[climbfuji]

These PRs reduce the optimization of a particular routine deep inside the stochastic physics code to obtain bit-for-bit identical results of the (fv3_control based) regression tests fv3_stochy / fv3_ccpp_stochy (static build) in PROD mode.

[climbfuji]

Associated PRs:

https://github.com/NCAR/FV3/pull/119
#205
https://github.com/NCAR/NEMSfv3gfs/pull/102

[climbfuji]

Standard regression tests (Theia, Intel 18, REPRO) all pass as expected.

rt_ccpp_hybrid.log
rt_ccpp_ref_create.log
rt_ccpp_standalone.log
rt_ccpp_static.log
rt_full.log

[climbfuji]

CCPP acceptance tests passed/failed as expected, list of failing tests (down one):

fv3_ccpp_stretched
fv3_ccpp_stretched_nest
fv3_ccpp_regional_control
fv3_ccpp_regional_restart
fv3_ccpp_regional_quilt
fv3_ccpp_regional_c768
fv3_ccpp_control_debug
fv3_ccpp_stretched_nest_debug
fv3_ccpp_gfdlmp
fv3_ccpp_csawmgshoc
fv3_ccpp_csawmg3shoc127
fv3_ccpp_csawmg
fv3_ccpp_gfdlmp_32bit
fv3_ccpp_cpt

rt_ccpp_ref_for_acceptance_create.log
rt_ccpp_static_for_acceptance.log

[climbfuji]

Ready to merge!

[llpcarson]

approved

[grantfirl]

All I could find were whitespace changes for get_stochy_pattern.F90 (correct?). So, the !DIR$ OPTIMIZE:1 is what lowers optimization for one subroutine within a file? I didn't even know that was possible. Cool.

[climbfuji]

Yes, !DIR$ OPTIMIZE:N is an Intel-specific compiler directive, ignored by other compilers. It lowers the optimization for this routine only, not for any of the following/preceding subroutines, and also not for any "contained" subroutines. N can be 0,1,2 and I believe even higher. You can also say !DIR$ NOOPTIMIZE instead of !DIR$ OPTIMIZE:0.
These directives are a little confusing, because they work differently. Another one is !DIR$ NOFMA, which disables fused-multiply-adds (FMAs) that come with AVX2 from the point in the file where the directive is found until a !DIR$ FMA is detected. (just for your info)