nrest=n+4 in mersenne_twister

[NickSzapiro-NOAA]

This fixes the compiler warning:

mersenne_twister.F90:256:38:
            get(n+3:nrest)=transfer(stat%gset,get,nrest-(n+3)+1)
Warning: Intrinsic TRANSFER has partly undefined result: source size 8 < result size 16 [-Wsurprising]

and was bit-for-bit in ufs-weather-model regression tests a couple weeks ago.

The solution is by @DusanJovic-NOAA, with this context:

gset is declared as :
$ grep ':: gset' stochastic_physics/mersenne_twister.F90 FV3/ccpp/physics/physics/Radiation/mersenne_twister.f
stochastic_physics/mersenne_twister.F90:          real(kind_dbl_prec):: gset
FV3/ccpp/physics/physics/Radiation/mersenne_twister.f:          real(kind_dbl_prec):: gset
kind_dbl_prec is 8 bytes (64-bits) so get array (integer) should also be 64-bits or 2 element 32-bits array, 
to match the bitwise length of the gset. 

So to be 2-element array, the
get(n+3:nrest) =nrest 
must be n+4, and it is now defined as:
integer,parameter:: nrest=n+6

The caveat is that we are not familiar with this code, so we're hoping somebody who really understands what's going on here should double check this

The need is that, currently, ufs-weather-model RTs show a pattern of a "ball of diffs" in cpld_debug_gefs_intel (and cpld_debug_gefsv13_intel), failing in run-to-run reproducibility (ufs-community/ufs-weather-model#2675).

A test changing this in both stochastic physics and ccpp improves reproducibility in these tests. The matching stochastic physics PR is at NOAA-PSL/stochastic_physics#85

[BryanFlynt]

Looking this over I believe the new implementation with nrest=n+4 is the correct way to go.
Using a value of n+6 does not align because your trying to copy two 32-bit integers between one 64-bit real on these lines.

ccpp-physics/physics/Radiation/mersenne_twister.f

Line 247 in 2574825

sstat%gset=transfer(put(n+3:nrest),sstat%gset)

and

ccpp-physics/physics/Radiation/mersenne_twister.f

Line 266 in 2574825

get(n+3:nrest)=transfer(sstat%gset,get,nrest-(n+3)+1)

A person could simplify the real64 to int32 copy by allowing the transfer() function to determine the size.

get(n+3:nrest)=transfer(sstat%gset,get)  ! gset is size of 2 get's

FWIW,
I get concerned any time complex bit manipulations are being done this way yet the types are not explicitly defined (i.e. int32, int64, real32, real64). The documentation in the header says all integer must be 32-bit so why not explicitly declare them as such and then inadvertent compiler flags (-fdefault-integer-8) won't break things.

[grantfirl]

I'm definitely not an expert in the mersenne twister code, but it's hard to argue with the change. This is a prime candidate to combine with other PRs if you're OK with it because it is small and (according to your tests) doesn't change RT baselines.

[NickSzapiro-NOAA]

Thanks @BryanFlynt @grantfirl . Maybe let me confirm next week that tests are bit-for-bit on several platforms, and then I'm happy for this to be combined elsewhere.

[NickSzapiro-NOAA]

ufs-weather-model RTs are bit-for-bit on hera, derecho, and wcoss2. afaik, this is ok to combine with other PRs

[grantfirl]

Combined into #295. Keeping open until merged.