(Towards #1960) put WHEREs containing reductions into CodeBlocks and fix non-unity lower-bound bug.

[arporter]

No description provided.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Comparison is base (699ceb2) 99.85% compared to head (c360dcd) 99.85%.

❗ Current head c360dcd differs from pull request most recent head b1d501b. Consider uploading reports for the commit b1d501b to get more accurate results

Additional details and impacted files

@@           Coverage Diff           @@
##           master    #2372   +/-   ##
=======================================
  Coverage   99.85%   99.85%           
=======================================
  Files         352      352           
  Lines       47334    47399   +65     
=======================================
+ Hits        47264    47330   +66     
+ Misses         70       69    -1     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[arporter]

I've asked Simon if he'd mind testing this branch with the files on master of NEMO that have been giving him trouble.

[arporter]

There's a separate bug that causes us to generate incorrect loop bounds if the array being iterated over has a non-default lower bound. We currently always generate, e.g.:

      do widx2 = 1, SIZE(sjk, 2), 1
        do widx1 = 1, SIZE(sjk, 1), 1
          if (sjk(widx1,widx2,jn) /= 0._wp) then
            r1_sjk(widx1,widx2,jn) = 1._wp / sjk(widx1,widx2,jn)
          end if
        enddo
      enddo

while we should generate e.g. do widx2 = LBOUND(sjk,2), UBOUND(sjk,2), 1

[arporter]

Just checked with SPITZ12 (as that includes ice) and, with master, we get a compilation failure:

 nemo/cfgs/SPITZ_TEST/BLD_SCT_PSYCLONE/obj/iceistate.f90
NVFORTRAN-S-0074-Illegal number or type of arguments to sum - keyword argument array (/home/aporter/NEMO/nemo/cfgs/SPITZ_TEST/BLD_SCT_PSYCLONE/obj/iceistate.f90: 297)

0 inform, 0 warnings, 1 severes, 0 fatal for ice_istate

This PR fixes that one at least :-)

[arporter]

I've realised that even moving to LBOUND/UBOUND isn't guaranteed to work because the mask expression only has to have the same shape as any arrays assigned to within a where-assignment-stmt:

and shape is defined as:

i.e. the shape does not include the lower bounds of the rank(s) of an array.

[arporter]

Therefore, the generated loops should be over the shape of the mask expression. Whenever we index into an array (in a where-assignment-statement), that indexing should take into account the lower bound of the array.

[arporter]

Tests are now happy (incl. with compilation). Need to test NEMO passthrough now.

[arporter]

NEMO integration tests are green and performance is good :-)

[arporter]

I can build GYRE_PISCES and SPITZ12 on NEMO main with the excludes that mention 'where' commented out of mk/sct_psyclone.sh. The GYRE_PISCES case also runs fine (single MPI process). I don't have input files for SPITZ12 for this version of NEMO.

[arporter]

I just need to check again with HEAD of nemo now...

[arporter]

GYRE builds and runs and SPITZ12 builds :-) Need to look at coverage now.

[arporter]

CI is happy now. Ready for another look from @sergisiso.

[sergisiso]

@arporter The WHERE implementation looks good now, the integration test pass and also the formatting checks. I made some final minor comments. The only bigger comment is on the array_mixin get_effective_shape, but this is getting a bit out of scope so I would be happy to moving them to another issues if they are not simple to solve.

[arporter]

SPITZ12 still builds OK, as does ORCA2_ICE_PISCES (in passthrough mode). Integration tests are running. Ready for another look now.

[sergisiso]

Thanks @arporter for addressing all comments and refactoring the ArrayMixin utilities. This is a good improvement. The formatting checks pass, the integration test pass with expected results and the remaining where issues are properly marked with TODOs, so this is ready to merge.