NEMO issue in ice_var_vremap

[sergisiso]

In ice_var_vremap psyclone creates

    !$omp target
    !$omp teams loop collapse(2) default(shared) private(jk0,jk1,zts_cum1)
    do jk0 = 1, loop_stop_1, 1
      do jk1 = 1, loop_stop_2, 1
        if (zh_cum1(jk1) <= zh_cum0(jk0) .AND. zh_cum1(jk1) > zh_cum0(jk0 - 1)) then
          zts_cum1(jk1) = (zts_cum0(jk0 - 1) * (zh_cum0(jk0) - zh_cum1(jk1)) + zts_cum0(jk0) * (zh_cum1(jk1) - zh_cum0(jk0 - 1))) &
&/ (zh_cum0(jk0) - zh_cum0(jk0 - 1))
        end if
      enddo
    enddo
    !$omp end teams loop
    !$omp end target
    zts_cum1(nlay_i) = zts_cum0(nlay_i + 2)
    !$omp target
    !$omp teams loop collapse(1) default(shared) private(jk1)
    do jk1 = 1, nlay_i, 1
      pts_i(jk1) = MAX(0._wp, zts_cum1(jk1) - zts_cum1(jk1 - 1)) / MAX(zhnew, epsi20)
    enddo
    !$omp end teams loop
    !$omp end target

which fails to compile with Could not find allocated-variable index for symbol - zts_cum1 (icevar.psycloned.f90: 1535)

It seems zts_cum1 should not be a private variable in the first loop?

[sergisiso]

This is marked as private because the DependencyAnalysis considers variables with WRITE_WRITE race conditions must be private (zts_cum1(jk1) is rewritten). Also we have 'privatise_arrays' enables, so this condition matters for arrays.

[LonelyCat124]

This loop seems particularly difficult to handle, is this with force=True? I'm surprised we parallelise the (expected) race condition anyway.

[sergisiso]

We have the "write_write are private" logic for scalars. But by default for arrays we refuse to parallelise such case.

The problem here is that this had also the "privatise_arrays" transformation option, which then also accepts "write_write", as we solve race conditons by privatising them. Removing the transformation option has parallelised only the inner loop, and the trip count of the outer is small, so that may be ok for now (*pending actual performance results)