Looking at program/ptimpl.f, it looks like it is going to accumulate over all elements.
The ghost elements should be tagged with region number 21. So you could add a check in your element to skip isw.eq.13
if ix(nen1-1).eq.21, this would then avoid double counting.
You should create a very simple problem (say, one row of elements and two partitions). Run in serial mode and then in parallel
to double check.