The code that needs to be migrated to a newer compiler version is used for safety relevant studies. That means any test matrix deviations caused by source changes, compile options or compiler version have to be scrutinized.
The trouble I have today is that I cannot achieve identical results between ifort 11 and ifort 17 even when I use -xSSE2 -O0 -fp-model strict
The test matrix is run on the same machine (Xeon v4) with both binaries.
Unfortunately the code is using single precision. A partial switch to double precision did reduce but not eliminate the numerical sensitivity. The code is time explicit which means that small deviations are propagated over time and pop up sooner or later.
Can anybody recommend further compile flags or other approaches that may help eliminate these deviations ?