Numerical discrepancies after upgrading Fortran compilers (Intel v11.1 to v13.0)

Question

Background:

We're just trying to upgrade our Intel Composer installation from 2011 to 2013, but we are finding that although most results are unchanged, a few change a bit, and some quite a lot.

I've read through the release notes, and what I understood would not explain these differences (we have no multi-threading, classes - polymorphic or otherwise).

Question:

Are there any difference between the Intel Fortran compiler v13.0 and its immediate predecessor v11.0 in the following areas:

• Are uninitialised variables treated differently?
• Are there any compiler flags (like floating point settings) whose defaults have changed?
• Has the implicit casting of numeric types changed?

Answer 1

I'm just posting Steve Lionel's comment as an answer, so that I can close this question:

"The answers to your three questions are: No, No, No. But there can be many possible causes of numerical differences, especially when the algorithm is unstable. Maybe more of the code vectorizes, improvements to math routines, different order of operations, etc., etc. As others have said, if you can provide a test case to Intel Premier Support, we'll help you figure it out. But you can also work on it by seeing where in the process the results start to diverge." – Steve Lionel Apr 3 '14 at 20:17