I found differences between the calculations resulted from optzAdd2.for from Intel Signal Processing library.
If I turn on the global optimization, optimize for speed, or full optimization then the release version sometimes deviate quite significantly from the debug version. If I use local or no optimization, then the results of the debug and the release versions are the same.
I have checked that the inputs to the routine optzAdd2 are the same for both release and debug. There's no memory leak problem that I can detect, everything is initialized.
Is there a possibility that the routine optzAdd2 can only be optimized locally ?
Are there any users who have had similar experienes with the routines from signal processing library ?
Many thanks in advance,