As CVF employs only x87 extended precision instructions, rounding to single precision occurs only when storing to memory. Optimization removes many of these roundings, in order to improve performance, usually improving accuracy as well. In most cases, use of extra precision is clearly acceptable within the Fortran standard.
Among the more satisfactory ways to eliminate the variations:
1. Assure that you have written your source code with expressions evaluated in an accurate manner. For example, A*(B-C) rather than A*B - A*C. The Fortran standard specifically encourages compilers to perform this optimization.
2. Employ the /fltconsistency switch, to eliminate some of those optimizations.
3. Promote the critical calculations to double precision in source code, so that the debug and optimized versions are effectively the same.
4. Employ a compiler such as IFL which can generate SSE single precision instructions.
