I work on a code that used the old Intel FFT routines and have since migrated to the new DFTI routines found in the more recent versions of the MKL (I am using 10.x). Everything worked find until I changed the dimensions of the input array from a square matrix to a rectangular one. For the rectangular matrix, the DFTI routines are generating the incorrect answer.
Once I found the problem, I did some digging and found the DFTI wrappers for the old FFT routines and rewrote my transform routines using the wrappers as a template (i.e. I copied it). Like my original code, the new one works fine for a square matrix, but generates the wrong solution for a rectangular one.
Has anyone else run across this problem before? Is there additional setup steps I need to do when m != n?
Could you be more specific about what vesion of MKL do you use and what wrappers do you use? There was a problem in fftw3 Fortran wrappers which would appear as you described. It has been fixed in MKL 10.2.4.