When I was trying to use a Fujitsu Fortran 6.2c compiled ARPACK/PARPACK with MKL, I found a problem with BLAS function dnrm2 (called by PARPACK's pdnorm2 function) returning a NaN. Running everything in a debugger showed that it seems thatBLAS's cutlo and cuthi are also NaN. Everything works fine when compiling ARPACK with Intel Fortran, though. Is it because calls from Fujitsu compiled code skip some kind of initialization in MKL? As I found, Netlib's BLAS dnrm2 initializes cutlo/cuthi with a DATA statement within the function.
Thanks and bye,