vasp.4.6.28 on Itanium 64 bit dual CPU: flush denormal results to zero?

[lahaye]

Hi,

I have compiled vasp 4.6.28 on an Intel Itanium (64 bits) dual CPU system,
using CentOS release 4.3 (the free clone of RedHat enterprise edition 4).

I also have installed the following free packages from Intel:

Code: Select all

intel-ifort9-9.0-031
intel-icc9-9.0-030
intel-mkl-8.0.2p-4
intel-isubh9-9.0-030

These are the fortran/C++ compilers and Math Kernel LIbraries.

When I run vasp, I get a lot of warnings from the kernel like this:

Code: Select all

centos kernel: vasp(15839): floating-point assist fault at ip 4000000000483a61, isr 0000020000000008

This comes from so called "denormal results", which means that a number has become too
small for hardware calculations, and it must be calculated by the kernel, which is a
considerable performance loss.

There is a "-ftz" option with the Intel fortran compiler (flush denormal results to zero)
to force these numbers to become zero and continue as a hardware calculation.
I recompiled vasp with the -ftz option, but I still get the same warnings from the kernel.

I would appreciate if someone has suggestions on this matter.

Thank you,
Rob.

[admin]

usually there are no floating point underflows if the input is reasonable (please check this first). I would suggest to compile with -fpe3 and have a look at which part of the program execution the fp-errors occur.