Intel Fortran Compiler

Axel akoh... at gmail.com
Fri Dec 19 16:31:54 CET 2008



On Dec 19, 9:27 am, ilya <ily... at gmail.com> wrote:
> Hi!
>
> For me intel 11 seem to work fine.
> Try compiling without optimizations. Intel says that building with
> such optims (-xW) on non intel cpus is not a good idea.

ilya,
that is just marketing BS. if the CPU supports the proper SSE
instruction set, it should work. however, i found that in many
cases, the optimizations invoked by -xW (and other SSE related
options) can result in miscompiled code (on _all_ architectures,
also on intel).

to give an example, throughout many patchlevels simple code like this.

val(1)=1.0d0*dat
do i=2,n
   val(i)=val(i-1)*dat
end do

would invariably result in a segfault when SSE was enabled.

if you add to this the fact, that there rarely is a noticible speed
advantage in using SSE for DFT and MD codes (apart from BLAS which is
automatically using SSE if you use MKL or ATLAS or GOTO), it is
actually
better to go for more compact and portable code and compile for
pentium3
architecture (which works quite fast and much more reliable on _all_
current cpus).

cheers,
   axel.

>
> Ilya.
>
> On Dec 19, 3:01 pm, Jörg Saßmannshausen
>
> <jorg.sassm... at strath.ac.uk> wrote:
> > Dear all,
>
> > I have tried to build cp2k with the latest intel fortran compiler
> > (11.0.074), using the suggested compiler flags (-O2 -xW -funroll-loops
> > -fpp -free -mtune=core2) (the -mtune=core2 is a suggestion from Axel).
> > Whereas I can build the binaries for the sopt and popt, running the
> > H20-32.inp failes with the popt binary:
>
> > GENERATE|  Number of 1-4 interactions generated:
> >        0
> >   *
> >   *** 13:13:27 ERRORL2 in topology_util:reorder_structure1d processor
> >      ***
> >   *** 0  err=-300  condition FAILED at line 892
> >      ***
> >   *
>
> >   ===== Routine Calling Stack =====
>
> >              5 topology_molecules_check
> >              4 connectivity_control
> >              3 topology_control
> >              2 quickstep_create_force_env
> >              1 CP2K
> >   CP2K| condition FAILED at line 892
> >   CP2K| Abnormal program termination, stopped by process number 0
> > [op22tut:19247] MPI_ABORT invoked on rank 0 in communicator
> > MPI_COMM_WORLD with errorcode 1
>
> > Going back to compiler version 10.1.021 solved that problem, so I am
> > sticking with that test version and run the test suite over xmas.
>
> > I thought I let you know about that so you are aware of that pitfall. My
> > szstem here: dual Dual-Core AMD Opteron(tm) Processor 2212, Suse Linux
> > (64bit) installed, with gcc4.1.
>
> > Best wishes from Glasgow
>
> > Jörg


More information about the CP2K-user mailing list