intel v10.1 compilers

ilya ily... at gmail.com
Mon Apr 28 17:39:03 CEST 2008


The full version was 9.1.052 and as I know it's one of the latest
(maybe THE latest) versions.
The problem was only with parallel version of CP2K.

I think the problem was with the combination of many factors
(compiler, MKL, fftlib, MPI, etc).

On Apr 25, 10:45 pm, Axel <akoh... at gmail.com> wrote:
> On Apr 25, 2:55 am, ilya <ily... at gmail.com> wrote:
>
> > Hi !!!
>
> > I remember that when I tried to compile the whole cp2k with intel 9.1
> > on opteron I've got a lot of strange problems with seg faults etc.
>
> v9.1 != v9.1
>
> the magic is in the patchlevels. if the compiler is around
> long enough and has been patched often enough, it will be
> working as good as possible without major rewrites.
>
> it is just annoying to have to go through the same circle
> over and over again.
>
> axel.
>
> > Upgrading to 10.1 allowed me to successfully compile the thing.
>
> > In both cases of course I turned off Intel specific optimizations (no -
> > xSomething key; -O2).
>
> > It's interesting whether intel provides some sort of 'official'
> > support for nonintel x86 and x86_64 cpus
> > (and for intel cpus too :-)  ).
>
> > I also know that they replaced vectorization module in intel 10.1
> > (compared to 9.1). And in some cases this new vectorizer generates
> > code that runs several times slower (i've made a simple synthetic test
> > -- mandelbrot fractal generator). Some people on their developer forum
> > also reported such performance problems on new compilers.
> > Of course for real-problem code I think there is no such degradation.
> > But anyway...
>
> > So it's interesting what are they doing there in intel???
>
> > PS To tell the truth, their performance suites are rather handy and
> > FREE for linux. So it's very unpleasant to see such problems.
> > Unfortunately...
>
> > On Apr 24, 8:58 pm, Axel <akoh... at gmail.com> wrote:
>
> > > hi everybody,
>
> > > it may perhaps of interest to several people here, that after
> > > extensive
> > > testing, i have to come to the conclusion that the current intel
> > > v10.1
> > > compilers invariably miscompile some modules of cp2k on all platforms
> > > (but different ones on each) and regardless of the optimization level
> > > and
> > > patchlevel of the compiler. most problems show only up when running
> > > in parallel.
>
> > > in all the problematic cases, using a g95 or intel v9.1
> > > compiled executable (with -O2 optimization only) was
> > > fixing the problems.
>
> > > cheers,
> > >      axel.


More information about the CP2K-user mailing list