[gmx-users] intel compiler + amd64: help needed.

[Jones de Andrade]

Hi all.

Of course I want SSE enabled. There is no sense in benching it against
different compilers if in one you enable the speed and on the other, and in
the other you remove almost every optimization. That's why in the last
message I've sent the eror I'm getting in the compilation:

*************************************************************************************************************************************
 /bin/sh ../../../../libtool --mode=compile /usr/bin/gcc  -O3 -ffast-math
-funroll-all-loops -fpeel-loops -c -o nb_kernel010_x86_64_sse.lo
nb_kernel010_x86_64_sse.s
libtool: compile: unable to infer tagged configuration
libtool: compile: specify a tag with `--tag'
make[5]: ** [nb_kernel010_x86_64_sse.lo] Erro 1
*************************************************************************************************************************************

That's ther error that arises from enabling SSE, and doesn't seems to be
related to it (despite it doesn't make too much sense, since it's what is
enabled), but only to some kind of libtool option not being corretly passed.
>From what I could get from the internet, it's related to the swithings in
the compiler being used during the compilation process, but I could not get
a proper way to get around it for gromacs. And I do think this is last error
in the compilations in the way to get a fully working executable gromacs
package compiled with the intel compilers on AMD64 machines. The others are
just warnings (which didn't look too nice for me anyway, but also can mean
nothing), which can be thought about later.

Does anybody knows how to go around this specific error? Is it a known issue
in gromacs, under some kind of different subject?

Thank you all for all the help in advance. :)

Jones

On 6/2/06, Erik Lindahl <lindahl at sbc.su.se> wrote:
>
> Hi,
> > Erik: I'm sure about the flags. Not using the ones for SSE3,
> > despite my processor having those. I'll just try to activate those
> > as soon as I get a compiling procedure. By now, I even using the --
> > disable-x86-64-sse flag for configure, just to be sure. ;) Also, I
> > tried the "CC_FOR_BUILD=gcc" environment, but despite it has gone
> > throw the invsqrt compilation, I looked in and it was compiling
> > everything with gcc, and nothing with icc. :(
> You so do NOT want to do disable SSE in Gromacs. To get the extra 2%
> performance from Intel icc you just killed the assembly loops which
> will reduce your performance by roughly 50%.
>
> The Gromacs SSE code is completely portable, and checks if the
> instructions are supported at runtime.
> > Yang: Yes, I'm using the intel compiler with emt64. I reinstalled
> > the software step by step specifically to be sure about this point.
> > And I understand that gromacs has an assembly core, and that this
> > is basically the "heart" of the program speed. But, even thow,
> > 5-10% means something, and every nano-second counts. Specially when
> > you have a big delay to deal with, and that this 10% mean
> > 0.25ns. ;)  Even then, after I successfully compile it we will be
> > able to say for sure how much if extra performance (if it yelds
> > any) does the intel compiler gives on AMD64 machines. Moreover, as
> > I said in a long time ago message: I want to use QM/MM with gromacs
> > +cpmd, and at this point I really dislike the idea of compiling one
> > thing with intel (cpmd) and the other with gcc (gromacs). Since I
> > can't make cpmd work on gcc, I'm trying to make gromacs work on
> > intel. ;)
> We have already tried the intel compiler multiple times (both on ia32
> and x86-64 platforms), and it gets you roughly 2% extra performance
> in Gromacs.
>
> A really nice thing with icc is that the object format is completely
> gcc-compatible, so there shouldn't be any problems mixing.
>
> Cheers,
>
> Erik
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://maillist.sys.kth.se/pipermail/gromacs.org_gmx-users/attachments/20060602/2812aa16/attachment.html>