RE: GFortran 44 optimizations and Intel SSE4

["Jacob Anderson" <jwa at beyond-ordinary dot com>]

I was able to build the mvapich2 library using my optimizations. When I
tried to build an executable with said optimizations, I got:

undefined reference to `std::ctype<char>::_M_widen_init() const

If I remove the optimization options:

  -march=native -msse4.2 -maes -O3 -funroll-loops --param max-unroll-times=4

Then it links the executable just fine.

It would see, without any credible evidence though, that using the native
architecture optimizations causes some incompatibilities....

I am experimenting with the different options to find out which one causes
the link error.

-- Jake


> -----Original Message-----
> From: fortran-owner@gcc.gnu.org [mailto:fortran-owner@gcc.gnu.org] On
> Behalf Of Tim Prince
> Sent: Tuesday, November 01, 2011 11:05 AM
> To: fortran@gcc.gnu.org
> Subject: Re: GFortran 44 optimizations and Intel SSE4
> 
> On 11/1/2011 12:42 PM, Jacob Anderson wrote:
> > Hi,
> >
> > Thanks for the responses! Yes, I did mean the "-Ox" generic options.
> >
> > We are compiling against Intel X5670 and X5620, so no desktop
> architectures.  The "SSE4.2" extensions are of interest for their vector
> capabilities. Is there any online documentation on the optimization
> algorithms used when the SSE4 optimizations are enabled?
> >
> > We are using gfortran44 version 4.4.4 20100726:
> >
> > GNU Fortran (GCC) 4.4.4 20100726 (Red Hat 4.4.4-13) Copyright (C) 2010
> > Free Software Foundation, Inc.
> >
> > Thanks
> >
> > -- jake
> >
> >
> >
> >> -----Original Message-----
> >> From: fortran-owner@gcc.gnu.org [mailto:fortran-owner@gcc.gnu.org] On
> >> Behalf Of Tim Prince
> >> Sent: Tuesday, November 01, 2011 9:28 AM
> >> To: fortran@gcc.gnu.org
> >> Subject: Re: GFortran 44 optimizations and Intel SSE4
> >>
> >> On 11/1/2011 11:42 AM, Jacob Anderson wrote:
> >>
> >>>
> >>> Does the general optimization option to gfortran44 include Intel
> >>> SSE4
> >> optimizations (4.1 AND 4.2)?
> >>>
> >>
> >> Perhaps that's not what you meant to ask? By general, do you mean
> >> generic (definitely not including -msse4)?
> >>
> >> http://gcc.gnu.org/onlinedocs/gcc-4.4.6/gcc/i386-and-x86_002d64-
> >> Options.html#i386-and-x86_002d64-Options
> >>
> >> -msse4.1 and -msse4.2 are listed there.
> >>
> >>    -mtune=barcelona , which you might use for Core I7 (with or
> >> without -msse4), if you don't want to use a current gfortran, is
> listed there.
> >>
> >> --
> 
> Oh, just in case, auto-vectorization is enabled either by setting -O3 or
> by adding -ftree-vectorize to the options list at -O2 or -O1.  For
> myself, I've got improved results on the Xeon 5500 and 5600 series by
> adding -funroll-loops --param max-unroll-times=4
> 
> 
> --
> Tim Prince