Re: [gentoo-dev] Re: [gentoo-commits] gentoo-x86 commit in sys-cluster/openmpi: ChangeLog openmpi-1.1.1.ebuild openmpi-1.2.4.ebuild

[Matthias Langer <mlangc@gmx.at>]

[-- Attachment #1: Type: text/plain, Size: 3011 bytes --]


On Fri, 2007-12-14 at 16:16 +0000, Sébastien Fabbro wrote:
> -----BEGIN PGP SIGNED MESSAGE-----
> Hash: SHA1
> 
> On 14/12/07 14:12, Matthias Langer wrote:
> 
> > F77="ifort" FC="ifort" FFLAGS="-O3 -xO" emerge -av openmpi
> 
> This how it should be. To make it automatically reproducible, specify
> environment variables in the configuration files.

Hmm, i know this isn't a support list, but as it fits quite well: can
you tell me what configuration files I have to look for?

> 
> > Maybe someone can explain to me what positive side effects the removal
> > of the ifc USE flag has - and why this flag is generally discouraged.
> 
> Positive side effect: avoid cluttering the tree. Why icc/ifc are
> discouraged: you can always try to compile every C/C++ package with
> CC=icc and fortran packages with F77=ifort or FC=ifort. Packages which
> do specify more options with e.g. --enable-icc and friends can be easily
> worked out with the toolchain-funcs and fortran eclass, and most of the
> time they do nothing more than specify the environment variables.
> 
> If we allow icc/ifc flags, at some point, we could allow a whole bunch
> of other compiler flags such as "sunstudio". New keywords for compilers
> could be a better idea, but I doubt we have the human resources to test
> them.

Well, I basically agree. However, it should be noted that fortran cannot
be compared with C/C++. The latter are the languages no gentoo box can
live without, while fortran is a rather exotic kind of beast, that for
mostly historical reasons, is still used in scientific computing.

Last but not least, ifc is in the tree, while "sunstudio" is not...

To cut a long story short: I'm not completely happy with your reasoning,
but you convinced me nonetheless ;-).

> 
> > The reason it disappeared is that it makes gfortran horribly slow when
> > compiling against mpi. This is not the case with ifc, and therefore the
> > old ebuild in bugzilla emitted a bold warning when emerging with
> > USE="-ifc f90-typesafe" but kept quiet if USE="ifc f90-typesave". Thus
> > it *did make sense* to control it with a USE flag, at least with the
> > "ifc" USE flag being around also.
> 
> If the f90-typesafe options always improve compilation time with
> gfortran only,  why not using something like this (modified from the
> openmpi bump bug):

To be exact, "f90-typesafe" slows down gfortran horribly, while ifc
seems to run as fast as normally with it...

> 
> if use fortran; then
> 	case ${FORTRANC} in
> 		g77) myconf="${myconf} --disable-mpi-f90" ;;
> 		gfortran) myconf="${myconf} --with-mpi-f90-size=medium" 				
> myconf="${myconf} --with-f90-max-array-dim=4" 				  ;;
> 		if*) myconf="${myconf} blah" ;;
> 		*) die "unsupported fortran compiler: ${FORTRANC}"
> 	esac
> else
> 	myconf="${myconf} --disable-mpi-f90 --disable-mpi-f77"
> fi
> 

Well, openmpi-1.2.4-r1 has just been commited by jsbronder and contains
code like this...

Matthias 

[-- Attachment #2: This is a digitally signed message part --]
[-- Type: application/pgp-signature, Size: 189 bytes --]