All,
Just back from Athens after Tim & Frossie's wedding, so I saw this late.
One of our postgrads has been doing alot of work with the intel compiler
so I asked him to comment.
Al.
---------- Forwarded message ----------
Date: Wed, 5 May 2004 14:05:36 +0100 (BST)
From: Neil Symington <neil>
To: Alasdair Allan <aa>
Subject: Re: Q229: fortran compilers (fwd)
> I wasn't sure that this was eligible as an "official" enquiry so i've sent
> it to your individual address. I've just installed the intel fortran [95]
> (ifort) compiler since i've been told, anecdotally, that it generates very
> fast linux i86 code. On trying it (on some f77 code!) i get various
> errors which i suspect relate to the old g77 "no-second-underscore"
> problem. There are some ifort options that look as though they're
> relevant (-nus) but don;t match no-second-underscore exactly (and don;t
> get rid of the error messages.
We link against starlink libraries without needing any options about
underscores. We do need to link with 'libg2c', but I'm not sure what it
does.
The Intel fortran compiler does seem to produce faster code than the NAG
compiler, sometimes much faster. The -ipo flag makes a big difference on
multi-file programs. The vectorization options (for processors supporting
MMX, SSE etc) can produce amazing results for code that contains a *lot*
of integer or floating point maths. Our program does many calculations but
they're embedded in lots of other code, so the vectorization doesn't make
much difference. The vectorization options probably work best on nice f90
code processing big arrays.
We're still using version 7 of the compiler because 8 is quite new and
still seems to have problems with our program. Intel are updating it
frequently, so we may be able to move soon. If you haven't already done
so, register for Intel Premier Support. It's free and it's the only way to
get the updates (and support if you find a bug).
Neil
|