Newsgroups: comp.unix.cray,comp.lang.fortran,comp.parallel
From: bds@ipp-garching.mpg.de (Bruce       Scott          TK  )
Subject: [CRAY] atexpert and source code
Organization: Max-Planck-Institut fuer Plasmaphysik
Date: Sun, 5 Feb 1995 01:58:13 GMT
Message-ID: <3gthneINN4mf@slcbds.aug.ipp-garching.mpg.de>

Hi All,

I am using "atexpert" to diagnose parallel (autotasked) Fortran code on an
8-processor Cray EL.  Regions where you wasted time by not being
parallelised are identified as "serial regions" and are said to precede a
loop beginning at some line number.  For example, in subroutine "xpsolv"
there is a region labelled

    xpsolv@131

What line number does this refer to?  The one in the xpsolv.m.l list file
(produced with the -r2 flag), or the source code displayed by atexpert,
which appears to be the original, unprocessed source?

I'm pretty sure the trouble is with dot products encountered during a
conjugate-gradient solve, but I would like to confirm this.

I haven't used the craylib sparse routine because this is a 2D solve which
gets vectorised over a third index (we need a "sparsemlt" much like the old
rfftmlt, where the "mlt" refers to the existence of this other index).

-- 
Gruss,
Dr Bruce Scott                             The deadliest bullshit is
Max-Planck-Institut fuer Plasmaphysik       odorless and transparent
bds@ipp-garching.mpg.de                               -- W Gibson


