The ABINIT Users Mailing List ( CLOSED )

[Masayoshi Mikami <mmikami@rc.m-kagaku.co.jp>]

Dear all,

"I'm back" again...(like "T3" ... ;-) )
I am happy to tell you that I have successfully tuned
the optimization levels for v.5.3.4 on Itanium2/ifort (8.1).

Summary:
In ~/.abinit/buiild/(hostname).ac, enable the following line
with_fc_optflags="-O3 -tpp2" # default for most routines
fcflags_opt_17suscep="-O2"  # see (a) below
fcflags_opt_13paw="-O2"   # see (b) below
fcflags_opt_13iovars="-O2"  # see (c) below

(a) as reported before, to avoid the runtime errors
with v1/t55, t56, t57, t58, t71, t72, t73;  v4/t46.
All the tests appears related to "SCF precondition" (iprcel).
"SCF precondition" may be sensitive to optimization ?

(b) to avoid the runtime error with v5/t09
(Marc, sorry that I had to lower the optimization level
 of the PAW directory only for v5/t09...) WIth this -O2,
 v5/09 and v5/10 seems OK.

(c) to avoid the numerical errors wtih v2/t26 (-t28)
and v5/t21. I am wondering why the optimization
of "13iovars" will affect RF computation with
fractional q-points (not always, sometimes !)...

I will see which src directories will need such "-O3"
optimization for high speed (as Gilles suggested....
when I have time, I would like to see ...)

Well, I think that the mission completed... (no "T4" ;-) )

"Hasta la vista, baby", and Bon weekend,
Masayoshi

P.S.
One question: sometimes I have the message on screen, like
t63
File abfiles.run created, starting abinit
abinis(12935): unaligned access to 0x6000000000171f7d,  
ip=0x40000000001bfa10
abinis(12935): unaligned access to 0x6000000000171f7d,  
ip=0x40000000001bfa10
abinis(12935): unaligned access to 0x6000000000171f7d,  
ip=0x40000000001bfa10
abinis(12935): unaligned access to 0x6000000000171f7d,  
ip=0x40000000001bfa10
abinis(12935): unaligned access to 0x6000000000171f7d,  
ip=0x40000000001bfa10
abinis(12935): unaligned access to 0x6000000000171f7d,  
ip=0x40000000001bfa10
abinis(12935): unaligned access to 0x6000000000171f7d,  
ip=0x40000000001bfa10
abinis(12935): unaligned access to 0x6000000000171f7d,  
ip=0x40000000001bfa10
Finished abinit, comparing t63.out with /home/mmikami/ABINIT/ 
abinit-5.3.4/tests/v5/Refs/t63.out
(the above example taken from "make tests_v5". From the fldiff.report,
 "Summary Case_63 : no significant difference has been found")

We might need worry about this ? This is a sign about something ?

On 2007/05/17, at 17:22, Masayoshi Mikami wrote:

> Dear all,
>
> "I'm back ... "
> (imagine  "Terminator" on a car crashing into a police station ;-)
>
> About the optimization tuning for v.5.3.4 on Itanium2/ifort(8.1),
> I think that I have eliminated the source of the runtime errors
> about v1/t55, t56, t57, t58, t71, t72, t73 ;  v4/t46.
> In ~/.abinit/buiild/(hostname).ac, enable the following line
> with_fc_optflags="-O3 -tpp2"
> fcflags_opt_17suscep="-O2"
>
> Marc, as you suggested, I looked into the test cases,
> and thought that "iprcel" seemed fishy.
> Then, I run the above tests with "# iprcel ..."(commented),
> and had no runtime errors this time.
> By grepping "iprcel", I decided to lower the optimization
> levels for the source directories that containing "iprcel"
> and the related keywords (i.e.  SCF preconditioning).
> Some trials and errors, I have noticed that
> only "17suscep" with "-O2" can be fine.
>
> The lest things about  v2/t26, t27, t28,
> v5/t09. t10. t21 will be checked soon...
> (Just << fcflags_opt_16response="-O2" >>
>  did not work ...)
>
> The numerical difference was, i.e. from diff.t26
> 5888,5889c5875,5876
> <     etotal4  -1.3302676449E-01
> <     etotal5   9.7086339501E-07
> ---
> >     etotal4   2.1145567276E-01
> >     etotal5   9.7086341588E-07
>
> "I'll be back" again,
> Masayoshi
>
> P.S. Marc, I had asked our administrator to install ifort 9.1...
> (and not yet..., unfortunately... )
>
> On 2007/05/17, at 6:34, marc.torrent@cea.fr wrote:
>
> > Dear Masayoshi,
> >
> > I encountered several times problems as yours: Abinit crashes when  
> > some agressive optimization levels are chosen, especially with  
> > pgui compiler.
> > It's generally complicated to find the source of the problem (at  
> > which line Abinit crashes).
> > I generally find ifort compiler more stable; did you test v9.1  
> > instead of v8.1 ?
> >
> > Reading the list of your crashed tests, I'm pleased to see that  
> > none of the PAW tests... except v5#t09 but it is testing the  
> > "conducti" PAW feature...
> > v5#t10 is supposed to use v5#09 output in the conducti post- 
> > processor...
> >
> > You could have a look at the tests/v?/README files in order to  
> > find what problematic tests have in common... and put "-O2" to  
> > selected directories.
> >
> > "Asta la vista, baby",
> > Marc