floating point overflow with gfortran compiler

[Vittorio]

I just compiled abinit 7.4.2 with the gfortran compiler using option -ffpe-trap=invalid,overflow,zero and I am getting lots
of floating point exceptions running the test suite.
This is in prttagm.F90 around line 280 and 313.
The array dprarr contains very big values, initialized in outvar_a_h.F90
"dprarr(1:3,idtset)=dtsets(idtset)%bs_eh_cutoff(1:3)"
This is on Linux Fedora 19, x86-64, gfortran 4.8.1.
Am I missing anything?
Vittorio

[jmbeuken (jean-michel-beuken)]

Hi Vittorio,

thank you for the suggestion to use this option !
I confirm that there actually has some problems... :-(
we are going to look in more detail...
we'll keep you informed ...

regards

jmb

[Vittorio (vitti570)]

You might use also gfortran option -finit-real=snan it initializes
local variables to
signaling NaN (Not a Number). Do not forget -g -fbacktrace.
Happy to be of help,
Vittorio

[jmbeuken (jean-michel-beuken)]

Hi,

thanks again for your suggestions...

I have already reviewed some problems and
some are already fixed in the development version 7.5
(to be backported in the next production version 7.4.3, next week... )

cheers

jmb

[Vittorio (vitti570)]

Will you please warn me when backporting to 7.4.3 is completed?

[jmbeuken (jean-michel-beuken)]

of course...
jmb

[jmbeuken (jean-michel-beuken)]

Hi,

The 7.4.3 is out ... (it's the last of 7.4 serie )

thanks of "-ffep-trap=invalid,overflow,zero" option,
we have corrected two bugs and fixed some minors problems...
unfortunately, we forget one but is only related to the anaddb tool...
fixed in the development serie 7.5

now, I create a new bot in our testfarm to test systematically with the option "-ffep-trap=invalid,overflow,zero"
for the time being, we test sequential abinit without fallbacks ( except netcdf + etsf_io )

regards

jmb

[Vittorio (vitti570)]

Just installed abinit 7.4.3 with gfortran
-ffpe-trap=invalid,overflow,zero -finit-real=snan but I still get many
floating point exceptions at prttagm.F90 lines 280 and 313 because
dprarr is so huge.
Is that correct?

[Vittorio (vitti570)]

This is because in outvar_a_h.F90 line 319 dprarr is set to
dtsets%bs_eh_cutoff which is
so huge. Later at line 321 prttagm is called.
Maybe this is to detect undefined variables?
Please advise.

[jmbeuken (jean-michel-beuken)]

Hi,

As I already said in my previous mail, I found later some other problems fixed in 7.5.4 but not in 7.4.3 : sorry !

concerning the bs_eh_cutoff problem, we solve the problem with :

57_iovars/indefo.F90: dtsets(idtset)%bs_eh_cutoff = (/smallest_real*tol6,greatest_real*tol6,greatest_real*tol6/)

regards

jmb

[Vittorio (vitti570)]

In symq3.F90 lines 107 to 109 the array QPT is undefined.
This array is declared in memory_eval as array QPHON but then it is
left undefined.
Can you confirm it? What is the fix?
Thanks, Vittorio