Compilations with Intel (XE 2013) for CP2K-trunk (2.7dev) & regtests errors

[Rolf David]

Hi Iain,

It is really helpful !
I'll respond/comment point by point if you allow me.

Points 1 & 4/5 :

I intended to use the regtest as a base test for optimising CP2K on our 
cluster. 
First goal was to have a working "default" build. And then add more 
enhancement: the two you mentioned (libgrid and libsmm), also testing ELPA. 
(and maybe later move on onto MPI/OpenMP mostly for HF and memory needed 
for integrals)
So if -O2 it's not the determining factor in speed vs -O3 with ifort, I 
will stick to -O2 and some -O1/O0

Point 2:

Since we have intel procs, intel mkl/mpi I though intel ifort was the 
choice against gfortran. Maybe I was also dupped by intel advertising: "on 
intel machine intel is the best". And you know. if it costs money it should 
be better (Guess I was very wrong in this case and also relieved they 
didn't bought it for that.) and maybe because I'm *just* a theoretical 
chemist and this a new world (compilation/optimisation vs utilisation) but 
I'm learning ! 

Maybe after, the OpenMP part, i'll go with gfortran (and still Intel 
MKL/MPI because they are better than other BLAS...ScaLAPACK stack/OpenMPI ?)

Point 3: 

Yes, I was confused to when I reread (what arch result in what regtest)  : 
"So I assume some files...-> Fail segfault", I ment some file has to be 
compiled with -O2 (for the -O3 to work) but since there is little 
difference in speed (what I see from benchmarking quickly and from your 
tests also),  I'll start for all in -O2 and test with the 4 files you 
provided, and one by one. I know that qs_vxc_atom.F is vital. And I'll 
start from here.

>From the tests I already ran, I've found the H2-BECKE-MD.inp test need 
et_coupling.F set to -O2 (instead of -O3) to avoid FAILED but if all is in 
-O2 I guess this problem is no more ! (And also find the others, and if 
they are dependant/independant)

Point 4:

Ok. But I was surprised about that. 

For exemple the 4 NEB tests :
Correct in   -O1
Wrong in     -O2 (or -O3)
compared to EPCC Hydra cluster popt
But
Wrong in    -O1
Correct in   -O2 (or -O3)
compared to Sheffield Iceberg cluster popt

So... I was confused. 

And back to point 5:

I'll test against that also. Thanks for the advice.

I have some long simulation (AIMD, good correlation with experience) to 
check after, but the reg-test was in a way of testing segfault/large error 
quickly to put a "Warning".
If no WRONG result be careful. If one WRONG be careful as hell for this 
type of calculation. But mostly to check if I don't break old thing by 
changing libs/optimisation and maybe even code.


So I'll stick with ifort -O2, maybe later test gfortran -O3.

Anyway thanks for throwing some (a lot of !) light on my problems/questions.

Kind regards

Rolf
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.cp2k.org/archives/cp2k-user/attachments/20150610/0ab7f0eb/attachment.htm>