KNOWN_PROBLEMS

Known problems

This file contains the list of known problems
at the time a version of ABINIT is released.
Note that bugs discovered later than the
release are NOT described here...

The present release 6.12.3 is a production version 

0. List of platforms
1. Batteries of tests
2. List of known problems

================================================================
0. Platforms

For version 6.12.3, list closed on 20120411
after installation and successful tests for the standard set of platforms,
as mentioned on 
http://www.abinit.org/developers/abinit-dev-status/nightly-builds/abinit-6.12/build-status-6.12.html
and
http://www.abinit.org/developers/abinit-dev-status/developers/dev-environment/buildbot/buildbot-slave-matrix
Corresponding configuration files are to be found in doc/config/build-examples

Versions of compilers known to be too buggy for ABINIT
PGI 7.2.3 compiler          (doc/config/build-examples/amd-pgi7.2-3_chum.ac)    
  problems related to PGI have been discarded in the list below.
gfortran 4.1.2    (=> use gfortran >= 4.2 instead)
intel10.0         (=> use intel 10.1 or 11.0 instead)
intel12.0 or intel 12.1   (=> use intel 10.1 or 11.0 instead)

=============================================================
1. Batteries of tests

Compilation of Netcdf and corresponding testing were tried. 
NetCDF is OK for production.

When NetCDF was OK, compilation of ETSFIO and corresponding testing were tried. 
ETSF_IO is OK for production.

Compilation of BigDFT, and corresponding testing were tried. 
Bigdft is OK for production.

Compilation of Wannier90, and corresponding testing were tried.
Wannier90 is OK for production.

Compilation of libXC, and corresponding testing were tried.
LibXC is OK for production.

Compilation of ATOMPAW and corresponding testing were tried.
ATOMPAW is OK for production.

Compilation of Fox, and corresponding testing were tried.
FOX is OK for production, although it does not work on all platforms, but this is not a ABINIT problem.

Use of MPIIO and corresponding testing were tried.
MPIIO is OK for production, although it does not work on all platforms, but this is not a ABINIT problem.

Use of Scalapack automatically was not tested, and should NOT be attempted by non-expert users with this version.

Memory leaks are now tested automatically. 

Speed was not tested. 

Memory was not tested. 


===============================================
2. List of known problems

All the problems are listed cumulatively.
Those specific to version 6.12 are found at the end of this file

The type of the problem is identified as follows :

NUMERICAL ERROR : the physical result is incorrect,
 but the job does not crash. Moreover,
 the use of the input variables is legitimate.
 This is a very dangerous type of problem. 
 There are two subclasses :
 - INTERNALLY INCOHERENT
  (the bug can be seen because two sets of
   input variables should give the same results
   and do not) 
 - INTRINSIC
  (the bug does not cause incoherency, it can
   only be seen when comparing results with 
   those of another code or with experimental results)
 The latter one is the most dangerous, of course ...

CRASH : the code does not produce the result, because
 it stops unduly.

SLOWING DOWN : the physical result is correct,
 but the code does not go as fast as it should.
 For example, the reading of a restart file might 
 not be correct, or the SCF convergence 
 might not be as fast as it should.

OTHER : some information (not numerical results)
might be erroneous ...

In each class of problem, one can identify
MACHINE-DEPENDENT problems (on some platform, this
problem does not occur, not on others) as well as 
GENERAL problems.

Some of the problems listed below
have been reported by user, but have not yet been
examined for confirmation, in which case they are 
labelled "TO BE CONFIRMED". 

=================================================

P2. (XG, 001223 + later) 
OTHER  (TO BE DONE)
GENERAL
The memory needs are not yet evaluated correctly
for the spin-orbit case and non-collinear magnetism
case.

P8. (GMR, November 2000)
SLOWING DOWN
GENERAL
restartxf for optcell/=0 might have a bug,
despite Test_v1#80
Identified in March 2001 by GMR : only happens when
restarting from a different, non-zero, optcell.
Should be corrected, but documented, now.


P15. (XG, 010103) 
OTHER
GENERAL
Should enforce coding rule related to length of file

P117. (Laurent Pizzagalli, 010622)
CRASH
Observed on PC AMD, atomic relaxation (ionmov==7),
no cell optimisation, stops in xfpack complaining
about ndim. See mail 010622.

P118. (XG, 010710)
OTHER
GENERAL (d/dk by finite differences)
Test_v3 #5 . Sensitivity to phase choice.

P121. (GMR, 010801)
NUMERICAL RESULT
MACHINE_DEPENDENT (Compaq ES40, Turing)
The use of fourdp , with fftalg 200 or 1xx
gives complex conjugate results. This was seen
in the process of merging the GW code with ABINIT.
To be examined in detail.

P204. (XG, 011020)
OTHER (TO BE DONE)
MACHINE_DEPENDENT
Test_v3#14 : the output file is machine-dependent,
because the spin-phase of spin-degenerate wavefunctions
has not been fixed...
Or, is there something more fundamental ??


P302. (XG, 020305; DDKlug, 020313)
NUMERICAL ERROR
MACHINE_DEPENDENT (IBM 44P; SGI)
RF GGA : Test_v3#8,16,18,60,61,62,66
Problem with Response Function GGA tests.
The accuracy of the RF GGA Test_v3#16
is very bad on these machines : the acoustic
modes have large frequencies...

P350. (XG030225)
OTHER
GENERAL
psp*cc.f the fifth derivative is not computed.

P404. (XG030103)
NUMERICAL ERROR
LIKELY GENERAL
fftalg 401 works, but gives a slightly
different answer than usual for Tfast#10.
To test this, replace the default value in indefo.f

P405. (XG030106)
CONVERGENCE PROBLEM
LIKELY GENERAL
Test_v3#15 : with very well non-SCF converged wavefunction (nnsclo=2),
the total energy at intermediate steps is lower than the
final energy !

P406. (XG030106)
CONVERGENCE PROBLEM
LIKELY GENERAL
Test_v3#85 : set iscf to 5 in the RF, one sees a very bad convergence...

P416. (XG030522)
CRASH
Seen on Intel/PC with PGI compiler, in parallel
Test_v1#75,80  do not work
Problem to read XFhistory part of a wavefunction file.

P41.9. (XG030719)
PORTABILITY PROBLEM
GENERAL
Test_v4#37 is not portable : the spin decomposition of the
spinor wavefunctions is not gauge invariant. This
test has been disabled.

P42.7. (DH030929)
OTHER
Should examine the d2sym3.f routine, giving extra
zeros in the output file. See mail
from DHamman, on 2003-9-29.

P44.2. (DHamann20040527)
NUMERICAL INACCURACY
For the treatment of pseudopotentials with pspcod=6 (FHI),
the non-linear XC core charge is not treated consistently
in psp6cc.f and in the rest of the code.
See Don's mail to the developers of 20040527 
See also input variable optnlxccc .

P45.5. (XG050220)
OTHER : PARALLEL TESTING
Test_v4#90 (CASINO output) crash in parallel (test_paral on grumpy).

P45.6. (XG050225) 
NUMERICAL PROBLEM (with ionmov=12)
Test_v4#97 has one strange output

P45.17 (B. Andriyevsky 050405)
Problem of normalisation of the epsilon'' from 
the "conducti" code ?!
A factor of pi (3.1416) might be missing.
See his mail of 5 April 2005 to the forum.
See also other mails exchanged on the forum.

P45.18 (D Hamann 050510)
NUMERICAL ERROR
With finite-difference ddk calculation
See his mail of 10 May, 2005.
(Excerpt : ... 

The problem is that 4.4.4 and 4.5.2 WILL run drf.in and produce incorrect
results, all related to the dataset 3 finite-differend ddk calculation.
The results are only somewhat incorrect, not obvious garbage. Here is a
small sample, Born charges from "2nd-order matrix" in the .out files:

Correct (from erf.out):
   1    1   1    4   -4.1367712586    0.0000000000
   1    1   2    4   -0.0863319602    0.0000000000
   1    1   3    4    0.0706059161    0.0000000000

Incorrect (from drf.out):
   1    1   1    4   -3.4782061716    0.0000000000
   1    1   2    4   -0.0832639192    0.0000000000
   1    1   3    4    0.0560776887    0.0000000000


Putting in some istwfk printing, I found that setting berryopt/=0 now forces
istwfk=1 for dataset 3.  The problem apparently is that dataset 2 has some
istwfk>1, and apparently dataset 3 is getting incorrect WFK's from dataset 2
and using them.  Why it doesn't re-converge them first isn't clear.

Setting istwfk=1 in dataset 2 (erf.in) fixes the problem and gets correct
results.  (My hack also had the effect of setting istwfk=1 in dataset 2 as
well as 3.)

Curiously, removing the dataset 2 calculation altogether and feeding the
kptopt1=1 WFK's from dataset 1 directly to the finite-difference ddk
calculation (nrf.in) also fixes the problem, despite the fact that dataset 1
has some istwfk>1.

Clearly something bad is going on in WFK readin, but only in certain cases.
I've never looked at this part of the code, and I'm afraid I can't divert
any more time now towards tracking this down.



P46.1. (XG050701)
PORTABILITY PROBLEM
Test tutorial Finite electric field 
alas_relaxdion.in (the last one)
does not work on my PC : it stops before the end.


P46.21. (XG051017)
Memory leaks
Using the g95 compiler with debugging option,
makefile_macros.g95_sleepy_debug  (-g)
Identified remaining memory leaks in
Test_v2#01 : hdr not cleaned (init in hdr_init routine)
Test_v3#30 : hdr not cleaned (init in hdr_io routine)
Test_v4#50 : in sortph.F90 


P50.16 (XG061215)
DOC
Many documentation files should be updated still ...
Although most of the files put on the Web (.html) are OK.
In particular :
doc/developers/dirs_and_files is to be updated


P51.22 (MGiantomassi060627)
The new value of rprimd is not stored in _WFK file .
So, restart cannot benefit from modifications of acell.
See his mail of June 27, 2006. See also P8.


P53.11 (XG070121)
Still have to get rid off old iscf value for tutorial and tutorespfn automatic 
testing, as well as updating tutorial html files.
Later, get rid off old iscf value in other automatic tests.

P53.22 (XG070302)
PARALLELISM : EFFICIENCY
outkss.F90 actually works k point by k point, sequentially ...
In the k point and spin loop, one should write locally
vkb,vkbd, en and wfg , without communication, then
transmit them in a separate spin and k point loop.
The k point loop only starts at 1020 , the write start at line 1506.

P53.27 (XG070401, modified XG070601)
NUMERICAL ERROR
GW parallelism over bands
Using two or four procs does not give the same macroscopic
dielectric constant as with one proc., at the level of the fourth digit .
See also test paral N .


P54.1 (XG070407)
NUMERICAL ERROR
With jellium slab, forces differed between PC + PGI4.0 or IFC
See test v5#56
However, in atmdata.F90, the znucl 0 has been defined,
and everything is now the same between the two platforms.
Still, should investigate what is the correct answer ...

P54.7 (XG070601 - actually, much earlier than that, but not documented)
CRASH : finite electric field with atomic displacements
Test ffield_7 does not work on my PC, and has been disabled .
The corresponding section of the ffield lesson of the tutorial has been commented ... 


P54.18 (XG070825)
Other
The computation of born Effective charges and dielectric tensor
under finite electric field should not be activated
with rfuser , but by the usual rfelfd=2, used with berryopt=4 .
See Tv5#23 .
Also, content of the doc file (now doc/input_variables/README_EFIELD_BEC_DIEL)
should be mentioned in the description of the variable rfelfd=2 .

P54.19 (XG070830)
Other
The decomposition of the total energy is not printed out in the case of RF under
electric field. Cases v4#80 and v5#23


P55.22 (XG071109)
Numerical errors, finite electric field
for phonons
Test v4 # 80
Non variational and variational expressions give different numbers ...
See line 1215 of t80.out, compare with line 1342 and line 1344.
Also, lines 2944-2946 should not be printed, as well as
lines 2954-2956, 2964-2966, etc ...


P55.26 (XG071123)
Numerical error
Automatic test v5#24 (Electric field response with spin-orbit)
On PC + g95 , large numerical error   (tolerance has been set to a large value)


P55.34 (XG080301)
Numerical problems
Parallel over band, FFT and k points
Test R : there are significant differences
between tR0.out and tR4.out .
(no problem for physical results, e.g. the energy difference 
during the convergence is less than 0.1mHa, but not much smaller).


P56.03 (XG080802)
PAW with lobpcg algorithm (and band/FFT parallelism), and istwf 0.
Test W0 : interestingly, it runs correctly in parallel...
The use of iprcel 45 for dataset 4 has been disabled. 
 dielmt : 15 largest eigenvalues of the hermitian RPA dielectric matrix
  1-5  : NaN         NaN         NaN         NaN         NaN 
  6-10 : NaN         NaN         NaN         NaN         NaN 
  11-15: NaN         NaN         NaN         NaN         NaN


P56.08 (XG080828)
PAW with lobpcg algorithm (and band/FFT parallelism), and istwf 0.
Test W : numerical results for dataset 5, 6 and 7 are wrong, it seems...
They should be identical to test R and are not.
Also, test W is not portable.
It was disabled for v5.6.0


P56.17 (XG080913)
Wannier90 test procedure is not yet complete :
only the .out are tested.


P56.22 (XG081202)
# (Perhaps solved, should be checked)
# PAW + Spin-orbit is still problematic.
# In the non-spin-polarized case (Kramers degeneracy) nspden=1 with spin-orbit gives different
# results than nspden=4. Also, the forces are not correct.
# See test v5#17
Apparently solved ! (MT feb. 09)
Forces obtained with nspden=1 are 1.d-20 because of symmtries (nsym=6);
With nspden=4, they are 1.d-6, but nsym=1 (due to spinat value).
If better convergence is achieved with nspden=4, forces decrease.
If nsym=1 with nspden=1, forces become 1.d-6.
WARNING : XG090823, the portability of the computation of forces is still
very low.

P56.23 (XG081203)
The Etsf_io support of density files from PAW is not yet working.
Thes test etsf_io #10 has been disabled.

P57.06 (XG090110)
the bin_seq and bin_seqpar tar.gz files contain Makefiles that do not allow to perform the 
seq tests. Only tests_speed ...

P57.08 (JZwanzig090112)
On Opteron+ifort10.1 : transferability problem with v3#97 and v4#68
(Interestingly, no problem with chum, that is also an Opteron+ifort10.1).


P59.01 (MG090604)
On fock(IBM PW5) + XLF9.1, GW tests were failing due to call to deep_copy_ch1d in module m_copy.F90
It seems that XLF doesn't handle correctly pointers to string arrays neither when the length is assumed
nor when it's explicitly passed as argument. For the time being, deep_copy_ch1d is declared as private procedure.


P59.03 (XG090713)
For Tv4#69 and Tv2#36 , the .cnf file has to use opt=-easy for chpit+intel11.0 .
Perhaps this will be corrected with a better version of intel 11 ? 
Otherwise, should examine why chpit+intel11.0 is so sensitive for these tests, and set back opt=-medium


P59.06 (XG090719)
Original Tetsf_io#08 crashed with chpit+intel11.1 and test+gcc44.0 , while it works
on many other platforms. This happened at the time of the switch between abinis and abinit (with MPI).
It was observed that with ndtset 2 (instead of ndtset 4), no crash occur. This is the current status
of the test, but one should find the problem with ndtset 4.

P59.07 (XG090720)
On bigmac + g95, several problems :
- v3#83 hang
- all libxc fail
- v1#07
- v5#11

P59.09 (XG090827)
Although the Berry phase calculation for norm-conserving case is behaving properly
in parallel, the finite-electric field calculation is not entirely correct :
the eigenenergies differ in the sequential and parallel cases.
See test paral#L

P59.10 (XG090905)
Test v5#40 has low portability. Had to enforce -ridiculous in tests/v5/tests.cnf .
Identified a potential source of low portability : when eigenenergies are degenerate,
but the local DOS contribution is different for degenerate states, there is
a large arbitrariness of the result.
In ABINITv6.0.0, tried to average over the local DOS contributions of degenerate wavefunctions,
writing a new routine dos_degeneratewfs.F90 . Unfortunately, the portability was not improved ?!
See Tv5#40 .
 
P59.11 (XG090907)
See mail of BBertrand 
L'algorithme de Broyden pour déterminer un acell optimal ne converge pas
dans ce réseau rhombohedral

P59.13 (XG090909)
The input variables bmass and vmass should be merged.
Moreover, the doc is confusing : pointing both to bmass ...

P59.14 (XG090912)
vprtrb does not behave properly, see Tv6#02

P59.15 (MG090914)
GW test procedure is not yet complete :
only the .out are tested. Other external files (_GW, _SIG, _SGR) are not tested.
Similar problem in t532 in which the output and the reference _BXSF files should be compared.

P59.16 (SLeroux091015)
CRASH
With gfortran compiler,
if the first dataset has kptopt1 = 0 and the second dataset has kptopt2 = 1, 
during the deallocation of variables before changing the dataset,
abinit crash during deallocation of the variable kg in the routine gstate.F90 (line 1434)
(see tests/v6/t06.in, the commented dataset, for an example of input leading to crash). 
There's no problem with ifort compiler. 

P59.17 (XG091024)
Not sure pawprtwf works correctly, see the normalisation of wavefunctions
in Tetsf_io#09 (file t09.log)

 norm =    1.2344680606460083
 norm =   0.97410639792764531
 norm =   0.98106849665753615
 norm =   0.97966490153644847
 norm =    1.0448355540211380
 norm =   0.98043908884436204

Close to one, but why bigger than one ?


P61.1 (MG100312)
Check on group closure in v4/t26 and v4/t27 fails if all possible
product of symmetries is considered (new version of chkgrp).
The rotational part belong to the input set of symmetries but 
the magnetic part differs. To be investigated in more detail.
For the time being the check on the magnetic part has been removed.

P61.2 (XG100315)
The tests v6#72-76 have big problems of portability, see the enormous tolerances ...

P61.3 (MG100312)
The values of dtsets% printed at the end of the run are not consistent
with the values used during the different runs.
driver.F90 indeed passes a local dtset datatype to its children and
this local datatype is not copied to dtsets(:) before returning.
I've already changed driver.F90 but this modification will affect
several output files. Therefore it should be done by Xavier during the
final merge.

P62.1 (XG100624)
In test libxc#08 , the first values of ETOT differ from platform to platform, by a few part
per million, while the final result converges to the same value. However, in the mean time,
the potential residual differs by a few percent, for absolute values that are very large.
So, the tolerances in report.in have to be enormous, while the test is actually portable
at the level of the final result. It should not be very difficult to find
what's happening for the initialisation ... Well, one day or so ...

P63.1 (XG100829)
In test tutorespfn#ffield_6, the first value of ETOT differs in the MPI and serial versions
on testf. As they are reference platforms, I had to use opt=-ridiculous. 
However, the test is actually portable
at the level of the final result. It should not be very difficult to find
what's happening for the initialisation. See also P62.1 .

P63.2 (XG100906)
make abirules
is problematic ... Problems should be solved for v6.4.0 in production ... !

P66.1 (XG110115)
On chum/pathscale, the test mpiio/Q4  fails 

-P-0000  rwwf.F90:372:BUG
-P-0000   Reading option of rwwf. One should have npw=npw1
-P-0000  However, npw=169, and npw1=0.
-P-0000
-P-0000  leave_new : decision taken to exit ...

So, mpiio testing has been desactivated on chum/pathscale .

P66.2 (XG110205)
Several input variables are not documented or tested, see tests/abirules/Refs/t01.out

P66.3 (XG110206)
The tutorial detu, PAW2 and PAW3 should be automatically tested, 
especially now that AtomPAW is a plugin of ABINIT.
A new test directory has been created, to contain
the tutorials in which a plugin is used !
Also, hyperlinks should be created in the text of the detu tutorial.
(e.g. to the input variables ...)

P66.4 (MG110312)
fftalg=412 produces wrong results when fourwf is called with option=2, istwfk=1, cplex=2.
Calling fourwf with fftalg=412 and option=3 (R-->G) causes a sigfault error.
For the time being, the corresponding automatic tests in fftprof have been disabled.

P66.5 (XG110401)
Enabling the GPU for littlebuda has a strange side effect, of making Tv6#66 crash with a segfault.
So, enable_gpu = no was set in build-example.conf

[bb_intel-gfortran4.5_littlebuda]
...
enable_gpu = no

P67.1 (XG110307)
The FFTW3 algorithm cannot be used with DMFT. The default 
setting of fftalg=312 in case the FFTW3 library is present has been
disactivated in indefo.F90 when usedmft/=0 . Should be reactivated
when DMFT and FFTW3 are compatible ...

P67.2 (XG110309)
There is a lack of portability for the output of Tv5#07, between
the reference platform and the FFTW3 calculation.
T5#07 being related to self-consistent electron-positron calculation,
with iterative self-consistency control that is less flexible than the 
usual one, it has not been possible to adjust the test such as to obtain a better
portability. However, the final results are OK.

P67.3 (XG110315)
There is a lack of portability of Tv5#15 .
The problem lies in the fact that when the density
becomes negative, due to Bragg oscillations, the xc kernel can be extremely large, and when
use for the prediction of the residual of the forces, the differences between machines
can be extremely large. In order to see this effect, put ecut 7 and npulayit 7 (its default value).

P67.4 (XG110331)
The test Tv5#76 gives wrong numerical answer on littlebuda_gcc45_gpu.
Likely due to numerical precision on the GPU card ?

P67.5 (XG110511)
Switching to optforces=2 makes Tparal#T4 crash.
Also leads to some small modifications, hard to understand,
 in Tparal #O4, #U4, AB4. 

P68.1 (XG110528)
Tests nlo7-10 should be out of tutorespfn ...
 They test ird1den and ird1wf, but are not used in the tutorial !

P68.2 (XG110604)
mGGAs from libxc raise several problems ...
(1) Tran-Blaha is hard to make converge. See Tlibxc#11 and 12
(2) TPSS and VSxc from Van Voorhis and Scuseria lead to divergence
when the gradient is zero, while the kinetic energy density is not zero.

P610.3 (JMB111201)
Due to open64 compiler bug, the compilation of  "79_seqpar_mpi/lobpcgwf.F90" crashs
The optimization is disabled for the folder 'src/79_seqpar_mpi' :
   fcflags_opt_79_seqpar_mpi = -O0 -m64 -align64 

P611.1 (XG111210)
Lot of errors messages from abilint...
But without consequences at the level of the compilation, on any machine.
The output of abilint should be placed under the control of buildbot ... like everything else.

P611.2 (closed)

P611.3 (XG111220)
Disabled Tv6#102, due to a problem when merged mverstra branch...

P611.4 (XG111224)
For fock, had to use -O2, while previously it was OK with -O3.
Problem of compilation (compiler internal error) wth the routine m_vcoul.F90 .

P611.5 (closed)