~abinit-maintainers/abinit/6.0-stable

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 6.0.0 is not suitable for production runs.

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

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

For version 6.0.0, list closed on 20100119
after installation and successful tests for: 	

paral     Xeon with GNU Fort 4.4 compiler                 (doc/config/build-examples/bb_xeon-gfortran4.4_testf.ac)
serial    Xeon with GNU Fort 4.4 compiler                 (doc/config/build-examples/bb_xeon-gfortran4.4_testf_woMPI.ac)
paral     Xeon with Intel Fortran 10.1 compiler           (doc/config/build-examples/bb_xeon-intel10.1_green.ac)
paral     Xeon with g95 compiler                          (doc/config/build-examples/bb_xeon-g95_green.ac)
paral     Opteron/Linux with Pathscale compiler           (doc/config/build-examples/bb_amd-psc_chum.ac)
paral     Intel + Mac OSX with GNU Fortran 4.3 compiler   (doc/config/build-examples/bb_intel-gfortran4.3_bigmac.ac)
paral     Intel + Mac OSX with GNU Fortran 4.3 compiler all a maximal number of debug options
                                                          (doc/config/build-examples/bb_intel-gfortran4.3_bigmac_debug.ac)
paral     Itanium2/Linux with Intel Fortran 11.0 compiler (doc/config/build-examples/bb_ia64-intel11_chpit.ac)
paral     IBM Power 5 with XLF compiler                   (doc/config/build-examples/bb_powerpc64-ibm_fock.ac)
paral     Xeon with GNU Fort 4.3 compiler + MPIIO         (doc/config/build-examples/bb_xeon-gfortran4.3_buda_ref.ac)
paral     Intel with GNU Fort 4.4 compiler + ATLAS lib    (doc/config/build-examples/bb_intel-gfortran4.4_inca_sdebug.ac)

Not tried on :
                  Opteron/Linux with Intel Fortran  9.1 compiler  (doc/config/build-examples/bb_amd-intel9_chum.ac)
                  Opteron/Linux with Intel Fortran 10.1 compiler  (doc/config/build-examples/bb_amd-intel10_chum.ac)
                  Opteron/Linux with GNU Fort 4.2 compiler        (doc/config/build-examples/bb_amd-gfortran4.2_chum.ac)
                  Opteron/Linux with GNU Fort 4.3 compiler        (doc/config/build-examples/bb_amd-gfortran4.3_chum.ac)
                  Opteron/Linux with g95 compiler + openmpi       (doc/config/build-examples/bb_amd-g95_chum.ac)
                  Itanium2/Linux with GNU Fort 4.4 compiler       (doc/config/build-examples/bb_ia64-gfortran4.4_chpit.ac)
                  Compaq/Dec EV67 OSF                             (doc/config/build-examples/alphaev67-compaq_deccint.ac)
                  PowerPC + Mac OSX with g95 compiler + openmpi   (doc/config/build-examples/bb_ppc-g95-openmpi_max.ac)
                  IBM Power 4 with XLF compiler                   (doc/config/build-examples/bb_powerpc64-ibm_dirac.ac)
                  Intel + Mac OSX with g95 compiler               (doc/config/build-examples/bb_x86-g95_bigmac.ac)
                  Xeon with g95 compiler                          (doc/config/build-examples/bb_xeon-gfortra4.2_green.ac)

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, 4.3, or 4.4 instead)
intel10.0         (=> use intel9.1, 10.1 or 11.0 instead)

Platforms temporary out of the pool

Platforms definitely out of the pool 
SGI Octane IRIX (Spinoza) - 081111 v5.6.3

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

Compilation of Netcdf and corresponding testing were tried. 
NetCDF is OK for production.
[If needed, one can also try to "make check" inside lib/netcdf/netcdf-3.6.3]

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 Fox  was not tested, and should NOT be attempted by non-expert users with this version.

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

Speed was not tested. 

Memory was not tested. 


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

All the problems are listed cumulatively.
Those specific to version 6.0 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.

P9. (Massimilio Stengel, October 2000) 
CRASH
MACHINE-DEPENDENT : Compaq/DEC
Problem with non-linear XC correction, it seems

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.

P46.6. (XG050716)
MISC.
Variables nqptdm and qptdm should be substituted with nqpt and qpt.


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.


P53.8 (XG070117)
bdgw should be made spin-dependent ...


P53.10 (XG070121)
During the application of the abirules script (patch-40 to patch-56
in the abinit-devel--merge--5.3 branch),
the numerical values obtained from test v5#04 has changed (1e-7).
There might be a complex(dpc) changed to complex ?!


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.

P57.01 (XG081220)
Test etsf_xc#02 (spin polarized)
has revealed that 
Wu-Cohen differs largely in the native ABINIT implementation (ixc=23)
and in the lib_XC implementation (ixc=-118130).


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).

P57.09 (closed)

P57.16 (XG090501)
I have strong doubts about the validity of the coding of the Berry phase in the PAW case.
Indeed, no special treatment of phase that should be associated with Umklapp folding
is provided in smatrix_k.F90 for the PAW contribution computed in smatrix_pawinit.F90.


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.

P59.02 (closed)

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

P59.04 (closed)

P59.05 (closed)

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

P59.08 (closed)

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

P59.12 (closed)

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 ?


P60.01 (closed)