Segmentation fault bug in hybrid functional HFX calculations

Tobias Binninger t... at zurich.ibm.com
Tue Nov 21 12:42:33 CET 2017


Hello,

I have a problem using the HFX in cp2k. During runtime, I get a 
segmentation fault during the bits2ints decompression. The problem is 
persistent on different platforms and using different compilers: On a 
POWER8 cluster (both compiled with GCC and with XL), and also on a MacBook 
Pro compiled with GCC. I am compiling cp2k using the libint-1.1.4 for the 
ERI. Also, the error is persistent both with the sopt and popt version, so 
both with and without MPI. Therefore, I believe, the bug must be buried in 
the cp2k HFX routines. The segmentation fault apparently happens during a 
decompression step in the bits2ints_Nbits subroutines. Depending on the 
detailed input settings, it occurs in different Nbits values, so 
bits2ints_1, bits2ints_3, bits2ints_11, etc.

I can circumvent the problem only if I switch off the 
compression/decompression completely by setting &FORCE_EVAL &DFT &XC &HF 
&MEMORY MAX_MEMORY 0. Also, the segmentation fault does not show up when I 
put SCREEN_ON_INITIAL_P F. In this case, however, the SCF loop doesn’t 
converge properly.

Below is an exemplary input showing the segmentation fault after the first 
SCF step. The required wave function TEST_HFX_BUG-RESTART.wfn is attached.

Best regards,
Tobias


 &GLOBAL
   PRINT_LEVEL  LOW
   PROJECT_NAME TEST_HFX_BUG
   RUN_TYPE  ENERGY_FORCE
 &END GLOBAL

 &FORCE_EVAL
   METHOD  QS
   STRESS_TENSOR  ANALYTICAL
   &DFT
     BASIS_SET_FILE_NAME BASIS_MOLOPT
     BASIS_SET_FILE_NAME BASIS_ADMM
     BASIS_SET_FILE_NAME BASIS_ADMM_MOLOPT
     POTENTIAL_FILE_NAME GTH_POTENTIALS
     WFN_RESTART_FILE_NAME TEST_HFX_BUG-RESTART.wfn
     UKS  T
     MULTIPLICITY  1

     &SCF
       MAX_SCF  30
       EPS_SCF     1.0E-06
       SCF_GUESS  RESTART
       &OT  T
         ALGORITHM  IRAC
         PRECONDITIONER  FULL_ALL
         ENERGY_GAP     1.0E-03
       &END OT
       &OUTER_SCF T
        EPS_SCF 1.0E-06
        MAX_SCF 10
       &END OUTER_SCF

     &END SCF

     &QS
       EPS_DEFAULT     1.0E-12
       METHOD  GPW
       EPS_PGF_ORB 1.0E-14
       EPS_FILTER_MATRIX 0.0
     &END QS

     &MGRID
       NGRIDS  5
       CUTOFF     3.00E+02
       REL_CUTOFF     6.00E+01
     &END MGRID

    &AUXILIARY_DENSITY_MATRIX_METHOD
     METHOD BASIS_PROJECTION
     ADMM_PURIFICATION_METHOD MO_DIAG
    &END AUXILIARY_DENSITY_MATRIX_METHOD

    &XC
      &XC_FUNCTIONAL
        &PBE
          SCALE_X 0.75
          SCALE_C 1.0
        &END
      &END XC_FUNCTIONAL
      &HF
        &SCREENING
          EPS_SCHWARZ 1.0E-2
          EPS_SCHWARZ_FORCES 1.0E-2
          SCREEN_ON_INITIAL_P T
          SCREEN_P_FORCES T
        &END
        &MEMORY
          MAX_MEMORY  1000
          EPS_STORAGE_SCALING 1.0
        &END
        &INTERACTION_POTENTIAL
          POTENTIAL_TYPE TRUNCATED
          CUTOFF_RADIUS 2.0
          T_C_G_DATA t_c_g.dat
        &END
        FRACTION 0.25
      &END
    &END XC
   
  &END DFT

   &SUBSYS
     &CELL
       A     5.5    0.0     0.0
       B     0.0    5.75    0.0
       C     0.0    0.0    7.11
       MULTIPLE_UNIT_CELL  1 1 1
     &END CELL
     &COORD
Fe       0.0    0.0    0.0
Fe        2.7    2.1    3.6
O        1.8    0.0    6.8
O        4.5       0.0        5.6
O       0.0    1.8     0.3
     &END COORD

     &KIND Fe
       BASIS_SET DZVP-MOLOPT-SR-GTH
       AUX_FIT_BASIS_SET cFIT13
       POTENTIAL GTH-PBE-q16
     &END KIND
     &KIND O
       BASIS_SET DZVP-MOLOPT-SR-GTH
       AUX_FIT_BASIS_SET cFIT3
       POTENTIAL GTH-PBE-q6
     &END KIND
   &END SUBSYS
 &END FORCE_EVAL

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.cp2k.org/archives/cp2k-user/attachments/20171121/10cce143/attachment.html>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: TEST_HFX_BUG-RESTART.wfn
Type: application/octet-stream
Size: 37916 bytes
Desc: not available
URL: <https://lists.cp2k.org/archives/cp2k-user/attachments/20171121/10cce143/attachment.obj>


More information about the CP2K-user mailing list