Known bugs
From Diracwiki
This page contains patches to known bugs.
DIRAC10
Automatic augmentation for larger basis sets
Automatic augmentation using s-aug, d-aug, t-aug, ... prefixes.
Only affects larger basis sets and could lead to a linearly dependent basis. What happens is this: larger basis sets get split into blocks to save memory. Automatic augmentation incorrectly augmented each block. After this patch it only augments the most diffuse block as it should.
Save the following to a file:
diff --git a/abacus/herbas.F b/abacus/herbas.F
index 1c82a11..d1c0a9d 100644
--- a/abacus/herbas.F
+++ b/abacus/herbas.F
@@ -253,6 +253,8 @@ C for backward compatibility the old notation is retained
BASTMP = BASNAM(1:ISTART - 1)
END IF
END IF
+
+ iaug_save = iaug
C
C Determine if there are any user added basis functions/polarization
C functions
@@ -368,6 +370,15 @@ C INTEXP = INTEXP_SAVE
ENDIF
END IF
+ if (split) then
+! block has been split and this is not the last part
+! augment only last part
+! do not augment this part
+ iaug = 0
+ else
+ iaug = iaug_save
+ end if
+
INTEXP_SAVE = INTEXP
INTORB_SAVE = INTORB
CALL READNU(INTEXP, INTORB, NBLOCK, KAOVEC, CPRIMU,ALPHA,
then apply the patch in the DIRAC root directory:
patch -p1 < file
finally, recompile DIRAC:
make
DFT geometry optimization is incorrect
A bug in the DFT geometry optimization was unfortunately released together with DIRAC10. The bug leads to non-convergent geometry steps, shows incorrectly integrated number of electrons and incorrect XC energies at displaced geometries. The following patch restores geometry optimization at the DFT level.
Save the following to a file:
diff --git a/abacus/abaopt.F b/abacus/abaopt.F
index b961e5d..0574a43 100644
--- a/abacus/abaopt.F
+++ b/abacus/abaopt.F
@@ -2260,6 +2260,7 @@ C (by running HERMIT and SIRIUS).
C
use dirac_cfg
use dft_cfg
+ use xc_driver
#include "implicit.h"
#include "priunit.h"
@@ -2290,11 +2291,6 @@ C
END IF
NEWGEO = .TRUE.
- if (dirac_cfg_dft_calculation) then
-! get a new grid at new geometry
- call reset_grid()
- end if
-
IF (.NOT. EXSIR) THEN
CALL EXESIR(WORK,LWORK,IWOFF,WRKDLM)
IF (DOCCSD) THEN
@@ -2356,6 +2352,11 @@ ckr#endif
ENERGY = EMCSCF
END IF
#else /* PRG_DIRAC */
+
+ if (dirac_cfg_dft_calculation) then
+! get a new grid at new geometry
+ call reset_grid()
+ end if
C
C Geometry optimization for Dirac
C
diff --git a/dft/basis_info.F90 b/dft/basis_info.F90
index b32c86a..499cffb 100644
--- a/dft/basis_info.F90
+++ b/dft/basis_info.F90
@@ -320,17 +320,21 @@ contains
inquire(file = file_name, exist = file_exists)
if (file_exists) then
- close(file_unit, status = 'keep')
- return
- else
open(file_unit, &
file = file_name, &
- status = 'new', &
+ status = 'old', &
form = 'formatted', &
access = 'sequential')
- rewind(file_unit)
+ close(file_unit, status = 'delete')
end if
+ open(file_unit, &
+ file = file_name, &
+ status = 'new', &
+ form = 'formatted', &
+ access = 'sequential')
+ rewind(file_unit)
+
only_large = .false.
if (mc == 1) then
only_large = .true.
diff --git a/dirac/dirscf.F b/dirac/dirscf.F
index 869d69f..4d0dc9d 100644
--- a/dirac/dirscf.F
+++ b/dirac/dirscf.F
@@ -1041,6 +1041,10 @@ C
C Called from PSIDHF
C
C*****************************************************************************
+
+ use basis_info
+ use occupation_info
+
#include "implicit.h"
#include "priunit.h"
#include "dummy.h"
@@ -1068,6 +1072,11 @@ C in ONEFCK ... no need to prepare LUFCK1 !!!!
CALL WRIFCK(LUFCK1,FOCK,1)
CLOSE(LUFCK1,STATUS='KEEP')
+! write info to files
+! this is then used in dft and visualization
+ call basis_info_write()
+ call occupation_info_write()
+
IF(ONESYS) THEN
C
C Solve one-electron system ...
then apply the patch in the DIRAC root directory:
patch -p1 < file
finally, recompile DIRAC:
make
Parallel DFT crashes with MPICH2
Save the following to a file:
Index: dft/xc_driver.F90
===================================================================
--- dft/xc_driver.F90 (revision 12324)
+++ dft/xc_driver.F90 (revision 12325)
@@ -1675,8 +1675,9 @@
#ifdef VAR_MPI
subroutine dft_collect()
integer :: ierr
- real(8) :: sum_r
- integer :: sum_i
+ real(8) :: sum_r
+ real(8) :: dummy
+ integer :: sum_i
call mpi_reduce(nr_electrons, &
sum_r, &
@@ -1729,43 +1730,51 @@
end if
if (nr_fmat > 0) then
+! suggestion for Radovan (no dcopy + no extra alloc needed) -
+! a simple test with this code piece works fine for me.
if (i_am_master) then
- call alloc(fmat_container, mat_dim, mat_dim, nz, nr_fmat)
+ call mpi_reduce(MPI_IN_PLACE, &
+ fmat, &
+ mat_dim*mat_dim*nz*nr_fmat, &
+ mpi_double_precision, &
+ mpi_sum, &
+ 0, &
+ mpi_comm_world, &
+ ierr)
+ else
+! declare dummy as --> real(8) :: dummy
+! alternatively pass MPI_IN_PLACE instead of dummy for the recvbuf.
+ call mpi_reduce(fmat, &
+ dummy, &
+ mat_dim*mat_dim*nz*nr_fmat, &
+ mpi_double_precision, &
+ mpi_sum, &
+ 0, &
+ mpi_comm_world, &
+ ierr)
end if
-
- call mpi_reduce(fmat, &
- fmat_container, &
- mat_dim*mat_dim*nz*nr_fmat, &
- mpi_double_precision, &
- mpi_sum, &
- 0, &
- mpi_comm_world, &
- ierr)
-
- if (i_am_master) then
- call dcopy(mat_dim*mat_dim*nz*nr_fmat, fmat_container, 1, fmat, 1)
- call dealloc(fmat_container)
- end if
end if
if (do_geo) then
if (i_am_master) then
- call alloc(property_gradient_container, nr_atoms*3)
+ call mpi_reduce(MPI_IN_PLACE, &
+ property_gradient, &
+ nr_atoms*3, &
+ mpi_double_precision, &
+ mpi_sum, &
+ 0, &
+ mpi_comm_world, &
+ ierr)
+ else
+ call mpi_reduce(property_gradient, &
+ dummy, &
+ nr_atoms*3, &
+ mpi_double_precision, &
+ mpi_sum, &
+ 0, &
+ mpi_comm_world, &
+ ierr)
end if
-
- call mpi_reduce(property_gradient, &
- property_gradient_container, &
- nr_atoms*3, &
- mpi_double_precision, &
- mpi_sum, &
- 0, &
- mpi_comm_world, &
- ierr)
-
- if (i_am_master) then
- call dcopy(nr_atoms*3, property_gradient_container, 1, property_gradient, 1)
- call dealloc(property_gradient_container)
- end if
end if
end subroutine
then apply the patch in the DIRAC root directory:
patch -p0 < file
finally, recompile DIRAC:
make
User-defined CAM functional line is cut after character 80
Save the following to a file:
Index: dirac/dirrdn.F
===================================================================
--- dirac/dirrdn.F (revision 12323)
+++ dirac/dirrdn.F (revision 12324)
@@ -4019,7 +4019,7 @@
LOGICAL SYMSTP,NOINPUT,SETFUN,SUCCESS,IOTC,GAUNT_SAVE,
& TWOCOMP_SAVE, TWOCOMPBSS_SAVE
CHARACTER PROMPT*1, WORD*7, TABDIR(NDIR)*7, TABLE(NTABLE)*7,
- & WORD1*7, LINE*80
+ & WORD1*7, LINE*80, LINE100*100
DIMENSION KSOP(0:7),WORK(LWORK)
C
SAVE SETFUN
@@ -4332,12 +4332,13 @@
! .DFT
117 CONTINUE
C&&& DFT : Kohn-Sham (DFT) Hamiltonian
- READ(LUCMD,'(A80)') LINE
+!radovan: line can be longer than 80
+ read(lucmd, '(a100)') line100
IF (.NOT. SETFUN) THEN
C ... hjaaj: DFTINPUT accumulates def. of functional
C so if we call more than once (as in geom.opt.)
C we get n*functional for n'th call !!!! /3-Jun-2003 hjaaj
- CALL DFTINPUT(LINE, INPERR)
+ CALL DFTINPUT(line100, INPERR)
if (inperr .gt. 0) then
call quit('unknown functional after .DFT')
end if
then apply the patch in the DIRAC root directory:
patch -p0 < file
finally, recompile DIRAC:
make
Parallel MCSCF/CI calculations on one node with more than 2 processes are wrong (multiple nodes calculations are fine)
to apply the patch, please save the following patch to file "file"
2188c2188 < IF( IIOMOD_REL .eq. 0 )THEN --- > ! IF( IIOMOD_REL .eq. 0 )THEN 2199c2199 < END IF --- > ! END IF
and type then in the luciarel/ subdirectory of DIRAC:
patch transform_r.F < file
finally, recompile DIRAC:
make
DIRAC08
xlf compiler
Necessary preprocessor statements are missing. Apply the following patch to the configure script if you use the xlf compiler:
Index: /Dirac/trunk/configure
===================================================================
--- Dirac/configure (revision 7775)
+++ Dirac/configure (revision 7790)
@@ -2194,4 +2194,5 @@
# IBM's xlf should be comma delimited, the others take spaces
IBMMOD=1
+ cpp_f90="$cpp_f90 `getmods`"
;;
*)
Parallel 64-bit integer code
The released parallel DIRAC08 code is set up to always link with a 32-bit integer MPI library (tested for both 32-bit integer code and for 64-bit integer code with Open MPI and with Intel MPI (MPICH2 derivate)). However, if you link a parallel 64-bit integer code with a 64-bit integer MPI library you must apply the following patches to get a functioning parallel code:
Index: Dirac/trunk/include/infpar.h
===================================================================
--- Dirac/include/infpar.h (rev 7816)
+++ Dirac/include/infpar.h (rev 7817)
@@ -56,8 +56,12 @@
- INTEGER*4 ISTAT
+ INTEGER ISTAT
Index: Dirac/trunk/gp/global_file.F
===================================================================
--- Dirac/gp/global_file.F (revision 7817)
+++ Dirac/gp/global_file.F (revision 7818)
@@ -313,7 +313,7 @@
INCLUDE "mpif.h"
#include "global_file.h"
- INTEGER*4 STATUS(MPI_STATUS_SIZE)
+ INTEGER STATUS(MPI_STATUS_SIZE)
Parallel direct MP2 calculation
If you run a parallel direct MP2 calculation using DIRAC08 with the .CNVINT flag present on the *DHFCAL deck the calculation will fail. However, if you restart your MP2 calculation from a DFCOEF file and deactivate the .CNVINT flag the calculation will run fine. Example:
**WAVE FUNCTIONS .DHF .MP2 *DHFCAL .CLOSED SHELL 112 .INTFLG 1 1 1 .ERGCNV 1.0E-8 .CNVINT 0.01 0.0001 *MP2CAL .OCCUP 48..56 .VIRTUAL all .VIRTHR 1000.0 **END OF
The above given example will fail if you run it in parallel.
Temporary solution: Stop after the SCF calculation and restart MP2:
**WAVE FUNCTIONS .DHF .MP2 *DHFCAL .CLOSED SHELL 112 .INTFLG 1 1 1 .ERGCNV 1.0E-8 #.CNVINT #0.01 0.0001 *MP2CAL .OCCUP 48..56 .VIRTUAL all .VIRTHR 1000.0 **END OF
