phys-gw-from-bareH: Segmentation fault in dipole_io for hf_and_locXC with covariant dipoles

[thge123]

Description

Segmentation fault occurs in hf_and_locXC part when using BareHfromScratch with covariant dipoles (screenshot 1).

"yambo_sc" received signal SIGSEGV, Segmentation fault.

Possible cause

Related to a call to DIPOLE_IO from Bare_Hamiltonian_from_Scratch. DIPOLE_IO tries accessing the unallocated variable PAR_K_scheme%IND_bz%element_1D(ik) (line 202 in DIPOLE_IO.F) (screenshot 2).

PAR_K_scheme%ind_bz%element_1D should be allocated by the following line inside PARALLEL_SETUP_K_scheme (line 88 in PARALLEL_SETUP_K_scheme.F)

call PAR_INDEX_copy(PAR_IND_Xk_bz, PAR_K_scheme%IND_bz )

but PAR_IND_Xk_bz%element_1D seems to be unallocated to begin with (screenshot 3).

Is there a missing allocation step for PAR_IND_Xk_bz%element_1D somewhere before?

Steps to reproduce

HF_and_locXC calculation with yambo_sc on Silicon with 12x12x12 k-mesh from Quantum Espresso.
Files for reproduction: https://drive.google.com/file/d/1_NuduVHULqlLnFidAjdwHeW_mBoGFTYa/view?usp=sharing
Happens in both serial and parallel mode.

Yambo input file:

dipoles
HF_and_locXC                     # [R] Hartree-Fock
BareHfromScratch                                                
BareHScaleFermi                                                 
WriteVxc                                                        
dipoles                                                         
% DipBands                                                      
   1 |  10 |                     # [DIP] Bands range for dipoles
%                                                               
DipBandsALL                       # [DIP] Compute all bands range, not only valence and conduction
DipApproach='Covariant'
DipComputed= "R P V P2"                 # [DIP] [default R P V; extra P2 Spin Orb]
#FFTGvecs=  7239            RL
#EXXRLvcs=  47183           RL    # [XX] Exchange    RL components
#VXCRLvcs=  47183           RL    # [XC] XCpotential RL components
FFTGvecs=  200           RL
EXXRLvcs=  200           RL    # [XX] Exchange    RL components
VXCRLvcs=  200           RL    # [XC] XCpotential RL components
%QPkrange                        # [GW] QP generalized Kpoint/Band indices
1|1|1|10|
%

Version used

Yambo(sc) - MPI+SLK+SLEPC+HDF5_MPI_IO - Ver. 5.3.0
Phys-gw-from-bareH branch with commit: 77ccae0
Compiler: GNU Fortran (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0
MPI: mpirun (Open MPI) 4.1.2

2.

3.

[sangallidavide]

Fixed a number of problems related to the code:

Parallelization (the one you mention). Now Bare_Hamiltonian_from_Scratch uses DIP_par_scheme="HAM" in "not l_sc_run". Possibly the same could be used in l_sc_run. (see change in PARALLEL_global_Self_Energy here 9bbc538 )

There were few other issues (use of errore function from QE, impose maximum cutoff, cutoff on G-hartree not defined). Also p2y was broken.

It runs in serial.

[sangallidavide]

Addendum. After this commit e3ff8db it seems to also runs in parallel, although without memory distribution. Everything to be checked.

[thge123]

Hi Davide. Thanks for your quick response. One question - did the HF_and_locXC part work for you? I tried the commit from your last comment. I didn't get the segmentation fault but got the following error with the same input as above (both serial and parallel):

At line 58 of file V_Hartree.f90
Fortran runtime error: Index '7240' of dimension 1 of array 'fft_g_table' above upper bound of 7239

If it worked for you, could you share the input you used? I might be missing something.

[sangallidavide]

If you lower VXCRLvcs you need lo lower HARRLcvs to the same value (or lower). This could be avoided by changing the call to WF_load in Bare_Hamiltonian as follows: replace QP_ng_Vxc with max(QP_ng_Vxc,QP_ng_Hartree) (Check the correct variable name). I didn't yet because we would need to decide what is better for the user (discussed today with Fulvio). Anyway, my suggestion is to keep both VXCRLvcs and HARRLvcs to the maximum, and lower instead EXXRLvcs, which is responsible for the expensive part (e.g the exchange)