phys-gw-from-bareH: Segmentation fault in dipole_io for hf_and_locXC with covariant dipoles #187

Activity

1. sangallidavide

   added a commit that references this issue on Feb 17, 2025

   Version 5.3.0, Revision 23978, Hash 01ce042

   ...

   9bbc538
2. sangallidavide commented on Feb 17, 2025

   

   sangallidavide

   on Feb 17, 2025

   Member

   More actions

   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.
3. sangallidavide commented on Feb 17, 2025

   

   sangallidavide

   on Feb 17, 2025

   Member

   More actions

   Addendum. After this commit e3ff8db it seems to also runs in parallel, although without memory distribution. Everything to be checked.
4. thge123 commented on Feb 18, 2025

   

   thge123

   on Feb 18, 2025

   Author

   More actions

   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.
5. sangallidavide commented on Feb 18, 2025

   

   sangallidavide

   on Feb 18, 2025 via email

   Last edited by sangallidavide

   Member

   More actions

   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)

   👍React with 👍1Reacted by thge123
6. mrf1g12 commented on Oct 25, 2026

   

   mrf1g12

   on Oct 25, 2026

   More actions

   Hi, following up on this issue: using the same phys-gw-from-bareH workflow I encountered a second reproducible segmentation fault, this time in V_Hartree.F.

   I am using Yambo 5.3.0, revision 23979, commit 9bbc5386b7, with:

   • BareHfromScratch
   • BareHScaleFermi
   • HF_and_locXC
   • DipApproach="Covariant"

   The calculation was crashing at:

   vhg(ig)=((4._DP*pi/DL_vol)*(rhog(fft_g_table(ig,1))))/bare_qpg(1,ig)**2

   inside V_Hartree.F.

   The FFT grid prepared before entering V_Hartree was not large enough for the Hartree G-space. In Bare_Hamiltonian_from_Scratch.F, WF_load is called as

   call WF_load(WF,QP_ng_Vxc,WF_Go_indx,(/1,WFbands(2)/), &
                (/1,k%nibz/),title='-vloc')

   while V_Hartree subsequently loops up to QP_ng_SH.

   In my case,

   QP_ng_Vxc = 24959
   QP_ng_SH  = 78599
   

   and fft_g_table contained indices outside the current fft_size, leading to the out-of-bounds access and segmentation fault.

   I changed the call in Bare_Hamiltonian_from_Scratch.F to

   call WF_load(WF,max(QP_ng_Vxc,QP_ng_SH),WF_Go_indx, &
                (/1,WFbands(2)/),(/1,k%nibz/),title='-vloc')

   and this fixes the problem.

   I then restored the original input cutoffs, i.e. I did not increase VXCRLvcs/EXXRLvcs to the Hartree cutoff, and the calculation now passes the Hartree step correctly and proceeds to the G0W0 calculation.

   It therefore seems that, in this workflow, the FFT grid prepared by WF_load before V_Hartree needs to account for both QP_ng_Vxc and QP_ng_SH.