BSE calculation of spin polarized system in 4.1.4

[wufeng]

Dear Developers,
I met a problem when performing BSE calculation of spin polarized system with yambo 4.1.4. It raises an error at Wave Function Phases step:

Code: Select all

   [06.04.02] Wave-Function Phases
   ===============================
   [ERROR] STOP signal received while in :[06.04.02] Wave-Function Phases
   [ERROR]EMPTY WF index @ b:1 k:1 spin:2 ID:0

This error occurs for multiple different systems in parallel. For serial run there is no error.

From the source code, in src/parallel/PARALLEL_WF_index.F I found only the first spin is set during initialization:

Code: Select all

   do ib=1,NB
     b_filling(ib)=0
     if (states_to_load(ib,ik,1)) b_filling(ib)=1
   enddo
   !
   call PP_redux_wait(b_filling,COMM=PAR_COM_density%COMM)
   !
   do ib=1,NB
     if (b_filling(ib)==0.and.PAR_COM_density%CPU_id==0)  states_to_load(ib,ik,1)=.TRUE.
   enddo

And from debugger, states_to_load(1,1,2) is always FALSE so the error will be raised later when accessed.

I managed to change the line to

Code: Select all

 states_to_load(ib,ik,:)=.TRUE.

and the code works again. Not sure if the result is correct, however.

Is spin 2 skipped here as intended or is it a bug?

Thanks very much!

[wufeng]

The input/output/LOG is attached.

[Daniele Varsano]

Dear Feng Wu,
we have already reproduced the error you reported,
we will fix it as soon as possible and let you know.

Best,

Daniele

[Davide Sangalli]

Dear Feng Wu,
if you mean you replaced in src/parallel/PARALEL_WF_index.F

Code: Select all

     if (b_filling(ib)==0.and.PAR_COM_density%CPU_id==0)  states_to_load(ib,ik,1)=.TRUE.

with

Code: Select all

     if (b_filling(ib)==0.and.PAR_COM_density%CPU_id==0)  states_to_load(ib,ik,:)=.TRUE.

then I would say that your solution is correct, while the previous version was buggy.

D.

[wufeng]

Dear Feng Wu,
if you mean you replaced in src/parallel/PARALEL_WF_index.F

Code: Select all

     if (b_filling(ib)==0.and.PAR_COM_density%CPU_id==0)  states_to_load(ib,ik,1)=.TRUE.

with

Code: Select all

     if (b_filling(ib)==0.and.PAR_COM_density%CPU_id==0)  states_to_load(ib,ik,:)=.TRUE.

then I would say that your solution is correct, while the previous version was buggy.

D.

Yes that is exactly what I did. Thanks very much.

[martinspenke]

Dear Davide,

I checked this line of code in 4.2.0, and the bug was present. However,
I never obtained the mentioned error above for a spin polarized system when performing BSE calculations!!!
Does this bug change the BSE results or only the parallelization efficiency?

Best,
Martin

[Davide Sangalli]

If the calculation were affected the code would have crashed, with the message

Code: Select all

       [06.04.02] Wave-Function Phases
       ===============================
       [ERROR] STOP signal received while in :[06.04.02] Wave-Function Phases
       [ERROR]EMPTY WF index @ b:1 k:1 spin:2 ID:0

However it only happens very seldom that such part of the code is activated.
In short, when computing the BSE kernel, yambo tries to get the matrix elements associated to k-points outside the IBZ by "rotating" the matrix elements in the IBZ.
To do that, it first check that there are no problems with the phase of the WFs. In some special cases there is a phase factor and the matrix element need to be recomputed.
In the unlucky case that the CPU which needs to compute the matrix element for k out of the IBZ has not the the corresponding WF in the IBZ among the list of WFs to be loaded in memory, the subroutine src/parallel/PARALEL_WF_index.F takes care of adding such WF to the list.

Best,
D.