Fix division by zero at geoopt #1170
Conversation
Signed-off-by: Igor S. Gerasimov <[email protected]>
Signed-off-by: Igor S. Gerasimov <[email protected]>
| @@ -850,7 +850,7 @@ subroutine relax(env,iter,mol,anc,restart,maxcycle,maxdispl,ethr,gthr, & | |||
| write(env%unit,'(5x,"change ",e18.7,1x,"Eh")') echng | |||
| write(env%unit,'(3x,"gradient norm :",f14.7,1x,"Eh/α")',advance='no') gnorm | |||
| write(env%unit,'(3x,"predicted",e18.7)',advance='no') depred | |||
| write(env%unit,'(1x,"("f7.2"%)")') (depred-echng)/echng*100 | |||
| write(env%unit,'(1x,"("f7.2"%)")') (depred-echng)/(echng+1e-34_wp)*100 | |||
There was a problem hiding this comment.
How did you determine this fixed #500? At least in the example they attached, the gradient norm is also NaN.
I would probably say to just not even write the percentage difference if the actual echng is effectively 0, rather than write an enormous value.
There was a problem hiding this comment.
I found since I am using the same technique as for other issues. :-)
| @@ -613,7 +613,11 @@ Subroutine dphidrPBC(mode,nat,xyz,i,j,k,l,vTrR,vTrB,vTrC,phi,& | |||
| dphidrj=0 | |||
| dphidrk=0 | |||
| dphidrl=0 | |||
| onenner=1.0d0/(nan*nbn) | |||
| if (abs(nan*nbn).gt.eps) then | |||
There was a problem hiding this comment.
It might be worth clarifying what the purpose of this check is. It seems like we would only reach this condition if for some reason the magnitude of one or both of these surface normals was very small. The earlier check already determines if the normals are nearly parallel.
Is there a specific example you have found that the code change here fixes?
There was a problem hiding this comment.
The earlier check already determines if the normals are nearly parallel.
You can be here if nan or nbn equal zero and then you divide by zero :)
Is there a specific example you have found that the code change here fixes?
Yep, you can detect this with input provided in #500.
There was a problem hiding this comment.
I'm surprised the example from #500 goes through this function. I assumed from the name it only applied for PBC structures, but the example from the issue is molecular.
Fixes #500