Differences after introduction of fftGrid type
I noticed that one of the Hubbard 1 tests (Gd_Hubbard1) was failing with slightly different energies and distances. Even though it can be argued that the test is too sensitive, I searched for the commit that caused the changes. I traced the changes back to 8d289938, which introduced the fftGrid type in pwden.
Here I show test calculations with/without SOC with 1 iteration for the commits a1283ec8 and 0f781f26 (the original commit did not compile with intel but I also tested it with only the bugfix added and got the same results). Since only 1 iteration is done there is no change in the total energy, but the distances are already different.
|distance before 8d289938||distance after 8d289938|
The out files below show a different interstitial charge for the two versions, which lead to changes in the total energy in consecutive iterations. I do not know if this difference is expected, but the changes seem to only occur for SOC (second variation, have not tested noco yet)
Input and output file:
|before 8d289938||after 8d289938|