Closed janhubicka closed 1 month ago
Disabling openMP helps, but I am not sure why.
Curiously, clang's openMP is fast, but yields wrong result (only first part of image is correctly finetuned). Seems there is some sort of race.
set OMP_NESTED=true solves the problem both with speed of gcc implementation and wrong output with clang. There are really nested OMP regions since finetune solver calls precompute from OMP controlled region.
now works for me (and no one else uses this anyway :)
This may be related to the stack usage or perhaps different behaviour of OpenMP on Windows