Open marcelloPuligheddu opened 2 weeks ago
Workaround from 30cb671 works fine. It creates a new handle each time integrate_4c is called.
This is not perfect since: we could save handle in master and share them not really sure what happened to the memory allocated from previous handles
Something strange is happening.
I think, at last in this test, openmp threads > 1 are getting destroyed and recreated between call to integrate_four_center.
Unfortunately this completely invalidates all libGint internals, expecially the libGint_handle.
This means we have to recreate all handles and all associate objects at every call to int_4_c.
This should also lead to non insignificant memory leaks (as in undestroyed objects) on the cpu and probably also on the gpu.
To reproduce, run the H20-32.inp file (reduced to contain a single molecule) on 4 openmp, single MPI.
Possibly due to the small size of the test Possibly limited to my machine