Skip to content

Optimize initialization

Ye Luo requested to merge ye-luo/q-e:opt-init-rebase into master

I noticed two issues in my calculations recently.

  1. make_pointlists takes minutes instead of seconds in the initialization. So I rewrote the calculation of distance computing. Only 8 corners instead of 27 needed and Intel compiler is able to vectorize the loop. This is the algorithm used in QMCPACK.
  2. ggen uses a lot of scratch memory and it doesn't decrease as more nodes are used. I changed the algorithms to bring down the prefactor and was able to save a few GB per MPI. I only checked 1&6 MPI and not gamma and need extra help to check the gamma trick case since I never used it.

Let me know if there is anything I need to do.

Merge request reports