Skip to content

Draft: MGGA contribution to the Hamiltonian performed on the GPU

Pietro requested to merge bonfus/q-e:mggagpu into develop

I want to run simulations using SCAN on Marconi100@CINECA.

Without this patch

 cegterg      :   1291.32s CPU   1307.31s WALL (     720 calls)
 h_psi        :   1299.52s CPU   1310.02s WALL (    2252 calls)
 Called by h_psi:
 h_psi:calbec :      0.32s CPU      0.94s WALL (    2252 calls)
 h_psi_meta   :   1288.80s CPU   1293.19s WALL (    2252 calls)

With this patch

 cegterg      :     58.12s CPU     75.03s WALL (     720 calls)
 h_psi        :     45.85s CPU     57.45s WALL (    2252 calls)
 Called by h_psi:
 h_psi:calbec :      0.36s CPU      1.01s WALL (    2252 calls)
 h_psi_meta   :     38.77s CPU     48.29s WALL (    2252 calls)

It's still a draft because

  • It's not using batched FFTs, that would make it much faster,
  • I didn't test the gamma case, but more importantly...
  • it's yet another duplicated subroutine and yet another file, so maybe you want to wait for the OpenACC version and close this.

Merge request reports