[Perf] hPsi consuming too much CALLS for cg compared to Davidson #6790
Description
Details
nonlocal_pw operator consumes up to 70% of total computation time in some of our PW examples(https://mcresearch.github.io/abacus-user-guide/test-10cases.html), and is the main performance bottleneck.
See 10cases for details.
- 003_12Pt111
- 004_3BaTiO3
- 006_27Fe
- 007_32H2O
- 008_Li27Ni9O54Mn9Co9
- 009_216Si
nonlocal_pw accounts for 53%-73% of total time in these cases for ks_solver cg.
Note
While nonlocal_pw only accounts for ~10% if ks_solver is set to dav_subspace.
Caution
Seems that hPsi, nonlocal_pw, etc. are called too much times in CG cases.
Note
In all cases the CALLS count from cg to hPsi and nonlocal_pw, etc. is significantly higher than that from dav_subspace, ~1000x more!