performance of 1-e integral codes

[jeffhammond]

I have been timing 1-e integrals using https://github.com/jeffhammond/nwchem/tree/int1e_time and find that Hondo is ~2x slower than other options, despite being the default, at least for a common basis set like cc-pVDZ, for an organic molecule with ~500 atoms.

The other issue is that NWChem computes the potential 1-e term significantly slower than any other code I've tested. In other codes, the V term is 2-3x slower than the T and S terms. With NWChem, it is 40-80x slower.

I know 1-e integrals are not the bottleneck in anything, but I wonder if there are better defaults.

I am using AMD Zen4 with the NVHPC Fortran compiler, if it matters.

 JEFF
  int_1e_time:         kinetic              0.768110460000000
  int_1e_time:         overlap              0.744332784000000
  int_1e_time:       potential             69.189020218000010

  int_init: cando_txs set to always be   F
 JEFF
  int_1e_time:         kinetic              0.823256014000000
  int_1e_time:         overlap              0.799803475000000
  int_1e_time:       potential             71.307078696999994

  int_init: cando_hnd set to always be   F
 JEFF
  int_1e_time:         kinetic              0.915090658000000
  int_1e_time:         overlap              0.790493348000005
  int_1e_time:       potential             37.916529371999999

  int_init: cando_txs set to always be   F
  int_init: cando_hnd set to always be   F
 JEFF
  int_1e_time:         kinetic              0.920166108000001
  int_1e_time:         overlap              0.788834708000000
  int_1e_time:       potential             38.704786812999998

[edoapra]

Thanks for these numbers.
It might make sense to extend the Simint interface to the 1e- integrals and see if V integral performance from Simint is better.