Deprecate Built-In ADC

[JonathonMisiewicz]

Description

This PR deprecates Psi's built-in ADC module, per discussion on the January developer conference call, issue #1033, and my own investigation.

Checklist

• Tested ADC still runs, but with well-deserved warnings

Status

• Ready for review
• Ready for merge

[mfherbst]

Cool! Just as a remark: I suggest to default to adcc in procrouting as well for ADC(2) in case it is available. Right now the default is still the builtin.

[JonathonMisiewicz]

Cool! Just as a remark: I suggest to default to adcc in procrouting as well for ADC(2) in case it is available. Right now the default is still the builtin.

I'm in favor, but this sounds like the sort of thing other developers may object to, so I'll defer to popular opinion before deciding to adopt or not.

[hokru]

Could we catch the case where users request multiple roots and stop the program? Fewer and fewer people look at the actual output files.

I would also support the more extreme solution of just fully disabling the internal adc. Better some potential inconvenience than wrong results.

[JonathonMisiewicz]

Could we catch the case where users request multiple roots and stop the program? Fewer and fewer people look at the actual output files.

I would also support the more extreme solution of just fully disabling the internal adc. Better some potential inconvenience than wrong results.

1. This is possible in principle, but I don't know exactly what conditions trigger the bug, so there may be some single-root cases that also trigger it.
2. Removing the buggy internal ADC was my original proposal on the January conference call, but @fevangelista disapproved without a deprecation period first. I was hoping to avoid an argument about "remove vs deprecate" as not a good use of my time, but if we're going to have the argument anyways, I'm on the remove side.

[maxscheurer]

Well since the argument came up again: I'm all for removal, because with adcc, we can guarantee a) correct results (several test cases!) b) all sorts of properties and c) up-to-date documentation.

The step to go from requesting a psi-internal ADC calc to having the same job running with adcc is a simple conda install.

IMHO there should not be a "deprecation phase" for a feature which is obviously broken and a full replacement is already there.

[PeterKraus]

I suppose if people want to get the "broken" version of ADC, they can install Psi4 between (at least) 1.3.2 and 1.5, right? I don't think we need to deprecate features that don't work, and if adcc is a "drop-in" replacement for ADC, I'd too vote for removal. Maybe in 1.6, and perhaps 1.7, calls to ADC should be shimmed to adcc with a deprecation warning, and afterwards the ADC interface removed (if it's different than adcc).

[loriab]

I can see 3 levels of adc deprecation:

1. warn on build-in adc invocation and leave adc as default
2. warn on built-in adc invocation and switch adcc to default (managed_adcc, as @mfherbst suggests)
3. rm -rf adc/

Right now this PR does (1). I don't remember if @fevangelista was against (2) or only against (3). I'm against (3) until at least after 1.6 release. On the whole, I favor (2) now, but this PR of (1) now is an excellent first step.

[maxscheurer]

Ok, (2) is a solution I can live with 😄 Then at least the ADC calculations from 1.6+ won't be wrong by default 😉:+1:

[JonathonMisiewicz]

(2) was never discussed on the conference call, so I'll take that option for this PR. The driver-side changes are straightforward, test changes should be as well.

[JonathonMisiewicz]

Default switched to adcc, if available. Old test cases for built-in code still use the built-in code. Fixed a bug where it was impossible to select the built-in code. Docs updated accordingly.

@loriab @maxscheurer @mfherbst

[loriab]

I would have done "if module in {ADCC, ''}: run_adcc; elif module in {BUiltin}: run_adc" to avoid the complication of mixed defaults. But as-proposed satisfies the continuity camp by keeping running. And it satisfies the correctness camp by using adcc and peppering the user with adc warnings. So, lgtm!

[JonathonMisiewicz]

Great, will merge on test pass.