Enable complex coefficients for SelfHealingOverlap estimator #5291
Conversation
|
Review comments addressed. |
|
134617d |
|
Test this please |
| { | ||
| RefVector<SlaterDet> refs; | ||
| for (auto& component : Z) | ||
| if (auto* comp_ptr = dynamic_cast<SlaterDet*>(component.get()); comp_ptr) |
There was a problem hiding this comment.
findSD is seems too brief, this is now part of the public API and it would be nice if it was recognizable as what it is from that. Imagine I want to search the codebase for SlaterDet definitions
auto my_det = twf.findSD()
SD is not a character combination of much specificity.
This smells a bit, dynamic_casts aren't zero cost
There was a problem hiding this comment.
Don't worry about that. I'd like to addressed it separately.
findSD/findMSD are too specific. I'd like to change them as mentioned #5291 (comment)
| auto nmsd = msd_refvec.size(); | ||
|
|
||
| size_t nparams; | ||
| if (nmsd == 1 and nsd == 0) |
There was a problem hiding this comment.
Per Coding specification &&, || etc. only
https://qmcpack.readthedocs.io/en/develop/developing.html#alternative-operator-representations
|
For the tests to matter, tests need to exist. As far as I can tell the CI is only checking this compiles. |
deterministic-LiH_ae_msdj_legacy-vmc-estimator-sh_coeff tests does run in real builds in CI. |
|
Test this please |
|
Ok but application level testing is only indirect testing, and not of these changes. It also implies that development is being done with unknown input and validated against unknown data to an unknown degree. Instead of developing that testing already owed for previously committed code features are being expanded with unclear testing. We have to eventually have a uniform set of standards and we can't expect anyone to follow them if some can get code committed while ignoring them. |
|
Some comments where we can do better. I am disappointed this PR was merged without better test coverage or at least comments to say that these are coming. I hope the latter since we are where we are.
The need for 1-1 is not obviously documented, so I'll state it here: besides ease of future debugging, we only do sanitizer builds without MPI, so this code could have an off-by one indexing error in it that we would otherwise have caught automatically (Hopefully not). Additionally another critical reason is that many of the nightly builds do not have MPI. This is due to the limited testing resources we have. e.g. There are currently no nightly builds on AMD MI210 GPUs with MPI. This code could break on MI200 series e.g. Frontier., and we would not have a chance to catch it without a 1-1 test in the nightlies.
|
Proposed changes
Small update to handle complex wavefunctions in SelfHealingOverlap estimator. Resolves #5229 (fixes the "plaid scarf" pattern seen in the coefficients for complex runs).
Includes light scaffolding to later include rotated Slater determinants as well (commented out).
What type(s) of changes does this code introduce?
Does this introduce a breaking change?
What systems has this change been tested on?
Laptop. Inti at ORNL
Checklist