dfh: fix direct_iaq memory estimate in AO_core module #1100
Conversation
| if(direct_iaQ_) { | ||
| // the direct_iaQ method does not use sparse storage | ||
| // if do_wK added to code, the following will need to be changed to match | ||
| required = naux_ * nao_ * nao_ ; |
There was a problem hiding this comment.
@schiebermc Can you review this?
There was a problem hiding this comment.
I don't think this changes anything.
Having dense integrals implies that big_skips_[nao_] == naux_ * nao_ * nao_.
Is there a reason this was added in? did something break?
There was a problem hiding this comment.
@CDSherrill can explain more fully, but something did break. In a direct_iaq (fisapt) test with a certain amount of memory, it was passing through this memory estimator cleanly, then at AO_core transform, it was failing that memory estimator and throwing an error. With this fix in place, it noticed memory was insufficient for AO_core so switched to disk and completed correctly. Perhaps you can post the triggering input file, @obrien951.
There was a problem hiding this comment.
Aha, upon further inspection this is probably the case. I thought that I had switched cutoff_ = 0.0 when direct_iaQ_, but this appears to not be the case. Easily confirmable if the mask sparsity printed to the output is not 0.0%. Therefore big_skips_[nao_] represents the total number of doubles for the sparse form, even when direct_iaQ_.
There was a problem hiding this comment.
In fact line 484 uses the correct memory estimate (direct from naux_, nao_). I think the issue was that AO_core() was added during the JK building push, so I did not consider the dense integrals from direct_iaQ. Probably the discrepancy between this and lines 605-607 causes the throw. I approve of the changes.
There was a problem hiding this comment.
Thanks for checking this, @schiebermc
Todos
Notable points that this PR has either accomplished or will accomplish.
Checklist
Status