Fixes to SCF_TYPE + SCF_SUBTYPE + SCREENING Combinations #3060
Conversation
davpoolechem
force-pushed
the
dpoole34/compositejk-screenfix
branch
from
7842220
to
9162453
Compare
psi4/src/psi4/lib3index/dfhelper.cc
Outdated
| if (rank) eri[rank] = std::shared_ptr<TwoBodyAOInt>(eri.front()->clone()); | ||
| if (rank) { | ||
| eri[rank] = std::shared_ptr<TwoBodyAOInt>(eri.front()->clone()); | ||
| if (!eri[rank]->initialized()) eri[rank]->initialize_sieve(); |
There was a problem hiding this comment.
Does clone clone the sieve as well? Should one instead first just make sure eri.front() has been initialized()?
There was a problem hiding this comment.
It should, but I will double check. If it does, your suggestion would be a better implementation than what I did here.
There was a problem hiding this comment.
Yeah, it does copy the sieve as well. So I have changed all initialization instances where a clone is involved, to what you suggested!
| @@ -68,7 +68,7 @@ LinK::LinK(std::shared_ptr<BasisSet> primary, Options& options) : SplitJK(primar | |||
| lr_symmetric_ = true; | |||
|
|
|||
| // set up LinK integral tolerance | |||
| if (options["LINK_INTS_TOLERANCE"].has_changed()) { | |||
| if (options["LINK_INTS_TOLERANCE"].has_changed() && options.get_str("SCREENING") != "NONE") { | |||
| linK_ints_cutoff_ = options.get_double("LINK_INTS_TOLERANCE"); | |||
There was a problem hiding this comment.
what does the cutoff default to?
There was a problem hiding this comment.
It defaults to 1e-12 if any SCREENING other than NONE is selected; and 0.0 if SCREENING = NONE.
|
|
||
| /// handle composite methods | ||
| } else if (is_composite) { | ||
| auto jk = std::make_shared<CompositeJK>(primary, auxiliary, options); | ||
|
|
||
| if (options["INTS_TOLERANCE"].has_changed()) jk->set_cutoff(options.get_double("INTS_TOLERANCE")); |
There was a problem hiding this comment.
This has been moved into CompositeJK::common_init() to better account for behavior when SCREENING=NONE.
davpoolechem
force-pushed
the
dpoole34/compositejk-screenfix
branch
from
c5835b3
to
f9d513d
Compare
davpoolechem
force-pushed
the
dpoole34/compositejk-screenfix
branch
from
f9d513d
to
6d401ef
Compare
davpoolechem
force-pushed
the
dpoole34/compositejk-screenfix
branch
2 times, most recently
from
9a97a83
to
a4ec5d0
Compare
davpoolechem
force-pushed
the
dpoole34/compositejk-screenfix
branch
from
a4ec5d0
to
8bf8248
Compare
davpoolechem
force-pushed
the
dpoole34/compositejk-screenfix
branch
from
8bf8248
to
4d375d4
Compare
davpoolechem
force-pushed
the
dpoole34/compositejk-screenfix
branch
from
4d375d4
to
8dde823
Compare
Description
This PR changes the code to fix the majority of combinations of
SCF_TYPE/SCF_SUBTYPE/SCREENINGkeywords which were previously broken.It was discovered that there were a lot of combinations of
SCF_TYPE,SCF_SUBTYPE, andSCREENINGkeywords (e.g.,SCF_TYPE = {CompositeJK}withSCREENING = NONE) that would break upon use, either with an exception or with a hard error such as a segfault. PR #2978 added testing to detect these broken combinations. This PR actually fixes the broken combinations detected through the added test.The most notable change as a result of fixing these issues is the handling of how certain variables within the
TwoBodyAOIntclass are initialized. To enable universal support of JK builds withSCREENING = NONE, an option has been added to manually call theTwoBodyAOInt::create_sieve_pair_info()function via the new wrapper functionTwoBodyAOInt::initialize_sieve(). Multiple JK builds rely on values of variables initialized viacreate_sieve_pair_info(), andinitialize_sieve()is now used to initialize the required variables ifSCREENING = NONEis set. In this way,SCREENING = NONEcan exist with the benefits it provides, while still working with all JK builds. Also, changes are made to ensure that no screening actually occurs in JK builds whenSCREENINGis set toNONE.In line with the above, some tests have been expanded.
test_comprehensive_jk_screening.pyhas been changed to account for theSCF_TYPE/SCF_SUBTYPE/SCREENINGcombinations that were fixed.test_erisieve.pyhas been updated to more thoroughly testSCREENING=NONE.It is worth noting that, unfortunately,
TwoBodyAOInt::initialize_sieve()must be defined per integral engine, and so the fixes of this PR only work with Libint2. The issues this PR tries to fix, still persist if Simint is used instead.User API & Changelog headlines
Dev notes & details
TwoBodyAOInt,initialize_sieve, which allows for manual initialization of screening variables viacreate_sieve_pair_info()needed for JK builds.SCF_TYPE,SCF_SUBTYPE, andSCREENINGwhich were broken.test_comprehensive_jk_screening.pyandtest_erisieve.pytest to account for the previously-described fixes.Questions
-[] Should we keep the
initialize_sievename? I named it as such, because it is basically a light wrapper to manually callcreate_sieve_pair_info. But the name of "sieve" might be objectionable.Checklist
Status