Test Br2_def2-QZVPP_MP2 is broken with 2022 version, works with 2019

Dear all,

I have tried to run the tests with the 2022 version, but unfortunately many tests are failing. One of the first tests to fail is Br2_def2-QZVPP_MP2. This test passes with the 2019 version. Outputs for both are attached.

The problems seems to be that the test requests "calc=MP2", which used to result in a DF-MP2 calculation by default, but in the 2022 version this actually does a non-DF MP2. Changing the input file to "calc=DF-MP2" makes the test pass (see third attachment), but based on the file name (and the fact that there is already a test for DF-MP2) that is not a correct way to resolve this.

 

Please note that it is an old test job. Please delete the MTEST directory before unpacking the tarball.

Thank you, that has reduced the number of tests to 1053 and fixed almost all of the test failures. (MRCC build from source with all patches, using the latest ifort 2021.7.0 from the OneAPI 2022.3.0 bundle)

There is one particular test though, that crashes intermittently: Propanol_ONIOM_ee_PBE0_PBE_HF_STO-3G_sp_qcorr_special

Often this test would pass, sometimes it would pass many times in a row, but after running it many times, it looks like it crashes with a low probability, maybe around 10%. I have attached the output from a crash.

I have tried to debug it, but could not make much progress. Intermittent crashes like this are often a sign of using an uninitialized variable. I have run a debug build with valgrind, and it suggests that there is an uninitialized variable (or array) somewhere in make_qcorr (compmod.f line 4356). Values derived from the uninitialized variable appear to be used in print_layerborders (compmod.f, line 1792), write_minp (compmod.f, line 1496) and some other functions, according to valgrind.

So at the moment that is my best guess, especially as the immediate reason for the crash is the MINP file previously written having corrupted point charges.

Please let me know if the full valgrind output would be useful for debugging, I could upload that as well.
 

Dear Tibor,
Unfortunately, I cannot reproduce the reported failures of the mtest, however, valgrind do complains a lot about procedures in compmod.f. I investigated these warnings, and it turned out that some of the variables are uninitialized, which might be the source of the mtest failures. A patch is uploaded, and there is no more warnings from valgrind with neither gnu and intel compaliers. I hope the updated version solves the problem.

Best regards,
Bence Hégely

Thank you for the patch, I will rebuild MRCC and post an update here when I have results.

I have retested the previously failing input by running it 100 times (as the crash was intermittent to begin with), and I can confirm that none of them failed.
On a related note, I have also ran the test suite (1053 tests) 3 times, and I can confirm that I have seen no more failures if MRCC is compiled with , so ifort 2021.7.0 seems to be OK to use. (with the caveat that I have not tested libraries like LibXC)