Integration test needed for the ACForce estimator

[tiihonej]

QMCPACK runs with the ACForce estimator crash at launch with the latest code (since around Jan 16). The error message is something like the following:

Resetting Properties of the walkers 1 x 50
ERROR You cannot resize a constant size matrix beyond its initial max dimensions
Fatal Error. Aborting at Unhandled Exception
*** The MPI_Abort() function was called after MPI_FINALIZE was invoked.
*** This is disallowed by the MPI standard.
*** Your MPI job will now abort.
[psi3:13270] Local abort after MPI_FINALIZE started completed successfully, but am not able to aggregate error messages, and not able to guarantee that all other processes were killed!

I take suggestions to better pinpoint where this is coming from. Sample inputs are attached:
vmc_force_jan2020.tar.gz

[jtkrogel]

Sounds like a good target for addtional unit/deterministic/integration tests.

[prckent]

Ack! A simple view of this would be that we need a simple integration test and preferably a deterministic integration test that uses ACForce. However, better would be a unit or small scale integration test that can catch this.

[PDoakORNL]

#2232 will be done today and should fix this

[tiihonej]

#2242 Did not fix it for me. I could write a simple integration test for ACForce, @prckent . Actually, the above inputs slightly modified would probably work?

[prckent]

@tiihonej Please modify an existing test, e.g. LiH dimer. This will avoid introducing new files for the wavefunctions and potentials. Of course, this assumes that LiH triggers the test. I suspect that parsing the force to check the numerical results will require writing some new python code in the test system, but for now that is not essential: the key test is whether the code crashes or not.

[tiihonej]

@prckent Modifying existing file should be alright and, and LiH already covers a great deal of force-related features (the only exception probably being the derivatives of some higher-order WF-components). Testing against the values of the Pulay terms should be straightforward and catch all there is to catch regarding the actual performance.

[PDoakORNL]

@tiihonej Do you mean you get the same resize error?

I'm running your example with this command line

OMP_NUM_THREADS=1 mpirun --mca btl tcp,self -np 16 -map-by numa:PE=1 ../bin/qmcpack vmc.force.xml

[tiihonej]

@PDoakORNL Nevermind that, there was something else wrong with my build. I did a fresh checkout and everything works. I'll push an integration test and then close this issue.

[tiihonej]

There are a few issues with adding an integration test for the current ACForce estimator.

First, none of the needed observables (ACForce_{hf,pulay,wfgrad, Ewfgrad}) are implemented in the checking macros or nexus. Adding them is perhaps a bit tedious, considering that the final observable format is perhaps subject to discussion (ie, is scalar.dat a proper place for tensorial quantities? Should QMCPACK sum up the pulay corrections internally and only/also output the final force?).

Second, LiH_dimer_pp is a good and simple platform for the test, but the wavefunction is too good! At equilibrium we get essentially zero forces and corrections, which is not so good for testing.

I think a test like this should only be added once we are done with streamlining the code, including nexus support.